API:
* Database::check(): Fix checking of replication changesets. This reverts a
change incorrectly made in 1.3.7.
* Database::locked(): Return false instead of true for a closed inmemory DB.
* Database::commit(): If commit() failed with an exception while trying to add
pending changes (e.g. InvalidArgumentError due to a long term containing zero
bytes) then a subsequent commit() on the same object would throw the same
exception. Now we clear the pending changes in this situation (like we
already did for failure at other stages in the commit). This bug remains
unfixed for the chert backend as it's harder to fix there and the effort to
fix it and extra risk of breakage don't seem justified for a backend we
recommend people migrate away from.
* QueryParser::parse_query(): Optimise parsing of multi-word synonyms.
testsuite:
* Use 50-word synonym for qp_scale1 "large" case. 50 divides exactly into the
number of repetitions we do for the "small" case, which 60 (as used before)
doesn't. This makes the two cases a little more comparable and should help
make this testcase less flaky (see #764).
* Adjust testcase matches1 to work with remote shards where the matcher can
return slightly better bounds on the number of matches in some cases.
Resolves 2 XFAILs.
* The testharness get_remote_database() method is now supported for sharded
databases. This is needed for keepalive1 to run successfully under multi
test backends. Resolves 2 XFAILs of keepalive1.
* Improved test coverage:
+ Test locked() on a closed WritableDatabase, which already returns false (as
expected) in 1.4.x (but was broken on master).
+ Check multi databases in testsuite - this has been supported by
Database::check() since 1.4.12.
+ Also test OP_SYNONYM and OP_MAX in emptydb1.
+ Backport testcases boolorbug1, emptynot1, emptymaybe1 and
phraseweightcheckbug1 from git master - these are regression tests for
fixed bugs which only affected git master, but it's useful to confirm that
these bugs don't currently affect 1.4, and ensure they don't get introduced.
* perftest: Store memory sizes as long long since on Microsoft Windows long is
only 32 bits, which is less than common memory sizes.
matcher:
* Hoist positional check above OP_FILTER.
* Handle OP_FILTER with more than two subqueries correctly. Previously we'd
only check the first two subqueries in some situations.
remote backend:
* For a remote WritableDatabase, the client now keeps track of whether there
are pending changes, and if there aren't then we now do nothing for commit()
or cancel() calls. In particular this saves a message exchange when the
WritableDatabase destructor is called when changes have already been
committed with an explicit call to commit() (which is what we recommend
doing, since with an explicit call to commit() you get to see any exception
which gets thrown).
* When closing a remote prog WritableDatabase, previously an exception could
leave the remote connection open with the remote server running, and we'd
then wait for the specified timeout before closing the connection. Now we
close the connection before letting the exception propagate.
* Don't swallow exceptions from Database::close() on a remote database. If
we aren't in a transaction and so try to commit() and that fails then
previously the caller would have no indication of the failure.
* Fix handling the reported term weight when remote shards are searched.
Fixes 5 XFAILs in the testsuite.
* Add missing space to mismatching protocol versions error message.
build system:
* Fix to build when configured with --disable-backend-remote, broken by changes
in 1.4.14. Fixes#797, reported by Дилян Палаузов.
* The clang and icc compilers both define __GNUC__, which led our ABI mismatch
message to report them as "g++" with a bogus version (the version of GCC that
these compilers advertise themselves as, which for clang is always 4.2.0) -
now we report clang++ or icc along with the actual version of that compiler.
documentation:
* AUTHORS: Apply missed update to the thankyou list for 1.4.14.
* INSTALL: Note that MSVC 2019 works.
* INSTALL: Note that Xapian can use the system uuid.h on AIX and OpenBSD.
portability:
* Simplify probes for snprintf. The broken snprintf in libbsd in Linux libc4
is from ~25 years ago so way too ancient to matter now, and all callers
already handle the pre-ISO semantics of returning -1 for an undersize buffer
so we don't need to run a test program to probe for this at configure time,
which is more cross-compile friendly.
* Don't quote messages in #error - the quotes aren't required and appear in the
compiler output (at least with GCC and clang) making it less readable.
* Use a different approach for getting a 64-bit capable stat() for mingw32.
This means we now use the same stat variant for mingw32 and MSVC, which
seems a better plan.
* Work around unhelpful config.status behaviour. It comments out any #undef
lines in config.h, even those added via AH_TOP and AH_BOTTOM. Splitting
these lines means they don't match the regex hammer config.status uses.
* Avoid -Wdeprecated-copy warnings from clang 10.
* Avoid deprecation warning on recent Linux. We were including sys/sysctl.h if
it existed, which it does on Linux but we don't actually use it there.
Including it now warns that it is deprecated, so skip including it under
Linux. Reported on IRC by kumaran.
* Suppress GCC -Wduplicated-branches warning from our API headers in a
different way which avoids needing a compiler-specific #pragma.
* Workaround closefrom1 failure on macOS. It seems under macOS our fd tracking
can end up using fd 10 so start from 13 when testing closefrom() so we don't
close the fd which our fd tracking is using internally.
debug code:
* Log RemoteConnection::read_at_least() return value.
API:
* Xapian::QueryParser: Handle "" inside a quoted phrase better. In a quoted
boolean term, "" is treated as an escaped ", so handle it in a compatible way
for quoted phrases. Previously we'd drop out of the phrase and start a new
phrase. Fixes#630, reported by Austin Clements.
* Xapian::Stem: The constructor which takes a stemmer name now takes an
optional second bool parameter - if this is true, then an unknown stemmer
name falls back to using the "none" stemmer instead of throwing an exception.
This allows simply constructing a stemmer from an ISO language code without
having to worry about whether there's a stemmer for that language, and
without having to handle an exception if there isn't.
* Xapian::Stem: Fix a bug with handling 4-byte UTF-8 sequences which
potentially affects most of the stemmers. None of the stemmers work in
languages where 4-byte UTF-8 sequences are part of the alphabet, but this
bug could result in invalid UTF-8 sequences in terms generated from text
containing high Unicode codepoints such as emoji, which can cause issues (for
example, in some language bindings). Fix synced from Snowball git post
2.0.0. Reported by Ilari Nieminen in
https://github.com/snowballstem/snowball/issues/89.
* Xapian::Stem: Add a new is_none() method which tests if this is a "none"
stemmer.
* Xapian::Weight: The total length of all documents is now made available to
Xapian::Weight subclasses, and this is now used by DLHWeight, DPHWeight and
LMWeight. To maintain ABI compatibility, internally this still fetches the
average length and the number of documents, multiplies them, then rounds the
result, but in the next release series this will be handled directly.
* Xapian::Database::locked() on an inmemory database used to always return
false, but an inmemory Database is always actually a WritableDatabase
underneath, so now we always report true in this case because it's really
always report being locked for writing.
* Fix write one past end of std::vector on certain QueryParser parser errors.
This is undefined behaviour, but the write was always into reserved space, so
in practice we'd actually get away with it (it was noticed because it
triggers an error when running under ubsan and using libc++). Reported by
Germán M. Bravo.
* MSet::get_matches_estimated(): Improve rounding of result - a bug meant we
would almost always round down.
* Optimise test for UTF-8 continuation character. Performing a signed char
comparison shaves an instruction or two on most architectures.
* Database::get_revision(): Return revision 0 for a Database with no shards
rather that throwing InvalidOperationError.
* DPHWeight: Avoid dividing by 0 when searching a sharded database when one
shard is empty. The result wasn't used in this case, but it's still
undefined behaviour. Detected by UBSan.
testsuite:
* Fix failing multi_glass_remoteprog_glass tests on x86. When the tests are
run under valgrind, remote servers should be run using the runsrv wrapper
script, but this wasn't happening for remote servers in multi-databases - now
it is. Also, previously runsrv only used valgrind for the remote for an x86
build that didn't use SSE, but it seems there are x87 instructions in libc
that are affected by valgrind not providing excess precision, so do this for
x86 builds which use SSE too. Together these changes fix failures of
topercent2, xor2, tradweight1 under backend multi_glass_remoteprog_glass on
x86.
* Fix C++ One-Definition Rule (ODR) violation in testsuite code. Two different
source files linked into apitest were each defining a different `struct
test`. Wrap each in an anonymous namespace to localise it to the file it is
defined and used in. This was probably harmless in practice, unless trying
to build with Link-Time Optimisation or similar (which is how it was
detected).
* Test all language codes in stemlangs1. The testsuite hardcodes a list of
supported language codes which hadn't been updated since 2008.
* Improve DateRangeProcessor test coverage.
* The "singlefile" test harness backend manager now creates databases by
compacting the corresponding underlying backend database (creating it first
if need be) rather than always creating a temporary database to compact.
* Enable compaction testcases for multi and singlefile test harness backends.
* Add generated database support for remoteprog and remotetcp test harness
backends. Implemented by Tanmay Sachan.
* Add test harness support for running testcases using a multi database
comprised of one local and one remote shard, or two remote shards.
Implemented by Tanmay Sachan.
* Check if removing existing multi stub failed. Previously if removing an
existing stub failed, the test harness would create a temporary new stub and
then try to rename it over the old one, which will always fail on Microsoft
Windows.
* Wait for xapian-tcpsrv processes to finish before moving on to the next
testcase under __WIN32__ like we already do on POSIX platforms.
matcher:
* Handle pruning under a positional check. This used to be impossible, but
since 1.4.13 it can happen as we now hoist AND_NOT to just below where we
hoist the positional checks. The code on master already handles pruning here
so this bug is specific to the RELEASE/1.4 branch. Fixes#796, reported by
Oliver Runge.
* When searching with collapsing over multiple shards, at least some of which
are remote, uncollapsed_upper_bound could be too low and
uncollapsed_lower_bound too high. This was causing assertion failures in
testcases msize1 and msize2 under test harness backends
multi_glass_remoteprog_glass and multi_remoteprog_glass.
* Internally we no longer calculate a bogus total_term_count as the sum of
total_length * doc_count for all shards. Instead we just use the sum of
total_length, which gives the total number of term occurrences. This change
should improve the estimated collection_freq values for synonyms.
* Several places where we might divide zero by zero in a database where wdf was
always zero have been fixed.
* Optimise OP_AND_NOT better. We now combine its left argument with other
connected and-like subqueries, and gather up and hoist the negated subqueries
and apply them together above the combined and-like subqueries, just below
any positional filters.
* Optimise OP_AND_MAYBE better. We now combine its left argument with other
connected and-like subqueries, and gather up and hoist the optional
subqueries and apply them together above the combined and-like subqueries and
any hoisted positional filters.
* Treat all BoolWeight queries as scaled by 0 - we can optimise better if we
know the query is unweighted.
build system:
* configure: Stop using AC_FUNC_MEMCMP. The autoconf manual marks it as
"obsolescent", and it seems clear that nobody's relying on it as we're
missing the "'AC_LIBOBJ' replacement for 'memcmp'" which it would try to
use if needed.
glass backend:
* Allow zlib compression to reduce size by one byte. We were specifying an
output buffer size one byte smaller than the input, but it appears zlib won't
use the final byte in the buffer, so we actually need to pass the input size
as the output buffer size.
* Only try to compress Btree item values > 18 bytes, which saves CPU time
without sacrificing any significant size savings.
remote backend:
* Fix match stats when searching with collapsing over multiple shards and at
least some shards are remote. Bug discovered by Tanmay Sachan's test harness
improvements.
* Ignore orphaned remote protocol replies which can happen when searching with
a remote shard if an exception is thrown by another shard. Bug discovered
by Tanmay Sachan's test harness improvements.
* Wait for xapian-progsrv child to exit when a remote Database or
WritableDatabase object is closed under __WIN32__ like we already do for
POSIX platforms.
documentation:
* HACKING: Replace release docs with pointer to the developer guide where they
are now maintained.
* Correct documentation of initial messages in replication protocol.
tools:
* quest: Report bounds and estimate of number of matches.
* xapian-delve: Improve output when database revision information is not
available. We now specially handle the cases of a DB with multiple shards
and a backend which doesn't support get_revision().
portability:
* Eliminate 2 uses of atoi(). These are potentially problematic in a
multithreaded application if setlocale() is called by another thread at the
same time. See #665.
* Don't check __GNUC__ in visibility.h as the configure probe before defining
XAPIAN_ENABLE_VISIBILITY checks that the visibility attributes work. This
probably makes no difference in practice, as all compilers we're aware of
which support symbol visibility also define __GNUC__.
* Document Sun C++ requires --disable-shared. Closes#631.
* Fix warning from GCC 9 with -Wdeprecated-copy (which is enabled by -Wextra)
if a reference to an Error object is thrown.
* Suppress GCC warning in our API headers when compiling code using Xapian with
GCC and -Wduplicated-branches.
* Mark some internal classes as final (following GCC -Wsuggest-final-types
suggestions to allow some method calls to be devirtualised).
* Fix to build with --enable-maintainer-mode and Perl < 5.10, which doesn't
have the `//=` operator. It's unlikely developers will have such an old
Perl, but the mingw environment on appveyor CI does. The use of `//=` was
introduced by changes in 1.4.10.
API:
* Xapian::PostingSource: When a PostingSource without a clone() method is used
with a Database containing multiple shards, the documented behaviour has
always been that Xapian::InvalidOperationError is thrown. However, since at
least 1.4.0, this exception hasn't been thrown, but instead a single
PostingSource object would get used for all the shards, typically leading to
incorrect results. The actual behaviour now matches what was documented.
* Xapian::Database: Add size() method which reports the number of shards.
* Xapian::Database::check(): You can now pass a stub database which will check
all the databases listed in it (or throw Xapian::UnimplementError for
backends which don't support checking).
* Xapian::Document: When updating a document use a emplace_hint() to make the
bulk insertion O(n) instead of O(n·log(n)), and use std::move() to avoid
copying OmDocumentTerm objects.
* Xapian::Query: Add missing get_unique_terms_end() method.
* Xapian::iterator_valid(): Implement for Utf8Iterator
testsuite:
* Fix keepalive1 failures on some platforms. On some platforms a timeout
gives NetworkTimeoutError and on others NetworkError - since 1.4.10 changed
to checking the exact exception type, keepalive1 has been failing on the
former set of platforms. We now just check for NetworkError or a subclass
here (since NetworkTimeoutError is a subclass of NetworkError).
* Run cursordelbug1 testcase with multi databases too.
matcher:
* Ownership of PostingSource objects during the match now makes use of the
optional reference-counting mechanism rather than a separate flag.
remote backend:
* Fix remote protocol design bug. Previously some messages didn't send a reply
but could result in an exception being sent over the link. That exception
would then get read as a response to the next message instead of its actual
response so we'd be out of step. Fixes#783, reported by Germán M. Bravo.
This fix necessitated a minor version bump in the remote protocol (to 39.1).
If you are upgrading a live system which uses the remote backend, upgrade the
servers before the clients.
* Fix socket leaks on errors during opening a database. Fixes
https://github.com/xapian/xapian/pull/237 and #781, reported by Germán M.
Bravo.
* Don't close remote DB socket on receiving EOF as the levels above won't
know it's been closed and may try to perform operations on it, which would be
problematic if that fd gets reused in the meantime. Leaving it open means
any further operations will also get EOF. Reported by Germán M. Bravo.
* We add a wrapper around the libc socket() function which deals with the
corner case where SOCK_CLOEXEC is defined but socket() fails if it is
specified (which can happen with a newer libc and older kernel).
Unfortunately, this wrapper wasn't checking the returned value from socket()
correctly, so when SOCK_CLOEXEC was specified and non-zero it would create
the socket() with SOCK_CLOEXEC, then leak that one and create it again
without SOCK_CLOEXEC. We now check the return value properly.
* Fix potential infinite loop in ValueCountMatchSpy::merge_results() if passed
serialised results with extra data appended (which shouldn't happen in normal
use).
build system:
* Current versions of valgrind result in false positives on current versions of
macOS, so on this platform configure now only enables use of valgrind if it's
specified explicitly. Fixes#713, reported by Germán M. Bravo.
* Refactor macros to probe for compiler flags so they automatically cache
their results and consistently report success/failure.
* Rename our custom TYPE_SOCKLEN_T macro to XAPIAN_TYPE_SOCKLEN_T. The
AX_TYPE_SOCKLEN_T macro defines an alias of TYPE_SOCKLEN_T for itself which
means it can get used instead in some situations, but it isn't compatible
with our macro. We can't just switch to AX_TYPE_SOCKLEN_T as it doesn't
handle cases we need, so just rename our macro to avoid potential problems.
documentation:
* Improve API documentation for Xapian::Query class. Add missing doc
comments and improve some of the existing ones. Problems highlighted by
Дилян Палаузов in #790.
* Add Unicode consortium names and codes for categories from Chapter 4, Version
11 of the Unicode standard. Patch from David Bremner.
* Improve configure --help output - drop "[default=no]" for --enable-*
options which default off. Fixes#791, reported by and patch from Дилян
Палаузов.
* Fix API documentation typo - Query::op (the type) not op_ (a parameter name).
* Note which version Document::remove_postings() was added in.
* In the remote protocol documentation, MSG_REPLACEDOCUMENTTERM was documented
as not having a reply, but actually REPLY_ADDDOCUMENT is sent.
* Update list of <xapian/iterator.h> users.
tools:
* copydatabase: A change in 1.4.6 which added support for \ as directory
separator on platforms where that's the norm broke the code in copydatabase
which removes a trailing slash from input databases. Bug reported and
culprit commit identified by Eric Wong.
portability:
* Resolve crash on Windows when using clang-cl and MSVC. Reported by Christian
Mollekopf in https://github.com/xapian/xapian/pull/256.
* Add missing '#include <cstring>'. Patch from Tanmay Sachan.
* Fix str() helper function when converting the most negative value
of a signed integer type.
