Skip to content

Wöchentlicher PostgreSQL Newsletter - 06. November 2011

Der Originalartikel befindet sich unter:

== Wöchentlicher PostgreSQL Newsletter - 06. November 2011 ==

== PostgreSQL Produkt Neuigkeiten ==

ezNcrypt für Datenbanken unterstützt jetzt PostgreSQL.

pgpool-II 3.0.5, ein Connection Pooler und mehr,
ist erschienen.

RHQ 4.2, ein Systemmanagement- und Monotoringwerkzeug welches auf
PostgreSQL läuft, ist erschienen.

== PostgreSQL Lokal ==

PGConf.DE 2011 ist die Deutschsprachige PostgreSQL Konferenz
und wird am 11. November 2011 im Rheinischen Industriemuseum
in Oberhausen, Deutschland, stattfinden. Das Vortragsprogramm
ist jetzt verfügbar und die Registrierung ist offen.

Die fünfte Edition des Italienischen PostgreSQL
Tags (PGDay.IT 2011) findet am 25. November in
Prato, Italien statt.

Der Call for Papers ist offen für PostgreSQL Session #3,
welche am 2. Februar 2012 in Paris stattfindet. Der Einsendeschluß
ist der 30. November 2011, die ausgewählten Sprecher werden bis
zum 14. Dezember 2011 informiert. Vorträge (in Französisch oder
Englisch) sollen an call-for-paper AT postgresql-sessions DOT org
geschickt werden. Mehr Informationen unter:

Der Call for Papers für die FLOSS UK ist offen, die Veranstaltung
findet in Edinburgh vom 20. bis 22. März 2012 statt. Der
Einsendeschluß für Einreichungen ist der 18. November 2011,
ausgewählte Sprecher werden bis zum 25. November benachrichtigt.
Einreichungen sollen an postgresql2012 AT flossuk DOT org gehen.
Mehr Informationen:

== PostgreSQL in den News ==

Planet PostgreSQL:

Dieser wöchentliche PostgreSQL Newsletter wurde erstellt von David Fetter.

Sende Neuigkeiten und Ankündigungen bis Sonntag, 15 Uhr Pazifischer
Zeit. Bitte sende englische Beiträge an, deutsche an, italienische an, spanische an

== Angewandte Patches ==

Tom Lane pushed:

- Stop btree indexscans upon reaching nulls in either direction.  The
  existing scan-direction-sensitive tests were overly complex, and
  failed to stop the scan in cases where it's perfectly legitimate to
  do so.  Per bug #6278 from Maksym Boguk.  Back-patch to 8.3, which
  is as far back as the patch applies easily.  Doesn't seem worth
  sweating over a relatively minor performance issue in 8.2 at this
  late date.  (But note that this was a performance regression from
  8.1 and before, so 8.2 is being left as an outlier.)

- Fix race condition with toast table access from a stale syscache
  entry.  If a tuple in a syscache contains an out-of-line toasted
  field, and we try to fetch that field shortly after some other
  transaction has committed an update or deletion of the tuple, there
  is a race condition: vacuum could come along and remove the toast
  tuples before we can fetch them.  This leads to transient failures
  like "missing chunk number 0 for toast value NNNNN in
  pg_toast_2619", as seen in recent reports from Andrew Hammond and
  Tim Uckun.  The design idea of syscache is that access to stale
  syscache entries should be prevented by relation-level locks, but
  that fails for at least two cases where toasted fields are possible:
  ANALYZE updates pg_statistic rows without locking out sessions that
  might want to plan queries on the same table, and CREATE OR REPLACE
  FUNCTION updates pg_proc rows without any meaningful lock at all.
  The least risky fix seems to be an idea that Heikki suggested when
  we were dealing with a related problem back in August: forcibly
  detoast any out-of-line fields before putting a tuple into syscache
  in the first place.  This avoids the problem because at the time we
  fetch the parent tuple from the catalog, we should be holding an
  MVCC snapshot that will prevent removal of the toast tuples, even if
  the parent tuple is outdated immediately after we fetch it.  (Note:
  I'm not convinced that this statement holds true at every instant
  where we could be fetching a syscache entry at all, but it does
  appear to hold true at the times where we could fetch an entry that
  could have a toasted field.  We will need to be a bit wary of adding
  toast tables to low-level catalogs that don't have them already.)
  An additional benefit is that subsequent uses of the syscache entry
  should be faster, since they won't have to detoast the field.
  Back-patch to all supported versions.  The problem is significantly
  harder to reproduce in pre-9.0 releases, because of their
  willingness to flush every entry in a syscache whenever the
  underlying catalog is vacuumed (cf CatalogCacheFlushRelation); but
  there is still a window for trouble.

