OSDN Git Service

pg-rex/syncrep.git
17 years agoSpecifying a dbname should override the default database, not add to it.
Magnus Hagander [Tue, 12 Jun 2007 13:54:58 +0000 (13:54 +0000)]
Specifying a dbname should override the default database, not add to it.
Fixes buildfarm failures on contribcheck.

17 years agoFix missing variable initialization.
Magnus Hagander [Tue, 12 Jun 2007 13:26:45 +0000 (13:26 +0000)]
Fix missing variable initialization.

17 years agoFix search/replace error in makefile for ecpg tests
Magnus Hagander [Tue, 12 Jun 2007 13:09:58 +0000 (13:09 +0000)]
Fix search/replace error in makefile for ecpg tests

17 years agoFixed reduce/recuce problem due to CURRENT_P.
Michael Meskes [Tue, 12 Jun 2007 11:32:30 +0000 (11:32 +0000)]
Fixed reduce/recuce problem due to CURRENT_P.

17 years agoRewrite ECPG regression test driver in C, by splitting the standard
Magnus Hagander [Tue, 12 Jun 2007 11:07:34 +0000 (11:07 +0000)]
Rewrite ECPG regression test driver in C, by splitting the standard
regression driver into two parts and reusing half of it. Required to
run ECPG tests without a shell on MSVC builds.

Fix ECPG thread tests for MSVC build (incl output files).

Joachim Wieland and Magnus Hagander

17 years agoSynced parser and keyword list.
Michael Meskes [Tue, 12 Jun 2007 07:55:56 +0000 (07:55 +0000)]
Synced parser and keyword list.

17 years agoImprove UPDATE/DELETE WHERE CURRENT OF so that they can be used from plpgsql
Tom Lane [Mon, 11 Jun 2007 22:22:42 +0000 (22:22 +0000)]
Improve UPDATE/DELETE WHERE CURRENT OF so that they can be used from plpgsql
with a plpgsql-defined cursor.  The underlying mechanism for this is that the
main SQL engine will now take "WHERE CURRENT OF $n" where $n is a refcursor
parameter.  Not sure if we should document that fact or consider it an
implementation detail.  Per discussion with Pavel Stehule.

17 years agoMinor wording improvement.
Tom Lane [Mon, 11 Jun 2007 15:08:32 +0000 (15:08 +0000)]
Minor wording improvement.

17 years agoShould not try to unlink stdout.
Michael Meskes [Mon, 11 Jun 2007 12:01:23 +0000 (12:01 +0000)]
Should not try to unlink stdout.

17 years agoFixed one memory leak in descriptor code.
Michael Meskes [Mon, 11 Jun 2007 11:52:08 +0000 (11:52 +0000)]
Fixed one memory leak in descriptor code.
Made sure ecpg deletes output file in case of an error.

17 years agoDone:
Bruce Momjian [Mon, 11 Jun 2007 01:51:50 +0000 (01:51 +0000)]
Done:

<  o Allow UPDATE/DELETE WHERE CURRENT OF cursor
<
<    This requires using the row ctid to map cursor rows back to the
<    original heap row. This become more complicated if WITH HOLD cursors
<    are to be supported because WITH HOLD cursors have a copy of the row
<    and no FOR UPDATE lock.
<    http://archives.postgresql.org/pgsql-hackers/2007-01/msg01014.php
<
>  o -Allow UPDATE/DELETE WHERE CURRENT OF cursor

17 years agoSupport UPDATE/DELETE WHERE CURRENT OF cursor_name, per SQL standard.
Tom Lane [Mon, 11 Jun 2007 01:16:30 +0000 (01:16 +0000)]
Support UPDATE/DELETE WHERE CURRENT OF cursor_name, per SQL standard.

Along the way, allow FOR UPDATE in non-WITH-HOLD cursors; there may once
have been a reason to disallow that, but it seems to work now, and it's
really rather necessary if you want to select a row via a cursor and then
update it in a concurrent-safe fashion.

Original patch by Arul Shaji, rather heavily editorialized by Tom Lane.

17 years agoTeach heapam code to know the difference between a real seqscan and the
Tom Lane [Sat, 9 Jun 2007 18:49:55 +0000 (18:49 +0000)]
Teach heapam code to know the difference between a real seqscan and the
pseudo HeapScanDesc created for a bitmap heap scan.  This avoids some useless
overhead during a bitmap scan startup, in particular invoking the syncscan
code.  (We might someday want to do that, but right now it's merely useless
contention for shared memory, to say nothing of possibly pushing useful
entries out of syncscan's small LRU list.)  This also allows elimination of
ugly pgstat_discount_heap_scan() kluge.

17 years agoInsert ORDER BY into a few regression test queries that now have unstable
Tom Lane [Sat, 9 Jun 2007 17:24:46 +0000 (17:24 +0000)]
Insert ORDER BY into a few regression test queries that now have unstable
results due to syncscan patch, when shared_buffers is small enough.  Per
buildfarm reports and some local testing with shared_buffers set to the
lowest value considered by initdb.

17 years agoAllow numeric_fac() to be interrupted, since it can take quite a while for
Tom Lane [Sat, 9 Jun 2007 15:52:30 +0000 (15:52 +0000)]
Allow numeric_fac() to be interrupted, since it can take quite a while for
large inputs.  Also cause it to error out immediately if the result will
overflow, instead of grinding through a lot of calculation first.
Per gripe from Jim Nasby.