* Avoid calling close() on fd we know must actually be a WIN32 SOCKET.
* Include <ios> not <iomanip> for std::boolalpha.
* Rework setenv() compatibility handling. Now that Solaris 9 is dead we can
assume setenv() is provided by Unix-like platforms (POSIX requires it). For
other platforms, provide a compatibility implementation of setenv() which
so the compatibility code is encapsulated in one place rather than replicated
at every use.
* Fix maintainer-mode builds on Solaris where /bin/tr is not POSIX compliant.
We now use the simple workaround suggested by the autoconf manual.
* Improve support for Sun C++ (see #631):
+ Suppress unhelpful warning for lambda with multiple return statements.
+ Enable reporting the tags corresponding to warnings, which we need
to know in order to suppress any new unhelpful warnings.
+ Adjust our workaround for bug with this compiler's <cmath> header to avoid
a compiler warning.
+ Use -xldscope=symbolic for Sun C++. This flag is roughly equivalent to
-Bsymbolic-functions which we've probed for with GNU ld since Xapian 1.0.0.
And from the changelog for the language bindings:
Documentation:
* Update bindings HACKING document. Reported as out of date by Niwesh Gupta.
CSharp:
* Work around mono terminfo parsing bug - older cli-sn fails with e.g.
TERM=xterm-256color due to: https://github.com/mono/mono/issues/6752
Encountered on Kubuntu 18.10 and debugged by Tejasvi Tomar. Seems to be
fixed in the mono version in Debian buster.
Perl:
* Suppress warnings from older Perl headers due to use of constructs which look
like C++11 user-defined literals. They're fixed in newer versions so they're
just noise in our build. We were working around these in the CI build, so
drop that workaround as we want to make the build warning-clean for users
too. Reported by daniel93 on #xapian.
Python3:
* Fix build for changes in Sphinx 2.0 (which drops support for sphinx.main()).
Fixes#778, reported by karolyi. Also reported by Gaurav Arora.
* We now throw UnicodeEncodeError for bad Unicode string input. Previously
cases such as a lone surrogate would be handled by quietly skipping the bad
codepoints when converting to UTF-8 to pass to Xapian.
* We no longer use the deprecated old-style Py_UNICODE API, which currently
gives deprecation warnings and is slated to be removed in Python 4.0.
Ruby:
* Add support for block iteration. All the iterator methods in the Ruby API
now accept an optional block. If no block is given an array is returned so
existing code will still work. Partly based on a patch in
https://github.com/xapian/xapian/pull/232 from Cong Ding.
* Add missing wrappers for all the C++ methods returning iterators which
weren't wrapped for Ruby. Fixes#777, reported by do.
* Suppress warnings from Ruby 2.3 headers due to use of register and
constructs which look like C++11 user-defined literals. They're fixed in
newer versions so they're just noise in our build. We were working around
these in the CI build, so drop that workaround as we want to make the build
warning-clean for users too. Reported by daniel93 on #xapian.
* smoketest.rb: Don't leave temporary databases behind in /tmp.
API:
* MSet::SNIPPET_CJK_NGRAM - new flag for use with MSet::snippet() to enable
support for selecting and highlighting snippets which works with the
QueryParser and TermGenerator FLAG_CJK_NGRAM flags. This mode can also be
enabled by setting environment variable XAPIAN_CJK_NGRAM to a non-empty
value. (There was nominally already support for XAPIAN_CJK_NGRAM in
MSet::snippet(), but it didn't work usefully - the highlighting added was all
empty start/end pairs at the end of the span of CJK characters containing the
CJK ngram terms, which to the user would typically look like it was selecting
the end of the text and not highlighting anything).
* Deprecate XAPIAN_CJK_NGRAM environment variable. There are now flags which
can be used instead in all cases, and there's sadly no portable thread-safe
way to read an environment variable so checking environment variables is
problematic in library code that may be used in multithreaded programs.
* Query::OP_ELITE_SET currently incorrectly recursively flattens any OP_OR (or
OP_OR-like) subqueries into the list of subqueries it selects from - until
that's fixed, we now select from the full exploded list rather than the last
n (where n is the number of direct subqueries of the OP_ELITE_SET).
testsuite:
* Testcases which need a generated database now get run with a sharded
database.
* Avoid using strerror() in the testsuite which removes an obstacle to running
tests in parallel in separate threads.
matcher:
* Extend detection of cases of OP_SYNONYM with disjoint subqueries (which means
we don't need document length) which was added in 1.4.8 - we now detect when
all subqueries are different terms, or when all subqueries are
non-overlapping wildcards. The second case is what QueryParser produces for
a wildcard or partial query with a query prefix which maps to more than one
term prefix.
glass backend:
* Handle an empty value slot lower bound gracefully. This shouldn't happen for
a non-empty slot, but has been reported by a notmuch user so it seems there
is (or perhaps was as the database was several years old) a way it can come
about. We now check for this situation and set the smallest possible valid
lower bound instead, so other code assuming a valid lower bound will work
correctly. Reported by jb55.
chert backend:
* Handle an empty value slot lower bound gracefully, equivalent to the change
made for glass.
documentation:
* HACKING: We no longer use auto_ptr<>.
* NEWS: Correct factual error in old entry - the 0.4.1 release was Open Muscat
not OmSee (the OmSee name was only applied after that final release was made,
and only used internally to BrightStation).
portability:
* Suppress more clang -Wself-assign-overloaded warnings in testcases which are
deliberately testing handling of self-assignment.
* Add missing includes of <cerrno>. Fixes#776, reported by Matthieu Gautier.
debug code:
* When configured with --enable-log, the O_SYNC flag was always specified when
opening the logfile, with the intention that the most recent log entries
wouldn't get lost if there was a crash, but O_SYNC can incur a significant
performance overhead and most debugging is not of such crashes. So we no
longer specify O_SYNC by default, but you can now request synchronous logging
by including %! anywhere in the filename specified with XAPIAN_DEBUG_LOG
(the %! is replaced with the empty string). We also now use O_DSYNC if
available in preference to O_SYNC, since the mtime of the log file isn't
important.
API:
* DatabaseClosedError: New exception class thrown instead of DatabaseError when
an operation is attempted which can't be completed because it involves a
database which close() was previously called on. DatabaseClosedError is a
subclass of DatabaseError so existing code catching DatabaseError will still
work as before. Fixes#772, reported by Germán M. Bravo. Patch from
Vaibhav Kansagara.
* DatabaseNotFoundError: New exception class thrown instead of
DatabaseOpeningError when the problem is the problem is "file not found" or
similar. DatabaseNotFoundError is a subclass of DatabaseOpeningError so
existing code catching DatabaseOpeningError will still work as before. Fixes
#773, reported by Germán M. Bravo. Patch from Vaibhav Kansagara.
* Query: Make &=, |= and ^= on Query objects opportunistically append to
an existing query with a matching query operator which has a reference
count of 1. This provides an easy way to incrementally build flatter query
trees.
* Query: Support `query &= ~query2` better - this now is handled exactly
equivalent to `query = query & ~query2` and gives `query AND_NOT query2`
instead of `query AND (<alldocuments> AND_NOT query2)`.
* QueryParser: Now uses &=, |= and ^= to produce flatter query trees. This
fixes problems with running out of stack space when handling Query object
trees built by abusing QueryParser to parse very large machine-generated
queries.
* Stopper: Fix incorrect accents in Hungarian stopword list. Patch from David
Corbett.
testsuite:
* Test MSet::snippet() with small and zero lengths. Fixes#759. Patch from
Vaibhav Kansagara.
* Fix testcase stubdb4 annotations - this testcase doesn't need a backend.
* Add PATH annotation for testcases needing get_database_path() to avoid having
to repeatedly list the backends where this is supported in testcase
annotations.
* TEST_EXCEPTION helper macro now checks that the exact specified exception
type is thrown. Previously it would allow a subclass of the specified
exception type, but in testcases we really want to be able to test for an
exact type. Issue noted by Vaibhav Kansagara on IRC.
matcher:
* Map OP_VALUE_GE/OP_VALUE_LE on an empty slot to EmptyPostList. We already do
this for OP_VALUE_RANGE, and it's a little more efficient than creating a
postlist object which checks the empty value slot.
glass backend:
* We no longer flush all pending positional changes when a postlist, termlist
or all-terms is opened on a modified WritableDatabase. Doing so was
incurring a significant performance cost, and the first of these happens
internally when `replace_document(term, doc)` is used, which is the usual way
to support non-numeric unique ids. We now only flush pending positional
changes when committing. Reported and diagnosed by Germán M. Bravo.
remote backend:
* Use poll() where available instead of select(). poll() is specified by
POSIX.1-2001 so should be widely available by now, and it allows watching any
fd (select() is limited to watching fds < FD_SETSIZE). For any platforms
which still lack poll() we now workaround this select() limitation when a
high numbered fd needs to be watched (for example, by trying a non-blocking
read or write and on EAGAIN sleeping for a bit before retrying).
* Stop watching fds for "exceptional conditions" - none of these are relevant
to our usage.
* Remove 0.1s timeout in ready_to_read(). The comment says this is to avoid a
busy loop, but that's out of date - the matcher first checks which remotes
are ready to read and then does a second pass to handle those which weren't
with a blocking read.
build system:
* Stop probing for header sys/errno.h which is no longer used - it was only
needed for Compaq C++, support for which was dropped in 1.4.8.
documentation:
* docs/valueranges.html: Update to document RangeProcessor instead of
ValueRangeProcessor - the latter is deprecated and will be gone in the next
release series.
* Document RangeProcessor::operator()() returns OP_INVALID to signal it doesn't
recognise a range.
* Update some URLs for pages which have moved.
* Use https for URLs where available.
* HACKING: Update "empty()" section for changes in C++11.
portability:
* Suppress clang warnings for self-assignment tests. Some testcases trigger
this new-ish clang warning while testing that self-assignment works, which
seems a useful thing to be testing - at least one of these is a regression
test.
* Add std::move to fix clang -Wreturn-std-move warning (which is enabled by
-Wall).
* Add casts to fix ubsan warnings. These cases aren't undefined behaviour, but
are reported by ubsan extra checks implicit-integer-truncation and/or
implicit-conversion which it is useful to be able to enable to catch
potential bugs.
* Fix check for when to use _byteswap_ulong() - in practice this would only
have caused a problem if a platform provided _byteswap_ushort() but not
_byteswap_ulong(), but we're not aware of any which do.
* Fix return values of do_bswap() helpers to match parameter types (previously
we always returned int and only supported swapping types up to 32 bits, so
this probably doesn't result in any behavioural changes).
* Only include <intrin.h> if we'll use it instead of always including it when
it exists. Including <intrin.h> can result in warnings about duplicate
declarations of builtin functions under mingw.
* Remove call to close()/closesocket() when the argument is always -1 (since
the change to use getaddrinfo() in 1.3.3).
API:
* Document::add_posting(): Fix bugs with the change in 1.4.8 to more
efficiently handle insertion of a batch of extra positions in ascending
order. These could lead to missing positions and corrupted encoded
positional data.
remote backend:
* Avoid hang if remote connection shutdown fails by not waiting for the
connection to close in this situation. Seems to fix occasional hangs seen on
macOS. Patch from Germán M. Bravo.
API:
* QueryParser,TermGenerator: Add new stemming mode STEM_SOME_FULL_POS.
This stores positional information for both stemmed and unstemmed terms,
allowing NEAR and ADJ to work with stemmed terms. The extra positional
information is likely to take up a significant amount of extra disk space so
the default STEM_SOME is likely to be a better choice for most users.
* Database::check(): Fetch and decompress the document data to catch problems
with the splitting of large data into multiple entries, corruption of the
compressed data, etc. Also check that empty document data isn't explicitly
stored for glass.
* Fix an incorrect type being used for term positions in the TermGenerator API.
These were Xapian::termcount but should be Xapian::termpos. Both are
typedefs for the same 32-bit unsigned integer type by default (almost always
"unsigned int") so this change is entirely compatible, except that if you
were configuring 1.4.7 or earlier with --enable-64bit-termcount you need to
also use the new --enable-64bit-termpos configure option with 1.4.8 and up or
rebuild your applications. This change was necessary to make
--enable-64bit-termpos actually useful.
* Add Document::remove_postings() method which removes all postings in a
specified term position range much more efficiently than by calling
remove_posting() repeatedly. It returns the number of postings removed.
* Fix bugs with handling term positions >= 0x80000000. Reported by Gaurav
Arora.
* Document::add_posting(): More efficiently handle insertion of a batch of
extra positions in ascending order.
* Query: Simplify OP_SYNONYM with single OP_WILDCARD subquery by converting to
OP_WILDCARD with combiner OP_SYNONYM, which means such cases can take
advantage of the new matcher optimisation in this release to avoid needing
document length for OP_WILDCARD with combiner OP_SYNONYM.
matcher:
* Avoid needing document length for an OP_WILDCARD with combiner OP_SYNONYM.
We know that we can't get any duplicate terms in the expansion of a wildcard
so the sum of the wdf from them can't possibly exceed the document length.
* OP_SYNONYM: No longer tries to initialise weights for its subquery, which
should reduce the time taken to set up a large wildcard query.
* OP_SYNONYM: Fix frequency estimates when OP_SYNONYM is used with a
subquery containing OP_XOR or OP_MAX - in such cases the frequency
estimates for the first subquery of the OP_XOR/OP_MAX were used for
all its subqueries. Also the estimated collection frequency is
now rounded to the nearest integer rather than always being rounded
down.
glass backend:
* Revert change made in 1.4.6:
Enable glass's "open_nearby_postlist" optimisation (which especially helps
large wildcard queries) for writable databases without any uncommitted
changes as well.
The amended check isn't conservative enough as there may be postlist changes
in the inverter while the table is unmodified. This breaks testcase
T150-tagging.sh in notmuch's testsuite, reported by David Bremner.
* When indexing a document without any terms we now avoid some unnecessary work
when storing its termlist.
tools:
* xapian-delve: Test for all docs empty using get_total_length() which is
slightly simpler internally than get_avlength(), and avoids an exact floating
point equality check.
examples:
* quest: Support --weight=coord.
* xapian-pos: New tool to show term position info to help debugging when using
positional information in more complex ways.
portability:
* Fix undefined behaviour from C++ ODR violation due to using the same name
two different non-static inline functions. It seems that with current GCC
versions the desired function always ends up being used, but with current
clang the other function is sometimes used, resulting in database corruption
when using value slots in docid 16384 or higher with the default glass
backend. Patch from Germán M. Bravo.
* Suppress alignment cast warning on sparc Linux. The pointer being cast is to
a record returned by getdirentries(), so it should be suitable aligned.
* Drop special handling for Compaq C++. We never actually achieved a working
build using it, and I can find no evidence that this compiler still exists,
let alone that it was updated for C++11 which we now require.
* Create new database directories in race-free way.
* Avoid throwing and handling an exception in replace_document() when
adding a document with a specified docid which is <= last_docid but currently
unused.
* Use our portable code for handling UUIDs on all platforms, and only use
platform-specific code for generating a new UUID. This fixes a bug with
converting UUIDs to and from string representation on FreeBSD, NetBSD and
OpenBSD on little-endian platforms which resulted in reversed byte order in
the first three components, so the same database would report a different
UUID on these platforms compared to other platforms. With this fix, the
UUIDs of existing databases will appear to change on these platforms
(except in rare "palindronic" cases). Reported by Germán M. Bravo.
* Fix to build with a C++17 compiler. Previously we used a "byte" type
internally which clashed with "std::byte" in source files which use
"using namespace std;". Fixes#768, reported by Laurent Stacul.
* Adjust apitest testcase stubdb2 to allow for NetBSD oddity: NetBSD's
getaddrinfo() in IPv4 mode seems to resolve ::1 to an IPv4 address on the
local network.
* Avoid timer_create() on OpenBSD and NetBSD. On OpenBSD it always fails with
ENOSYS (and there's no prototype in the libc headers), while on NetBSD it
seems to work, but the timer never seems to fire, so it's useless to us (see
#770).
* Use SOCK_NONBLOCK if available to avoid a call to fcntl(). It's supported by
at least Linux, FreeBSD, NetBSD and OpenBSD.
* Use O_NOINHERIT for O_CLOEXEC on Windows. This flag has essentially the same
effect, and it's common in other codebases to do this.
* On AIX O_CLOEXEC may be a 64-bit constant which won't fit in an int. To
workaround this stupidity we now call the non-standard open64x() instead
of open() when the flags don't fit in an int.
* Add functions to add/multiply with overflow check. These are implemented
with compiler builtins or equivalent where possible, so the overflow check
will typically just require a check of the processor's overflow or carry
flag.
API:
* Database::check(): Fix bogus error reports for documents with length zero
due to a new check added in 1.4.6 that the doclength was between the stored
upper and lower bounds, which failed to allow for the lower bound ignoring
documents with length zero (since documents indexed only by boolean terms
aren't involved in weighted searches). Reported by David Bremner.
* Query: Use of Query::MatchAll in multithreaded code causes problems because
the reference counting gets messed up by concurrent updates. Document that
Query(string()) should be used instead of MatchAll in multithreaded code, and
avoid using it in library code. Reported by Germán M. Bravo.
* Stem:
+ Stemming algorithms added for Irish, Lithuanian, Nepali and Tamil.
+ Merge Snowball compiler changes which improve code generation.
+ Merge optimisations to the Arabic and Turkish stemmers.
glass backend:
* A long-lived cursor on a table in a WritableDatabase could get into
an invalid state, which typically resulted in a DatabaseCorruptError
being thrown with the message:
Db block overwritten - are there multiple writers?
But in fact the on-disk database is not corrupted - it's just that
the cursor in memory has got into an inconsistent state. It looks
like we'll always detect the inconsistency before it can cause on-disk
corruption but it's hard to be completely certain.
