on-the-fly, and thereby avoid blowing out memory when the planner has
underestimated the hash table size. Hash join will now obey the
work_mem limit with some faithfulness. Per my recent proposal
(hash aggregate part isn't done yet though).
the freelist, plus per-buffer spinlocks that protect access to individual
shared buffer headers. This requires abandoning a global freelist (since
the freelist is a global contention point), which shoots down ARC and 2Q
as well as plain LRU management. Adopt a clock sweep algorithm instead.
Preliminary results show substantial improvement in multi-backend situations.
of AND and OR clauses. The key point here is that an OR on the
predicate side has to be treated gingerly: we may be able to prove
that the OR is implied even when no one of its components is implied.
For example (x OR y) implies (x OR y OR z) even though no one of x,
y, or z can be individually proven. This code handles both the
example shown recently by Sergey Koshcheyev and the one shown last
October by Dawid Kuroczko.
no held locks. This maintains the invariant that proclocks are present
only for procs that are holding or awaiting a lock; when this is not
true, LockRelease will fail. Per report from Stephen Clouse.
indexscans involving partial indexes. These would always be dominated
by a simple indexscan on such an index, so there's no point in considering
them. Fixes overoptimism in a patch I applied last October.
it was in 7.4, and add some comments explaining why it has to be this way.
I broke it for OR'd index predicates in a fit of code cleanup last summer.
Per example from Sergey Koshcheyev.
in favor of looking at the flat file copy of pg_database during backend
startup. This should finally eliminate the various corner cases in which
backend startup fails unexpectedly because it isn't able to distinguish
live and dead tuples in pg_database. Simplify locking on pg_database
to be similar to the rules used with pg_shadow and pg_group, and eliminate
FlushRelationBuffers operations that were used only to reduce the odds
of failure of GetRawDatabaseInfo.
initdb forced due to addition of a trigger to pg_database.
implement the md5() SQL-level function). The old code did the
following:
1. de-toast the datum
2. convert it to a cstring via textout()
3. get the length of the cstring via strlen()
Since we are treating the datum context as a blob of binary data,
the latter two steps are unnecessary. Once the data has been
detoasted, we can just use it as-is, and derive its length from
the varlena metadata.
This patch improves some run-of-the-mill md5() computations by
just under 10% in my limited tests, and passes the regression tests.
I also noticed that md5_text() wasn't checking the return value
of md5_hash(); encountering OOM at precisely the right moment
could result in returning a random md5 hash. This patch corrects
that. A better fix would be to make md5_hash() only return on
success (and/or allocate via palloc()), but since it's used in
the frontend as well I don't see an easy way to do that.
during flat-file writing. The only difference is that SnapshotSelf
would consider tuples of the 'current command' within the current
transaction as valid, where SnapshotNow wouldn't. We can eliminate
the need for this with one extra CommandCounterIncrement call before
we start reading the catalogs.
the AMI_OVERRIDE flag. The fact that TransactionLogFetch treats
BootstrapTransactionId as always committed is sufficient to make
bootstrap work, and getting rid of extra tests in heavily used code
paths seems like a win. The files produced by initdb are demonstrably
the same after this change.
file now identifies group members by usesysid not name; this avoids
needing to depend on SearchSysCache which we can't use during startup.
(The old representation was entirely broken anyway, since we did not
regenerate the file following RENAME USER.) It's only a 95% solution
because if the group membership list is big enough to be toasted out
of line, we cannot read it during startup. I think this will do for
the moment, until we have time to implement the planned pg_role
replacement for pg_group.
in GetNewTransactionId(). Since the limit value has to be computed
before we run any real transactions, this requires adding code to database
startup to scan pg_database and determine the oldest datfrozenxid.
This can conveniently be combined with the first stage of an attack on
the problem that the 'flat file' copies of pg_shadow and pg_group are
not properly updated during WAL recovery. The code I've added to
startup resides in a new file src/backend/utils/init/flatfiles.c, and
it is responsible for rewriting the flat files as well as initializing
the XID wraparound limit value. This will eventually allow us to get
rid of GetRawDatabaseInfo too, but we'll need an initdb so we can add
a trigger to pg_database.
that return tuples (such as EXPLAIN). Per gripe from Michael Fuhr.