17 years agoDisallow the cost balancing code from resulting in a zero cost limit, which
Alvaro Herrera [Fri, 8 Jun 2007 21:21:28 +0000 (21:21 +0000)]
Disallow the cost balancing code from resulting in a zero cost limit, which
causes a division-by-zero error in the vacuum code.  This can happen when there
are more workers than cost limit units.

Per report from Galy Lee in
<200705310914.l4V9E6JA094603@wwwmaster.postgresql.org>.

17 years agoAvoid passing zero as a value for vacuum_cost_limit, because it's not a valid
Alvaro Herrera [Fri, 8 Jun 2007 21:09:49 +0000 (21:09 +0000)]
Avoid passing zero as a value for vacuum_cost_limit, because it's not a valid
value for the vacuum code.  Instead, make zero signify getting the value from a
higher level configuration facility, just like -1 in the original coding.  We
still document that -1 is the value that disables the feature, to avoid
confusing the user unnecessarily.

Reported by Galy Lee in <200705310914.l4V9E6JA094603@wwwmaster.postgresql.org>;
per subsequent discussion.

17 years agoAdd note that LIMIT without ORDER BY can produce outright nondeterministic
Tom Lane [Fri, 8 Jun 2007 20:26:18 +0000 (20:26 +0000)]
Add note that LIMIT without ORDER BY can produce outright nondeterministic
results.  Necessary due to introduction of syncscan patch.

17 years agoDone:
Bruce Momjian [Fri, 8 Jun 2007 18:45:22 +0000 (18:45 +0000)]
Done:

< * Allow sequential scans to take advantage of other concurrent
> * -Allow sequential scans to take advantage of other concurrent
<
<   One possible implementation is to start sequential scans from the lowest
<   numbered buffer in the shared cache, and when reaching the end wrap
<   around to the beginning, rather than always starting sequential scans
<   at the start of the table.
<
<   http://archives.postgresql.org/pgsql-patches/2006-12/msg00076.php
<   http://archives.postgresql.org/pgsql-hackers/2006-12/msg00408.php
<   http://archives.postgresql.org/pgsql-hackers/2006-12/msg00784.php
<   http://archives.postgresql.org/pgsql-hackers/2007-03/msg00415.php
<

17 years agoArrange for large sequential scans to synchronize with each other, so that
Tom Lane [Fri, 8 Jun 2007 18:23:53 +0000 (18:23 +0000)]
Arrange for large sequential scans to synchronize with each other, so that
when multiple backends are scanning the same relation concurrently, each page
is (ideally) read only once.

Jeff Davis, with review by Heikki and Tom.

17 years agoRedefine IsTransactionState() to only return true for TRANS_INPROGRESS state,
Tom Lane [Thu, 7 Jun 2007 21:45:59 +0000 (21:45 +0000)]
Redefine IsTransactionState() to only return true for TRANS_INPROGRESS state,
which is the only state in which it's safe to initiate database queries.
It turns out that all but two of the callers thought that's what it meant;
and the other two were using it as a proxy for "will GetTopTransactionId()
return a nonzero XID"?  Since it was in fact an unreliable guide to that,
make those two just invoke GetTopTransactionId() always, then deal with a
zero result if they get one.

17 years agoRework temp_tablespaces patch so that temp tablespaces are assigned separately
Tom Lane [Thu, 7 Jun 2007 19:19:57 +0000 (19:19 +0000)]
Rework temp_tablespaces patch so that temp tablespaces are assigned separately
for each temp file, rather than once per sort or hashjoin; this allows
spreading the data of a large sort or join across multiple tablespaces.
(I remain dubious that this will make any difference in practice, but certain
people insisted.)  Arrange to cache the results of parsing the GUC variable
instead of recomputing from scratch on every demand, and push usage of the
cache down to the bottommost fd.c level.

17 years agoAvoid losing track of data for shared tables in pgstats. Report by Michael
Alvaro Herrera [Thu, 7 Jun 2007 18:53:17 +0000 (18:53 +0000)]
Avoid losing track of data for shared tables in pgstats.  Report by Michael
Fuhr, patch from Tom Lane after a messier suggestion by me.

17 years agoFix array_dims() example to be consistent with the data previously shown.
Tom Lane [Thu, 7 Jun 2007 14:49:56 +0000 (14:49 +0000)]
Fix array_dims() example to be consistent with the data previously shown.
Christian Rossow

17 years agoThe functions bt_metap, bt_page_stats and bt_page_items had moved
Magnus Hagander [Thu, 7 Jun 2007 09:56:25 +0000 (09:56 +0000)]
The functions bt_metap, bt_page_stats and bt_page_items had moved
from contrib/pgstattuple to pageinspect. We've already fixed English
documentation, but Japanese version does not catch up.