The bug is in code to rebuild the cursor when the underlying table
changes in ways which require that, which is a fairly rare occurrence
to start with, and only triggers when a block in the cursor has been
released, reallocated, and we tried to load it in the cursor at the
same level - the cursor wrongly assumes it has the current version
of the block.
Reported with a reproducer by Sylvain Taverne. Confirmed by David
Bremner as also fixing a problem in notmuch for which he hadn't managed
to find a reduced reproducer.
API:
* API classes now support C++11 move semantics when using a compiler which
we are confident supports them (currently compilers which define
__cplusplus >= 201103 plus a special check for MSVC 2015 or later).
C++11 move semantics provide a clean and efficient way for threaded code to
hand-off Xapian objects to worker threads, but in this case it's very
unhelpful for availability of these semantics to vary by compiler as it
quietly leads to a build with non-threadsafe behaviour. To address this,
user code can #define XAPIAN_MOVE_SEMANTICS before #include <xapian.h> to
force this on, and will then get a compilation failure if the compiler lacks
suitable support.
* MSet::snippet():
+ We were only escaping output for HTML/XML in some cases, which would
potentially allow HTML to be injected into output (this has been assigned
CVE-2018-0499).
+ Include certain leading non-word characters in snippets. Previously we
started the snippet at the start of the first actual word, but there are
various cases where including non-word characters in front of the actual
word adds useful context or otherwise aids comprehension. Reported by
Robert Stepanek in https://github.com/xapian/xapian/pull/180
* Add MSetIterator::get_sort_key() method. The sort key has always been
available internally, but wasn't exposed via the public API before, which
seems like an oversight as the collapse key has long been available.
Reported by 张少华 on xapian-discuss.
* Database::compact():
+ Allow Compactor::resolve_duplicate_metadata() implementations to delete
entries. Previously if an implementation returned an empty string this
would result in a user meta-data entry with an empty value, which isn't
normally achievable (empty meta-data values aren't stored), and so will
cause odd behaviour. We now handle an empty returned value by interpreting
it in the natural way - it means that the merged result is to not set a
value for that key in the output database.
+ Since 1.3.5 compacting a WritableDatabase with uncommitted changes throws
Xapian::InvalidOperationError when compacting to a single-file glass
database. This release adds similar checks for chert and when compacting
to a multiple-file glass database.
+ In the unlikely event that the total number of documents or the total
length of all documents overflow when trying to compact a multi-database,
we throw an exception. This is now a DatabaseError exception instead of a
const char* exception (a hang-over from before this code was turned into a
public API in the library).
* Document::remove_term(): Handle removing term at current TermIterator
position - previously the underlying iterator was invalidated, leading to
undefined behaviour (typically a segmentation fault). Reported by Gaurav
Arora.
* TermIterator::get_termfreq() now always returns an exact answer. Previously
for multi-databases we approximated the result, which is probably either a
hang-over from when this method was used during Enquire::get_eset(), or else
due to a thinking that this method would be used in that situation (it
certainly is not now). If the user creates a TermIterator object and asks it
for term frequencies then we really should give them the correct answer - it
isn't hugely costly and the documentation doesn't warn that it might be
approximated.
* QueryParser::parse_query():
+ Now adds a colon after the prefix when prefixing a boolean term which
starts with a colon. This means the mapping is reversible, and matches
what omega actually does in this case when it tries to reverse the mapping.
Thanks to Andy Chilton for pointing out this corner case.
+ The parser now makes use of newer features in the lemon parser generator to
make parsing faster and use less memory.
* Enquire::get_mset(): Fix bug with get_mset(0, 0, X) when X > 0 which was
causing an attempt to access an element in an empty vector. Reported by
sielicki in #xapian.
* Stem:
+ Add Indonesian stemming algorithm.
+ Small optimisations to almost all stemming algorithms.
* Stopper:
+ Add Indonesian stopword list.
+ The installed version of the Finnish stopword list now has one word per
line. Previously it had several space-separated words on some lines, which
works with C++'s std::istream_iterator but may be inconvenient for use from
some other languages.
+ The installed versions of stopword lists are now sorted in byte order
rather than whatever collation order is specified by LC_COLLATE or similar
at build time. This makes the build more reproducible, and also may be
more efficient for loading into some data structures.
* WritableDatabase::replace_document(term, doc): Check for last_docid wrapping
when used on a sharded database.
* Database::locked(): Consistently throw FeatureUnavailableError on platforms
where we can't test for a database lock without trying to take it.
Previously GNU Hurd threw DatabaseLockError while platforms where we don't
use fcntl() locking at all threw UnimplementedError.
* Database and WritableDatabase constructors: Fix handling of entries for
disabled backends in stub database files to throw FeatureUnavailableError
instead of DatabaseError.
* Database::get_value_lower_bound() now works correctly for sharded databases.
Previously it returned the empty string if any shard had no values in the
specified slot.
* PostingIterator was failing to keep an internal reference to the parent
Database object for sharded databases.
* ValueIterator::skip_to() and check() had an off-by-one error in their docid
calculations in some cases with sharded databases.
* Add Database::get_total_length() method. Previously you had to calculate
this from get_avlength() and get_doccount(), taking into account rounding
issues. But even then you couldn't reliably get the exact value when total
length is large since a double's mantissa has more limited precision than an
unsigned long long.
* Add Xapian::iterator_rewound() for bidirectional iterators, to test if the
iterator is at the start (useful for testing whether we're done when
iterating backwards).
* DatabaseOpeningError exceptions now provide errno via get_error_string()
rather than turning it into a string and including it in the exception
message.
* WritableDatabase::replace_document(): when passed a Document object which
came from a database and has unmodified values, we used to always read
those values into a memory structure. Now we only do this if the document
is being replaced to the same document ID which it came from, which should
make other cases a bit more efficient.
* Enquire::get_eset(): When approximating term frequencies we now round to the
nearest integer - previously we always rounded down.
matcher:
* OP_VALUE_*: When a value slot's lower and upper bound are equal, we know
that exactly how many documents the subquery can match (either 0 or those
bounds). This also avoids a division by zero which previously happened
when trying to calculate the estimate.
* Speed up sorting by keys. Use string::compare() to avoid having to call
operator< if operator> returns false.
* Fix clamping of maxitems argument to get_mset() - it was being clamped
to db.get_doccount(), now it's clamped to db.get_doccount() - first. In
practice this doesn't actually seem to cause any issues.
* If a match time limit is in effect, when it expires we now clamp
check_at_least to first + maxitems instead of to maxitems. In practice this
also doesn't seem to actually cause any issues (at least we've failed to
construct a testcase where it actually makes an observable difference).
* Fix percentages when only some shards have positions. If the final shard
didn't have positions this would lead to under-counting the total number leaf
of subqueries which would lead to incorrect positional calculations (and a
division by zero if the top level of the query was positional. This bug was
introduced in 1.4.3.
* OP_NEAR: Fix "phantom positions", where OP_NEAR would think a term without
positional information occurred at position 1 if it had the lowest term
frequency amongst the OP_NEAR's subqueries.
* Fix termfreq used in weight calculations for a term occurring more than once
in the query. Previously the termfreq for such terms was multiplied by the
number of different query positions they appeared at.
* OP_SYNONYM: We use the doclength upper bound for the wdf upper bound of a
synonym - now we avoid fetching it twice when the doclength upper bound is
explicitly needed.
* Short-cut init() when factor is 0 in most Weight subclasses. This indicates
the object is for the term-independent weight contribution, which is always 0
for most schemes, so there's no point fetching any stats or doing any
calculations. This fixes a divide by zero for TfIdfWeight, detected by
UBSan.
* OP_OR: Fix bug which caused orcheck1 to fail once hooked up to run with the
inmemory backend.
* Iterating of positions has been sped up, which means phrase matching is now
faster (by a little over 5% in some simple tests).
* Fix use after free of QueryOptimiser hint in certain cases involving
multiple databases only some of which have positional information.
This bug was introduced by changes in xapian-core 1.4.3. Fixes#752,
reported and analysed by Robert Stepanek.
* An unweighted OP_AND_MAYBE is now optimised to just its left branch - the
other branch or branches only contribute weight, so can be completely ignored
when the operator is unweighted.
glass backend:
* Fix glass freelist bug when changes to a new database which didn't modify the
termlist table were committed. In this corner case, a block which had been
allocated to be the root block in the termlist table was leaked. This was
largely harmless, except that it was detected by Database::check() and caused
it to report an error. Reported by Antoine Beaupré and David Bremner.
* Fix glass freelist bug with cancel_transaction(). The freelist wasn't
reset to how it was before the transaction, resulting in leaked blocks.
This was largely harmless, except that it was detected by Database::check()
and caused it to report an error.
* Improve the per-term wdf upper bound. Previously we used min(cf(term),
wdf_upper_bound(db)) which is tight for any terms which attain that
upper bound, and also for terms with termfreq == 1 (the latter are common
in the database (e.g. 66% for a database of wikipedia), but probably
much less common in searches). When termfreq > 1 we now use
max(first_wdf(term), cf(term) - first_wdf(term)), which means terms with
termfreq == 2 will also attain their bound (another 11% for the same
database) while terms with higher termfreq but below the global bound will
get a tighter bound.
* Fix Database::locked() on single-file glass db to just return false (such
databases can't be opened as a WritableDatabase so there can't be a write
lock). Previously this failed with: "DatabaseLockError: Unable to get write
lock on /flintlock: Testing lock"
* Fix compaction when both the input and output are specified as a file
descriptor. Previously this threw an exception due to an overeager check
that destination != source.
* Use O_TRUNC when compacting to single file. If the output already exists but
is larger than our output we don't want to just overwrite the start of it.
This case also used to result in confusing compaction percentages.
* Enable glass's "open_nearby_postlist" optimisation (which especially helps
large wildcard queries) for writable databases without any uncommitted
changes as well.
* Make get_unique_terms() more efficient for glass. We approximate
get_unique_terms() by the length of the termlist (which counts boolean terms
too) but clamp this to be no larger than the document length. Since we need
to open the termlist to get its length, it makes more sense to get the
document length from that termlist for no extra cost rather than looking it
up in the postlist table.
* Database::check() now checks document lengths against the stored document
length lower and upper bounds. Patch from Uppinder Chugh. Fixes
https://trac.xapian.org/ticket/617.
* Fix bogus handling of most-recently-read value slot statistics. It seems
that we get lucky and this can't actually cause a problem in practice due
to another layer of caching above, but if nothing else it's a bug waiting to
happen.
* If we fail to create the directory for a new database because the path
already exists, the exception now reports EEXIST as the errno value rather
than whatever errno value happened to be set from an earlier library call.
remote backend:
* xapian-tcpsrv --one-shot no longer forks. We need fork to handle multiple
concurrent connections, but when handling a single connection forking just
adds overhead and potentially complicates process management for our caller.
This aligns with the behaviour under __WIN32__ where we use threads instead
of forking, and service the connection from the main thread with --one-shot.
* Fix repeat call to ValueIterator::check() on the same docid to not always
set valid to true for remote backend.
inmemory backend:
* Fix repeat call to ValueIterator::check() on the same docid to not always
set valid to true for inmemory backend.
* Use binary chop instead of linear search in all places where we're searching
for a term or document - we weren't taking advantage of the sorted order
everywhere.
tools:
* xapian-delve:
+ Document values can contain binary data, so escape them by default for
output. Other options now supported are to decode as a packed integer
(like omindex uses for last modified), decode using
Xapian::sortable_unserialise(), and to show the raw form (which was the
previous behaviour).
+ Report current database revision.
* xapian-inspect:
+ Report entry count when opening table
+ Support inspecting single file DBs via a new --table option (which can also
be used with a non-single-file DB instead of specifying the path to the
table).
+ Add "first" and "last" commands which jump to the first/last entry in the
current table respectively.
+ "until" now counts and reports the number of entries advanced by.
+ Document "until" with no arguments - this advances to the end of the table,
but wasn't mentioned in the help.
+ Commands "goto" and "until" which take a key as an argument now expect the
key in the same escaped form that's used for display. This makes it much
simpler to interact with tables with binary keys.
+ Fix to expect .glass not .DB extension of glass tables.
API:
* Database::check():
+ Fix checking a single table - changes in 1.4.2 broke such checks unless you
specified the table without any extension.
+ Errors from failing to find the file specified are now thrown as
DatabaseOpeningError (was DatabaseError, of which DatabaseOpeningError is
a subclass so existing code should continue to work). Also improved the
error message when the file doesn't exist is better.
* Drop OP_SCALE_WEIGHT over OP_VALUE_RANGE, OP_VALUE_GE and OP_VALUE_LE in the
Query constructor. These operators always return weight 0 so OP_SCALE_WEIGHT
over them has no effect. Eliminating it at query construction time is cheap
(we only need to check the type of the subquery), eliminates the confusing
"0 * " from the query description, and means the OP_SCALE_WEIGHT Query object
can be released sooner. Inspired by Shivanshu Chauhan asking about the query
description on IRC.
* Drop OP_SCALE_WEIGHT on the right side of OP_AND_NOT in the Query
constructor. OP_AND_NOT takes no weight from the right so OP_SCALE_WEIGHT
has no effect there. Eliminating it at query construction time is cheap
(just need to check the subquery's type), eliminates the confusing "0 * "
from the query description, and means the OP_SCALE_WEIGHT object can be
released sooner.
* MSet::snippet(): Favour candidate snippets which contain more of a diversity
of matching terms by discounting the relevance of repeated terms using an
exponential decay. A snippet which contains more terms from the query is
likely to be better than one which contains the same term or terms multiple
times, but a repeated term is still interesting, just less with each
additional appearance. Diversity issue highlighted by Robert Stepanek's
patch in https://github.com/xapian/xapian/pull/117 - testcases taken from his
patch.
* MSet::snippet(): New flag SNIPPET_EMPTY_WITHOUT_MATCH to get an empty snippet
if there are no matches in the text passed in. Implemented by Robert
Stepanek.
* Round MSet::get_matches_estimated() to an appropriate number of significant
figures. The algorithm used looks at the lower and upper bound and where the
estimate sits between them, and then picks an appropriate number of
significant figures. Thanks to Sébastien Le Callonnec for help sorting out a
portability issue on OS X.
* Add Database::locked() method - where possible this non-invasively checks if
the database is currently open for writing, which can be useful for
dashboards and other status reporting tools.
testsuite:
* Add more tests of Database::check(). Fixes#238, reported by Richard
Boulton.
* Make apitest testcase nosuchdb1 fail if we manage to open the DB.
* Skip testcases which throw NetworkError with errno value ECHILD - this
indicates system resource starvation rather than a Xapian bug. Such failures
are seen on Debian buildds from time to time, see:
https://bugs.debian.org/681941
* Use terms that exist in the database for most snippet tests. It's good to
test that snippet highlighting works for terms that aren't in the database,
but it's not good for all our snippet tests to feature such terms - it's
not the common usage.
matcher:
* Fix incorrect results due to uninitialised memory. The array holding max
weight values in MultiAndPostList is never initialised if the operator is
unweighted, but the values are still used to calculate the max weight to pass
to subqueries, leading to incorrect results. This can be observed with an OR
under an unweighted AND (e.g. OR under AND on the right side of AND_NOT).
The fix applied is to simply default initialise this array, which should lead
to a max weight of 0.0 being passed on to subqueries. Bug reported in
notmuch by Kirill A. Shutemov, and forwarded by David Bremner.
* Improve value range upper bound and estimated matches. The value slot
frequency provides a tighter upper bound than Database::get_doccount().
The estimate is now calculated by working out the proportion of possible
values between the slot lower and upper bounds which the range covers
(assuming a uniform distribution). This seems to work fairly well in
practice, and is certainly better than the crude estimate we were using:
Database::get_doccount() / 2
* Handle arbitrary combinations of OP_OR under OP_NEAR/OP_PHRASE, partly
addressing #508. Thanks to Jean-Francois Dockes for motivation and testing.
* Only convert OP_PHRASE to OP_AND if full DB has no positions. Until now the
conversion was done independently for each sub-database, but being consistent
with the results from a database containing all the same documents seems more
useful.
* Avoid double get_wdf() call for first subquery of OP_NEAR and OP_PHRASE,
which will speed them up by a small amount.
documentation:
* Correct "Query::feature_flag" -> "QueryParser::feature_flag". Fixes#747,
reported by James Aylett.
* Rename set_metadata() `value` parameter to `metadata`. This change is
particularly motivated by making it easier to map this case specially in SWIG
bindings, but the new name is also clearer and better documents its purpose.
* Rename value range parameters. The new names (`range_limit` instead of
`limit`, `range_lower` instead of `begin` and `range_upper` instead of `end`)
are particularly motivated by making it easier to map them specially in SWIG
bindings, but they're also clearer names which better document their
purposes.
* Change "(key, tag)" to "(key, value)" in user metadata docs. The user
metadata is essentially what's often called a "key-value store" so users
are likely to be familiar with that terminology.
* Consistently name parameter of Weight::unserialise() overridden forms.
In xapian/weight.h it was almost always named `serialised`, but LMWeight
named it `s` and CoordWeight omitted the name.
* Fix various minor documentation comment typos.
* INSTALL: Update section about -Bsymbolic-functions which is not a new
GNU ld feature at this point.
tools:
* xapian-delve: Uses new Database::locked() method to report if the database
is currently locked.
portability:
* Fix configure probe for __builtin_exp10() to work around bug on mingw - there
GCC generates a call to exp10() for __builtin_exp10() but there is no exp10()
function in the C library, so we get a link failure. Use a full link test
instead to avoid this issue. Reported by Mario Emmenlauer on xapian-devel.