Side effect: fix an old bug that unintentionally disabled backward scans
for all SPI-created cursors.
column with a default expression. In that situation, we need to rewrite
the heap relation. To evaluate the new default expression, we use
ExecEvalExpr(); however, this can allocate memory in the current memory
context, and ATRewriteTable() does not switch out of the active portal's
heap memory context. The end result is a rather large memory leak (on
the order of gigabytes for a reasonably sized table).
This patch changes ATRewriteTable() to switch to the per-tuple memory
context before beginning the per-tuple loop. It also removes an explicit
heap_freetuple() in the loop, since that is no longer needed.
In an unrelated change, I noticed the code was scanning through the
attributes of the new tuple descriptor for each tuple of the old table.
I changed this to use precomputation, which should slightly speed up
the loop.
Thanks to steve@deefs.net for reporting the leak.
there are corner cases involving dropping toasted columns in which the
previous coding would fail, too: the new version of the table might not
have any TOAST table, but we'd still propagate possibly-wide values of
dropped columns forward.
form of CASE (eg, CASE 0 WHEN 1 THEN ...) can be constant-folded as it
was in 7.4. Also, avoid constant-folding result expressions that are
certainly unreachable --- the former coding was a bit cavalier about this
and could generate unexpected results for all-constant CASE expressions.
Add regression test cases. Per report from Vlad Marchenko.
tests. Contributed by Koju Iijima, review from Neil Conway, Gavin Sherry
and Tom Lane.
Also, fix error in description of WITH CHECK OPTION clause in the CREATE
VIEW reference page: it should be "CASCADED", not "CASCADE".
estimate to less than the number of values estimated for any one grouping
Var, as suggested by Manfred. This is intuitively right, and what's
more it puts the plan choices in the subselect regression test back the
way they were before ...
clamp the estimated number of groups to table row count over 10, instead
of table row count; this reflects a heuristic that people probably won't
group over a near-unique set of columns, and the knowledge that we don't
currently have any way to estimate the correlation of the columns better
than guessing. This change creates a trivial plan change in one of the
regression tests.
look at the actual aggregate transition datatypes and the actual overhead
needed by nodeAgg.c, instead of using pessimistic round numbers.
Per a discussion with Michael Tiemann.
command. This is useful because we can allow truncation of tables
referenced by foreign keys, so long as the referencing table is
truncated in the same command.
Alvaro Herrera
to avoid problems when a cursor depends on objects created or changed in
the same subtransaction. We'd like to do better someday, but this seems
the only workable answer for 8.0.1.
(1) Keep a pin on the scan's current buffer and mark buffer. This
avoids the need to do a ReadBuffer() for each tuple produced by the
scan. Since ReadBuffer() is expensive, this is a significant win.
(2) Convert a ReleaseBuffer(); ReadBuffer() pair into
ReleaseAndReadBuffer(). Surely not a huge win, but it saves a lock
acquire/release...
(3) Remove a bunch of duplicated code in rtget.c; make rtnext() handle
both the "initial result" and "subsequent result" cases.
(4) Add support for index tuple killing
(5) Remove rtscancache(): it is dead code, for the same reason that
gistscancache() is dead code (an index scan ought not be invoked with
NoMovementScanDirection).
The end result is about a 10% improvement in rtree index scan perf,
according to contrib/rtree_gist/bench.
got it wrong when the JOIN was in an outer query level. Per example from
Laurie Burrow. Also fix same issue in markTargetListOrigin. I think the
latter is only a latent bug since we currently don't apply markTargetListOrigin
except at the outer level ... but should do it right anyway.