- Preserve Var location information during flatten_join_alias_vars.
  This allows us to give correct syntax error pointers when
  complaining about ungrouped variables in a join query with
  aggregates or GROUP BY.  It's pretty much irrelevant for the
  planner's use of the function, though perhaps it might aid debugging

- Revert "Stop btree indexscans upon reaching nulls in either
  direction." This reverts commit
  048fffed55ff1d6d346130e4a6b7be434e81e82c.  As pointed out by Naoya
  Anzai, we need to do more work to make that idea handle end-of-index
  cases, and it is looking like too much risk for a back-patch.  So
  bug #6278 is only going to be fixed in HEAD.

- Fix btree stop-at-nulls logic properly.  As pointed out by Naoya
  Anzai, my previous try at this was a few bricks shy of a load,
  because I had forgotten that the initial-positioning logic might not
  try to skip over nulls at the end of the index the scan will start
  from.  We ought to fix that, because it represents an unnecessary
  inefficiency, but first let's get the scan-stop logic back to a safe
  state.  With this patch, we preserve the performance benefit
  requested in bug #6278 for the case of scanning forward into NULLs
  (in a NULLS LAST index), but the reverse case of scanning backward
  across NULLs when there's no suitable initial-positioning qual is
  still inefficient.

- Avoid scanning nulls at the beginning of a btree index scan.  If we
  have an inequality key that constrains the other end of the index,
  it doesn't directly help us in doing the initial positioning ... but
  it does imply a NOT NULL constraint on the index column.  If the
  index stores nulls at this end, we can use the implied NOT NULL
  condition for initial positioning, just as if it had been stated
  explicitly.  This avoids wasting time when there are a lot of nulls
  in the column.  This is the reverse of the examples given in bugs
  #6278 and #6283, which were about failing to stop early when we
  encounter nulls at the end of the indexscan.

- Fix handling of PlaceHolderVars in nestloop parameter management.
  If we use a PlaceHolderVar from the outer relation in an inner
  indexscan, we need to reference the PlaceHolderVar as such as the
  value to be passed in from the outer relation.  The previous code
  effectively tried to reconstruct the PHV from its component
  expression, which doesn't work since (a) the Vars therein aren't
  necessarily bubbled up far enough, and (b) it would be the wrong
  semantics anyway because of the possibility that the PHV is supposed
  to have gone to null at some point before the current join.  Point
  (a) led to "variable not found in subplan target list" planner
  errors, but point (b) would have led to silently wrong answers.  Per
  report from Roger Niederland.

- Fix inline_set_returning_function() to allow multiple OUT
  parameters.  inline_set_returning_function failed to distinguish
  functions returning generic RECORD (which require a column list in
  the RTE, as well as run-time type checking) from those with multiple
  OUT parameters (which do not).  This prevented inlining from
  happening.  Per complaint from Jay Levitt.  Back-patch to 8.4 where
  this capability was introduced.

- Improve comments for TSLexeme data structure.  Mostly, clean up
  long-ago pgindent damage.

- Fix bogus code in contrib/ tsearch dictionary examples.  Both
  dict_int and dict_xsyn were blithely assuming that whatever memory
  palloc gives back will be pre-zeroed.  This would typically work for
  just about long enough to run their regression tests, and no longer
  :-(.  The pre-9.0 code in dict_xsyn was even lamer than that, as it
  would happily give back a pointer to the result of palloc(0),
  encouraging its caller to access off the end of memory.  Again, this
  would just barely fail to fail as long as memory contained nothing
  but zeroes.  Per a report from Rodrigo Hjort that code based on
  these examples didn't work reliably.