* Fix configure probe for log2() which was failing on at least some platforms
due to ambiguity between overloaded forms of log2(). Make the probe
explicitly check for log2(double) to avoid this problem.
* Workaround the unhelpful semantics of AI_ADDRCONFIG on platforms which follow
the old RFC instead of POSIX (such as Linux) - if only loopback networking is
configured, localhost won't resolve by name or IP address, which causes
testsuites using the remote backend over localhost to fail in auto-build
environments which deliberately disable networking during builds. The
workaround implemented is to check if the hostname is "::1", "127.0.0.1" or
"localhost" and disable AI_ADDRCONFIG for these. This doesn't catch all
possible ways to specify localhost, but should catch all the ways these might
be specified in a testsuite. Fixes https://bugs.debian.org/853107, reported
by Daniel Schepler and the root cause uncovered by James Clarke.
* Fix build failure cross-compiling for android due to not pulling in header
for errno.
* Fix compiler warnings.
debug code:
* Adjust assertion in InMemoryPostList. Calling skip_to() is fine when the
postlist hasn't been started yet (but the assertion was failing for a term
not in the database). Latent bug, triggered by testcases complexphrase1 and
complexnear1 as updated for addition of support for OP_OR subqueries of
OP_PHRASE/OP_NEAR.
API:
* Add XAPIAN_AT_LEAST(A,B,C) macro.
* MSet::snippet(): Optimise snippet generation - it's now ~46% faster in a
simple test.
* Add Xapian::DOC_ASSUME_VALID flag which tells Database::get_document() that
it doesn't need to check that the passed docid is valid. Fixes#739,
reported by Germán M. Bravo.
* TfIdfWeight: Add support for the L wdf normalisation. Patch from Vivek Pal.
* BB2Weight: Fix weights when database has just one document. Our existing
attempt to clamp N to be at least 2 was ineffective due to computing
N - 2 < 0 in an unsigned type.
* DPHWeight: Fix reversed sign in quadratic formula, making the upper bound a
tiny amount higher.
* DLHWeight: Correct upper bound which was a bit too low, due to flawed logic
in its derivation. The new bound is slightly less tight (by a few percent).
* DLHWeight,DPHWeight: Avoid calculating log(0) when wdf is equal to the
document length.
* TermGenerator: Handle stemmer returning empty string - the Arabic stemmer
can currently do this (e.g. for a single tatweel) and user stemmers can too.
Fixes#741, reported by Emmanuel Engelhart.
* Database::check(): Fix check that the first docid in each doclength chunk is
more than the last docid in the previous chunk - this code was in the wrong
place so didn't actually work.
* Database::get_unique_terms(): Clamp returned value to be <= document length.
Ideally get_unique_terms() ought to only count terms with wdf > 0, but that's
expensive to calculate on demand.
glass backend:
* When compacting we now only write the iamglass file out once, and we write it
before we sync the tables but sync it after, which is more I/O friendly.
* Database::check(): Fix in SEGV when out == NULL and opts != 0.
* Fix potential SEGV with corrupt value stats.
chert backend:
* Fix potential SEGV with corrupt value stats.
build system:
* Add XO_REQUIRE autoconf macro to provide an easy way to handle version checks
in user configure scripts.
tools:
* quest: Support BM25+, LM and PL2+ weighting schemes.
* xapian-check: Fix when ellipses are shown in 't' mode. They were being shown
when there were exactly 6 entries, but we only start omitting entries when
there are *more* than 6. Fix applies to both glass and chert.
portability:
* Avoid using opendir()/readdir() in our closefrom() implementation as these
functions can call malloc(), which isn't safe to do between fork() and exec()
in a multi-threaded program, but after fork() is exactly where we want to
use closefrom(). Instead we now use getdirentries() on Linux and
getdirentriesattr() on OS X (OS X support bugs shaken out with help from
Germán M. Bravo).
* Support reading UUIDs from /proc/sys/kernel/random/uuid which is especially
useful when building for Android, as it avoids having to cross-build a UUID
library.
* Disable volatile workaround for excess precision SEGV for SSE - previously it
was only being disabled for SSE2.
* When building for x86 using a compiler where we don't know how to disable
use of 387 FP instructions, we now run remote servers for the testsuite under
valgrind --tool=none, like we do when --disable-sse is explicitly specified.
* Add alignment_cast<T> which has the same effect as reinterpret_cast<T> but
avoids warnings about alignment issues.
* Suppress warnings about unused private members. DLHWeight and DPHWeight
have an unused lower_bound member, which clang warns about, but we need to
keep them there in 1.4.x to preserve ABI compatibility.
* Remove workaround for g++ 2.95 bug as we require at least 4.7 now.
* configure: Probe for <cxxabi.h>. GCC added this header in GCC 3.1, which
is much older than we support, so we've just assumed it was available if
__GNUC__ was defined. However, clang lies and defines __GNUC__ yet doesn't
seem to reliably provide <cxxabi.h>, so we need to probe for it.
* Fix "unused assignment" warning.
* configure: Probe for __builtin_* functions. Previously we just checked for
__GNUC__ being defined, but it's cleaner to probe for them properly -
compilers other than GCC and those that pretend to be GCC might provide these
too.
* Use __builtin_clz() with compilers which support it to speed up encoding
and especially decoding of positional data. This speed up phrase searching
by ~0.5% in a simple test.
* Check signed right shift behaviour at compile time - we can use a test on a
constant expression which should optimise away to just the required version
of the code, which means that on platforms which perform sign-extension
(pretty much everything current it seems) we don't have to rely on the
compiler optimising a portable idiom down to the appropriate right shift
instruction.
* Improve configure check for log2(). We include <cmath> so the check really
should succeed if only std::log2() is declared.
* Enable win32-dll option to LT_INIT.
debug code:
* xapian-inspect:
+ Support glass instead of chert.
+ Allow control of showing keys/tags.
+ Use more mnemonic letters than X for command arguments in help.
API:
* Constructing a Query for a non-reference counted PostingSource object will
now try to clone the PostingSource object (as happened in 1.3.4 and
earlier). This clone code was removed as part of the changes in 1.3.5 to
support optional reference counting of PostingSource objects, but that breaks
the case when the PostingSource object is on the stack and goes out of scope
before the Query object is used. Issue reported by Till Schäfer and analysed
by Daniel Vrátil in a bug report against Akonadi:
https://bugs.kde.org/show_bug.cgi?id=363741
* Add BM25PlusWeight class implementing the BM25+ weighting scheme, implemented
by Vivek Pal (https://github.com/xapian/xapian/pull/104).
* Add PL2PlusWeight class implementing the PL2+ weighting scheme, implemented
by Vivek Pal (https://github.com/xapian/xapian/pull/108).
* LMWeight: Implement Dir+ weighting scheme as DIRICHLET_PLUS_SMOOTHING.
Patch from Vivek Pal.
* Add CoordWeight class implementing coordinate matching. This can be useful
for specialised uses - e.g. to implement sorting by the number of matching
filters.
* DLHWeight,DPHWeight,PL2Weight: With these weighting schemes, the formulae
can give a negative weight contribution for a term in extreme cases. We
used to try to handle this by calculating a per-term lower bound on the
contribution and subtracting this from the contribution, but this idea
is fundamentally flawed as the total offset it adds to a document depends on
what combination of terms that document matches, meaning in general the
offset isn't the same for every matching document. So instead we now clamp
each term's weight contribution to be >= 0.
* TfIdfWeight: Always scale term weight by wqf - this seems the logical
approach as it matches the weighting we'd get if we weighted every non-unique
term in the query, as well as being explicit in the Piv+ formula.
* Fix OP_SCALE_WEIGHT to work with all weighting schemes - previously it was
ignored when using PL2Weight and LMWeight.
* PL2Weight: Greatly improve upper bound on weight:
+ Split the weight equation into two parts and maximise each separately as
that gives an easily solvable problem, and in common cases the maximum is
at the same value of wdfn for both parts. In a simple test, the upper
bounds are now just over double the highest weight actually achieved -
previously they were several hundred times. This approach was suggested by
Aarsh Shah in: https://github.com/xapian/xapian/pull/48
+ Improve upper bound on normalised wdf (wdfn) - when wdf_upper_bound >
doclength_lower_bound, we get a tighter bound by evaluating at
wdf=wdf_upper_bound. In a simple test, this reduces the upper bound on
wdfn by 36-64%, and the upper bound on the weight by 9-33%.
* PL2Weight: Fix calculation of upper_bound when P2>0. P2 is typically
negative, but for a very common term it can be positive and then we should
use wdfn_lower not wdfn_upper to adjust P_max.
* Weight::unserialise(): Check serialised form is empty when unserialising
parameter-free schemes BoolWeight, DLHWeight and DPHWeight.
* TermGenerator::set_stopper_strategy(): New method to control how the Stopper
object is used. Patch from Arnav Jain.
* QueryParser: Fix handling of CJK query over multiple prefixes. Previously
all the n-gram terms were AND-ed together - now we AND together for each
prefix, then OR the results. Fixes#719, reported by Aaron Li.
* Add Database::get_revision() method which provides access to the database
revision number for chert and glass, intended for use by xapiand. Marked
as experimental, so we don't have to go through the usual deprecation cycle
if this proves not to be the approach we want to take. Fixes#709,
reported by German M. Bravo.
* Mark RangeProcessor constructor as `explicit`.
* Update to Unicode 9.0.0.
* Reimplement ESet and ESetIterator as we did for MSet and MSetIterator in
1.3.5. ESetIterator internally now counts down to the end of the ESet, so
the end test is now against 0, rather than against eset.size(). And more of
the trivial methods are now inlined, which reduces the number of relocations
needed to load the library, and should give faster code which is a very
similar size to before.
* MSetIterator and ESetIterator are now STL-compatible random_access_iterators
(previously they were only bidirectional_iterators).
* TfIdfWeight: Support freq and squared IDF normalisations. Patch from Vivek
Pal.
* New Xapian::Query::OP_INVALID to provide an "invalid" query object.
* Reject OP_NEAR/OP_PHRASE with non-leaf subqueries early to avoid a
potential segmentation fault if the non-leaf subquery decayed at
just the wrong moment. See #508.
* Reduce positional queries with a MatchAll or PostingSource subquery to
MatchNothing (since these subqueries have no positional information, so
the query can't match).
* Deprecate ValueRangeProcessor and introduce new RangeProcessor class as
a replacement. RangeProcessor()::operator()() method returns Xapian::Query,
so a range can expand to any query. OP_INVALID is used to signal that
a range is not recognised. Fixes#663.
* Combining of ranges over the same quantity with OP_OR is now handled by
an explicit "grouping" parameter, with a sensible default which works
for value range queries. Boolean term prefixes and FieldProcessor now
support "grouping" too, so ranges and other filters can now be grouped
together.
* Formally deprecate WritableDatabase::flush(). The replacement commit()
method was added in 1.1.0, so code can be switched to use this and still
work with 1.2.x.
* Fix handling of a self-initialised PIMPL object (e.g. Xapian::Query q(q);).
Previously the uninitialised pointer was copied to itself, resulting in
undefined behaviour when the object was used to destroyed. This isn't
something you'd see in normal code, but it's a cheap check which can probably
be optimised away by the compiler (GCC 6 does).
* The Snipper class has been replaced with a new MSet::snippet() method.
The implementation has also been redone - the existing implementation was
slower than ideal, and didn't directly consider the query so would sometimes
selects a snippet which doesn't contain any of the query terms (which users
quite reasonably found surprising). The new implementation is faster, will
always prefer snippets containing query terms, and also understands exact
phrases and wildcards. Fixes#211.
* Add optional reference counting support for ErrorHandler, ExpandDecider,
KeyMaker, PostingSource, Stopper and TermGenerator. Fixes#186, reported
by Richard Boulton. (ErrorHandler's reference counting isn't actually used
anywhere in xapian-core currently, but means we can hook it up in 1.4.x if
ticket #3 gets addressed).
* Deprecate public member variables of PostingSource. The new getters and/or
setters added in 1.2.23 and 1.3.5 are preferred. Fixes#499, reported by
Joost Cassee.
* Reimplement MSet and MSetIterator. MSetIterator internally now counts down
to the end of the MSet, so the end test is now against 0, rather than against
mset.size(). And more of the trivial methods are now inlined, which reduces
the number of relocations needed to load the library, and should give faster
code which is a very similar size to before.
* Only issue prefetch hints for documents if MSet::fetch() is called. It's not
useful to send the prefetch hint right before the actual read, which was
happening since the implementation of prefetch hints in 1.3.4. Fixes#671,
reported by Will Greenberg.
* Fix OP_ELITE_SET selection in multi-database case - we were selecting
different sets for each subdatabase, but removing the special case check for
termfreq_max == 0 solves that.
* Remove "experimental" marker from FieldProcessor, since we're happy with the
API as-is. Reported by David Bremner on xapian-discuss.
* Remove "experimental" marker from Database::check(). We've not had any
negative feedback on the current API.
* Databse::check() now checks that doccount <= last_docid.
* Database::compact() on a WritableDatabase with uncommitted changes could
produce a corrupted output. We now throw Xapian::InvalidOperationError in
this case, with a message suggesting you either commit() or open the database
from disk to compact from. Reported by Will Greenberg on #xapian-discuss
* Add Arabic stemmer. Patch from Assem Chelli in
https://github.com/xapian/xapian/pull/45
* Improve the Arabic stopword list. Patch from Assem Chelli.
* Make functions defined in xapian/iterator.h 'inline'.
* Don't force the user to specify the metric in the geospatial API -
GreatCircleMetric is probably what most users will want, so a sensible
default.
* Xapian::DBCHECK_SHOW_BITMAP: This was added in 1.3.0 (so has never been in
a stable release) and was superseded by Xapian::DBCHECK_SHOW_FREELIST in
1.3.2, so just remove it.
* Make setting an ErrorHandler a no-op - this feature is deprecated and we're
not aware of anyone using it. We're hoping to rework ErrorHandler in 1.4.x,
which will be simpler without having to support the current behaviour as well
as the new. See #3.
* Update to Unicode 8.0.0. Fixes#680.
* Overhaul database compaction API. Add a Xapian::Database::compact() method,
with the Database object specifying the source database(s).
Xapian::Compactor is now just a functor to use if you want to control
progress reporting and/or the merging of user metadata. The existing API
has been reimplemented using the new one, but is marked as deprecated.
* Add support for a default value when sorting. Fixes#452, patch from
Richard Boulton.
* Make all functor objects non-copyable. Previously some were, some weren't,
but it's hard to correctly make use of this ability. Fixes#681.
* Fix use after free with WILDCARD_LIMIT_MOST_FREQUENT. If we tried to open a
postlist after processing such a wildcard, the postlist hint could be
pointing to a PostList object which had been deleted. Fixes#696, reported
by coventry.
* Add support for optional reference counting of MatchSpy objects.
* Improve Document::get_description() - the output is now always valid UTF-8,
doesn't contain implementation details like "Document::Internal", and more
clearly reports if the document is linked to a database.
* Remove XAPIAN_CONST_FUNCTION marker from sortable_serialise_() helper, as it
writes to the passed in buffer, so it isn't const or pure. Fixes
decvalwtsource2 testcase failure when compiled with clang.
* Make PostingSource::set_maxweight() public - it's hard to wrap for the
bindings as a protected method. Fixes#498, reported by Richard Boulton.
* Database:
+ Add new flag Xapian::DB_RETRY_LOCK which allows opening a database for
writing to wait until it can get a write lock. (Fixes#275, reported by
Richard Boulton).
+ Fix Database::get_doclength_lower_bound() over multiple databases when some
are empty or consist only of zero-length documents. Previously this would
report a lower bound of zero, now it reports the same lowest bound as a
single database containing all the same documents.
+ Database::check(): When checking a single table, handle the ".glass"
extension on glass database tables, and use the extension to guide the
decision of which backend the table is from.
* Query:
+ Add new OP_WILDCARD query operator, which expands wildcards lazily, so now
we create the PostList tree for a wildcard directly, rather than creating
an intermediate Query tree. OP_WILDCARD offers a choice of ways to limit
wildcard expansion (no limit, throw an exception, use the first N by term
name, or use the most frequent N). (See tickets #48 and #608).
* QueryParser:
+ Add new set_max_expansion() method which provides access to OP_WILDCARD's
choice of ways to limit expansion and can set limits for partial terms as
well as for wildcards. Partial terms now default to the 100 most frequent
matching terms. (Completes #608, reported by boomboo).
+ Deprecate set_max_wildcard_expansion() in favour of set_max_expansion().
* Add support for optional reference counting of FieldProcessor and
ValueRangeProcessor objects.
* Update Unicode character database to Unicode 7.0.0.
* New Xapian::Snipper class from Mihai Bivol's GSOC 2012 project. (mostly
fixes#211)
* Fix all get_description() methods to always return UTF-8 text. (fixes#620)
* Database::check():
+ Alter to take its "out" parameter as a pointer to std::ostream instead of a
reference, and make passing NULL mean "do not produce output", and make
the second and third parameters optional, defaulting to a quiet check.
+ Escape invalid UTF-8 data in keys and tags reported by xapian-check, using
the same code we use to clean up strings returned by get_description()
methods.
+ Correct failure message which talks above the root block when it's actually
testing a leaf key.
+ Rename DBCHECK_SHOW_BITMAP to DBCHECK_SHOW_FREELIST (old name still
provided for now, but flagged as deprecated - DBCHECK_SHOW_BITMAP was new
in 1.3.0, so will likely be removed before 1.4.0).
* Methods and functions which take a string to unserialise now consistently
call that parameter "serialised".
* Weight: Make number of distinct terms indexing each document and the
collection frequency of the term available to subclasses. Patch from
Gaurav Arora's Language Modelling branch.
* WritableDatabase: Add support for multiple subdatabases, and support opening
a stub database containing multiple subdatabases as a WritableDatabase.
