Previously, the right to split a bucket was represented by a
heavyweight lock on the page number of the primary bucket page.
Unfortunately, this meant that every scan needed to take a heavyweight
lock on that bucket also, which was bad for concurrency. Instead, use
a cleanup lock on the primary bucket page to indicate the right to
begin a split, so that scans only need to retain a pin on that page,
which is they would have to acquire anyway, and which is also much
cheaper.
In addition to reducing the locking cost, this also avoids locking out
scans and inserts for the entire lifetime of the split: while the new
bucket is being populated with copies of the appropriate tuples from
the old bucket, scans and inserts can happen in parallel. There are
minor concurrency improvements for vacuum operations as well, though
the situation there is still far from ideal.
This patch also removes the unworldly assumption that a split will
never be interrupted. With the new code, a split is done in a series
of small steps and the system can pick up where it left off if it is
interrupted prior to completion. While this patch does not itself add
write-ahead logging for hash indexes, it is clearly a necessary first
step, since one of the things that could interrupt a split is the
removal of electrical power from the machine performing it.
Amit Kapila. I wrote the original design on which this patch is
based, and did a good bit of work on the comments and README through
multiple rounds of review, but all of the code is Amit's. Also
reviewed by Jesper Pedersen, Jeff Janes, and others.
Discussion: http://postgr.es/m/CAA4eK1LfzcZYxLoXS874Ad0+S-ZM60U9bwcyiUZx9mHZ-KCWhw@mail.gmail.com
Instead of confusingly stating platform-dependent defaults for these
parameters in the comments in postgresql.conf.sample (with the main
entry being a lie on Linux), teach initdb to install the correct
platform-dependent value in postgresql.conf, similarly to the way
we handle other platform-dependent defaults. This won't do anything
for existing 9.6 installations, but since it's effectively only a
documentation improvement, that seems OK.
Since this requires initdb to have access to the default values,
move the #define's for those to pg_config_manual.h; the original
placement in bufmgr.h is unworkable because that file can't be
included by frontend programs.
Adjust the default value for wal_writer_flush_after so that it is 1MB
regardless of XLOG_BLCKSZ, conforming to what is stated in both the
SGML docs and postgresql.conf. (We could alternatively make it scale
with XLOG_BLCKSZ, but I'm not sure I see the point.)
Copy-edit related SGML documentation.
Fabien Coelho and Tom Lane, per a gripe from Tomas Vondra.
Discussion: <30ebc6e3-8358-09cf-44a8-578252938424@2ndquadrant.com>
Previously, requesting commit timestamp for transactions
FrozenTransactionId and BootstrapTransactionId resulted in an error.
But since those values can validly appear in committed tuples' Xmin,
this behavior is unhelpful and error prone: each caller would have to
special-case those values before requesting timestamp data for an Xid.
We already have a perfectly good interface for returning "the Xid you
requested is too old for us to have commit TS data for it", so let's use
that instead.
Backpatch to 9.5, where commit timestamps appeared.
Author: Craig Ringer
Discussion: https://www.postgresql.org/message-id/CAMsr+YFM5Q=+ry3mKvWEqRTxrB0iU3qUSRnS28nz6FJYtBwhJg@mail.gmail.com
A new thing also called a "barrier" is proposed, but whether we decide
to take that patch or not, this file seems to have outlived its
usefulness.
Thomas Munro
Condition variables provide a flexible way to sleep until a
cooperating process causes an arbitrary condition to become true. In
simple cases, this can be accomplished with a WaitLatch/ResetLatch
loop; the cooperating process can call SetLatch after performing work
that might cause the condition to be satisfied, and the waiting
process can recheck the condition each time. However, if the process
performing the work doesn't have an easy way to identify which
processes might be waiting, this doesn't work, because it can't
identify which latches to set. Condition variables solve that problem
by internally maintaining a list of waiters; a process that may have
caused some waiter's condition to be satisfied must "signal" or
"broadcast" on the condition variable.
Robert Haas and Thomas Munro
The debug messages that merely print StartTransactionCommand,
CommitTransactionCommand, ProcessUtilty, or ProcessQuery with no
additional details seem to be useless. Get rid of them.
The transaction status messages produced by ShowTransactionState are
occasionally useful, but they are extremely verbose, producing
multiple lines of log output every time they fire, which can happens
multiple times per transaction. So, reduce the level to DEBUG5; avoid
emitting an extra line just to explain which debug point is at issue;
and tighten up the rest of the message so it doesn't use quite so much
horizontal space.
With these changes, it's possible to run a somewhat busy system with a
log level even as high as DEBUG4, whereas previously anything above
DEBUG2 would flood the log with output that probably wasn't really all
that useful.
Replay of XLOG_BTREE_VACUUM during Hot Standby was previously thought to
require complex interlocking that matched the requirements on the
master. This required an O(N) operation that became a significant
problem with large indexes, causing replication delays of seconds or in
some cases minutes while the XLOG_BTREE_VACUUM was replayed.
This commit skips the “pin scan” that was previously required, by
observing in detail when and how it is safe to do so, with full
documentation. The pin scan is skipped only in replay; the VACUUM code
path on master is not touched here.
No tests included. Manual tests using an additional patch to view WAL records
and their timing have shown the change in WAL records and their handling has
successfully reduced replication delay.
This is a back-patch of commits 687f2cd7a0, 3e4b7d8798, b602842613
by Simon Riggs, to branches 9.4 and 9.5. No further backpatch is
possible because this depends on catalog scans being MVCC. I (Álvaro)
additionally updated a slight problem in the README, which explains why
this touches the 9.6 and master branches.
Invent a new function heap_modify_tuple_by_cols() that is functionally
equivalent to SPI_modifytuple except that it always allocates its result
by simple palloc. I chose however to make the API details a bit more
like heap_modify_tuple: pass a tupdesc rather than a Relation, and use
bool convention for the isnull array.
Use this function in place of SPI_modifytuple at all call sites where the
intended behavior is to allocate in current context. (There actually are
only two call sites left that depend on the old behavior, which makes me
wonder if we should just drop this function rather than keep it.)
This new function is easier to use than heap_modify_tuple() for purposes
of replacing a single column (or, really, any fixed number of columns).
There are a number of places where it would simplify the code to change
over, but I resisted that temptation for the moment ... everywhere except
in plpgsql's exec_assign_value(); changing that might offer some small
performance benefit, so I did it.
This is on the way to removing SPI_push/SPI_pop, but it seems like
good code cleanup in its own right.
Discussion: <9633.1478552022@sss.pgh.pa.us>
When squeezing a bucket during vacuum, it's not necessary to retain
any tuples already marked as dead, so ignore them when deciding which
tuples must be moved in order to empty a bucket page. Similarly, when
splitting a bucket, relocating dead tuples to the new bucket is a
waste of effort; instead, just ignore them.
Amit Kapila, reviewed by me. Testing help provided by Ashutosh
Sharma.
Generally, WAL resource managers are only supposed to examine the
top 4 bits of a WAL record's xl_info; the rest are reserved for
the WAL mechanism itself. A few places were not consistent about
doing this with respect to XLOG_CHECKPOINT and XLOG_SWITCH records.
There's no bug currently, since no additional bits ever get set in
these specific record types, but that might not be true forever.