ITAGAKI Takahiro

17 years agoFix up text concatenation so that it accepts all the reasonable cases that
Tom Lane [Wed, 6 Jun 2007 23:00:50 +0000 (23:00 +0000)]
Fix up text concatenation so that it accepts all the reasonable cases that
were accepted by prior Postgres releases.  This takes care of the loose end
left by the preceding patch to downgrade implicit casts-to-text.  To avoid
breaking desirable behavior for array concatenation, introduce a new
polymorphic pseudo-type "anynonarray" --- the added concatenation operators
are actually text || anynonarray and anynonarray || text.

17 years agoMinor editorialization: don't flush plan cache without need.
Tom Lane [Tue, 5 Jun 2007 21:50:19 +0000 (21:50 +0000)]
Minor editorialization: don't flush plan cache without need.

17 years agoDowngrade implicit casts to text to be assignment-only, except for the ones
Tom Lane [Tue, 5 Jun 2007 21:31:09 +0000 (21:31 +0000)]
Downgrade implicit casts to text to be assignment-only, except for the ones
from the other string-category types; this eliminates a lot of surprising
interpretations that the parser could formerly make when there was no directly
applicable operator.

Create a general mechanism that supports casts to and from the standard string
types (text,varchar,bpchar) for *every* datatype, by invoking the datatype's
I/O functions.  These new casts are assignment-only in the to-string direction,
explicit-only in the other, and therefore should create no surprising behavior.
Remove a bunch of thereby-obsoleted datatype-specific casting functions.

The "general mechanism" is a new expression node type CoerceViaIO that can
actually convert between *any* two datatypes if their external text
representations are compatible.  This is more general than needed for the
immediate feature, but might be useful in plpgsql or other places in future.

This commit does nothing about the issue that applying the concatenation
operator || to non-text types will now fail, often with strange error messages
due to misinterpreting the operator as array concatenation.  Since it often
(not always) worked before, we should either make it succeed or at least give
a more user-friendly error; but details are still under debate.

Peter Eisentraut and Tom Lane

17 years agoThe session_replication_role actually can be changed at will during
Jan Wieck [Tue, 5 Jun 2007 20:00:41 +0000 (20:00 +0000)]
The session_replication_role actually can be changed at will during
a session regardless of the existence of cached plans. The plancache
only needs to be invalidated so that rules affected by the new setting
will be reflected in the new query plans.

Jan

17 years agoMove call of MarkBufferDirty() before XLogInsert() as required.
Teodor Sigaev [Tue, 5 Jun 2007 12:47:49 +0000 (12:47 +0000)]
Move call of MarkBufferDirty() before XLogInsert() as required.
Many thanks to Heikki Linnakangas <heikki@enterprisedb.com> for his
sharp eyes.

17 years agoRemove ill-conceived CRLF translation for Windows in syslogger.
Andrew Dunstan [Mon, 4 Jun 2007 22:21:42 +0000 (22:21 +0000)]
Remove ill-conceived CRLF translation for Windows in syslogger.

17 years agoTidy overly long configure help lines.
Andrew Dunstan [Mon, 4 Jun 2007 21:55:43 +0000 (21:55 +0000)]
Tidy overly long configure help lines.

17 years agoFix bundle bugs of GIN:
Teodor Sigaev [Mon, 4 Jun 2007 15:56:28 +0000 (15:56 +0000)]
Fix bundle bugs of GIN:
- Fix possible deadlock between UPDATE and VACUUM queries. Bug never was
  observed in 8.2, but it still exist there. HEAD is more sensitive to
  bug after recent "ring" of buffer improvements.
- Fix WAL creation: if parent page is stored as is after split then
  incomplete split isn't removed during replay. This happens rather rare, only
  on large tables with a lot of updates/inserts.
- Fix WAL replay: there was wrong test of XLR_BKP_BLOCK_* for left
  page after deletion of page. That causes wrong rightlink field: it pointed
  to deleted page.
- add checking of match of clearing incomplete split
- cleanup incomplete split list after proceeding

All of this chages doesn't change on-disk storage, so backpatch...
But second point may be an issue for replaying logs from previous version.

17 years agoOn win32, retry reading when WSARecv returns WSAEWOULDBLOCK. There seem
Magnus Hagander [Mon, 4 Jun 2007 13:39:28 +0000 (13:39 +0000)]
On win32, retry reading when WSARecv returns WSAEWOULDBLOCK. There seem
to be cases when at least Windows 2000 can do this even though select
just indicated that the socket is readable.

Per report and analysis from Cyril VELTER.

17 years agoOn win32, don't use SO_REUSEADDR for TCP sockets.
Magnus Hagander [Mon, 4 Jun 2007 11:59:20 +0000 (11:59 +0000)]
On win32, don't use SO_REUSEADDR for TCP sockets.

Per failure on buildfarm member baiji and subsequent discussion.

17 years agoRemove gratuitous response messages from utility programs.
Peter Eisentraut [Mon, 4 Jun 2007 10:02:40 +0000 (10:02 +0000)]
Remove gratuitous response messages from utility programs.

(Possibly release notes material, lest users be confused.)

The --quiet option is now obsolete and without effect in createdb,
createuser, dropdb, dropuser; kept for compatibility but marked for
removal in 8.4.

Progress messages when acting on all databases now go to stdout instead
of stderr, since they are not in fact errors.

Ordered options in reindexdb reference page alphabetically, like in
other programs' pages.