* WritableDatabase can now be constructed from just a pathname (defaulting to
opening the database with DB_CREATE_OR_OPEN).
* WritableDatabase: Add flags which can be bitwise OR-ed into the second
argument when constructing:
+ Xapian::DB_NO_SYNC: to disable use of fsync, etc
+ Xapian::DB_DANGEROUS: to enable in-place updates
+ Xapian::DB_BACKEND_CHERT: if creating, create a chert database
+ Xapian::DB_BACKEND_GLASS: if creating, create a glass database
+ Xapian::DB_NO_TERMLIST: create a database without a termlist (see #181)
+ Xapian::DB_FULL_SYNC flag - if this is set for a database, we use the Mac
OS X F_FULL_SYNC instead of fdatasync()/fsync()/etc on the version file
when committing.
* Database: Add optional flags argument to constructor - the following can be
bitwise OR-ed into it:
+ Xapian::DB_BACKEND_CHERT (only open a chert database)
+ Xapian::DB_BACKEND_GLASS (only open a glass database)
+ Xapian::DB_BACKEND_STUB (only open a stub database)
* Xapian::Auto::open_stub() and Xapian::Chert::open() are now deprecated in
favour of these new flags.
* Add LMWeight class, which implements the Unigram Language Modelling weighting
scheme. Patch from Gaurav Arora.
* Add implementations of a number of DfR weighting schemes (BB2, DLH, DPH,
IfB2, IneB2, InL2, PL2). Patches from Aarsh Shah.
* Add support for the Bo1 query expansion scheme. Patch from Aarsh Shah.
* Add Enquire::set_time_limit() method which sets a timelimit after which
check_at_least will be disabled.
* Database: Trying to perform operations on a database with no subdatabases now
throws InvalidOperationError not DocNotFoundError.
* Query: Implement new OP_MAX query operator, which returns the maximum weight
of any of its subqueries. (see #360)
* Query: Add methods to allow introspection on Query objects - currently you
can read the leaf type/operator, how many subqueries there are, and get a
particular subquery. For a query which is a term, Query::get_terms_begin()
allows you to get the term. (see #159)
* Query: Only simplify OP_SYNONYM with a single subquery if that subquery is a
term or MatchAll.
* Avoid two vector copies when storing term positions in most common cases.
* Reimplement version functions to use a single function in libxapian which
returns a pointer to a static const struct containing the version
information, with inline wrappers in the API header which call this. This
means we only need one relocation instead of 4, reducing library load time a
little.
* Make TermGenerator flags an anonymous enum, and typedef TermGenerator::flags
to int for backward compatibility with existing user code which uses it.
* Stem: Fix incorrect Unicode codepoints for o-double-acute and u-double-acute
in the Hungarian Snowball stemmer. Reported by Tom Lane to snowball-discuss.
* Stem: Add an early english stemmer.
* Provide the stopword lists from Snowball plus an Arabic one, installed in
${prefix}/share/xapian-core/stopwords/. Patch from Assem Chelli, fixes#269.
* Improve check for direct inclusion of Xapian subheaders in user code to
catch more cases.
* Add simple API to help with creating language-idiomatic iterator wrappers
in <xapian/iterator.h>.
* Give an compilation error if user code tries to include API headers other
than xapian.h directly - these other headers are an internal implementation
detail, but experience has shown that some people try to include them
directly. Please just use '#include <xapian.h>' instead.
* Update Unicode character database to Unicode 6.2.0.
* Add FieldProcessor class (ticket#128) - currently marked as an experimental
API while we sort out how best to sort out exactly how it interacts with
other QueryParser features.
* Add implementation of several TF-IDF weighting schemes via a new TfIdfWeight
class.
* Add ExpandDeciderFilterPrefix class which only return terms with a particular
prefix. (fixes#467)
* QueryParser: Adjust handling of Unicode opening/closing double quotes - if a
quoted boolean term was started with ASCII double quote, then only ASCII
double quote can end it, as otherwise it's impossible to quote a term
containing Unicode double quotes.
* Database::check(): If the database can't be opened, don't emit a bogus
warning about there being too many documents to cross-check doclens.
* TradWeight,BM25Weight: Throw SerialisationError instead of NetworkError if
unserialise() fails.
* QueryParser: Change the default stemming strategy to STEM_SOME, to eliminate
the API gotcha that setting a stemmer is ignored until you also set a
strategy.
* Deprecate Xapian::ErrorHandler. (ticket#3)
* Stem: Generate a compact and efficient table to decode language names. This
is both faster and smaller than the approach we were using, with the added
benefit that the table is auto-generated.
* xapian.h:
+ Add check for Qt headers being included before us and defining
'slots' as a macro - if they are, give a clear error advising how to work
around this (previously compilation would fail with a confusing error).
+ Add a similar check for Wt headers which also define 'slots' as a macro
by default.
* Update Unicode character database to Unicode 6.1.0. (ticket#497)
* TermIterator returned by Enquire::get_matching_terms_begin(),
Query::get_terms_begin(), Database::synonyms_begin(),
QueryParser::stoplist_begin(), and QueryParser::unstem_begin() now stores the
list of terms to iterate much more compactly.
* QueryParser:
+ Allow Unicode curly double quote characters to start and/or end phrases.
+ The set_default_op() method will now reject operators which don't make
sense to set. The operators which are allowed are now explicitly
documented in the API docs.
* Query: The internals have been completely reimplemented (ticket#280). The
notable changes are:
+ Query objects are smaller and should be faster.
+ More readable format for Query::get_description().
+ More compact serialisation format for Query objects.
+ Query operators are no longer flattened as you build up a tree (but the
query optimiser still combines groups of the same operator). This means
that Query objects are truly immutable, and so we don't need to copy Query
objects when composing them. This should also fix a few O(n*n) cases when
building up an n-way query pair-wise. (ticket#273)
+ The Query optimiser can do a few extra optimisations.
* There's now explicit support for geospatial search (this API is currently
marked as experimental). (ticket#481)
* There's now an API (currently experimental) for checking the integrity of
databases (partly addresses ticket#238).
* Database::reopen() now returns true if the database may have been reopened
(previously it returned void). (ticket#548)
* Deprecate Xapian::timeout in favour of POSIX type useconds_t.
* Deprecate Xapian::percent and use int instead in the API and our own code.
* Deprecate Xapian::weight typedef in favour of just using double and change
all uses in the API and our own code. (ticket#560)
* Rearrange members of Xapian::Error to reduce its size (from 48 to 40 bytes on
x86-64 Linux).
* Assignment operators for PositionIterator and TermIterator now return *this
rather than void.
* PositionIterator, PostingIterator, TermIterator and ValueIterator now
handle their reference counts in hand-crafted code rather than using
intrusive_ptr/RefCntPtr, which means the compiler can inline the destructor
and default constructor, so a comparison to an end iterator should now
optimise to a simple NULL pointer check, but without the issues which the
ValueIteratorEnd_ proxy class approach had (such as not working in templates
or some cases of overload resolution).
* Enquire:
+ Previously, Enquire::get_matching_terms_begin() threw InvalidArgumentError
if the query was empty. Now we just return an end iterator, which is more
consistent with how empty queries behave elsewhere.
+ Remove the deprecated old-style match spy approach of using a MatchDecider.
* Remove deprecated Sorter class and MultiValueSorter subclass.
* Xapian::Stem:
+ Add stemmers for Armenian (hy), Basque (eu), and Catalan (ca).
+ Stem::operator= now returns a reference to the assigned-to object.
testsuite:
* OP_SCALE_WEIGHT: Check top weight is non-zero - if it is zero, tests which
try to check that OP_SCALE_WEIGHT works will always pass.
* testsuite: Check SerialisationError descriptions from Xapian::Weight
subclasses mention the weighting scheme name.
* Merge queryparsertest and termgentest into apitest. Their testcases now use
the backend manager machinery in the testharness, so we don't have to
hard-code use of inmemory and chert backends, but instead run them under all
backends which support the required features. This fixes some test failures
when both chert and glass are disabled due to trying to run spelling tests
with the inmemory backend.
* Avoid overflowing collection frequency in totaldoclen1. We're trying to test
total document length doesn't wrap, so avoid collection freq overflowing in
the process, as that triggers errors when running the testsuite under ubsan.
We should handle collection frequency overflow better, but that's a separate
issue.
* Add some test coverage for ESet::get_ebound().
* Fix testcase notermlist1 to check correct table extension - ".glass" not
".DB" (chert doesn't support DB_NO_TERMLIST).
* unittest: We can't use Assert() to unit test noexcept code as it throws an
exception if it fails. Instead set up macros to set a variable and return if
an assertion fails in a unittest testcase, and check that variable in the
harness.
* Add unit test for internal C_isupper(), etc functions.
* If command line option --verbose/-v isn't specified, set the verbosity level
from environmental variable VERBOSE.
* Re-enable replicate3 for glass, as it no longer fails.
* Add more test coverage for get_unique_terms().
* Don't leave an extra fd open when starting xapian-tcpsrv for remotetcp tests.
* Extend checkstatsweight1 to check that Weight::get_collection_freq() returns
the same number as Database::get_collection_freq().
* queryparsertest: Add testcase for FieldProcessor on boolean prefix with
quoted contents.
* queryparsertest: Enable some disabled cases which actually work (in some
cases with slightly tweaked expected answers which are equivalent to those
that were shown).
* Make use of the new writable multidatabase feature to simplify the
multi-database handling in the test harness.
* Change querypairwise1_helper to repeat the query build 100 times, as with a
fast modern machine we were sometimes trying with so many subqueries that we
would run out of stack.
* apitest: Use Xapian::Database::check() in cursordelbug1. (partly addresses
#238)
* apitest: Test Query ops with a single MatchAll subquery.
* apitest: New testcase readonlyparentdir1 to ensure that commit works with a
read-only parent directory.
* tests/generate-api_generated: Test that the string returned by a
get_description() method isn't empty.
* Use git commit hash in title of test coverage reports generated from a git
tree.
* Make unittest use the test harness, so it gets all the valgrind and fd leak
checks, and other handy features all the other tests have.
* Improve test coverage in several places.
* Compress generated HTML files in coverage report.
matcher:
* Fix stats passed to Weight with OP_SYNONYM. Previously the number of
unique terms was never calculated, and a term which matched all documents
would be optimised to an all-docs postlist, which fails to supply the
correct wdf info.
* Use floating point calculation for OR synonym freq estimates. The division
was being done as an integer division, which means the result was always
getting rounded down rather than rounded to the nearest integer.
* Fix upper bound on matches for OP_XOR. Due to a reversed conditional, the
estimate could be one too low in some cases where the XOR matched all the
documents in the database.
* Improve lower bound on matches for OP_XOR. Previously the lower bound was
always set to 0, which is valid, but we can often do better.
* Optimise value range which is a superset of the bounds. If the value
frequency is equal to the doccount, such a range is equivalent to MatchAll,
and we now avoid having to read the valuestream at all.
* Optimise OP_VALUE_RANGE when the upper bound can't be exceeded. In this
case, we now use ValueGePostList instead of ValueRangePostList.
* Streamline collation of statistics for use by weighting schemes - tests show
a 2% or so increase in speed in some cases.
* If a term matches all documents and its weight doesn't depend on its wdf, we
can optimise it to MatchAll (the previous requirement that maxpart == 0 was
unnecessarily strict).
* Fix the check for a term which matches all documents to use the sub-db
termfreq, not the combined db termfreq.
* When we optimise a postlist for a term which matches all documents to use
MatchAll, we still need to set a weight object on it to get percentages
calculated correctly.
* Drop MatchNothing subqueries in OR-like situations in add_subquery() rather
than adding them and then handling it later.
* Handle the left side of AND_NOT and AND_MAYBE being MatchNothing in
add_subquery() rather than in done().
* Handle QueryAndLike with a MatchNothing subquery in add_subquery() rather
than done().
* Query: Multi-way operators now store their subquery pointers in a custom
class rather than std::vector<Xapian::Query>. The custom class take the
same amount of space, or often less. It's particularly efficient when
there are two subqueries, which is very desirable as we no longer flatten a
subtree of the same operator as we build the query.
* Optimise an unweighted query term which matches all the documents in a
subdatabase to use the "MatchAll" postlist. (ticket#387)
glass backend:
* Fix allterms with prefix on glass with uncommitted changes. Glass aims to
flush just the relevant postlist changes in this case but the end of the
range to flush was wrong, so we'd only actually flush changes for a term
exactly matching the prefix. Fixes#721.
* Fix Database::check() parsing of glass changes file header. In practice this
was unlikely to actually cause problems.
* Make glass the default backend. The format should now be stable, except
perhaps in the unlikely event that a bug emerges which requires a format
change to address.
* Don't explicitly store the 2 byte "component_of" counter for the first
component of every Btree entry in leaf blocks - instead use one of the upper
bits of the length to store a "first component" flag. This directly saves 2
bytes per entry in the Btree, plus additional space due to fewer blocks and
fewer levels being needed as a result. This particularly helps the position
table, which has a lot of entries, many of them very small. The saving would
be expected to be a little less than the saving from the change which shaved
2 bytes of every Btree item in 1.3.4 (since that saved 2 bytes multiple times
for large entries which get split into multiple items). A simple test
suggests a saving of several percent in total DB size, which fits that. This
change reduces the maximum component size to 8194, which affects tables
with a 64KB blocksize in normal use and tables with >= 16KB blocksize with
full compaction.
* Refactor glass backend key comparison - == and < operations are replaced by
a compare() function returns negative, 0 or positive (like strcmp(), memcmp()
and std::string::compare()). This allows us to avoid a final compare to
check for equality when binary chopping, and to terminate early if the binary
chop hits the exact entry.
* If a cursor is moved to an entry which doesn't exist, we need to step back to
the first component of previous entry before we can read its tag. However we
often don't actually read its tag (e.g. if we only wanted the key), so make
this stepping back lazy so we can avoid doing it when we don't want to read
the tag.
* Avoid creating std::string objects to hold data when compressing and
decompressing tags with zlib.
* Store minimum compression length per table in the version file, with 0
meaning "don't compress". Currently you can only change this setting with a
hex editor on the file, but now it is there we can later make use of it
without needing a database format change.
* Database::check() now performs additional consistency checks for glass.
Reported by Jean-Francois Dockes and Bob Cargill via xapian-discuss.
* Database::check(): check docids don't exceed db_last_docid when checking
a single glass table.
* We now throw DatabaseCorruptError in a few cases where it's appropriate
but we didn't previously, in particular in the case where all the files in a
DB have been truncated to zero size (which makes handling of this case
consistent with chert).
* Fix compaction to a single file which already exists. This was hanging.
Noted by Will Greenberg on #xapian.
* Shave 2 bytes of every Btree item (which will probably typically reduce
database size by several percent).
* More compact item format for branch blocks - 2 bytes per item smaller. This
means each branch block can branch more ways, reducing the number of Btree
levels needed, which is especially helpful for cold-cache search times.
* Track an upper bound on spelling word frequency. This isn't currently used,
but will be useful for improving the spelling algorithm, and we want to
stabilise the glass backend format. See #225, reported by Philip Neustrom.
* Support 64-bit docids in the glass backend on-disk format. This changes the
encoding used by pack_uint_preserving_sort() to one which supports 64 bit
values, and is a byte smaller for values 16384-32767, and the same size for
all other 32 bit values. Fixes#686, from original report by James Aylett.
* Use memcpy() not memmove() when no risk of overlap.
* Store length of just the key data itself, allowing keys to be up to 255 bytes
long - the previous limit was 252.
* Change glass to store DB stats in the version file. Previously we stored
them in a special item in the postlist table, but putting them in the version
file reduces the number of block reads required to open the database, is
simpler to deal with, and means we can potentially recalculate tight upper
and lower bounds for an existing database without having to commit a new
revision.
* Add support for a single-file variant for glass. Currently such databases
can only be opened for reading - to create one you need to use
xapian-compact (or its API equivalent). You can embed such databases within
another file, and open them by passing in a file descriptor open on that file
and positioned at the offset the database starts at). Database::check() also
supports them. Fixes#666, reported by Will Greenberg (and previously
suggested on xapian-discuss by Emmanuel Engelhart).
* Avoid potential DB corruption with full-compaction when using 64K blocks.
* Where posix_fadvise() is available, use it to prefetch postlist Btree blocks
from the level below the root block which will be needed for postlists of
terms in the query, and similarly for the docdata table when MSet::fetch() is
called. Based on patch by Will Greenberg in #671.
* When reporting freelist errors during a database check, distinguish between a
block in use and in the freelist, and a block in the freelist more than once.
* Fix compaction and database checking for the change to the format of keys
in the positionlist table which happened in 1.3.2.
* After splitting a block, we always insert the new block in the parent right
after the block it was split from - there's no need to binary chop.
* Avoid infinite recursion when we hit the end of the freelist block we're
reading and the end of the block we're writing at the same time.
* Fix freelist handling to allow for the newly loaded first block of the
freelist being already used up.
* 'brass' backend renamed to 'glass' - we decided to use names in ascending
alphabetical order to make it easier to understand which backend is newest,
and since 'flint' was used recently, we skipped over 'd', 'e' and 'f'.
* Change positionlist keys to be ordered by term first rather than docid first,
which helps phrase searching significantly. For more efficient indexing,
positionlist changes are now batched up in memory and written out in key
order.
* Use a separate cursor for each position list - now we're ordering the
position B-tree by term first, phrase matching would cause a single cursor
to cycle between disparate areas of the B-tree and reread the same blocks
repeatedly.
* Reference count blocks in the btree cursor, so cursors can cheaply share
blocks. This can significantly reduce the amount of memory used by cursors
for queries which contain a lot of terms (e.g. wildcards which expand to a
lot of terms).