- Don't assume that a tuple's header size is unchanged during
  toasting.  This assumption can be wrong when the toaster is passed a
  raw on-disk tuple, because the tuple might pre-date an ALTER TABLE
  ADD COLUMN operation that added columns without rewriting the table.
  In such a case the tuple's natts value is smaller than what we
  expect from the tuple descriptor, and so its t_hoff value could be
  smaller too.  In fact, the tuple might not have a null bitmap at
  all, and yet our current opinion of it is that it contains some
  trailing nulls.  In such a situation, toast_insert_or_update did the
  wrong thing, because to save a few lines of code it would use the
  old t_hoff value as the offset where heap_fill_tuple should start
  filling data.  This did not leave enough room for the new nulls
  bitmap, with the result that the first few bytes of data could be
  overwritten with null flag bits, as in a recent report from Hubert
  Depesz Lubaczewski.  The particular case reported requires ALTER
  INSERT ... SELECT * FROM ..., and further requires that there be
  some out-of-line toasted fields in one of the tuples to be copied;
  else we'll not reach the troublesome code.  The problem can only
  manifest in this form in 8.4 and later, because before commit
  a77eaa6a95009a3441e0d475d1980259d45da072, CREATE TABLE AS or
  INSERT/SELECT wouldn't result in raw disk tuples getting passed
  directly to heap_insert --- there would always have been at least a
  junkfilter in between, and that would reconstitute the tuple header
  with an up-to-date t_natts and hence t_hoff.  But I'm backpatching
  the tuptoaster change all the way anyway, because I'm not convinced
  there are no older code paths that present a similar risk.

- Un-break horology regression test.  Adjust ill-considered
  timezone-dependent tests added in commit
  8a3d33c8e6c681d512f79af4a521ee0c02befcef so that they won't fail on
  DST transition days.  Per all-pink buildfarm.

Magnus Hagander pushed:

- Document that multiple LDAP servers can be specified

- Pre-pad WAL files when streaming transaction log.  Instead of
  filling files as they appear, pre-pad the WAL files received when
  streaming xlog the same way that the server does. Data is streamed
  into a .partial file which is then renamed()d into palce when it's
  complete, but it will always be 16MB.  This also means that the
  starting position for pg_receivexlog is now simply right after the
  last complete segment, and we never need to deal with partial
  segments there.  Patch by me, review by Fujii Masao

- Properly close replication connection in pg_receivexlog

- Add missing space in comment

- Make psql \d on a sequence show the table/column owning it

- Show statistics target for columns in \d+ on a table

- Update regression tests for \d+ modification.  Noted by Tom

Simon Riggs pushed:

- Split work of bgwriter between 2 processes: bgwriter and
  checkpointer.  bgwriter is now a much less important process,
  responsible for page cleaning duties only. checkpointer is now
  responsible for checkpoints and so has a key role in shutdown. Later
  patches will correct doc references to the now old idea that
  bgwriter performs checkpoints.  Has beneficial effect on performance
  at high write rates, but mainly refactoring to more easily allow
  changes for power reduction by simplifying previously tortuous code
  around required to allow page cleaning and checkpointing to time
  slice in the same process.  Patch by me, Review by Dickson Guedes

- Add new file for checkpointer.c

- Have checkpointer send stats once each processing loop.  Noted by
  Fujii Masao

- Comment changes to show bgwriter no longer performs checkpoints.

- Fix timing of Startup CLOG and MultiXact during Hot Standby.  Patch
  by me, bug report by Chris Redekop, analysis by Florian Pflug

- Start Hot Standby faster when initial snapshot is incomplete.  If
  the initial snapshot had overflowed then we can start whenever the
  latest snapshot is empty, not overflowed or as we did already, start
  when the xmin on primary was higher than xmax of our starting
  snapshot, which proves we have full snapshot data.  Bug report by
  Chris Redekop

- Remove spurious entry from missed catch while patch juggling

- Derive oldestActiveXid at correct time for Hot Standby.  There was a
  timing window between when oldestActiveXid was derived and when it
  should have been derived that only shows itself under heavy load.
  Move code around to ensure correct timing of derivation.  No change
  to StartupSUBTRANS() code, which is where this failed.  Bug report
  by Chris Redekop

- Refactor xlog.c to create src/backend/postmaster/startup.c.  Startup
  process now has its own dedicated file, just like all other
  special/background processes. Reduces role and size of xlog.c

- Reduce checkpoints and WAL traffic on low activity database server
  Previously, we skipped a checkpoint if no WAL had been written since
  last checkpoint, though this does not appear in user documentation.
  As of now, we skip a checkpoint until we have written at least one
  enough WAL to switch the next WAL file. This greatly reduces the
  level of activity and number of WAL messages generated by a very low
  activity server. This is safe because the purpose of a checkpoint is
  to act as a starting place for a recovery, in case of crash.  This
  patch maintains minimal WAL volume for replay in case of crash, thus
  maintaining very low crash recovery time.