17 years agoUpdate expected files for textual changes
Peter Eisentraut [Mon, 4 Jun 2007 07:48:17 +0000 (07:48 +0000)]
Update expected files for textual changes

17 years agoClarify some error messages about duplicate things.
Peter Eisentraut [Sun, 3 Jun 2007 22:16:03 +0000 (22:16 +0000)]
Clarify some error messages about duplicate things.

17 years agoRemove description for:
Bruce Momjian [Sun, 3 Jun 2007 18:49:28 +0000 (18:49 +0000)]
Remove description for:

        o -Add a GUC variable to control the tablespace for temporary objects
          and sort files
<
<    It could start with a random tablespace from a supplied list and
<    cycle through the list.
<

17 years agoCreate a GUC parameter temp_tablespaces that allows selection of the
Tom Lane [Sun, 3 Jun 2007 17:08:34 +0000 (17:08 +0000)]
Create a GUC parameter temp_tablespaces that allows selection of the
tablespace(s) in which to store temp tables and temporary files.  This is a
list to allow spreading the load across multiple tablespaces (a random list
element is chosen each time a temp object is to be created).  Temp files are
not stored in per-database pgsql_tmp/ directories anymore, but per-tablespace
directories.

Jaime Casanova and Albert Cervera, with review by Bernd Helmle and Tom Lane.

17 years agoMinimal message corrections found by spell checker.
Peter Eisentraut [Sat, 2 Jun 2007 23:36:35 +0000 (23:36 +0000)]
Minimal message corrections found by spell checker.

17 years agoFix erroneous error reporting for overlength input in text_date(),
Tom Lane [Sat, 2 Jun 2007 16:41:09 +0000 (16:41 +0000)]
Fix erroneous error reporting for overlength input in text_date(),
text_time(), and text_timetz().  7.4-vintage bug found by Greg Stark.

17 years agoRe-add TODO and clarify it is for the kernel cache:
Bruce Momjian [Sat, 2 Jun 2007 11:28:01 +0000 (11:28 +0000)]
Re-add TODO and clarify it is for the kernel cache:

< * Allow free-behind capability for large sequential scans, perhaps using
<   posix_fadvise()
> * Allow free-behind capability for large sequential scans to avoid
>   kernel cache spoiling

17 years agoTODO item not needed anymore now that the buffer cache is
Bruce Momjian [Sat, 2 Jun 2007 02:46:38 +0000 (02:46 +0000)]
TODO item not needed anymore now that the buffer cache is
scan-resistant:

<
< * Allow free-behind capability for large sequential scans, perhaps using
<   posix_fadvise()
<
<   Posix_fadvise() can control both sequential/random file caching and
<   free-behind behavior, but it is unclear how the setting affects other
<   backends that also have the file open, and the feature is not supported
<   on all operating systems.

17 years agoImprove efficiency of LIKE/ILIKE code, especially for multi-byte charsets,
Andrew Dunstan [Sat, 2 Jun 2007 02:03:42 +0000 (02:03 +0000)]
Improve efficiency of LIKE/ILIKE code, especially for multi-byte charsets,
and most especially for UTF8. Remove unnecessary special cases for bytea
processing and single-byte charset ILIKE.  a ILIKE b is now processed as
lower(a) LIKE lower(b) in all cases. The code is now considerably simpler. All
comparisons are now performed byte-wise, and the text and pattern are also
advanced byte-wise where it is safe to do so - essentially where a wildcard is
not being matched.
Andrew Dunstan, from an original patch by ITAGAKI Takahiro, with ideas from
Tom Lane and Mark Mielke.

17 years agoFix aboriginal bug in BufFileDumpBuffer that would cause it to write the
Tom Lane [Fri, 1 Jun 2007 23:43:11 +0000 (23:43 +0000)]
Fix aboriginal bug in BufFileDumpBuffer that would cause it to write the
wrong data when dumping a bufferload that crosses a component-file boundary.
This probably has not been seen in the wild because (a) component files are
normally 1GB apiece and (b) non-block-aligned buffer usage is relatively
rare.  But it's fairly easy to reproduce a problem if one reduces RELSEG_SIZE
in a test build.  Kudos to Kurt Harriman for spotting the bug.

17 years agoAllow leading and trailing whitespace in the input to the boolean
Neil Conway [Fri, 1 Jun 2007 23:40:19 +0000 (23:40 +0000)]
Allow leading and trailing whitespace in the input to the boolean
type. Also, add explicit casts between boolean and text/varchar. Both
of these changes are for conformance with SQL:2003.

Update the regression tests, bump the catversion.

17 years agoMake CREATE/DROP/RENAME DATABASE wait a little bit to see if other backends
Tom Lane [Fri, 1 Jun 2007 19:38:07 +0000 (19:38 +0000)]
Make CREATE/DROP/RENAME DATABASE wait a little bit to see if other backends
will exit before failing because of conflicting DB usage.  Per discussion,
this seems a good idea to help mask the fact that backend exit takes nonzero
time.  Remove a couple of thereby-obsoleted sleeps in contrib and PL
regression test sequences.