* Under glass, optimise the turning of a query into a postlist to reuse the
cursor blocks which are the same as the previous term's postlist. This is
particularly effective for a wildcard query which expands to a lot of terms.
* Keep track of unused blocks in the Btrees using freelists rather than
bitmaps. (fixes#40)
* Eliminate the base files, and instead store the root block and freelist
pointers in the "iamglass" file.
* When compacting, sync all the tables together at the end.
* In DB_DANGEROUS mode, update the version file in-place.
* Only actually store the document data if it is non-empty. The table which
holds the document data is now lazily created, so won't exist if you never
set the document data.
* Iterating positional data now decodes it lazily, which should speed up
phrases which include common words.
* Compress changesets in brass replication. Increments the changeset version.
Ticket #348
* Restore two missing lines in database checking where we report a block with
the wrong level.
* When checking if a block was newly allocated in this revision, just look
at its revision number rather than consulting the base file's bitmap.
remote backend:
* Improve handling of invalid remote stub entries: Entries without a colon now
give an error rather than being quietly skipped; IPv6 isn't yet supported,
but entries with IPv6 addresses now result in saner errors (previously the
colons confused the code which looks for a port number).
* Fix hook for remote support of user weighting schemes. The commented-out
code used entirely the wrong class - now we use the server object we have
access to, and forward the method to the class which needs it.
* Avoid dividing zero by zero when calculating the average length for an empty
database.
* Bump remote protocol version to 38.0, due to extra statistics being tracked
for weighting.
* Make Weight::Internal track if any max_part values are set, so we don't need
to serialise them when they've not been set.
* Prefix compress list of terms and metadata keys in the remote protocol.
This requires a remote protocol major version bump.
* When propagating exceptions from a remote backend server, the protocol now
sends a numeric code to represent which exception is being propagated, rather
than the name of the type, as a number can be turned back into an exception
with a simple switch statement and is also less data to transfer.
(ticket#471)
* Remote protocol (these changes require a protocol major version bump):
+ Unify REPLY_GREETING and REPLY_UPDATE.
+ Send (last_docid - doccount) instead of last_docid and (doclen_ubound -
doclen_lbound) instead of doclen_ubound.
* Remove special check which gives a more helpful error message when a modern
client is used against a remote server running Xapian <= 0.9.6.
chert backend:
* When using 64-bit Xapian::docid, consistently use the actual maximum valid
docid value rather instead of the maximum value the type can hold.
* Where posix_fadvise() is available, use it to prefetch postlist Btree blocks
from the level below the root block which will be needed for postlists of
terms in the query, and similarly for the record table when MSet::fetch() is
called. Based on patch by Will Greenberg in #671.
* Fix problems with get_unique_terms() on a modified chert database.
* Fix xapian-check on a single chert table, which seg faulted in 1.3.2.
* Improve DBCHECK_FIX:
+ if fixing a whole database, we now take the revision from the first table
we successfully look at, which should be correct in most cases, and is
definitely better than trying to determine the revision of each broken
table independently.
+ handle a zero-sized .DB file.
+ After we successfully regenerate baseA, remove any empty baseB file to
prevent it causing problems. Tracked down with help from Phil Hands.
* Iterating positional data now decodes it lazily, which should speed up
phrases which include common words.
flint backend:
* Remove flint backend.
API:
* PostingSource: Public member variables are now wrapped by methods (mostly
getters and/or setters, depending on whether they should be readable,
writable or both). In 1.3.5, the public members variables have been
deprecated - we've added the replacement methods in 1.2.23 as well to make
it easier for people to migrate over.
chert backend:
* xapian-check now performs additional consistency checks for chert. Reported
by Jean-Francois Dockes and Bob Cargill via xapian-discuss.
documentation:
* Update links to Xapian website and trac to use https, which is now supported,
thanks to James Aylett.
portability:
* On older Linux kernels, rename() of a file within a directory on NFS can
sometimes erroneously fail with EXDEV. This should only happen if you
try to rename a file across filing systems, so workaround this issue by
retrying up to 5 times on EXDEV (which should be plenty to avoid this
bug, and we don't want to risk looping forever). Fixes#698, reported by
Mark Dufour.
API:
* Add FLAG_CJK_NGRAM for QueryParser and TermGenerator. Has the same effect as
setting the environment variable XAPIAN_CJK_NGRAM. Fixes#180, reported by
Richard Boulton, with contributions from Pavel Strashkin, Mikkel Kamstrup
Erlandsen and Brandon Schaefer.
* Fix bug parsing multiple non-exclusive filter terms - previously this could
result in such filters effectively being ignored.
* Fix Database::get_doclength_lower_bound() over multiple databases when some
are empty or consist only of zero-length documents. Previously this would
report a lower bound of zero, now it reports the same lowest bound as a
single database containing all the same documents.
* Make Database::get_wdf_upper_bound("") return 0.
* Mark constructors taking a single argument as "explicit" to avoid unwanted
implicit conversions.
testsuite:
* If command line option --verbose/-v isn't specified, set the verbosity level
from environmental variable VERBOSE.
* Skip timed tests if $AUTOMATED_TESTING is set. Fixes#553, reported by
Dagobert Michelsen.
* Don't leave an extra fd open when starting xapian-tcpsrv for remotetcp tests.
* apitest: Revert disabling of part of adddoc5 for clang - the test failure was
in fact due to a bug in 1.3.x, and 1.2.x was never affected.
* apitest: Tweak bounds checks in dbstats1 testcase - multi backends should
give tight bounds.
brass backend:
* Format limit on docid now correctly imposed when sizeof(int) > 4.
* Avoid potential DB corruption with full-compaction when using 64K blocks.
chert backend:
* Format limit on docid now correctly imposed when sizeof(int) > 4.
* Avoid potential DB corruption with full-compaction when using 64K blocks.
flint backend:
* Format limit on docid now correctly imposed when sizeof(int) > 4.
* Avoid potential DB corruption with full-compaction when using 64K blocks.
remote backend:
* Fix to handle total document length exceeding 34,359,738,368. (Fixes#678,
reported by matf).
* Avoid dividing by zero when getting the average length for an empty database.
* Stop apparent error from remote server when read-only client disconnects. A
read-only client just closes the connection when done, but the server
previously reported "Got exception NetworkError: Received EOF", which sounds
like there was a problem. Now we just say "Connection closed" here, and
"Connection closed unexpectedly" if the client connects in the middle of an
exchange. Possibly fixes#654, reported by German M. Bravo.
* Give a clearer error message when the client and server remote protocol
versions aren't compatible.
* Check length of key in MSG_SETMETADATA.
build system:
* pkg-config: Fix library name in .pc file to say "xapian" not "xapian-core".
Reported by Eric Lindblad to the xapian-devel list.
* Private symbol decode_length() is no longer visible outside the library.
documentation:
* Stop maintaining ChangeLog files. They make merging patches harder, and stop
'git cherry-pick' from working as it should. The git repo history should be
sufficient for complying with GPLv2 2(a).
* Strip out "quickstart" examples which are out of date and rather redundant
with the "simple" examples.
* Correct documentation of Enquire::get_query(). If no query has been set,
the documentation said Xapian::InvalidArgumentError was thrown, but in
fact we just return a default initialised Query object (i.e. Query()). This
seems reasonable behaviour and has been the case since Xapian 0.9.0.
* Document xapian-compact --blocksize takes an argument.
* Update snowball website link to snowballstem.org.
tools:
* xapian-replicate: Fix replication for files > 4GB on 32-bit platforms.
Previously replication would fail to copy a file whose size didn't fit in
size_t. Fixes#685, reported by Josh Elsasser.
* xapian-tcpsrv: Better error if -p/--port not specified
* quest: Support `-f cjk_ngram`.
examples:
* xapian-metadata: Extend "list" subcommand to take optional key prefix.
portability:
* Fix new warnings from recent versions of GCC and clang.
* Add spaces between literal strings and macros which expand to literal strings
for C++11 compatibility in __WIN32__-specific code.
* Need <unistd.h> for unlink() on FreeBSD, reported by Germán M. Bravo via
github PR 72.
* Fix testsuite to build when S_ISSOCK() isn't defined.
* Don't provide our own implementation of sleep() under __WIN32__ if there
already is one - mingw provides one, and in some situations it seems to clash
with ours. Reported to xapian-discuss by John Alveris.
* Add missing '#include <arpa/inet.h>' to htons(). Seems to be implicitly
included on most platforms, but Interix needs it. Reported by Eric Lindblad
on xapian-discuss.
* Disable "<FUNCTION> is expected to return a value" warning from Sun's C++
compiler, as it fires for functions ending in a "throw" statement. Genuine
instances will be caught by compilers with superior warning machinery.
* Prefer scalbn() to ldexp() where possible, since the former doesn't ever set
errno.
* '#include <config.h>' in the "simple" examples, as when compiling with xlC on
AIX, _LARGE_FILES gets defined by AC_SYS_LARGEFILE to enable large file
support, and defining this changes the ABI of std::string, so it also needs
to be defined when compiling code using Xapian.
* On cygwin, include <arpa/inet.h> instead of winsock headers for htons() and
htonl().
* Include <cygwin/version.h> for CYGWIN_VERSION_API_MAJOR.
* Avoid referencing static members via an object in that object's own
definition, as this doesn't work with all compilers (noted with GCC 3.3), and
is a bit of an odd construct anyway. Reported by Eric Lindblad on
xapian-discuss.
* GCC < 3.4.2 lacks operator<< overloads for unsigned long long on some
platforms, so simply work around this by using str(), as this isn't
performance sensitive code. Reported by Eric Lindblad on xapian-discuss.
* Fix delete which should be delete[] in brass backend cursor code.
Problems found locating distfiles:
Package cabocha: missing distfile cabocha-0.68.tar.bz2
Package convertlit: missing distfile clit18src.zip
Package php-enchant: missing distfile php-enchant/enchant-1.1.0.tgz
Otherwise, existing SHA1 digests verified and found to be the same on
the machine holding the existing distfiles (morden). All existing
SHA1 digests retained for now as an audit trail.
API:
* QueryParser: Extend the set of characters allowed in the start of a range to
be anything except for '(' and characters <= ' '. This better matches what's
accepted for a range end (anything except for ')' and characters <= ' ').
Reported by Jani Nikula.
matcher:
* Reimplement OP_PHRASE for non-exact phrases. The previous implementation was
buggy, giving both false positives and false negatives in rare cases when
three or more terms were involved. Fixes#653, reported by Jean-Francois
Dockes.
* Reimplement OP_NEAR - the new implementation consistently requires the terms
to occur at different positions, and fixes some previously missed matches.
* Fix a reversed check for picking the shorter position list for an exact
phrase of two terms. The difference this makes isn't dramatic, but can be
measured (at least with cachegrind). Thanks to kbwt for spotting this.
* When matching an exact phrase, if a term doesn't occur where we want, use
its actual position to advance the anchor term, rather than just checking
the next position of the anchor term.
brass backend:
* Fix cursor versioning to consider cancel() and reopen() as events where
the cursor version may need incrementing, and flag the current cursor version
as used when a cursor is rebuilt. Fixes#675, reported by Germán M. Bravo.
* Avoid using file descriptions < 3 for writable database tables, as it risks
corruption if some code in the same process tries to write to stdout or
stderr without realising it is closed. (Partly addresses #651)
chert backend:
* Fix cursor versioning to consider cancel() and reopen() as events where
the cursor version may need incrementing, and flag the current cursor version
as used when a cursor is rebuilt. Fixes#675, reported by Germán M. Bravo.
* Avoid using file descriptions < 3 for writable database tables, as it risks
corruption if some code in the same process tries to write to stdout or
stderr without realising it is closed. (Partly addresses #651)
* After splitting a block, we always insert the new block in the parent right
after the block it was split from - there's no need to binary chop.
flint backend:
* Fix cursor versioning to consider cancel() and reopen() as events where
the cursor version may need incrementing, and flag the current cursor version
as used when a cursor is rebuilt. Fixes#675, reported by Germán M. Bravo.
remote backend:
* Fix sort by value when multiple databases are in use and one or more are
remote. This change necessitated a minor version bump in the remote
protocol. Fixes#674, reported by Dylan Griffith. If you are upgrading a
live system which uses the remote backend, upgrade the servers before the
clients.
build system:
* The compiler ABI check in the public API headers now issues a warning
(instead of an error) for an ABI mismatch for ABI versions 2 and later
(which means GCC >= 3.4). The changes in these ABI versions are bug fixes
for corner cases, so there's a good chance of things working - e.g. building
xapian-bindings with GCC 5.1 (which defaults to ABI version 8) against
xapian-core built with GCC 4.9 (which defaults to ABI version 2) seems to
work OK. A warning is still useful as a clue to what is going on if linking
fails due to a missing symbol.
* xapian-config,xapian-core.pc: When compiling with xlC on AIX, the reported
--cxxflags/--cflags now include -D_LARGE_FILES=1 as this is defined for the
library, and defining it changes the ABI of std::string with this compiler,
so it must also be defined when building code using the Xapian API.
* xapian-core.pc: Include --enable-runtime-pseudo-reloc in --libs output for
mingw and cygwin, like xapian-config does.
* xapian-core.pc: Fix include directory reported by `pkg-config --cflags`.
This bug was harmless if xapian-core was installed to a directory which was
on the default header search path (such as /usr/include).
* xapian-config: Fix typo so cached result of test in is_uninstalled() is
actually used on subsequent calls. Fixes#676, reported (with patch) by Ryan
Schmidt.
* configure: Changes in 1.2.19 broke the custom macro we use to probe for
supported compiler flags such that the flags never got used. This release
fixes this problem.
* configure: Set default value for AUTOM4TE before AC_OUTPUT so the default
will actually get used. Only relevant when building in maintainer mode
(e.g. from git).
* soaktest: Link with libtool's '-no-install' or '-no-fast-install', like we
already do for other test programs, which means that libtool doesn't need to
generate shell script wrappers for them on most platforms.
* Generate and install a file for pkg-config. (Fixes#540)
* configure: Update link to cygwin FAQ in error message.
documentation:
* API documentation: Minor wording tweaks and formatting improvements.
* docs/deprecation.rst: Add deprecation of omindex --preserve-nonduplicates
which happened in 1.2.4.
* HACKING: Update URL.
* HACKING: libtool 2.4.6 is now used for bootstrapping snapshots and releases.
* include/xapian/weight.h: Document the enum stat_flags values.
* docs/postingsource.rst: Use a modern class in postingsource example. (Noted
by James Aylett)
* docs/deprecation.rst,docs/replication.rst: Fix typos.
* Update doxygen configuration files to avoid warnings about obsolete tags from
newer doxygen versions.
* HACKING: Update details of building Xapian packages.
tools:
* xapian-compact: Make sure we open all the tables of input databases at the
same revision. (Fixes#649)
* xapian-metadata: Add 'list' subcommand to list all the metadata keys.
* xapian-replicate: Fix connection timeout to be 10 seconds rather than 10000
seconds (the incorrect timeout has been the case since 1.2.3).
* xapian-replicate: Set SO_KEEPALIVE for xapian-replicate's connection to the
master, and add command line option to allow setting socket-level timeouts
(SO_RCVTIMEO and SO_SNDTIMEO) on platforms that support them. Fixes#546,
reported by nkvoll.
* xapian-replicate-server: Avoid potentially reading uninitialised data if a
changeset file is truncated.
* xapian-check: For chert and brass, cross-check the position and postlist
tables to detect positional data for non-existent documents.
portability:
* Add spaces between literal strings and macros which expand to literal strings
for C++11 compatibility.
* ValueCountMatchSpy::top_values_begin(): Fix the comparison function not to
return true for two equal elements, which manifests as incorrect sorting in
some cases when using clang's libc++ (which recent OS X versions do).
* apitest: The adddoc5 testcase fails under clang due to an exception handling
bug, so just #ifdef out the problematic part of the testcase when building
with clang for now.
* Fix clang warnings on OS X. Reported by Germán M. Bravo.
* Fix examples to build with IBM's xlC compiler on AIX - they were failing due
to _LARGE_FILES being defined for the library build but not for the examples,
and defining this changes the ABI of std::string with this compiler.
* configure: Improve the probe for whether the test harness can use RTTI to
work for IBM's xlC compiler (which defaults to not generating RTTI).
* Fix to build with Sun's C++ compiler.
* Use F_DUPFD where available to dup to a file descriptor which is >= 2, rather
than calling dup() until we get one.
* When unserialising a double, avoid reading one byte past the end of the
serialised value. In practice this was harmless on most platforms, as
dbl_max_mantissa is 255 for IEEE-754 format doubles, and at least GCC's
std::string keeps the buffer nul-terminated. Reported by Germán M. Bravo in
github PR#67.
* When unserialising a double, add missing cast to unsigned char when we check
if the value will fit in the double type. On machines with IEEE-754 doubles
(which is most current platforms) this happened to work OK before. It would
also have been fine on machines where char is unsigned by default.
* Fix incorrect use of "delete" which should be "delete []". This is
undefined behaviour in C++, though the type is POD, so in practice this
probably worked OK on many platforms.
* When locking a database for writing, use F_OFD_SETLK where available, which
avoids having to fork() a child process to hold the lock. This currently
requires Linux kernel >= 3.15, but it has been submitted to POSIX so
hopefully will be widely supported eventually. Thanks to Austin Clements for
pointing out this now exists.
* Fix detection of fdatasync(), which appears to have been broken practically
forever - this means we've probably been using fsync() instead, which
probably isn't a big additional overhead. Thanks to Vlad Shablinsky for
helping with Mac OS X portability of this fix.
* configure: Define MINGW_HAS_SECURE_API under mingw to get _putenv_s()
declared in stdlib.h.