Let's follow the generic coding rule here too.
Michael Paquier
If a restartpoint flushed no dirty buffers, it could fail to update
the minimum recovery point, leading to a minimum recovery point prior
to the starting REDO location. perform_base_backup() would interpret
that as meaning that no WAL files at all needed to be included in the
backup, failing an internal sanity check. To fix, have restartpoints
always update the minimum recovery point to just after the checkpoint
record itself, so that the file (or files) containing the checkpoint
record will always be included in the backup.
Code by Amit Kapila, per a design suggestion by me, with some
additional work on the code comment by me. Test case by Michael
Paquier. Report by Kyotaro Horiguchi.
An oversight in setting the boundaries of known commit timestamps during
startup caused old commit timestamps to become inaccessible after a
server restart.
Author and reporter: Julien Rouhaud
Review, test code: Craig Ringer
"xlog" is not a particularly clear abbreviation for "write-ahead log",
and it sometimes confuses users into believe that the contents of the
"pg_xlog" directory are not critical data, leading to unpleasant
consequences. So, rename the directory to "pg_wal".
This patch modifies pg_upgrade and pg_basebackup to understand both
the old and new directory layouts; the former is necessary given the
purpose of the tool, while the latter merely avoids an unnecessary
backward-compatibility break.
We may wish to consider renaming other programs, switches, and
functions which still use the old "xlog" naming to also refer to
"wal". However, that's still under discussion, so let's do just this
much for now.
Discussion: CAB7nPqTeC-8+zux8_-4ZD46V7YPwooeFxgndfsq5Rg8ibLVm1A@mail.gmail.com
Michael Paquier
When a relation is truncated, it is important that the FSM is truncated as
well. Otherwise, after recovery, the FSM can return a page that has been
truncated away, leading to errors like:
ERROR: could not read block 28991 in file "base/16390/572026": read only 0
of 8192 bytes
We were using MarkBufferDirtyHint() to dirty the buffer holding the last
remaining page of the FSM, but during recovery, that might in fact not
dirty the page, and the FSM update might be lost.
To fix, use the stronger MarkBufferDirty() function. MarkBufferDirty()
requires us to do WAL-logging ourselves, to protect from a torn page, if
checksumming is enabled.
Also fix an oversight in visibilitymap_truncate: it also needs to WAL-log
when checksumming is enabled.
Analysis by Pavan Deolasee.
Discussion: <CABOikdNr5vKucqyZH9s1Mh0XebLs_jRhKv6eJfNnD2wxTn=_9A@mail.gmail.com>
These functions were originally added in commit d8cedf67a to support
use of int2vector columns as catcache lookup keys. However, there are
no catcaches that use such columns. (Indeed I now think it must always
have been dead code: a catcache with such a key column would need an
underlying unique index on the column, but we've never had an int2vector
btree opclass.)
Getting rid of the int2vector-specific operator and function does not
lose any functionality, because operations on int2vectors will now fall
back to the generic anyarray support. This avoids a wart that a btree
index on an int2vector column (made using anyarray_ops) would fail to
match equality searches, because int2vectoreq wasn't a member of the
opclass. We don't really care much about that, since int2vector is not
meant as a type for users to use, but it's silly to have extra code and
less functionality.
If we ever do want a catcache to be indexed by an int2vector column,
we'd need to put back full btree and hash opclasses for int2vector,
comparable to the support for oidvector. (The anyarray code can't be
used at such a low level, because it needs to do catcache lookups.)
But we'll deal with that if/when the need arises.
Also worth noting is that removal of the hash int2vector_ops opclass will
break any user-created hash indexes on int2vector columns. While hash
anyarray_ops would serve the same purpose, it would probably not compute
the same hash values and thus wouldn't be on-disk-compatible. Given that
int2vector isn't a user-facing type and we're planning other incompatible
changes in hash indexes for v10 anyway, this doesn't seem like something
to worry about, but it's probably worth mentioning here.
Amit Langote
Discussion: <d9bb74f8-b194-7307-9ebd-90645d377e45@lab.ntt.co.jp>
While this isn't a lot of code, it's been essentially untestable for
a very long time, because libpq doesn't support anything older than
protocol 2.0, and has not since release 6.3. There's no reason to
believe any other client-side code still uses that protocol, either.
Discussion: <2661.1475849167@sss.pgh.pa.us>
WaitLatch, WaitLatchOrSocket, and WaitEventSetWait now taken an
additional wait_event_info parameter; legal values are defined in
pgstat.h. This makes it possible to uniquely identify every point in
the core code where we are waiting for a latch; extensions can pass
WAIT_EXTENSION.
Because latches were the major wait primitive not previously covered
by this patch, it is now possible to see information in
pg_stat_activity on a large number of important wait events not
previously addressed, such as ClientRead, ClientWrite, and SyncRep.
Unfortunately, many of the wait events added by this patch will fail
to appear in pg_stat_activity because they're only used in background
processes which don't currently appear in pg_stat_activity. We should
fix this either by creating a separate view for such information, or
else by deciding to include them in pg_stat_activity after all.
Michael Paquier and Robert Haas, reviewed by Alexander Korotkov and
Thomas Munro.
We had thirty different GIN array opclasses sharing the same operators and
support functions. That still didn't cover all the built-in types, nor
did it cover arrays of extension-added types. What we want is a single
polymorphic opclass for "anyarray". There were two missing features needed
to make this possible:
1. We have to be able to declare the index storage type as ANYELEMENT
when the opclass is declared to index ANYARRAY. This just takes a few
more lines in index_create(). Although this currently seems of use only
for GIN, there's no reason to make index_create() restrict it to that.
2. We have to be able to identify the proper GIN compare function for
the index storage type. This patch proceeds by making the compare function
optional in GIN opclass definitions, and specifying that the default btree
comparison function for the index storage type will be looked up when the
opclass omits it. Again, that seems pretty generically useful.
Since the comparison function lookup is done in initGinState(), making
use of the second feature adds an additional cache lookup to GIN index
access setup. It seems unlikely that that would be very noticeable given
the other costs involved, but maybe at some point we should consider
making GinState data persist longer than it now does --- we could keep it
in the index relcache entry, perhaps.
Rather fortuitously, we don't seem to need to do anything to get this
change to play nice with dump/reload or pg_upgrade scenarios: the new
opclass definition is automatically selected to replace existing index
definitions, and the on-disk data remains compatible. Also, if a user has
created a custom opclass definition for a non-builtin type, this doesn't
break that, since CREATE INDEX will prefer an exact match to opcintype
over a match to ANYARRAY. However, if there's anyone out there with
handwritten DDL that explicitly specifies _bool_ops or one of the other
replaced opclass names, they'll need to adjust that.
Tom Lane, reviewed by Enrique Meneses
Discussion: <14436.1470940379@sss.pgh.pa.us>
Apparent copy-and-pasteo in standby_desc_invalidations() had two
entries for msg->id == SHAREDINVALRELMAP_ID.
Aleksander Alekseev
Discussion: <20160923090814.GB1238@e733>
Move the updating of the control file to "in production" status until
the point where WAL writes are allowed. Before, there could be a
significant gap between the control file update and write transactions
actually being allowed. This makes it more reliable to use the control
status to verify the end of a promotion.