- Update more comments about checkpoints being done by bgwriter

- Improve docs for timing and skipping of checkpoints.  Greg Smith

- Move user functions related to WAL into xlogfuncs.c

Bruce Momjian pushed:

- Allow pg_upgrade to upgrade an old cluster that doesn't have a
  'postgres' database.

- Update pg_upgrade comment on missing 'postgres' database.

- Adjust pg_upgrade "new database skip" code, e.g. 'postgres', to more
  cleanly handle old/new database mismatches.

Peter Eisentraut pushed:

- Clean up whitespace and indentation in parser and scanner files.
  These are not touched by pgindent, so clean them up a bit manually.

- Add note about using GNU tar warning options for base backups

- Fix archive_command example.  The given archive_command example
  didn't use %p or %f, which wouldn't really work in practice.

Robert Haas pushed:

- Initialize myProcLocks queues just once, at postmaster startup.  In
  assert-enabled builds, we assert during the shutdown sequence that
  the queues have been properly emptied, and during process startup
  that we are inheriting empty queues.  In non-assert enabled builds,
  we just save a few cycles.

- Check the return value of getcwd(), instead of assuming success.
  Kevin Grittner

- Silence bogus compiler warning.

Heikki Linnakangas pushed:

- Support range data types.  Selectivity estimation functions are
  missing for some range type operators, which is a TODO.  Jeff Davis

- Oops, forgot to fix the catversion when I committed the range types
  patch.  It was inadvertently changed to 201111111, which is a wrong
  date.  Change it to current date, and remove the comment that was
  supposed to remind me to fix it before committing.

Andrew Dunstan pushed:

- Do not treat a superuser as a member of every role for HBA purposes.
  This makes it possible to use reject lines with group roles.  Andrew
  Dunstan, reviewed by Robert Haas.

- Role membership of superusers is only by explicit membership for
  HBA.  Document that this rule applies to 'samerole' as well as to
  named roles.  Per gripe from Tom Lane.

Alvaro Herrera pushed:

- Implement a dry-run mode for isolationtester.  This mode prints out
  the permutations that would be run by the given spec file, in the
  same format used by the permutation lines in spec files.  This helps
  in building new spec files.  Author: Alexander Shulgin, with some
  tweaks by me

- Unbreak isolationtester on Win32.  I broke it in a previous commit
  because I neglected to install the necessary incantations to have
  getopt() work on Windows.  Per red blots in buildfarm.

== Abgelehnte Patches (bis jetzt) ==

No one was disappointed this week :-)

== Eingesandte Patches ==

Scott Mead sent in two revisions of a patch to see some context around

Shigeru HANADA sent in another revision of the patch to add a
PostgreSQL FDW.

Peter Eisentraut sent in another revision of the patch to enable psql
to switch automatically between normal and \x mode depending on the
width of the output.

Robert Haas sent in three revisions of a patch to drop the "=>"
notation from hstore.

Andrew Dunstan sent in another revision of the patch to add an
--exclude-table-data option to pg_dump.

KaiGai Kohei sent in two more revisions of the patch to fix certain
types of information leaks in VIEWs.

Andrew Dunstan sent in another revision of the patch to add a \setenv
command to psql.

KaiGai Kohei sent in a patch to add checks for INSERT permission on
new tables constructed by SELECT INTO or CREATE TABLE AS.

Simon Riggs and Robert Haas traded revisions of a patch to skip busy
pages during VACUUM.

Alvaro Herrera sent in another revision of the patch to add foreign
key locks.

Pavan Deolasee sent in a patch to store hot members of PGPROC out of
band, a performance optimization.

Gabriele Bartolini sent in a WIP patch to allow arrays to be foreign
keys to scalar primary keys.

Tomas Vondra sent in a patch that would allow optional "cleaning" of
queries tracked in pg_stat_statements, compressing the result and
making it more readable.

Greg Smith sent in a patch adds a new function to the pageinspect
extension for measuring total free space, in either tables or indexes.
It returns the free space as a percentage, so higher numbers mean more

J Smith sent in a fix to some corner-case bugs in the unaccent module.


No Trackbacks


Display comments as Linear | Threaded

No comments

Add Comment

Enclosing asterisks marks text as bold (*word*), underscore are made via _word_.
E-Mail addresses will not be displayed and will only be used for E-Mail notifications.
To leave a comment you must approve it via e-mail, which will be sent to your address after submission.
Form options