* Use POSIX O_NONBLOCK in preference to O_NDELAY - the semantics of the latter
differ between BSD and System V.
* According to POSIX, strerror() may not be thread safe, so use alternative
thread-safe ways to translate errno values where possible.
* On Microsoft Windows, avoid defining EADDRINUSE, etc if they're already
defined, and use WSAE* constants un-negated - they start from a high value
so won't collide with E* constants.
debug code:
* Fix some overly strict assertions in flint, which caused apitest's
cursordelbug1 to fail with assertions on.
* Add more assertions to the chert backend code.
API:
* Xapian::BM5Weight:
+ Improve BM25 upper bound in the case when our wdf upper bound > our
document length lower bound. Thanks to Craig Macdonald for pointing out
this trick.
+ Pre-multiply termweight by (param_k1 + 1) rather than doing it for
every weighted term in every document considered.
testsuite:
* Don't report apparent leaks of fds opened on /dev/urandom - at least on
Linux, something in the C library seems to lazily open it, and the report of
a possible leak followed by assurance that it's OK really is just noise we
can do without.
matcher:
* Fix false matches reported for non-exact phrases in some cases. Fixes the
reduced testcase in #657, reported by Jean-Francois Dockes.
brass backend:
* Only full sync after writing the final base file (only affects Max OS X).
chert backend:
* Only full sync after writing the final base file (only affects Max OS X).
flint backend:
* Only full sync after writing the final base file (only affects Max OS X).
build system:
* For Sun's C++ compiler, pass -library=Crun separately since libtool looks for
" -library=stlport4 " (with the spaces). (fixes#650)
* Remove .replicatmp (created by the test suite) upon "make clean".
documentation:
* include/xapian/compactor.h: Fix formatting of doxygen comment.
* HACKING: freecode no longer accepts updates, so drop that item from the
release checklist.
* docs/overview.rst: Add missing database path to example of using
xapian-progsrv in a stub database file.
portability:
* Suppress unused typedef warnings from debugging logging macros, which occur
in functions which always exit via throwing an exception when compiling with
recent versions of GCC or clang.
* Fix debug logging code to compile with clang. (fixes#657, reported by
Germán M. Bravo)
debug code:
* Add missing RETURN() markup for debug logging in a few places, highlighted by
warnings from recent GCC.
* Fix incorrect return types in debug logging annotations so that code compiles
when configured with --enable-log.
API:
* Document: Fix get_docid() to return the docid for the sub-database (as it
is explicitly documented to) for Document objects passed to functors like
KeyMaker during the match. (fixes#636, reported by Jeff Rand).
* Document: Don't store the termname in OmDocumentTerm - we were only using it
in get_description() output and an exception message. Speeds up indexing
etext.txt using simpleindex by 0.4%, and should reduce memory usage a bit
too. (Change inspired by comments from Vishesh Handa on xapian-devel).
* Database: Iterating the values in a particular slot is now a bit more
efficient for inmemory and remote backends (but still slow compared to
flint, chert and brass).
testsuite:
* apitest: Expand crashrecovery1 to check that the expected base files exist
and ones which shouldn't exist don't.
* queryparsertest: Fix testcase for empty wildcard followed by negation to
enable FLAG_LOVEHATE so the negation is actually parsed. Fortunately the
fixed testcase passes.
matcher:
* OP_SYNONYM: avoid fetching the doclength if the weighting scheme doesn't
need it and the calculated wdf for the synonym is <= doclength_lower_bound
for the current subdatabase. (fixes#360)
build system:
* Releases are now bootstrapped with libtool 2.4.2 instead of 2.4, and with
config.guess and config.sub updated to the latest versions.
documentation:
* Add an example of initializing SimpleStopper using a file listing stopwords.
(Patch from Assem Chelli)
* Improve the descriptions of the stem_strategy values in the API docs.
(Reported by "oilap" on #xapian)
* docs/sorting.rst: Fix incorrect parameter types in Xapian::Weight
subclass example.
* docs/glossary.rst: Add definition of "collection frequency".
* HACKING:
+ makeindex is now in Debian package texlive-binaries.
+ Replace a link to the outdated autotools "goat book" with a link to the
"Portable Shell" chapter of the autoconf manual.
* include/xapian/base.h: Remove very out of date comments talking about atomic
assignment and locking - since 0.5.0 we've adopted a "user locks" policy.
(Reported by Jean-Francois Dockes)
examples:
* delve:
+ Add -A <prefix> option to list all terms with a particular prefix.
+ Send errors to stderr not stdout.
+ If -v is specified more than once, show even more info in some cases.
(NEWS file claimed this was backported in 1.2.15, but it actually wasn't).
* quest:
+ Add --default-op option.
+ Add --weight option to allow the weighting scheme to be specified.
portability:
* Explicitly '#include <algorithm>' for std::max(), fixing build with VS2013.
(Fixes#641, reported by "boomboo").
* Fix testcase blocksize1 not to try to delete an open database, which isn't
possible under Windows. (Fixes#643, reported by Chris Olds)
* docs/quickstart.rst: Split --cxxflags and --libs for portability (noted by
"Hurricane Tong" on xapian-devel).
* Fix warnings with clang 5.0.
debug code:
* Add assertions that weighting scheme upper bounds aren't exceeded.
API:
* Enquire::set_sort_by_relevance_then_value() and
Enquire::set_sort_by_relevance_then_key(): Fix sense of reverse parameter.
Reported by "boomboo" on IRC.
* BM25Weight: Fix case where (k1 == 0 || b == 0) but k2 != 0. Reported by
"boomboo" on IRC.
* Unicode::tolower(): Fix to give correct results for U+01C5, U+01C8, U+01CB,
and U+01F2 (previously these were left unchanged).
* PositionIterator,PostingIterator,TermIterator,ValueIterator: Don't segfault
if skip_to() or check() is called on an iterator which is already at_end().
Reported by David Bremner.
* ValueCountMatchSpy: get_description() on a default-constructed
ValueCountMatchSpy object no longer fails when xapian-core is built with
--enable-log.
* ValueCountMatchSpy: get_total() on a default-constructed ValueCountMatchSpy
object now returns 0 rather than segfaulting.
testsuite:
* Automatically probe for and hook in eatmydata to the testsuite using the
wrapper script it now includes.
* Fix apitest to build when brass, chert or flint are disabled.
* If -v/--verbose is specified more than once to a test program, show the
diagnostic output for passing tests as well as failing/skipped ones.
* queryparsertest: Change qp_scale1 to time 5 repetitions of the large query to
help average out variations.
* queryparsertest: Add test coverage for explicit synonym of a term with a
prefix (e.g. ~foo:search).
* apitest: Remove code from registry* testcases which tries to test the
consequences of throwing an exception from a destructor - it's complex to
ensure we don't leak memory while doing this (it seems GCC doesn't release
the object in this case, but clang does), and it's generally frowned upon,
plus C++11 makes destructors noexcept by default.
* Fix "make check" to actually removed cached databases first, as is
intended.
brass backend:
* Fix handling of invalid block sizes passed to Xapian::Brass::open() - the
size gets fixed as documented, but the uncorrected size was passed to the
base file (and abort() was called if 0 was passed).
* Validate "dir_end" when reading a block. (fixes#592)
* When moving a cursor on a read-only table, check if the block we want is in
the internal cursor. We already do this for a writable table, as it is
necessary for correctness, but it's a cheap check and may avoid asking the
OS for a block we actually already have.
* Correctly report the database as closed rather than 'Bad file descriptor'
in certain cases.
* Reuse a cursor for reading values from valuestreams rather than creating
a new one each time. This can dramatically reduce the number of blocks
redundantly reread when sorting by value. The rereads will generally get
served from VM cache, but there's still an overhead to that.
chert backend:
* Fix handling of invalid block sizes passed to Xapian::Chert::open() - the
size gets fixed as documented, but the uncorrected size was passed to the
base file (and abort() was called if 0 was passed).
* Validate "dir_end" when reading a block. (fixes#592)
* When moving a cursor on a read-only table, check if the block we want is in
the internal cursor. We already do this for a writable table, as it is
necessary for correctness, but it's a cheap check and may avoid asking the
OS for a block we actually already have.
* Correctly report the database as closed rather than 'Bad file descriptor'
in certain cases.
* Reuse a cursor for reading values from valuestreams rather than creating
a new one each time. This can dramatically reduce the number of blocks
redundantly reread when sorting by value. The rereads will generally get
served from VM cache, but there's still an overhead to that.
flint backend:
* Fix handling of invalid block sizes passed to Xapian::Flint::open() - the
size gets fixed as documented, but the uncorrected size was passed to the
base file (and abort() was called if 0 was passed).
* Validate "dir_end" when reading a block. (fixes#592)
* When moving a cursor on a read-only table, check if the block we want is in
the internal cursor. We already do this for a writable table, as it is
necessary for correctness, but it's a cheap check and may avoid asking the
OS for a block we actually already have.
* Correctly report the database as closed rather than 'Bad file descriptor'
in certain cases.
build system:
* configure: Improve reporting of GCC version.
* Use -no-fast-install on platforms where -no-install causes libtool to emit a
warning.
* docs/Makefile.am: Fix handling of MAINTAINER_NO_DOCS.
* Include UnicodeData.txt and the script to generate the unicode tables from
it.
* Compress source tarballs with xz instead of gzip.
* Split XAPIAN_LIBS out of XAPIAN_LDFLAGS so that -l flags for libraries
configure detects are needed appear after -L flags specified by the user
that may be needed to find such libraries. (fixes#626)
* XO_LIB_XAPIAN now handles the user specifying a relative path in
XAPIAN_CONFIG, e.g.: "./configure XAPIAN_CONFIG=../xapian-core/xapian-config"
* Adjust XO_LIB_XAPIAN to strip _gitNNN suffix from snapshot versions.
* configure: Handle git snapshot naming when calculating REVISION.
* configure: Enable -fshow-column for GCC - things like vim's quickfix mode
will then jump to the appropriate column for a compiler error or warning, not
just the appropriate line.
* configure: Report GCC version in configure output.
documentation:
* postingsource.rst: Clarify a couple of points (reported by "vHanda" on IRC).
* The API documentation shipped with the release is now generated with
doxygen 1.8.5 instead of 1.5.9, which is most evident in the different
HTML styling newer doxygen uses.
* Document how Utf8Iterator handles invalid UTF-8 in API documentation.
* Improve how descriptions of deprecated features appear in the API
documentation.
* docs/remote_protocol.rst: Correct error in documentation of REPLY_DOCDATA
message.
* docs/overview.rst: Correct documentation for how to specify "prog" remote
databases in stub files.
* Direct users to git in preference to SVN - we'll be switching entirely in
the near future.
portability:
* Protect the ValueIterator::check() method against Mac OS X SDK headers
which define a check() macro.
* Fix warning from xlC compiler.
* Avoid use of grep -e in configure, as /usr/bin/grep on Solaris doesn't
support -e.
* Fix check for flags which might be needed for ANSI mode for compilers called
'cxx'.
* configure: Improve handling of Sun's C++ compiler - trick libtool into not
adding -library=Cstd, and prefer -library=stdcxx4 if supported. Explicitly
add -library=Crun which seems to be required, even though the documentation
suggests otherwise.
* configure: clang doesn't support -Wstrict-null-sentinel or -Wlogical-op, so
don't pass it these options.
* Fix build errors and warnings with mingw.
* Suppress "unused local typedef" warnings from GCC 4.8.
* If the compiler supports C++11, use static_assert to implement
CompileTimeAssert.
* tests/zlib-vg.c: Fix two warnings when compiled with clang.
* Fix failure when built with -D_GLIBCXX_DEBUG - we were modifying the top()
element of a heap before calling pop(), such that the heap comparison
operation (which is called when -D_GLIBCXX_DEBUG is on to verify the heap is
valid) would read off the end of the data. In a normal build, this issue
would likely never manifest.
* configure: When generating ABI compatibility checks in xapian/version.h, pass
$CXXFLAGS and $CPPFLAGS to $CXXCPP as they could contain options which affect
the ABI (such as -fabi-version for GCC). (Fixes#622)
* Microsoft GUIDs in binary form have reversed byte order in the first three
components compared to standard UUIDs, so the same database would report a
different UUID on Windows to on other platforms. We now swap the bytes to
match the standard order. With this fix, the UUIDs of existing databases
will appear to change on Windows (except in rare "palindronic" cases).
* Fix a couple of issues to get Xapian to build and work on AIX.
* common/safeuuid.h: Remove bogus take-address-of from uuid handling code for
NetBSD and OpenBSD.
* Under cygwin, use cygwin_conv_path() if using a new enough cygwin version,
rather than the now deprecated cygwin_conv_to_win32_path(). Reported by
"Haroogan" on the xapian-devel mailing list.
* common/safeuuid.h: Add missing '#include <cstdlib>' and qualify free with std.
* Fix 'unused label' warning when chert backend is disabled.
* xapian.h: Add check for Wt headers being included before us and defining
'slots' as a macro - if they are, give a clear error advising how to work
around this (previously compilation would fail with a confusing error).
tools:
* xapian-chert-update: Fix -b to work rather than always segfaulting (reported
in http://bugs.debian.org/716484).
* xapian-chert-update: The documented alias --blocksize for -b has never
actually been supported, so just drop mentions of it from --help and the man
page.
* xapian-check:
+ Fix chert database check that first docid in each doclength chunk is more
than the last docid in the previous chunk - previously this didn't actually
work.
+ Fix database check not to falsely report "position table: Junk after
position data" whenever there are 7 unused bits (7 is OK, *more* than 7
isn't).
+ Fix to report block numbers correctly for links within the B-tree.
+ If the METAINFO key is missing, only report it once per table.
+ Fix database consistency checking to always open all the tables at the same
revision - not doing this could lead to false errors being reported after a
commit interrupted by the process being killed or the machine crashing.
Reported by Joey Hess in http://bugs.debian.org/724610
examples:
* quest: Add --check-at-least option.
debug code:
* Fix assertion failure for when an OrPostList decays to an AndPostList - the
ordering of the subqueries by estimated termfreq may not be the same as it
was when the OrPostList was constructed, as the subqueries may themselves
have decayed. Reported by Michel Pelletier.
* Fix -Wcast-qual warning from GCC 4.7 when configured with --enable-log.
Xapian-core 1.2.15 (2013-04-16):
API:
* QueryParser/TermGenerator: Don't include CJK codepoints which are
punctuation in N-grams.
* TermGenerator: Fix bug where we failed to generate the first bigram
from the second sequence of N-grammable CJK characters in a piece of text.
brass backend:
* Call fdatasync()/fsync() when creating the "iambrass" file.
chert backend:
* Call fdatasync()/fsync() when creating the "iamchert" file.
flint backend:
* Call fdatasync()/fsync() when creating the "iamflint" file.
tools:
* delve: If -v is specified more than once, show even more info in some cases.
Xapian-core 1.2.14 (2013-03-14):
API:
* MSet::get_document(): Don't cache retrieved Document objects unless they
were requested with fetch(). This avoids using a lot of memory when many
MSet entries are retrieved. (Fixes#604)
matcher:
* Check if a candidate document has at least the minimum weight needed
before checking positional information, which speeds up slow phrase
searches (partly addresses #394).
brass backend:
* Fix multipass compaction not to damage document values, and to merge the
database stats correctly. (fixes#615)
chert backend:
* Fix multipass compaction not to damage document values, and to merge the
database stats correctly. (fixes#615)
flint backend:
* Fix multipass compaction bug. (fixes#615)
tools:
* xapian-replicate:
+ Fix handling of delays between replication events - the subtraction of the
target time and the current time was reversed, so we wouldn't sleep when
before the deadline, but would sleep after it for the amount we'd missed it
by.
+ On Microsoft Windows, we no longer sleep for more than 43 years if the
target time for a replication event had already passed. (Fixes#472)
Xapian-core 1.2.13 (2013-01-09):
API:
* TermGenerator: Add new method TermGenerator::set_max_word_length() to allow
this limit to be adjusted by the user.
* QueryParser: Implicitly close any unclosed brackets at the end of the query
string. Patch from Sehaj Singh Kalra.
* DateValueRangeProcessor: Add extra constructor overloaded form so that in
DateValueRangeProcessor(1, "date:"), the const char * gets interpreted as
std::string rather than bool.
matcher:
* Improved fix for #590 - count all matching LeafPostList objects with a Weight
object rather than trying to prune at the MultiAndPostList level based on
max_wt (if wdf is always zero for a term, BM25 gives max_wt of 0, which lead
to us never counting that subquery.
* Fix calculation of 0.0/0.0 in some cases. This then got used as a minimum
weight, but it seems this gives -nan (at least on x86-64 Linux) so it may
have been harmless in practice.
* We no longer use the highest weighted MSet entry to calculate percentages, so
remove code which finds it.
brass backend:
* Close excess file handles before we get the fcntl lock, which avoids the
lock being released again if one is open on the lock file. Notably this
avoids a situation where multiple threads in the same process could succeed
in locking a database concurrently.
chert backend:
* Close excess file handles before we get the fcntl lock, which avoids the
lock being released again if one is open on the lock file. Notably this
avoids a situation where multiple threads in the same process could succeed
in locking a database concurrently.
flint backend:
* Close excess file handles before we get the fcntl lock, which avoids the
lock being released again if one is open on the lock file. Notably this
avoids a situation where multiple threads in the same process could succeed
in locking a database concurrently.
remote backend:
* Improve the UnimplementedError message for a MatchSpy subclass which doesn't
implement name() so it's clearer that it is this particular subclass which
can't be used remotely, rather than all MatchSpy objects.
documentation:
* valueranges.html: Update documentation to reflect change in Xapian 1.1.2 -
DateValueRangeProcessor and StringValueRangeProcessor now support a prefix or
suffix.