17 years agoAdd URL for:
Bruce Momjian [Fri, 1 Jun 2007 18:41:55 +0000 (18:41 +0000)]
Add URL for:

        o Research self-referential UPDATEs that see inconsistent row versions
          in read-committed mode
<
>    http://archives.postgresql.org/pgsql-hackers/2007-06/msg00016.php

17 years agoUpdate wording:
Bruce Momjian [Fri, 1 Jun 2007 18:40:42 +0000 (18:40 +0000)]
Update wording:

        o Research self-referential UPDATEs that see inconsistent row versions
          in read-committed mode

http://archives.postgresql.org/pgsql-hackers/2007-05/msg00507.php

17 years agoBuy back some of the cycles spent in more-expensive hash functions by
Tom Lane [Fri, 1 Jun 2007 17:38:44 +0000 (17:38 +0000)]
Buy back some of the cycles spent in more-expensive hash functions by
selecting power-of-2, rather than prime, numbers of buckets in hash joins.
If the hash functions are doing their jobs properly by making all hash bits
equally random, this is good enough, and it saves expensive integer division
and modulus operations.

17 years agoFix several hash functions that were taking chintzy shortcuts instead of
Tom Lane [Fri, 1 Jun 2007 15:33:19 +0000 (15:33 +0000)]
Fix several hash functions that were taking chintzy shortcuts instead of
delivering a well-randomized hash value.  I got religion on this after
observing that performance of multi-batch hash join degrades terribly if the
higher-order bits of hash values aren't random, as indeed was true for say
hashes of small integer values.  It's now expected and documented that hash
functions should use hash_any or some comparable method to ensure that all
bits of their output are about equally random.

initdb forced because this change invalidates existing hash indexes.  For the
same reason, this isn't back-patchable; the hash join performance problem
will get a band-aid fix in the back branches.

17 years agoWording improvement.
Bruce Momjian [Fri, 1 Jun 2007 04:11:43 +0000 (04:11 +0000)]
Wording improvement.

17 years agoUpdate FAQ_DEV URL to output for text format.
Bruce Momjian [Fri, 1 Jun 2007 04:05:36 +0000 (04:05 +0000)]
Update FAQ_DEV URL to output for text format.

17 years agoAdd URL for code comments to developer's FAQ:
Bruce Momjian [Fri, 1 Jun 2007 00:28:35 +0000 (00:28 +0000)]
Add URL for code comments to developer's FAQ:

http://www.ibm.com/developerworks/linux/library/l-clear-code/?ca=dgr-FClnxw01linuxcodetips

17 years agoThe shortcut exit that I recently added to ExecInitIndexScan() for
Tom Lane [Thu, 31 May 2007 20:45:26 +0000 (20:45 +0000)]
The shortcut exit that I recently added to ExecInitIndexScan() for
EXPLAIN-only operation was a little too short; it skipped initializing the
node's result tuple type, which may be needed depending on what's above the
indexscan node.  Call ExecAssignResultTypeFromTL before exiting.  (For good
luck I moved up the ExecAssignScanProjectionInfo call as well, so that
everything except indexscan-specific initialization will still be done.)
Per example from Grant Finnemore.

17 years agoChange build_index_pathkeys() so that the expressions it builds to represent
Tom Lane [Thu, 31 May 2007 16:57:34 +0000 (16:57 +0000)]
Change build_index_pathkeys() so that the expressions it builds to represent
index key columns always have the type expected by the index's associated
operators, ie, we add RelabelType nodes when dealing with binary-compatible
index opclasses.  This is needed to get varchar indexes to play nicely with
the new EquivalenceClass machinery, as per recent gripe from Josh Berkus that
CVS HEAD was failing to match a varchar index column to a constant restriction
in the query.

It seems likely that this change will allow removal of a lot of ugly ad-hoc
RelabelType-stripping that the planner has traditionally done while matching
expressions to other expressions, but I'll worry about that some other day.

17 years agoMake some messages more consistent
Peter Eisentraut [Thu, 31 May 2007 15:13:06 +0000 (15:13 +0000)]
Make some messages more consistent

17 years agoReplace ReadBuffer to ReadBufferWithStrategy in all vacuum-involved places
Teodor Sigaev [Thu, 31 May 2007 14:03:09 +0000 (14:03 +0000)]
Replace ReadBuffer to ReadBufferWithStrategy in all vacuum-involved places
to implement limited-size "ring" of buffers for VACUUM for GIN & GIST

17 years agoDowngrade some low-level startup messages to DEBUG1.
Peter Eisentraut [Thu, 31 May 2007 07:36:12 +0000 (07:36 +0000)]
Downgrade some low-level startup messages to DEBUG1.

17 years agoFix overly-strict sanity check in BeginInternalSubTransaction that made it
Tom Lane [Wed, 30 May 2007 21:01:39 +0000 (21:01 +0000)]
Fix overly-strict sanity check in BeginInternalSubTransaction that made it
fail when used in a deferred trigger.  Bug goes back to 8.0; no doubt the
reason it hadn't been noticed is that we've been discouraging use of
user-defined constraint triggers.  Per report from Frank van Vugt.