From: Michael Paquier <michael.paquier@gmail.com>
The GiST search queue is implemented as a pairing heap rather than as
Red-Black Tree, since 9.5 (commit e7032610). I neglected these comments
in that commit.
lockdefs.h was only split from lock.h relatively recently, and
represents a minimal subset of the old lock.h. heapam.h only needs
that smaller subset, so adjust it to include only that. This requires
some corresponding adjustments elsewhere.
Peter Geoghegan
The full generality of deleting an arbitrary number of tuples is no longer
needed, so let's save some code and cycles by replacing the original coding
with an implementation based on PageIndexTupleDelete.
We can always get back the old code from git if we need it again for new
callers (though I don't care for its willingness to mess with line pointers
it wasn't told to mess with).
Discussion: <552.1473445163@sss.pgh.pa.us>
PageIndexTupleOverwrite performs approximately the same function as
PageIndexTupleDelete (or PageIndexDeleteNoCompact) followed by PageAddItem
targeting the same item pointer offset. But in the case where the new
tuple is the same size as the old, it avoids shuffling other data around on
the page, because the new tuple is placed where the old one was rather than
being appended to the end of the page. This has been shown to provide a
substantial speedup for some GiST use-cases.
Also, this change allows some API simplifications: we can get rid of
the rather klugy and error-prone PAI_ALLOW_FAR_OFFSET flag for
PageAddItemExtended, since that was used only to cover a corner case
for BRIN that's better expressed by using PageIndexTupleOverwrite.
Note that this patch causes a rather subtle WAL incompatibility: the
physical page content change represented by certain WAL records is now
different than it was before, because while the tuples have the same
itempointer line numbers, the tuples themselves are in different places.
I have not bumped the WAL version number because I think it doesn't matter
unless you are trying to do bitwise comparisons of original and replayed
pages, and in any case we're early in a devel cycle and there will probably
be more WAL changes before v10 gets out the door.
There is probably room to make use of PageIndexTupleOverwrite in SP-GiST
and GIN too, but that is left for a future patch.
Andrey Borodin, reviewed by Anastasia Lubennikova, whacked around a bit
by me
Discussion: <CAJEAwVGQjGGOj6mMSgMwGvtFd5Kwe6VFAxY=uEPZWMDjzbn4VQ@mail.gmail.com>
When heap_lock_tuple decides to follow the update chain, it tried to
also lock any version of the tuple that was created by an update that
was subsequently rolled back. This is pointless, since for all intents
and purposes that tuple exists no more; and moreover it causes
misbehavior, as reported independently by Marko Tiikkaja and Marti
Raudsepp: some SELECT FOR UPDATE/SHARE queries may fail to return
the tuples, and assertion-enabled builds crash.
Fix by having heap_lock_updated_tuple test the xmin and return success
immediately if the tuple was created by an aborted transaction.
The condition where tuples become invisible occurs when an updated tuple
chain is followed by heap_lock_updated_tuple, which reports the problem
as HeapTupleSelfUpdated to its caller heap_lock_tuple, which in turn
propagates that code outwards possibly leading the calling code
(ExecLockRows) to believe that the tuple exists no longer.
Backpatch to 9.3. Only on 9.5 and newer this leads to a visible
failure, because of commit 27846f02c176; before that, heap_lock_tuple
skips the whole dance when the tuple is already locked by the same
transaction, because of the ancient HeapTupleSatisfiesUpdate behavior.
Still, the buggy condition may also exist in more convoluted scenarios
involving concurrent transactions, so it seems safer to fix the bug in
the old branches too.
Discussion:
https://www.postgresql.org/message-id/CABRT9RC81YUf1=jsmWopcKJEro=VoeG2ou6sPwyOUTx_qteRsg@mail.gmail.comhttps://www.postgresql.org/message-id/48d3eade-98d3-8b9a-477e-1a8dc32a724d@joh.to
Reading 2PC state files during recovery was borked, causing corruptions during
recovery. Effect limited to servers with 2PC, subtransactions and
recovery/replication.
Stas Kelvich, reviewed by Michael Paquier and Pavan Deolasee
StandbyRecoverPreparedTransactions() leaked the buffer
used for two phase state file. This was leaked once
at startup and at every shutdown checkpoint seen.
Backpatch to 9.6
Stas Kelvich
Add a location field to the DefElem struct, used to parse many utility
commands. Update various error messages to supply error position
information.
To propogate the error position information in a more systematic way,
create a ParseState in standard_ProcessUtility() and pass that to
interested functions implementing the utility commands. This seems
better than passing the query string and then reassembling a parse state
ad hoc, which violates the encapsulation of the ParseState type.
Reviewed-by: Pavel Stehule <pavel.stehule@gmail.com>
computeLeafRecompressWALData() tried to produce a uint16 WAL log field by
memcpy'ing the first two bytes of an int-sized variable. That accidentally
works on little-endian hardware, but not at all on big-endian. Replay then
thinks it's looking at an ADDITEMS action with zero entries, and reads the
first two bytes of the first TID therein as the next segno/action,
typically leading to "unexpected GIN leaf action" errors during replay.
Even if replay failed to crash, the resulting GIN index page would surely
be incorrect. To fix, just declare the variable as uint16 instead.
Per bug #14295 from Spencer Thomason (much thanks to Spencer for turning
his problem into a self-contained test case). This likely also explains
a previous report of the same symptom from Bernd Helmle.
Back-patch to 9.4 where the problem was introduced (by commit 14d02f0bb).
Discussion: <20160826072658.15676.7628@wrigleys.postgresql.org>
Possible-Report: <2DA7350F7296B2A142272901@eje.land.credativ.lan>
While we don't need multiple iterators at the moment, the interface is
nicer and less dangerous this way.
Aleksander Alekseev, with some changes by me.
This can't really work because standby_mode expects there to be more
WAL arriving, which there will not ever be because there's no WAL
receiver process to fetch it. Moreover, if standby_mode is on then
hot standby might also be turned on, causing even more strangeness
because that expects read-only sessions to be executing in parallel.
Bernd Helmle reported a case where btree_xlog_delete_get_latestRemovedXid
got confused, but rather than band-aiding individual problems it seems
best to prevent getting anywhere near this state in the first place.
Back-patch to all supported branches.
In passing, also fix some omissions of errcodes in other ereport's in
readRecoveryCommandFile().
Michael Paquier (errcode hacking by me)
Discussion: <00F0B2CEF6D0CEF8A90119D4@eje.credativ.lan>
Since the hash AM is going to be revamped to have WAL, this is a good
opportunity to clean up the include file a little bit to avoid including
a lot of extra stuff in the future.
Author: Amit Kapila
Previously pg_xlogdump failed to dump the contents of the WAL file
if the file starts with the continuation WAL record which spans
more than one pages. Since pg_xlogdump assumed that the continuation
record always fits on a page, it could not find the valid WAL record to
start reading from in that case.
This patch changes pg_xlogdump so that it can handle a continuation
WAL record which crosses a page boundary and find the valid record
to start reading from.
Back-patch to 9.3 where pg_xlogdump was introduced.