* Clarify that the "reverse" parameter of set_sort_by_relevance_then_value()
and set_sort_by_relevance_then_key() only affects the ordering of the
value/key part of the sort.
* docs/quickstart.html: Fix seriously outdated statement that Xapian doesn't
create the database directory - that changed in 0.7.2 (released 2003-07-11).
* HACKING: Try to make it clearer we're looking for a dual-licence on submitted
patches.
tools:
* xapian-replicate:
+ Add a --full-copy option to force a full copy to be sent. (ticket#436)
+ Add --quiet option, and be a little more verbose by default.
+ Allow files > 32G to be be copied by replication.
+ Fix "if (fd > 0)" tests in some replication code to be "if (fd >= 0)".
In practice this is unlikely to actually have caused problems since
stdin is typically still open and using fd 0.
+ Simplify how we open the .DB file on the replication slave to just call
open() once with O_CREAT, rather than once without, than stat() if that
fails, and then again with O_CREAT|O_TRUNC if stat() doesn't show an
ordinary file exists.
examples:
* quest:
+ New --flags command line option to allow setting arbitrary QueryParser
flags.
+ Align option descriptions in --help output, and make the initial letter of
such descriptions consistently lowercase.
Xapian-core 1.2.12 (2012-06-27):
build system:
* 1.2.11 had its library version information incorrectly set. This resulted in
the shared library having an incorrect SONAME - e.g. on Linux,
libxapian.so.21 instead of libxapian.so.22. This release has been made to
fix this problem.
documentation:
* AUTHORS: Add the GSoC students.
Xapian-core 1.2.11 (2012-06-26):
API:
* Add new QueryParser::STEM_ALL_Z stemming strategy, which stems all terms and
adds a Z prefix. (Patch from Sehaj Singh Kalra, fixes ticket#562)
* Add TermGenerator::set_stemming_strategy() method, with strategies which
correspond to those of QueryParser. Based on patch from Sehaj Singh Kalra,
with some tweaks for adding term positions in more cases. (Fixes ticket#563)
* Correct "BM25Weight" to "TradWeight" in exception message from TradWeight.
* We were failing to call init() for user-defined Weight objects providing the
term-independent weight. These now get called with init(0.0).
* Xapian::Auto::open_stub() now throws a Xapian::DatabaseOpeningError exception
if the stub file can't be opened. Previously we failed to check for this
condition, which resulted in us treating the file as empty.
testsuite:
* When the testsuite is using valgrind, we used to run remote servers under
valgrind too (but with --tool=none) to get consistent behaviour as valgrind's
emulation of x87 excess precision isn't exact. Now we only do this if x87 FP
instructions are actually in use (which means x86 architecture and configure
run with --disable-sse).
* Make sure XAPIAN_MAX_CHANGESETS gets unset after replication testcases which
set it, so further testcases don't waste time generating changesets.
* Improved test coverage (including more tests for closed databases -
ticket#337).
brass backend:
* After closing the database, methods which try to use the termlist would throw
FeatureUnavailableError with message "Database has no termlist", assuming
that the termlist table not being open meant it wasn't present. Fix to check
if the postlist_table is open to determine which case we're in.
chert backend:
* After closing the database, methods which try to use the termlist would throw
FeatureUnavailableError with message "Database has no termlist", assuming
that the termlist table not being open meant it wasn't present. Fix to check
if the postlist_table is open to determine which case we're in.
inmemory backend:
* Check if the database is closed in metadata_keys_begin() for InMemory
Databases.
build system:
* xapian-config: Don't interpret a missing .la file as meaning that we only
have static libraries.
documentation:
* Fix API documentation for Query constructors - both XOR and ELITE_SET can
take any number of subqueries, not only exactly two.
* Backport missing API documentation comments for operator++ and operator*
methods or PositionIterator, PostingIterator and TermGenerator.
* docs/replication.rst: Update documentation - since 1.2.5, the value of
XAPIAN_MAX_CHANGESETS determines how many changesets we keep.
* docs/admin_notes.rst: Correction - we don't "create a lock file", we "lock a
file".
* Fix API documentation for TradWeight constructor - "k1" should be "k".
portability:
* configure: Overhaul handling of compilers which pretend to be GCC. Clang
is now detected, and we only pass it warning flags it actually understands.
And we now check for symbol visibility support with Intel's compiler.
* configure: Solaris automatically pulls in library dependencies, so set
link_all_deplibs_CXX=no there.
* configure: We now check -Bsymbolic-functions for all compilers.
* configure: Enable -Wdouble-promotion for GCC >= 4.6.
* Pass -ldl last when compiling zlib-vg.so, as that seems to be needed on
Ubuntu 12.04.
* Fix incorrect use of "delete" which should be "delete []". This is
undefined behaviour in C++, though the type is POD, so in practice this
probably worked OK on many platforms.
* In BM25Weight when k1 or b is zero (not the default), we used to multiply
an uninitialised double by zero, which is undefined behaviour, but in
practice will often give zero, leading to the desired results.
* xapian.h: Add check for Qt headers being included before us and defining
'slots' as a macro - if they are, give a clear error advising how to work
around this (previously compilation would fail with a confusing error).
Xapian-core 1.2.10 (2012-05-09):
API:
testsuite:
* apitest: Extend tradweight1 to test that TradWeight(0) means that wdf and
document length don't affect the weight of a term.
* termgentest: Check that TermGenerator discards words > 64 bytes.
matcher:
* Don't count unweighted subqueries of MultiAndPostList in percentage
calculations, as OP_FILTER maps to MultiAndPostList now. (ticket#590)
brass backend:
* When compacting, if the output database is empty, don't write out a metainfo
tag. Take care not to divide by zero when computing the percentage size
change for a table.
chert backend:
* When compacting, if the output database is empty, don't write out a metainfo
tag. Take care not to divide by zero when computing the percentage size
change for a table.
documentation:
* API documentation:
+ Note version when Database::close() was added.
+ Fix switched lower and upper in API documentation for Weight methods
get_doclength_lower_bound() and get_doclength_upper_bound(). Correct
maximum to minimum in get_doclength_lower_bound() comment and note that this
excludes zero length documents. Fix "An lower" to "A lower".
* docs/admin_notes.html: Mention that postlist and termlist tables also hold
value info for chert. Mention that xapian-chert-update was removed in 1.3.0.
Mention that you need to use copydatabase from 1.2.x to convert flint to
chert.
* HACKING: Update section on patches to mention git (git diff and git
format-patch), and using "-r" with normal diff, and also that ptardiff offers
a nice way to diff against an unpacked tarball.
debug code:
* Fix use of AssertEq() on NULL, which doesn't compile, at least with recent
GCC.
Xapian-core 1.2.9 (2012-03-08):
API:
* QueryParser: Fix FLAG_AUTO_SYNONYMS not to enable auto multi-word synonyms
too (but in a different way to trunk so as to not break the ABI).
matcher:
* Fix issue with running AND, OR and XOR queries against a database with no
documents in it - this was leading to a divide by zero, which led to
MSet::get_matches_estimated() reporting 2147483648 on i386.
build system:
* Remove configure's --with-stlport and --with-stlport-compiler options, as
they don't allow you to actually specify what you need to (at least to use
the Debian STLport package), and instead document what to pass to configure
to enable building with STLport (though it seems to no longer be actively
maintained, and the debug mode (which is probably the most interesting
feature now) doesn't seem to work on Debian stable).
documentation:
* Document that OP_ELITE_SET with non-term subqueries might pick subqueries
which don't match anything. Closes ticket#49.
* Document that you can define a static operator delete method in your subclass
if deallocation needs to be handled specially. (Closes ticket#554)
* Assorted minor documentation improvements.
portability:
* Address new warnings from GCC 4.6.
* Fix argument order when linking xapian-check to fix mingw build.
(ticket#567)
* Add some missing explicit header includes to fix build with STLport.
1.2.8:
API:
* Add support to TermGenerator and QueryParser for indexing and searching CJK
text using n-grams. Currently this is only enabled when the environmental
variable XAPIAN_CJK_NGRAM is set to a non-empty value.
portability:
+ Some fixes for warnings when cross-compiling to mingw.
* tests/soaktest/soaktest.cc: With Sun's compiler, random() and srandom()
aren't in <cstdlib> so we need to use <stdlib.h> instead.
1.2.7:
API:
* Document objects now track whether any document positions have been modified
so that replacing a modified document can completely skip considering
updating positions if none have changed. Currently the flint, chert, and
brass backends implement this optimisation. A common case this speeds up is
adding and/or removing boolean filter terms to/from existing documents - for
example this gives an 18% speedup for adding tags in notmuch.
portability:
* Fix -Wshadow warnings from GCC 4.6.
* Fix warning from GCC 3.3.
1.2.6:
API:
* QueryParser:
+ Add new set_max_wildcard_expansion() method to allow limiting the number of
terms a wildcard can expand to. (ticket#350)
+ If default_op is OP_NEAR or OP_PHRASE then disable stemming of the terms,
since we don't index positional information for stemmed terms by default.
* Spelling correction was failing to correctly handle words which had the same
trigram in an even number of times.
portability:
* Fix to build for mingw.
1.2.5:
API:
* Enquire::get_eset() now accepts a min_wt argument to allow the minimum wanted
weight to be specified. Default is 0, which gives the previous behaviour.
* QueryParser: Handle NEAR/<offset> and ADJ/<offset> where offset isn't an
integer the same way at the end of the query as in the middle.
* Replication:
+ Only keep $XAPIAN_MAX_CHANGESETS changeset files when generating a new one
(previously this variable only controlled if we generated changesets or
not). Closes ticket#278.
+ $XAPIAN_MAX_CHANGESETS is reread each time, rather than only when the
database is opened.
+ If you build Xapian with DANGEROUS mode enabled, changeset files now
actually have the appropriate flag set (the reader will currently throw an
exception, but that's better than quietly handling them incorrectly).
portability:
* api/compactor.cc: Add missing header <ctime> for time() (ticket#530).
* api/compactor.cc: Use msvc_posix_rename() under __WIN32__ to atomically
update stub file after compaction (ticket#525).
* Fix uninitialised variable warnings with gcc -O3.
* Eliminate std::string member of global static object used when compiled with
--enable-log which was causes problems on Mac OS X.
* Fix some issues highlighted by clang++ warnings.
1.2.4:
API:
* QueryParser:
+ Avoid a double free if Query construction throws an exception in a
particular case. Fixes ticket#515.
+ Allow phrase generators between a probabilistic prefix and the term itself
(e.g. path:/usr/local).
+ The correct window size wasn't being set in some cases when default_op was
set to OP_PHRASE.
* Enquire::get_mset():
+ Avoid pointlessly trying to allocate lots of memory if the first document
requested is larger than the size of the database.
+ An empty query now returns an MSet with firstitem set correctly -
previously firstitem was always 0 in this case.
* Document: Initialise docid to 0 when creating a document from
scratch, as documented.
* Compactor:
+ Move the database compaction and merging functionality into this new class,
and make xapian-compact a simple wrapper around this class. (ticket#175)
+ Inputs can now be stub database directories or files, in which case the
databases in the stub are used as inputs.
+ Add support for compacting to a stub database, which can be one of the
inputs (for atomic update).
+ If spellings and/or synonyms were only present in some source databases,
they weren't copied to the output database, but now they are.
portability:
* configure: Add support for --enable-sse=sse and --enable-sse=sse2 to allow
control of which SSE instructions to use.
* configure: Enable use of SSE maths on x86 by default with Sun's compiler.
* configure: Beef up the test for whether -lm is required and add a special
case to force it to be for Sun's C++ compiler - there's some interaction with
libtool and/or shared objects which means that the previous configure test
didn't think -lm is needed here when it is.
* Fix to build on OpenBSD 4.5 with GCC 3.3.5.
* Need to avoid excess precision on m68k when targeting models 68010, 68020,
68030 as well as 68000.
* Fix compilation with Sun's C++ compiler.
* Fix testsuite to build on Solaris < 10.
1.2.3:
API:
* Database::get_spelling_suggestion() will now suggest a correction even if the
passed word is in the dictionary, provided the correction has at least the
same frequency. Partly addresses #225.
* QueryParser:
+ Fix handling of groups of terms which are all stopwords - in situations
where this causes a problem we now disable stopword checks for such groups.
(ticket#245)
+ Fix to be smarter about handling a boolean filter term containing ".." in
the presence of valuerangeprocessors.
portability:
* configure: Don't pass -mtune=generic unless GCC >= 4.2 is in use
(ticket#492).
* Fix handling of some obscure cases of resolving relative paths on Microsoft
Windows. (ticket#243).
* Optimise closing of all unwanted file descriptors after forking by using
closefrom() if available, and otherwise providing our own implementation
(optimised to some extent for many platforms).
* Fix test harness to build under Microsoft Windows (ticket#495).
1.2.2:
portability:
* Revert 1.2.1 change to visibility of Xapian::Weight's copy constructor as
it making it private broke compilation with GCC 4.1 (which seems to be a
bug in this compiler version).
* tests/harness/testsuite.cc: Need <cstdio> for sprintf(). Fixes compilation
error which was masked if valgrind was installed. (ticket#489)
pkgsrc changes:
* Remove options (the "quartz" backend was unrelated to Darwin and
no longer exists).
* Unconditionally buildlink libuuid. If that's overzealous, improve
its builtin detection.
Xapian-core 1.0.18 (2009-02-14):
API:
* Document: Add new add_boolean_term() method, which is an alias for add_term()
with wdfinc=0.
* QueryParser:
+ Add support for quoting boolean terms so they can contain arbitrary
characters (partly addresses ticket#128).
+ Add ENCLOSING_MARK and COMBINING_SPACING_MARK categories, plus several
zero-width space characters, as phrase generators. This mirrors a better
fix in 1.1.4, but without losing compatibility with existing databases.
+ Fix handling of an explicit AND before a hated term (foo AND -bar).
(ticket#447)
* TermIterator: Only include trailing '+' or '#' on a term if it isn't followed
by a word character (makes more sense and matches QueryParser's behaviour).
(ticket#446)
* Database: Fix many methods to behave better on a database with no
subdatabases, such as is constructed by Database(). Fixes ticket#415.
testsuite:
* Add test coverage for xapian-compact, and improve coverage for
WritableDatabase::replace_document().
* apitest: Rename matchfunctor<n> to matchdecider<n> to match current
terminology.
flint backend:
* When updating documents, don't update posting entries which haven't changed.
Largely fixes ticket #250.
* If the number of entries in the position table happened to be 4294967296 or
an exact multiple, Xapian would ignore positional data for that table when
running queries.
* Iterating all the terms in the database with a prefix is now slightly more
efficient.
* Fix locking code to work if stdin and/or stdout have been closed.
* If a document is replaced with itself unmodified, we no longer increase the
automatic flush counter.
* When iterating a posting list modified since the last flush(), the reported
wdf is now correct (previously it was too high by its old value).
* Replacing a document deleted since the last flush failed to update the
collection frequency and wdf, and caused an assertion failure when assertions
were enabled.
* WritableDatabase::replace_document() didn't always remove old positional
data (the only effect is that the position table was bloated by unwanted
entries).
* xapian-inspect:
+ New "until" command which shows entries until a specified key is reached.
+ New "open" command which allows easy switching between tables.
* xapian-compact: Fix typos in --help output.
quartz backend:
* Replacing a document deleted since the last flush failed to update the
collection frequency and wdf, and caused an assertion failure when assertions
were enabled.
* WritableDatabase::replace_document() didn't always remove old positional
data (the only effect is that the position table was bloated by unwanted
entries).
remote backend:
* Throw UnimplementedError if a MatchDecider is used with the remote backend.
Previously Xapian returned incorrect results in this case.
build system:
* configure: With --enable-maintainer-mode, enable -Werror for GCC >= 4.1
rather than >= 4.0 as Apple's GCC 4.0 gives bogus uninitialised variable
warnings.
documentation:
* The API documentation now includes Xapian::Error and subclasses, and doesn't
mention Xapian::Query::Internal.
* Make clear in the Xapian::Document API documentation that this class is a
lazy handle and discuss the issues this can cause.
* INSTALL: Improve text about zlib dependency.
* HACKING: Add details of our licensing policy for accepting patches.
examples:
* quest: If no database is specified, still parse the query and report
Query::get_description() to provide an easy way to check how a query parses.
portability:
* Fix GCC 4.2 warning.
xapian-core 1.0.17 (2009-11-18):
API:
* QueryParser:
+ Fix handling of a group of two or more terms which are all stopwords which
notably caused issues when default_op was OP_AND, but could probably
manifest in other cases too. Fixes ticket#406.
+ Fix interaction of FLAG_PARTIAL and FLAG_SYNONYM. (ticket#407)
* Database: A database created via the default constructor no longer causes a
segfault when the methods get_metadata() or metadata_keys_begin() are called.
flint backend:
* Don't try to close the fd one more than the maximum allowable when locking
the database. Harmless, except it causes a warning when running under
valgrind. (ticket#408)
remote backend:
* Xapian::Sorter isn't supported with the remote backend so throw
UnimplementedError rather than giving incorrect results. (ticket#384)
* Fix potential reading off the end of the MSet which is returned internally
by the remote server.
documentation:
* Various documentation comment improvements for the Database class.
examples:
* examples/quest.cc: Tighten up the type of the error we catch to detect an
unknown stemming language.
portability:
* xapian-config: Need to quote ^ for Solaris /bin/sh.
* configure: Actually use any flags we determine are needed to switch the
compiler to proper ANSI C++ mode, when building xapian-core - this stopped
working in 1.0.12, breaking support for HP's aCC, Compaq's cxx, Sun's CC, and
SGI's CC.
schmonz
joerg
ryoon
wiz
wiedi
agc
jaapb
asau
dholland
sbd
abs