17 years agoUpdate:
Bruce Momjian [Wed, 30 May 2007 20:26:06 +0000 (20:26 +0000)]
Update:

< * Consider allowing 64-bit integers to be passed by value on 64-bit
<   platforms
> * Consider allowing 64-bit integers and floats to be passed by value on
>   64-bit platforms
>
>   Also change 32-bit floats (float4) to be passed by value at the same
>   time.
>

17 years agoMake large sequential scans and VACUUMs work in a limited-size "ring" of
Tom Lane [Wed, 30 May 2007 20:12:03 +0000 (20:12 +0000)]
Make large sequential scans and VACUUMs work in a limited-size "ring" of
buffers, rather than blowing out the whole shared-buffer arena.  Aside from
avoiding cache spoliation, this fixes the problem that VACUUM formerly tended
to cause a WAL flush for every page it modified, because we had it hacked to
use only a single buffer.  Those flushes will now occur only once per
ring-ful.  The exact ring size, and the threshold for seqscans to switch into
the ring usage pattern, remain under debate; but the infrastructure seems
done.  The key bit of infrastructure is a new optional BufferAccessStrategy
object that can be passed to ReadBuffer operations; this replaces the former
StrategyHintVacuum API.

This patch also changes the buffer usage-count methodology a bit: we now
advance usage_count when first pinning a buffer, rather than when last
unpinning it.  To preserve the behavior that a buffer's lifetime starts to
decrease when it's released, the clock sweep code is modified to not decrement
usage_count of pinned buffers.

Work not done in this commit: teach GiST and GIN indexes to use the vacuum
BufferAccessStrategy for vacuum-driven fetches.

Original patch by Simon, reworked by Heikki and again by Tom.

17 years agoFix wording:
Bruce Momjian [Wed, 30 May 2007 20:02:41 +0000 (20:02 +0000)]
Fix wording:

< * Consider allowing 64-bit integers to be passed by reference on 64-bit
> * Consider allowing 64-bit integers to be passed by value on 64-bit

17 years agoAdd:
Bruce Momjian [Wed, 30 May 2007 20:00:35 +0000 (20:00 +0000)]
Add:

>
> * Consider allowing 64-bit integers to be passed by reference on 64-bit
>   platforms

17 years agoAdd standard error redirection for OS/X & darwin startup script.
Bruce Momjian [Wed, 30 May 2007 19:58:34 +0000 (19:58 +0000)]
Add standard error redirection for OS/X & darwin startup script.

Les Hill

17 years agoUpdate documentation mention of VACUUM FULL and CLUSTER where
Bruce Momjian [Wed, 30 May 2007 19:45:01 +0000 (19:45 +0000)]
Update documentation mention of VACUUM FULL and CLUSTER where
appropriate.

Guillaume Cottenceau

17 years agoTweak: use memcpy() in text_time(), rather than manually copying bytes
Neil Conway [Wed, 30 May 2007 19:38:05 +0000 (19:38 +0000)]
Tweak: use memcpy() in text_time(), rather than manually copying bytes
in a loop.

17 years agoAdd URL for:
Bruce Momjian [Wed, 30 May 2007 19:07:20 +0000 (19:07 +0000)]
Add URL for:

* Improve speed with indexes

  For large table adjustments during VACUUM FULL, it is faster to cluster
  or reindex rather than update the index.  Also, index updates can bloat
  the index.

17 years agoAdd documentation reference to statistical functions from func.sgml.
Bruce Momjian [Wed, 30 May 2007 18:13:29 +0000 (18:13 +0000)]
Add documentation reference to statistical functions from func.sgml.

Jim Nasby

17 years agoUpdate /contrib OS/X startup files, and move to a separate OS/X
Bruce Momjian [Wed, 30 May 2007 18:10:38 +0000 (18:10 +0000)]
Update /contrib OS/X startup files, and move to a separate OS/X
directory.

Mark Cotner and David Fetter

17 years agoUpdate cvsutils documentation description.
Bruce Momjian [Wed, 30 May 2007 18:00:08 +0000 (18:00 +0000)]
Update cvsutils documentation description.

17 years agoUpdate cvs instructions to suggest cvsutils.
Bruce Momjian [Wed, 30 May 2007 17:41:43 +0000 (17:41 +0000)]
Update cvs instructions to suggest cvsutils.

David Fetter

17 years agoFix trivial misspelling in comment.
Tom Lane [Wed, 30 May 2007 16:16:32 +0000 (16:16 +0000)]
Fix trivial misspelling in comment.

17 years agoFix a bug in input processing for the "interval" type. Previously,
Neil Conway [Tue, 29 May 2007 04:58:43 +0000 (04:58 +0000)]
Fix a bug in input processing for the "interval" type. Previously,
"microsecond" and "millisecond" units were not considered valid input
by themselves, which caused inputs like "1 millisecond" to be rejected
erroneously.

Update the docs, add regression tests, and backport to 8.2 and 8.1

17 years agommgr README tweak: "either" is no longer correct. The previous wording
Neil Conway [Tue, 29 May 2007 04:19:35 +0000 (04:19 +0000)]
mmgr README tweak: "either" is no longer correct. The previous wording
compared PortalContext with QueryContext, but the latter no longer exists.

17 years agoStop a few regression tests from needlessly disabling GEQO. This was
Neil Conway [Tue, 29 May 2007 04:17:08 +0000 (04:17 +0000)]
Stop a few regression tests from needlessly disabling GEQO. This was
necessary in 1997, when geqo_threshold did not exist, but it is no
longer needed.