CASE 'a' WHEN 'a' THEN 1 ELSE 2 END. This worked in 7.4 and before
but had been broken due to premature freezing of the type of the test
expression. Per gripe from GÄbor SzÃcs.
so that we can get the size of a shared inval message back down to what it
was in 7.4 (and simplify the logic too). Phase 2 of fixing the
'SMgrRelation hashtable corrupted' problem.
is the minimum required fix. I want to look next at taking advantage of
it by simplifying the message semantics in the shared inval message queue,
but that part can be held over for 8.1 if it turns out too ugly.
releases, a nonzero 'c' argument meant that the input string could be
terminated by either that character or \0. Recent refactoring broke
that, causing the thing to scan for 'c' only. This went undetected
because no part of the main code actually passes nonzero 'c'. However
it broke tsearch2 and possibly other user-written code that assumed
the old definition. Per report from Tom Hebbron.
discussion on pgsql-hackers-win32 list. Documentation still needs to
be tweaked --- I'm not sure how to refer to the APPDATA folder in
user documentation.
share lock on a buffer being written out before releasing BufMgrLock in
the BufferAlloc code path; if we do it later we might block on someone
who's re-pinned the buffer. I believe this is only an issue for BufferAlloc
and not the other places that call FlushBuffer. BufferSync must continue
to do it the old way since it may well be trying to write buffers that
other backends have pinned; but it should not be holding any conflicting
locks. FlushRelationBuffers is okay since it's got exclusive lock at the
relation level.
Also performed an initial run through of upgrading our Copyright date to
extend to 2005 ... first run here was very simple ... change everything
where: grep 1996-2004 && the word 'Copyright' ... scanned through the
generated list with 'less' first, and after, to make sure that I only
picked up the right entries ...
to shared memory as soon as possible, ie, right after read_backend_variables.
The effective difference from the original code is that this happens
before instead of after read_nondefault_variables(), which loads GUC
information and is apparently capable of expanding the backend's memory
allocation more than you'd think it should. This should fix the
failure-to-attach-to-shared-memory reports we've been seeing on Windows.
Also clean up a few bits of unnecessarily grotty EXEC_BACKEND code.
that is, files are sought in the same directory as the referencing file.
Also allow absolute paths in @file constructs. Improve documentation
to actually say what is allowed in an included file.
the remainder of the current clog page during system startup. While
this was a good idea, it turns out the code fails if nextXid is
exactly at a page boundary, because we won't have created the "current"
clog page yet in that case. Since the page will be correctly zeroed
when we execute the first transaction on it, the solution is just to
do nothing when exactly at a page boundary. Per trouble report from
Dave Hartwig.
its presence. This amounts to desupporting Kerberos 5 releases 1.0.*,
which is small loss, and simplifies use of our Kerberos code on platforms
with Red-Hat-style include file layouts. Per gripe from John Gray and
followup discussion.
silently ignored, allowing one to write bizarre things like
DECLARE x setof int;
in plpgsql. This has misled at least one novice into thinking that
plpgsql variables could be sets ...
had to do in DECLARE CURSOR. AFAICS these are all the places affected.
PREPARE case per example from Michael Fuhr, EXPLAIN case located by
grepping for planner calls ...
(rd_att) field of a nailed-in-cache relcache entry. This fixes the bug
reported by Alvaro 8-Dec-2004; I believe it probably also explains
Grant Finnemore's report of 10-Sep-2004.
In an unrelated change in the same file, put back 7.4's response to
failure to rename() the relcache init file, ie, unlink the useless
temp file. I did not put back the warning message, since there might
actually be some reason not to have that.
of an inheritance child table is binary-compatible with the rowtype of
its parent, invent an expression node type that does the conversion
correctly. Fixes the new bug exhibited by Kris Shannon as well as a
lot of old bugs that would only show up when using multiple inheritance
or after altering the parent table.
is null-terminated. I think this is not a real bug because the parser
would always have truncated the identifier to NAMEDATALEN-1 already,
but let's be safe. Per report from Klocwork.
> throughout to the spellings suggested by your book.
Great.
A follow-up patch for current CVS HEAD is attached, and available at
http://troels.arvin.dk/db/pgsql/conformance/pgsql-sql-conformance-
followup.patch
The patch
- includes a core feature ID that had been left
out by mistake (C011)
- updates the sql_feature_packages.txt table to
reflect changes in SQL:2003 which were not
covered properly in my last patch
Troels Arvin
> seconds to 10 seconds. The original number was plucked from thin air
> some months ago, and I'd like to review that now based upon further
> thought, observation and experience.