Author: Pavan Deolasee
Reviewed-By: Michael Paquier and Craig Ringer
Discussion: CABOikdPsPByMiG6J01DKq6om2+BNkxHTPkOyqHM2a4oYwGKsqQ@mail.gmail.com
I found that half a dozen (nearly 5%) of our AllocSetContextCreate calls
had typos in the context-sizing parameters. While none of these led to
especially significant problems, they did create minor inefficiencies,
and it's now clear that expecting people to copy-and-paste those calls
accurately is not a great idea. Let's reduce the risk of future errors
by introducing single macros that encapsulate the common use-cases.
Three such macros are enough to cover all but two special-purpose contexts;
those two calls can be left as-is, I think.
While this patch doesn't in itself improve matters for third-party
extensions, it doesn't break anything for them either, and they can
gradually adopt the simplified notation over time.
In passing, change TopMemoryContext to use the default allocation
parameters. Formerly it could only be extended 8K at a time. That was
probably reasonable when this code was written; but nowadays we create
many more contexts than we did then, so that it's not unusual to have a
couple hundred K in TopMemoryContext, even without considering various
dubious code that sticks other things there. There seems no good reason
not to let it use growing blocks like most other contexts.
Back-patch to 9.6, mostly because that's still close enough to HEAD that
it's easy to do so, and keeping the branches in sync can be expected to
avoid some future back-patching pain. The bugs fixed by these changes
don't seem to be significant enough to justify fixing them further back.
Discussion: <21072.1472321324@sss.pgh.pa.us>
HandleParallelMessages leaked memory into the caller's context. Since it's
called from ProcessInterrupts, there is basically zero certainty as to what
CurrentMemoryContext is, which means we could be leaking into long-lived
contexts. Over the processing of many worker messages that would grow to
be a problem. Things could be even worse than just a leak, if we happened
to service the interrupt while ErrorContext is current: elog.c thinks it
can reset that on its own whim, possibly yanking storage out from under
HandleParallelMessages.
Give HandleParallelMessages its own dedicated context instead, which we can
reset during each call to ensure there's no accumulation of wasted memory.
Discussion: <16610.1472222135@sss.pgh.pa.us>
The previous coding here was capable of adding a "parallel worker" context
line to errors that were not, in fact, returned from a parallel worker.
Instead of using an errcontext callback to add that annotation, just paste
it onto the message by hand; this looks uglier but is more reliable.
Discussion: <19757.1472151987@sss.pgh.pa.us>
When there is an identifiable REPLICA IDENTITY index on the target table,
heap_update leaks the id_attrs bitmapset. That's not many bytes, but it
adds up over enough rows, since the code typically runs in a query-lifespan
context. Bug introduced in commit e55704d8b, which did a rather poor job
of cloning the existing use-pattern for RelationGetIndexAttrBitmap().
Per bug #14293 from Zhou Digoal. Back-patch to 9.4 where the bug was
introduced.
Report: <20160824114320.15676.45171@wrigleys.postgresql.org>
With Asserts off, these variables are set but never used, resulting
in warnings from pickier compilers. Fix that with our standard solution.
Per report from Jeff Janes.
Previously, the spgSplitTuple action could only create a new upper tuple
containing a single labeled node. This made it useless for opclasses
that prefer to work with fixed sets of nodes (labeled or otherwise),
which meant that restrictive prefixes could not be used with such
node definitions. Change the output field set for the choose() method
to allow it to specify any valid node set for the new upper tuple,
and to specify which of these nodes to place the modified lower tuple in.
In addition to its primary use for fixed node sets, this feature could
allow existing opclasses that use variable node sets to skip a separate
spgAddNode action when splitting a tuple, by setting up the node needed
for the incoming value as part of the spgSplitTuple action. However, care
would have to be taken to add the extra node only when it would not make
the tuple bigger than before. (spgAddNode can enlarge the tuple,
spgSplitTuple can't.)
This is a prerequisite for an upcoming SP-GiST inet opclass, but is
being committed separately to increase the visibility of the API change.
In passing, improve the documentation about the traverse-values feature
that was added by commit ccd6eb49a.
Emre Hasegeli, with cosmetic adjustments and documentation rework by me
Discussion: <CAE2gYzxtth9qatW_OAqdOjykS0bxq7AYHLuyAQLPgT7H9ZU0Cw@mail.gmail.com>
INSERT .. ON CONFLICT runs a pre-check of the possible conflicting
constraints before performing the actual speculative insertion. In case
the inserted tuple included TOASTed columns the ON CONFLICT condition
would be handled correctly in case the conflict was caught by the
pre-check, but if two transactions entered the speculative insertion
phase at the same time, one would have to re-try, and the code for
aborting a speculative insertion did not handle deleting the
speculatively inserted TOAST datums correctly.
TOAST deletion would fail with "ERROR: attempted to delete invisible
tuple" as we attempted to remove the TOAST tuples using
simple_heap_delete which reasoned that the given tuples should not be
visible to the command that wrote them.
This commit updates the heap_abort_speculative() function which aborts
the conflicting tuple to use itself, via toast_delete, for deleting
associated TOAST datums. Like before, the inserted toast rows are not
marked as being speculative.
This commit also adds a isolationtester spec test, exercising the
relevant code path. Unfortunately 9.5 cannot handle two waiting
sessions, and thus cannot execute this test.
Reported-By: Viren Negi, Oskari Saarenmaa
Author: Oskari Saarenmaa, edited a bit by me
Bug: #14150
Discussion: <20160519123338.12513.20271@wrigleys.postgresql.org>
Backpatch: 9.5, where ON CONFLICT was introduced
Per discussion, we should provide such functions to replace the lost
ability to discover AM properties by inspecting pg_am (cf commit
65c5fcd35). The added functionality is also meant to displace any code
that was looking directly at pg_index.indoption, since we'd rather not
believe that the bit meanings in that field are part of any client API
contract.
As future-proofing, define the SQL API to not assume that properties that
are currently AM-wide or index-wide will remain so unless they logically
must be; instead, expose them only when inquiring about a specific index
or even specific index column. Also provide the ability for an index
AM to override the behavior.
In passing, document pg_am.amtype, overlooked in commit 473b93287.
Andrew Gierth, with kibitzing by me and others
Discussion: <87mvl5on7n.fsf@news-spur.riddles.org.uk>
In _bt_unlink_halfdead_page(), we might fail to find an immediate left
sibling of the target page, perhaps because of corruption of the page
sibling links. The code intends to cope with this by just abandoning
the deletion attempt; but what actually happens is that it fails outright
due to releasing the same buffer lock twice. (And error recovery masks
a second problem, which is possible leakage of a pin on another page.)
Seems to have been introduced by careless refactoring in commit efada2b8e.
Since there are multiple cases to consider, let's make releasing the buffer
lock in the failure case the responsibility of _bt_unlink_halfdead_page()
not its caller.
Also, avoid fetching the leaf page's left-link again after we've dropped
lock on the page. This is probably harmless, but it's not exactly good
coding practice.
Per report from Kyotaro Horiguchi. Back-patch to 9.4 where the faulty code
was introduced.
Discussion: <20160803.173116.111915228.horiguchi.kyotaro@lab.ntt.co.jp>
tqual.h is included in some front-end compiles, and a static inline
breaks on buildfarm member castoroides. Since the macro is never
referenced, it should dodge that problem, although this doesn't
seem like the cleanest way of hiding things from front-end compiles.