17 years agoUpdate wording:
Bruce Momjian [Mon, 28 May 2007 23:56:29 +0000 (23:56 +0000)]
Update wording:

< * Fix self-referential UPDATEs seeing inconsistent row versions in
> * Fix self-referential UPDATEs that see inconsistent row versions in

17 years agoAdd:
Bruce Momjian [Mon, 28 May 2007 23:56:00 +0000 (23:56 +0000)]
Add:

>
> * Fix self-referential UPDATEs seeing inconsistent row versions in
>   read-committed mode
>
>   http://archives.postgresql.org/pgsql-hackers/2007-05/msg00507.php

17 years agoTweak the code in a couple of places to try to deliver more user-friendly
Tom Lane [Mon, 28 May 2007 16:43:24 +0000 (16:43 +0000)]
Tweak the code in a couple of places to try to deliver more user-friendly
error messages when a single COPY line is too long for us to handle.  Per
example from Johann Spies.

17 years agoApplied patch send by Joachim Wieland to fix INTEGER_DATETIMES under MSVC.
Michael Meskes [Mon, 28 May 2007 09:46:47 +0000 (09:46 +0000)]
Applied patch send by Joachim Wieland to fix INTEGER_DATETIMES under MSVC.

17 years agoCode cleanup: use "bool" for Boolean variables, rather than "int".
Neil Conway [Sun, 27 May 2007 20:32:16 +0000 (20:32 +0000)]
Code cleanup: use "bool" for Boolean variables, rather than "int".

17 years agoOoops, I was too busy worrying about getting the transactional infrastructure
Tom Lane [Sun, 27 May 2007 17:28:36 +0000 (17:28 +0000)]
Ooops, I was too busy worrying about getting the transactional infrastructure
right to think carefully about how insert and delete counts map to
n_live_tuples.  Of course a deletion should reduce n_live_tuples.

17 years agoSorry guys, I committed the file from my development snapshot instead the one from...
Michael Meskes [Sun, 27 May 2007 15:59:57 +0000 (15:59 +0000)]
Sorry guys, I committed the file from my development snapshot instead the one from HEAD. Fixing it now.

17 years agoThe source code changed, so the expected file changed too.
Michael Meskes [Sun, 27 May 2007 11:32:59 +0000 (11:32 +0000)]
The source code changed, so the expected file changed too.

17 years agoApplied Joachim's patch for ecpg_config.h creation on Vista.
Michael Meskes [Sun, 27 May 2007 11:16:41 +0000 (11:16 +0000)]
Applied Joachim's patch for ecpg_config.h creation on Vista.
Changed variable test to not run into infinite loops on backend errors.

17 years agopgstat's on-proc-exit hook has to execute after the last transaction commit
Tom Lane [Sun, 27 May 2007 05:37:50 +0000 (05:37 +0000)]
pgstat's on-proc-exit hook has to execute after the last transaction commit
or abort within a backend; rearrange InitPostgres processing to make it so.
Revealed by just-added Asserts along with ECPG regression tests (hm, I wonder
why the core regression tests didn't expose it?).  This possibly is another
reason for missing stats updates ...

17 years agoFix up pgstats counting of live and dead tuples to recognize that committed
Tom Lane [Sun, 27 May 2007 03:50:39 +0000 (03:50 +0000)]
Fix up pgstats counting of live and dead tuples to recognize that committed
and aborted transactions have different effects; also teach it not to assume
that prepared transactions are always committed.

Along the way, simplify the pgstats API by tying counting directly to
Relations; I cannot detect any redeeming social value in having stats
pointers in HeapScanDesc and IndexScanDesc structures.  And fix a few
corner cases in which counts might be missed because the relation's
pgstat_info pointer hadn't been set.

17 years agoRepair two constraint-exclusion corner cases triggered by proving that an
Tom Lane [Sat, 26 May 2007 18:23:02 +0000 (18:23 +0000)]
Repair two constraint-exclusion corner cases triggered by proving that an
inheritance child of an UPDATE/DELETE target relation can be excluded by
constraints.  I had rearranged some code in set_append_rel_pathlist() to
avoid "useless" work when a child is excluded, but overdid it and left
the child with no cheapest_path entry, causing possible failure later
if the appendrel was involved in a join.  Also, it seems that the dummy
plan generated by inheritance_planner() when all branches are excluded
has to be a bit less dummy now than was required in 8.2.
Per report from Jan Wieck.  Add his test case to the regression tests.

17 years agoCreate hooks to let a loadable plugin monitor (or even replace) the planner
Tom Lane [Fri, 25 May 2007 17:54:25 +0000 (17:54 +0000)]
Create hooks to let a loadable plugin monitor (or even replace) the planner
and/or create plans for hypothetical situations; in particular, investigate
plans that would be generated using hypothetical indexes.  This is a
heavily-rewritten version of the hooks proposed by Gurjeet Singh for his
Index Advisor project.  In this formulation, the index advisor can be
entirely a loadable module instead of requiring a significant part to be
in the core backend, and plans can be generated for hypothetical indexes
without requiring the creation and rolling-back of system catalog entries.

The index advisor patch as-submitted is not compatible with these hooks,
but it needs significant work anyway due to other 8.2-to-8.3 planner
changes.  With these hooks in the core backend, development of the advisor
can proceed as a pgfoundry project.