>
> This change has little or no effect on performance, since the interval
> is there mainly to avoid repeated respawn attempts if archiver fails at
> startup. Archiver start-up time is very quick, so there is little danger
> of exceeding 10 seconds.
>
> On a busy system, if the archiver does die, then many files can build up
> in the 60 seconds before respawning. That xlog file backlog could take
> some time to clear. This then leaves a larger than normal window of data
> loss for a possibly long period.
>
> It's a minor change only, with no other effect on function.
Simon Riggs
the "ps" argument list on Unix - meaning that there is no way to
identify for example the stats processors or the bgwriter.
This patch adds this functionality, in a bit of a crufty way. It creates
a kernel Event object with the name of what would be in the title. This
can be viewed using for example Process Explorer.
It's been very handy for me during both debugging and using. I haven't
figured a better way, but perhaps someone has one that's less crufty? If
not, here is at least a working patch :-)
Magnus Hagander
reasons I outlined in pghackers a few days ago.
Also, undo someone's overly optimistic decision to reduce tuple state
checks from if (...) elog() to Asserts. If I trusted this code more,
I might think it was a good idea to disable these checks in production
installations. But I don't.
escapes --- they aren't simply quoted characters. Problem noted by
Antti Salmela. Also fix problem with incorrect handling of multibyte
characters when followed by a quantifier.
In particular, there was a mathematical tie between the two possible
nestloop-with-materialized-inner-scan plans for a join (ie, we computed
the same cost with either input on the inside), resulting in a roundoff
error driven choice, if the relations were both small enough to fit in
sort_mem. Add a small cost factor to ensure we prefer materializing the
smaller input. This changes several regression test plans, but with any
luck we will now have more stability across platforms.
a relation's number of blocks, rather than the possibly-obsolete value
in pg_class.relpages. Scale the value in pg_class.reltuples correspondingly
to arrive at a hopefully more accurate number of rows. When pg_class
contains 0/0, estimate a tuple width from the column datatypes and divide
that into current file size to estimate number of rows. This improved
methodology allows us to jettison the ancient hacks that put bogus default
values into pg_class when a table is first created. Also, per a suggestion
from Simon, make VACUUM (but not VACUUM FULL or ANALYZE) adjust the value
it puts into pg_class.reltuples to try to represent the mean tuple density
instead of the minimal density that actually prevails just after VACUUM.
These changes alter the plans selected for certain regression tests, so
update the expected files accordingly. (I removed join_1.out because
it's not clear if it still applies; we can add back any variant versions
as they are shown to be needed.)
prevents problems when the DECLARE is in a portal and is executed
repeatedly, as is possible in v3 protocol. Per analysis by Oliver
Jowett, though I didn't use his patch exactly.
error conditions during regexp compile, but not during regexp execution;
any sort of "can't happen" errors would be treated as no-match instead
of being reported as they should be. Noticed while trying to duplicate
a reported Tcl bug.
to be processed by GUC before InitPostgres, because any required lookup
of the encoding conversion function has to be done during InitializeClientEncoding.
So, I broke this last week by moving GUC processing to after InitPostgres :-(.
What we can do as a compromise is process non-SUSET variables during
command line scanning (the same as before), and postpone the processing
of only SUSET variables. None of the SUSET variables need to be set
before InitPostgres.
fill factor has been exceeded. We usually run with ffactor == 1, but
the way the test was coded, it wouldn't split a bucket until the actual
fill factor reached 2.0, because of use of integer division. Change
from > to >= so that it will split more aggressively when the table
starts to get full.
few cycles during transaction exit. A typical session probably
wouldn't have as many as half a dozen portals open at once, so the
original value of 64 seems far larger than needed.
no need for it to be nearly as big as the global hash table, and since
it's not in shared memory it can grow if it does need to be bigger.
By reducing the size, we speed up hash_seq_search(), which saves a
significant fraction of subtransaction entry/exit overhead.
to the original List; per report from Sebastian BÎck. I think this is
the last such bug --- I examined every lcons() call in the backend and
the rest seem OK --- but it's nervous-making that we're still finding
'em so many months after the List rewrite went in.
collector until the transaction commits. Per recent discussion, this
should avoid confusing autovacuum when an updating transaction runs for
a long time.
this is to avoid scenarios where incoming backends find no live copies
of a database's row because the only live copy is in an as-yet-unwritten
shared buffer, which they can't see. Also, use FlushRelationBuffers()
for forcing out pg_database, instead of the much more expensive BufferSync().