Report and review by Tom Lane; patch by me.
As mentioned in its commit message, eca0f1db left open a race condition,
where a page could be marked all-visible, after the code checked
PageIsAllVisible() to pin the VM, but before the page is locked. Plug
that hole.
Reviewed-By: Robert Haas, Andres Freund
Author: Amit Kapila
Discussion: CAEepm=3fWAbWryVW9swHyLTY4sXVf0xbLvXqOwUoDiNCx9mBjQ@mail.gmail.com
Backpatch: -
Previously, we tested for MVCC snapshots to see whether they were too
old, but not TOAST snapshots, which can lead to complaints about missing
TOAST chunks if those chunks are subject to early pruning. Ideally,
the threshold lsn and timestamp for a TOAST snapshot would be that of
the corresponding MVCC snapshot, but since we have no way of deciding
which MVCC snapshot was used to fetch the TOAST pointer, use the oldest
active or registered snapshot instead.
Reported by Andres Freund, who also sketched out what the fix should
look like. Patch by me, reviewed by Amit Kapila.
As noted by Alvaro, there are CHECK_FOR_INTERRUPTS() calls in the shm_mq.c
functions called by HandleParallelMessages(). I believe they're all
unreachable since we always pass nowait = true, but it doesn't seem like
a great idea to assume that no such call will ever be reachable from
HandleParallelMessages(). If that did happen, there would be a risk of a
recursive call to HandleParallelMessages(), which it does not appear to be
designed for --- for example, there's nothing that would prevent
out-of-order processing of received messages. And certainly such cases
cannot easily be tested. So let's prevent it by holding off interrupts for
the duration of the function. Back-patch to 9.5 which contains identical
code.
Discussion: <14869.1470083848@sss.pgh.pa.us>
ParallelMessagePending *must* be marked volatile, because it's set
by a signal handler. On the other hand, it's pointless for
HandleParallelMessageInterrupt to save/restore errno; that must be,
and is, done at the outer level of the SIGUSR1 signal handler.
Calling CHECK_FOR_INTERRUPTS() inside HandleParallelMessages, which itself
is called from CHECK_FOR_INTERRUPTS(), seems both useless and hazardous.
The comment claiming that this is needed to handle the error queue going
away is certainly misguided, in any case.
Improve a couple of error message texts, and use
ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE to report loss of parallel worker
connection, since that's what's used in e.g. tqueue.c. (Maybe it would be
worth inventing a dedicated ERRCODE for this type of failure? But I do not
think ERRCODE_INTERNAL_ERROR is appropriate.)
Minor stylistic cleanups.
Since a892234 & fd31cd265 the visibilitymap's freeze bit is used to
avoid vacuuming the whole relation in anti-wraparound vacuums. Doing so
correctly relies on not adding xids to the heap without also unsetting
the visibilitymap flag. Tuple locking related code has not done so.
To allow selectively resetting all-frozen - to avoid pessimizing
heap_lock_tuple - allow to selectively reset the all-frozen with
visibilitymap_clear(). To avoid having to use
visibilitymap_get_status (e.g. via VM_ALL_FROZEN) inside a critical
section, have visibilitymap_clear() return whether any bits have been
reset.
There's a remaining issue (denoted by XXX): After the PageIsAllVisible()
check in heap_lock_tuple() and heap_lock_updated_tuple_rec() the page
status could theoretically change. Practically that currently seems
impossible, because updaters will hold a page level pin already. Due to
the next beta coming up, it seems better to get the required WAL magic
bump done before resolving this issue.
The added flags field fields to xl_heap_lock and xl_heap_lock_updated
require bumping the WAL magic. Since there's already been a catversion
bump since the last beta, that's not an issue.
Reviewed-By: Robert Haas, Amit Kapila and Andres Freund
Author: Masahiko Sawada, heavily revised by Andres Freund
Discussion: CAEepm=3fWAbWryVW9swHyLTY4sXVf0xbLvXqOwUoDiNCx9mBjQ@mail.gmail.com
Backpatch: -
We've broken this code path at least twice in the past, so it's prudent
to have a test case that covers it. To allow exercising the code path
without creating a very large (and slow to run) test case, redefine the
sort threshold to be bounded by maintenance_work_mem as well as the number
of available buffers. While at it, fix an ancient oversight that when
building a temp index, the number of available buffers is not NBuffers but
NLocBuffer. Also, if assertions are enabled, apply a direct test that the
sort actually does return the tuples in the expected order.
Peter Geoghegan
Patch: <CAM3SWZTBAo4hjbBd780+MrOKiKp_TMo1N3A0Rw9_im8gbD7fQA@mail.gmail.com>
When heap_update needs to look for a page for the new tuple version,
because the current one doesn't have sufficient free space, or when
columns have to be processed by the tuple toaster, it has to release the
lock on the old page during that. Otherwise there'd be lock ordering and
lock nesting issues.
To avoid concurrent sessions from trying to update / delete / lock the
tuple while the page's content lock is released, the tuple's xmax is set
to the current session's xid.
That unfortunately was done without any WAL logging, thereby violating
the rule that no XIDs may appear on disk, without an according WAL
record. If the database were to crash / fail over when the page level
lock is released, and some activity lead to the page being written out
to disk, the xid could end up being reused; potentially leading to the
row becoming invisible.
There might be additional risks by not having t_ctid point at the tuple
itself, without having set the appropriate lock infomask fields.
To fix, compute the appropriate xmax/infomask combination for locking
the tuple, and perform WAL logging using the existing XLOG_HEAP_LOCK
record. That allows the fix to be backpatched.
This issue has existed for a long time. There appears to have been
partial attempts at preventing dangers, but these never have fully been
implemented, and were removed a long time ago, in
11919160 (cf. HEAP_XMAX_UNLOGGED).
In master / 9.6, there's an additional issue, namely that the
visibilitymap's freeze bit isn't reset at that point yet. Since that's a
new issue, introduced only in a892234f83, that'll be fixed in a
separate commit.
Author: Masahiko Sawada and Andres Freund
Reported-By: Different aspects by Thomas Munro, Noah Misch, and others
Discussion: CAEepm=3fWAbWryVW9swHyLTY4sXVf0xbLvXqOwUoDiNCx9mBjQ@mail.gmail.com
Backpatch: 9.1/all supported versions
0ac5ad5 started to compress infomask bits in WAL records. Unfortunately
the replay routines for XLOG_HEAP_LOCK/XLOG_HEAP2_LOCK_UPDATED forgot to
reset the HEAP_XMAX_INVALID (and some other) hint bits.
Luckily that's not problematic in the majority of cases, because after a
crash/on a standby row locks aren't meaningful. Unfortunately that does
not hold true in the presence of prepared transactions. This means that
after a crash, or after promotion, row level locks held by a prepared,
but not yet committed, prepared transaction might not be enforced.
Discussion: 20160715192319.ubfuzim4zv3rqnxv@alap3.anarazel.de
Backpatch: 9.3, the oldest branch on which 0ac5ad5 is present.