17 years agoRemove ruleutils.c's use of varnoold/varoattno as a shortcut for determining
Tom Lane [Thu, 24 May 2007 18:58:42 +0000 (18:58 +0000)]
Remove ruleutils.c's use of varnoold/varoattno as a shortcut for determining
what a Var node refers to.  This is no longer necessary because the new
flat-range-table representation of plan trees makes it relatively easy to dig
down through child plan levels to find the original reference; and to keep
doing it that way, we'd have to store joinaliasvars lists in flattened RTEs,
as demonstrated by bug report from Leszek Trenkner.  This change makes
varnoold/varoattno truly just debug aids, which wasn't quite the case before.
Perhaps we should drop them, or only have them in assert-enabled builds?

17 years agoAvoid assuming that the fields of struct timeval have exactly type long.
Tom Lane [Thu, 24 May 2007 18:54:10 +0000 (18:54 +0000)]
Avoid assuming that the fields of struct timeval have exactly type long.
This is probably incorrect on some platforms, and definitely draws a
compiler warning on Darwin.

17 years agoAdd deprecation notice about planned removal in 8.4 and advice to try using
Peter Eisentraut [Thu, 24 May 2007 11:13:08 +0000 (11:13 +0000)]
Add deprecation notice about planned removal in 8.4 and advice to try using
the newer XML stuff in core.

(This should probably also be referred to in the release notes.)

17 years agoRepair planner bug introduced in 8.2 by ability to rearrange outer joins:
Tom Lane [Tue, 22 May 2007 23:23:58 +0000 (23:23 +0000)]
Repair planner bug introduced in 8.2 by ability to rearrange outer joins:
in cases where a sub-SELECT inserts a WHERE clause between two outer joins,
that clause may prevent us from re-ordering the two outer joins.  The code
was considering only the joins' own ON-conditions in determining reordering
safety, which is not good enough.  Add a "delay_upper_joins" flag to
OuterJoinInfo to flag that we have detected such a clause and higher-level
outer joins shouldn't be permitted to commute with this one.  (This might
seem overly coarse, but given the current rules for OJ reordering, it's
sufficient AFAICT.)

The failure case is actually pretty narrow: it needs a WHERE clause within
the RHS of a left join that checks the RHS of a lower left join, but is not
strict for that RHS (else we'd have simplified the lower join to a plain
join).  Even then no failure will be manifest unless the planner chooses to
rearrange the join order.

Per bug report from Adam Terrey.

17 years agoRemove unsupported -u option from pg_standby example usage.
Alvaro Herrera [Tue, 22 May 2007 15:31:44 +0000 (15:31 +0000)]
Remove unsupported -u option from pg_standby example usage.
From Pavel Stehule.

17 years agoFix best_inner_indexscan to return both the cheapest-total-cost and
Tom Lane [Tue, 22 May 2007 01:40:33 +0000 (01:40 +0000)]
Fix best_inner_indexscan to return both the cheapest-total-cost and
cheapest-startup-cost innerjoin indexscans, and make joinpath.c consider
both of these (when different) as the inside of a nestloop join.  The
original design was based on the assumption that indexscan paths always
have negligible startup cost, and so total cost is the only important
figure of merit; an assumption that's obviously broken by bitmap
indexscans.  This oversight could lead to choosing poor plans in cases
where fast-start behavior is more important than total cost, such as
LIMIT and IN queries.  8.1-vintage brain fade exposed by an example from
Chuck D.

17 years agoTeach tuplestore.c to throw away data before the "mark" point when the caller
Tom Lane [Mon, 21 May 2007 17:57:35 +0000 (17:57 +0000)]
Teach tuplestore.c to throw away data before the "mark" point when the caller
is using mark/restore but not rewind or backward-scan capability.  Insert a
materialize plan node between a mergejoin and its inner child if the inner
child is a sort that is expected to spill to disk.  The materialize shields
the sort from the need to do mark/restore and thereby allows it to perform
its final merge pass on-the-fly; while the materialize itself is normally
cheap since it won't spill to disk unless the number of tuples with equal
key values exceeds work_mem.

Greg Stark, with some kibitzing from Tom Lane.

17 years agoXPath fixes:
Peter Eisentraut [Mon, 21 May 2007 17:10:29 +0000 (17:10 +0000)]
XPath fixes:
 - Function renamed to "xpath".
 - Function is now strict, per discussion.
 - Return empty array in case when XPath expression detects nothing
   (previously, NULL was returned in such case), per discussion.
 - (bugfix) Work with fragments with prologue: select xpath('/a',
   '<?xml version="1.0"?><a /><b />'); // now XML datum is always wrapped
   with dummy <x>...</x>, XML prologue simply goes away (if any).
 - Some cleanup.

Nikolay Samokhvalov

Some code cleanup and documentation work by myself.

17 years agoFix spurious German index entry
Peter Eisentraut [Mon, 21 May 2007 15:11:12 +0000 (15:11 +0000)]
Fix spurious German index entry

17 years agoAlso removed the function not just the call, sorry.
Michael Meskes [Mon, 21 May 2007 07:07:48 +0000 (07:07 +0000)]
Also removed the function not just the call, sorry.