There's no need to write out pages belonging to other relations.
Rather than using ReadBuffer() to increment the reference count on an
already-pinned buffer, we should use IncrBufferRefCount() as it is
faster and does not require acquiring the BufMgrLock.
even uglier than it was already :-(. Also, on Windows only, use temporary
shared memory segments instead of ordinary files to pass over critical
variable values from postmaster to child processes. Magnus Hagander
more than 65K columns, or when the created table has more than 65K columns
due to adding inherited columns from parent relations. Fix a similar
crash when processing SELECT queries with more than 65K target list
entries. In all three cases we would eventually detect the error and
elog, but the check was being made too late.
We don't really want to start a new SPI connection, just keep using the old
one; otherwise we have memory management problems as illustrated by
John Kennedy's bug report of today. This requires a bit of a hack to
ensure the SPI stack state is properly restored, but then again what we
were doing before was a hack too, strictly speaking. Add a regression
test to cover this case.
plain SUSET instead. Also delay processing of options received in
client connection request until after we know if the user is a superuser,
so that SUSET values can be set that way by legitimate superusers.
Per recent discussion.
buffer is valid, as ReadBuffer() will elog on error. Most of the call
sites of ReadBuffer() got this right, but this patch fixes those call
sites that did not.
> pg specific, like "PostgreSQL.1". I have not done this since a new compile
> would not detect a running old beta. But now would be the time (or never).
Zeugswetter Andreas
shared memory segment ID. If we can't access the existing shmem segment,
it must not be relevant to our data directory. If we can access it,
then attach to it and check for an actual match to the data directory.
This should avoid some cases of failure-to-restart-after-boot without
introducing any significant risk of failing to detect a still-running
old backend.
estimates when combining the estimates for a range query. As pointed out
by Miquel van Smoorenburg, the existing check for an impossible combined
result would quite possibly fail to detect one default and one non-default
input. It seems better to use the default range query estimate in such
cases. To do so, add a check for an estimate of exactly DEFAULT_INEQ_SEL.
This is a bit ugly because it introduces additional coupling between
clauselist_selectivity and scalarltsel/scalargtsel, but it's not like
there wasn't plenty already...
working as intended --- for some reason, FROM a.b.c was getting
parsed as if it were a function name and not a qualified name.
I think there must be a bug in bison, because it should have
complained that the grammar was ambiguous. Anyway, fix it along
the same lines previously used for func_name vs columnref, and get
rid of the right-recursion in attrs that seems to have confused
bison.
type-and-length coercion function, make sure that the coercion function
is told the correct typmod. Fixes Kris Jurka's example of a domain
over bit(N).
clause implicitly whenever one is not given explicitly. Remove concept
of a schema having an associated tablespace, and simplify the rules for
selecting a default tablespace for a table or index. It's now just
(a) explicit TABLESPACE clause; (b) default_tablespace if that's not an
empty string; (c) database's default. This will allow pg_dump to use
SET commands instead of tablespace clauses to determine object locations
(but I didn't actually make it do so). All per recent discussions.
to DAY precision or coarser; leave the timezone alone when precision is
HOUR or less. This avoids surprises for inputs near a DST transition
time, as per example from Matthew Gabeler-Lee. (The only reason we
recalculate at all is so that outputs that are supposed to represent
days will come out as local midnight, and that's not relevant for sub-day
precision.)
use it, as per my proposal of yesterday. This gives us a means of
determining the zone offset to impute to an unlabeled timestamp that
is both efficient and reliable, unlike all our previous tries involving
mktime() and localtime(). The behavior for invalid or ambiguous times
at a DST transition is fixed to be really and truly "assume standard
time", fixing a bug that has come and gone repeatedly but was back
again in 7.4. (There is some ongoing discussion about whether we should
raise an error instead, but for the moment I'll make it do what it was
previously intended to do.)
Tom Lane
Neil Conway
Bruce Momjian
Peter Eisentraut
PostgreSQL Daemon