When key-share locking a tuple that has been not-key-updated, and the
update is a committed transaction, in some cases we raised
serializability errors:
ERROR: could not serialize access due to concurrent update
Because the key-share doesn't conflict with the update, the error is
unnecessary and inconsistent with the case that the update hasn't
committed yet. This causes problems for some usage patterns, even if it
can be claimed that it's sufficient to retry the aborted transaction:
given a steady stream of updating transactions and a long locking
transaction, the long transaction can be starved indefinitely despite
multiple retries.
To fix, we recognize that HeapTupleSatisfiesUpdate can return
HeapTupleUpdated when an updating transaction has committed, and that we
need to deal with that case exactly as if it were a non-committed
update: verify whether the two operations conflict, and if not, carry on
normally. If they do conflict, however, there is a difference: in the
HeapTupleBeingUpdated case we can just sleep until the concurrent
transaction is gone, while in the HeapTupleUpdated case this is not
possible and we must raise an error instead.
Per trouble report from Olivier Dony.
In addition to a couple of test cases that verify the changed behavior,
I added a test case to verify the behavior that remains unchanged,
namely that errors are raised when a update that modifies the key is
used. That must still generate serializability errors. One
pre-existing test case changes behavior; per discussion, the new
behavior is actually the desired one.
Discussion: https://www.postgresql.org/message-id/560AA479.4080807@odoo.comhttps://www.postgresql.org/message-id/20151014164844.3019.25750@wrigleys.postgresql.org
Backpatch to 9.3, where the problem appeared.
GiST index build could go into an infinite loop when presented with boxes
(or points, circles or polygons) containing NaN component values. This
happened essentially because the code assumed that x == x is true for any
"double" value x; but it's not true for NaNs. The looping behavior was not
the only problem though: we also attempted to sort the items using simple
double comparisons. Since NaNs violate the trichotomy law, qsort could
(in principle at least) get arbitrarily confused and mess up the sorting of
ordinary values as well as NaNs. And we based splitting choices on box size
calculations that could produce NaNs, again resulting in undesirable
behavior.
To fix, replace all comparisons of doubles in this logic with
float8_cmp_internal, which is NaN-aware and is careful to sort NaNs
consistently, higher than any non-NaN. Also rearrange the box size
calculation to not produce NaNs; instead it should produce an infinity
for a box with NaN on one side and not-NaN on the other.
I don't by any means claim that this solves all problems with NaNs in
geometric values, but it should at least make GiST index insertion work
reliably with such data. It's likely that the index search side of things
still needs some work, and probably regular geometric operations too.
But with this patch we're laying down a convention for how such cases
ought to behave.
Per bug #14238 from Guang-Dih Lei. Back-patch to 9.2; the code used before
commit 7f3bd86843 is quite different and doesn't lock up on my simple
test case, nor on the submitter's dataset.
Report: <20160708151747.1426.60150@wrigleys.postgresql.org>
Discussion: <28685.1468246504@sss.pgh.pa.us>
On a standby, ThisTimelineID is always 0, so we would generate a
filename in timeline 0 even for other timelines. Instead, use starttli
which we have retreived from the controlfile.
Report by: Francesco Canovai in bug #14230
Author: Marco Nenciarini
Reviewed by: Michael Paquier and Amit Kapila
Previously, workers sent data to the leader using the client encoding.
That mostly worked, but the leader the converted the data back to the
server encoding. Since not all encoding conversions are reversible,
that could provoke failures. Fix by using the database encoding for
all communication between worker and leader.
Also, while temporary changes to GUC settings, as from the SET clause
of a function, are in general OK for parallel query, changing
client_encoding this way inside of a parallel worker is not OK.
Previously, that would have confused the leader; with these changes,
it would not confuse the leader, but it wouldn't do anything either.
So refuse such changes in parallel workers.
Also, the previous code naively assumed that when it received a
NotifyResonse from the worker, it could pass that directly back to the
user. But now that worker-to-leader communication always uses the
database encoding, that's clearly no longer correct - though,
actually, the old way was always broken for V2 clients. So
disassemble and reconstitute the message instead.
Issues reported by Peter Eisentraut. Patch by me, reviewed by
Peter Eisentraut.
After pg_upgrade, it is possible that some tuples' Xmax have multixacts
corresponding to the old installation; such multixacts cannot have
running members anymore. In many code sites we already know not to read
them and clobber them silently, but at least when VACUUM tries to freeze
a multixact or determine whether one needs freezing, there's an attempt
to resolve it to its member transactions by calling GetMultiXactIdMembers,
and if the multixact value is "in the future" with regards to the
current valid multixact range, an error like this is raised:
ERROR: MultiXactId 123 has not been created yet -- apparent wraparound
and vacuuming fails. Per discussion with Andrew Gierth, it is completely
bogus to try to resolve multixacts coming from before a pg_upgrade,
regardless of where they stand with regards to the current valid
multixact range.
It's possible to get from under this problem by doing SELECT FOR UPDATE
of the problem tuples, but if tables are large, this is slow and
tedious, so a more thorough solution is desirable.
To fix, we realize that multixacts in xmax created in 9.2 and previous
have a specific bit pattern that is never used in 9.3 and later (we
already knew this, per comments and infomask tests sprinkled in various
places, but we weren't leveraging this knowledge appropriately).
Whenever the infomask of the tuple matches that bit pattern, we just
ignore the multixact completely as if Xmax wasn't set; or, in the case
of tuple freezing, we act as if an unwanted value is set and clobber it
without decoding. This guarantees that no errors will be raised, and
that the values will be progressively removed until all tables are
clean. Most callers of GetMultiXactIdMembers are patched to recognize
directly that the value is a removable "empty" multixact and avoid
calling GetMultiXactIdMembers altogether.
To avoid changing the signature of GetMultiXactIdMembers() in back
branches, we keep the "allow_old" boolean flag but rename it to
"from_pgupgrade"; if the flag is true, we always return an empty set
instead of looking up the multixact. (I suppose we could remove the
argument in the master branch, but I chose not to do so in this commit).
This was broken all along, but the error-facing message appeared first
because of commit 8e9a16ab8f and was partially fixed in a25c2b7c4d.
This fix, backpatched all the way back to 9.3, goes approximately in the
same direction as a25c2b7c4d but should cover all cases.
Bug analysis by Andrew Gierth and Álvaro Herrera.
A number of public reports match this bug:
https://www.postgresql.org/message-id/20140330040029.GY4582@tamriel.snowman.nethttps://www.postgresql.org/message-id/538F3D70.6080902@publicrelay.comhttps://www.postgresql.org/message-id/556439CF.7070109@pscs.co.ukhttps://www.postgresql.org/message-id/SG2PR06MB0760098A111C88E31BD4D96FB3540@SG2PR06MB0760.apcprd06.prod.outlook.comhttps://www.postgresql.org/message-id/20160615203829.5798.4594@wrigleys.postgresql.org
When the index is predicted to need more than NBuffers buckets,
CREATE INDEX attempts to sort the index entries by hash key before
insertion, so as to reduce thrashing. This code path got broken by
commit 9f03ca9151, which overlooked that _hash_form_tuple() is not
just an alias for index_form_tuple(). The index got built anyway, but
with garbage data, so that searches for pre-existing tuples always
failed. Fix by refactoring to separate construction of the indexable
data from calling index_form_tuple().
Per bug #14210 from Daniel Newman. Back-patch to 9.5 where the
bug was introduced.
Report: <20160623162507.17237.39471@wrigleys.postgresql.org>
This requires some core changes as well so that we can properly
WAL-log the truncation. Specifically, it changes the format of the
XLOG_SMGR_TRUNCATE WAL record, so bump XLOG_PAGE_MAGIC.
Patch by me, reviewed but not fully endorsed by Andres Freund.
The fact that no workers were successfully launched in the previous
iteration does not excuse us from setting up properly to try again.
This appears to explain crashes I saw in parallel regression testing
due to error_mqh being NULL when it shouldn't be.
Minor other cosmetic fixes too.
Commit a892234f83 added a new bit per
page to the visibility map fork indicating whether the page is
all-frozen, but incorrectly assumed that if lazy_scan_heap chose to
freeze a tuple then that tuple would not need to later be frozen
again. This turns out to be false, because xmin and xmax (and
conceivably xvac, if dealing with tuples from very old releases) could
be frozen at separate times.
Thanks to Andres Freund for help in uncovering and tracking down this
issue.
Transmit the leader's temp-namespace state to workers. This is important
because without it, the workers do not really have the same search path
as the leader. For example, there is no good reason (and no extant code
either) to prevent a worker from executing a temp function that the
leader created previously; but as things stood it would fail to find the
temp function, and then either fail or execute the wrong function entirely.
We still prohibit a worker from creating a temp namespace on its own.
In effect, a worker can only see the session's temp namespace if the leader
had created it before starting the worker, which seems like the right
semantics.
Also, transmit the leader's BackendId to workers, and arrange for workers
to use that when determining the physical file path of a temp relation
belonging to their session. While the original intent was to prevent such
accesses entirely, there were a number of holes in that, notably in places
like dbsize.c which assume they can safely access temp rels of other
sessions anyway. We might as well get this right, as a small down payment
on someday allowing workers to access the leader's temp tables. (With
this change, directly using "MyBackendId" as a relation or buffer backend
ID is deprecated; you should use BackendIdForTempRelations() instead.
I left a couple of such uses alone though, as they're not going to be
reachable in parallel workers until we do something about localbuf.c.)
Move the thou-shalt-not-access-thy-leader's-temp-tables prohibition down
into localbuf.c, which is where it actually matters, instead of having it
in relation_open(). This amounts to recognizing that access to temp
tables' catalog entries is perfectly safe in a worker, it's only the data
in local buffers that is problematic.
Having done all that, we can get rid of the test in has_parallel_hazard()
that says that use of a temp table's rowtype is unsafe in parallel workers.
That test was unduly expensive, and if we really did need such a
prohibition, that was not even close to being a bulletproof guard for it.
(For example, any user-defined function executed in a parallel worker
might have attempted such access.)
This terminology provoked widespread complaints. So, instead, rename
the GUC max_parallel_degree to max_parallel_workers_per_gather
(leaving room for a possible future GUC max_parallel_workers that acts
as a system-wide limit), and rename the parallel_degree reloption to
parallel_workers. Rename structure members to match.
These changes create a dump/restore hazard for users of PostgreSQL
9.6beta1 who have set the reloption (or applied the GUC using ALTER
USER or ALTER DATABASE).
If a Gather node has read as many tuples as it needs (for example, due
to Limit) it may detach the queue connecting it to the worker before
reading all of the worker's tuples. Rather than let the worker
continue to generate and send all of the results, have it stop after
sending the next tuple.
More could be done here to stop the worker even quicker, but this is
about as well as we can hope to do for 9.6.
This is in response to a problem report from Andreas Seltenreich.
Commit 44339b892a should be actually be
sufficient to fix that example even without this change, but it seems
better to do this, too, since we might otherwise waste quite a large
amount of effort in one or more workers.
Discussion: CAA4eK1KOKGqmz9bGu+Z42qhRwMbm4R5rfnqsLCNqFs9j14jzEA@mail.gmail.com
Amit Kapila
Mostly these are just comments but there are a few in documentation
and a handful in code and tests. Hopefully this doesn't cause too much
unnecessary pain for backpatching. I relented from some of the most
common like "thru" for that reason. The rest don't seem numerous
enough to cause problems.
Thanks to Kevin Lyda's tool https://pypi.python.org/pypi/misspellings
Per post-commit review comments from Andres Freund, improve variable
names, comments, and in one place, slightly improve the code structure.
Masahiko Sawada
Commit 2ed5b87f96 introduced a bug in
mark/restore, in an attempt to optimize repeated restores to the
same page. This caused an assertion failure during a merge join
which fed directly from an index scan, although the impact would
not be limited to that case. Revert the bad chunk of code from
that commit.
While investigating this bug it was discovered that a particular
"paranoia" set of the mark position field would not prevent bad
behavior; it would just make it harder to diagnose. Change that
into an assertion, which will draw attention to any future problem
in that area more directly.
Backpatch to 9.5, where the bug was introduced.
Bug #14169 reported by Shinta Koyanagi.
Preliminary analysis by Tom Lane identified which commit caused
the bug.
BRIN was relying on the ability to remove a tuple from an index page,
then putting another tuple in the same line pointer. But PageAddItem
refuses to add a tuple beyond the first free item past the last used
item, and in particular, it rejects an attempt to add an item to an
empty page anywhere other than the first line pointer. PageAddItem
issues a WARNING and indicates to the caller that it failed, which in
turn causes the BRIN calling code to issue a PANIC, so the whole
sequence looks like this:
WARNING: specified item offset is too large
PANIC: failed to add BRIN tuple
To fix, create a new function PageAddItemExtended which is like
PageAddItem except that the two boolean arguments become a flags bitmap;
the "overwrite" and "is_heap" boolean flags in PageAddItem become
PAI_OVERWITE and PAI_IS_HEAP flags in the new function, and a new flag
PAI_ALLOW_FAR_OFFSET enables the behavior required by BRIN.
PageAddItem() retains its original signature, for compatibility with
third-party modules (other callers in core code are not modified,
either).
Also, in the belt-and-suspenders spirit, I added a new sanity check in
brinGetTupleForHeapBlock to raise an error if an TID found in the revmap
is not marked as live by the page header. This causes it to react with
"ERROR: corrupted BRIN index" to the bug at hand, rather than a hard
crash.
Backpatch to 9.5.
Bug reported by Andreas Seltenreich as detected by his handy sqlsmith
fuzzer.
Discussion: https://www.postgresql.org/message-id/87mvni77jh.fsf@elite.ansel.ydns.eu
Commit 65c5fcd353 broke this by removing a
header include directive that is conditionally required. Add that back
to nbtree.c, with annotation to keep pgrminclude from re-breaking it.
Peter Geoghegan
Report: <CAM3SWZTNjHFYW_UG8bu0BnogqQ2HfsTgkzXLueuUhfTcYbu5HA@mail.gmail.com>
Page image should be MAXALIGN'ed because existing code could directly align
pointers in page instead of align offset from beginning of page.
Found during play with indexes as extenstion, Alexander Korotkov and me
Some comments mentioned XLogReplayBuffer, but there's no such function:
that was an interim name for a function that got renamed to
XLogReadBufferForRedo, before commit 2c03216d83 was pushed.
This new limit affects both the max_parallel_degree GUC and the
parallel_degree reloption. There may some day be a use case for using
more than 1024 CPUs for a single query, but that's surely not the case
right now. Not only do not very many people have that many CPUs, but
the code hasn't been tested at that kind of scale and is very unlikely
to perform well, or even work at all, without a lot more work. The
issue addressed by commit 06bd458cb8 is
probably just one problem of many.
The idea of a more reasonable limit here was suggested by Tom Lane;
the value of 1024 was suggested by Amit Kapila.
That way, if the result overflows size_t, you'll get an error instead
of undefined behavior, which seems like a plus. This also has the
effect of casting the number of workers from int to Size, which is
better because it's harder to overflow int than size_t.
Dilip Kumar reported this issue and provided a patch upon which this
patch is based, but his version did use mul_size.
Hash indexes are not WAL-logged, and so do not maintain the LSN of
index pages. Since the "snapshot too old" feature counts on
detecting error conditions using the LSN of a table and all indexes
on it, this makes it impossible to safely do early vacuuming on any
table with a hash index, so add this to the tests for whether the
xid used to vacuum a table can be adjusted based on
old_snapshot_threshold.
While at it, add a paragraph to the docs for old_snapshot_threshold
which specifically mentions this and other aspects of the feature
which may otherwise surprise users.
Problem reported and patch reviewed by Amit Kapila
This reverts commits f07d18b6e9, 82c83b3372, 3a3b309041, and
24c5f1a103.
This feature has shown enough immaturity that it was deemed better to
rip it out before rushing some more fixes at the last minute. There are
discussions on larger changes in this area for the next release.
Previously, ginInsertCleanup could exit early if it detects that someone else
is cleaning up the pending list, without waiting for that someone else to
finish the job. But in this case vacuum could miss tuples to be deleted.
Cleanup process now locks metapage with a help of heavyweight
LockPage(ExclusiveLock), and it guarantees that there is no another cleanup
process at the same time. Lock is taken differently depending on caller of
cleanup process: any vacuums and gin_clean_pending_list() will be blocked
until lock becomes available, ordinary insert uses conditional lock to
prevent indefinite waiting on lock.
Insert into pending list doesn't use this lock, so insertion isn't blocked.
Also, patch adds stopping of cleanup process when at-start-cleanup-tail is
reached in order to prevent infinite cleanup in case of massive insertion. But
it will stop only for automatic maintenance tasks like autovacuum.
Patch introduces choice of limit of memory to use: autovacuum_work_mem,
maintenance_work_mem or work_mem depending on call path.
Patch for previous releases should be reworked due to changes between 9.6 and
previous ones in this area.
Discover and diagnostics by Jeff Janes and Tomas Vondra
Patch by me with some ideas of Jeff Janes
So far, when a transaction with pending invalidations, but without an
assigned xid, committed, we simply ignored those invalidation
messages. That's problematic, because those are actually sent for a
reason.
Known symptoms of this include that existing sessions on a hot-standby
replica sometimes fail to notice new concurrently built indexes and
visibility map updates.
The solution is to WAL log such invalidations in transactions without an
xid. We considered to alternatively force-assign an xid, but that'd be
problematic for vacuum, which might be run in systems with few xids.
Important: This adds a new WAL record, but as the patch has to be
back-patched, we can't bump the WAL page magic. This means that standbys
have to be updated before primaries; otherwise
"PANIC: standby_redo: unknown op code 32" errors can be encountered.
XXX:
Reported-By: Васильев Дмитрий, Masahiko Sawada
Discussion:
CAB-SwXY6oH=9twBkXJtgR4UC1NqT-vpYAtxCseME62ADwyK5OA@mail.gmail.comCAD21AoDpZ6Xjg=gFrGPnSn4oTRRcwK1EBrWCq9OqOHuAcMMC=w@mail.gmail.com
Commit 36a35c550a turned the interface between ginPlaceToPage and
its subroutines in gindatapage.c and ginentrypage.c into a royal mess:
page-update critical sections were started in one place and finished in
another place not even in the same file, and the very same subroutine
might return having started a critical section or not. Subsequent patches
band-aided over some of the problems with this design by making things
even messier.
One user-visible resulting problem is memory leaks caused by the need for
the subroutines to allocate storage that would survive until ginPlaceToPage
calls XLogInsert (as reported by Julien Rouhaud). This would not typically
be noticeable during retail index updates. It could be visible in a GIN
index build, in the form of memory consumption swelling to several times
the commanded maintenance_work_mem.
Another rather nasty problem is that in the internal-page-splitting code
path, we would clear the child page's GIN_INCOMPLETE_SPLIT flag well before
entering the critical section that it's supposed to be cleared in; a
failure in between would leave the index in a corrupt state. There were
also assorted coding-rule violations with little immediate consequence but
possible long-term hazards, such as beginning an XLogInsert sequence before
entering a critical section, or calling elog(DEBUG) inside a critical
section.
To fix, redefine the API between ginPlaceToPage() and its subroutines
by splitting the subroutines into two parts. The "beginPlaceToPage"
subroutine does what can be done outside a critical section, including
full computation of the result pages into temporary storage when we're
going to split the target page. The "execPlaceToPage" subroutine is called
within a critical section established by ginPlaceToPage(), and it handles
the actual page update in the non-split code path. The critical section,
as well as the XLOG insertion call sequence, are both now always started
and finished in ginPlaceToPage(). Also, make ginPlaceToPage() create and
work in a short-lived memory context to eliminate the leakage problem.
(Since a short-lived memory context had been getting created in the most
common code path in the subroutines, this shouldn't cause any noticeable
performance penalty; we're just moving the overhead up one call level.)
In passing, fix a bunch of comments that had gone unmaintained throughout
all this klugery.
Report: <571276DD.5050303@dalibo.com>
The reverted changes were intended to force a choice of whether any
newly-added BufferGetPage() calls needed to be accompanied by a
test of the snapshot age, to support the "snapshot too old"
feature. Such an accompanying test is needed in about 7% of the
cases, where the page is being used as part of a scan rather than
positioning for other purposes (such as DML or vacuuming). The
additional effort required for back-patching, and the doubt whether
the intended benefit would really be there, have indicated it is
best just to rely on developers to do the right thing based on
comments and existing usage, as we do with many other conventions.
This change should have little or no effect on generated executable
code.
Motivated by the back-patching pain of Tom Lane and Robert Haas
The code had a query-lifespan memory leak when encountering GIN entries
that have posting lists (rather than posting trees, ie, there are a
relatively small number of heap tuples containing this index key value).
With a suitable data distribution this could add up to a lot of leakage.
Problem seems to have been introduced by commit 36a35c550, so back-patch
to 9.4.
Julien Rouhaud
Rename this function to GenericXLogRegisterBuffer() to make it clearer
what it does, and leave room for other sorts of "register" actions in
future. Also, replace its "bool isNew" argument with an integer flags
argument, so as to allow adding more flags in future without an API
break.
Alexander Korotkov, adjusted slightly by me
The previous coding could allow the contents of the "hole" between pd_lower
and pd_upper to diverge during replay from what it had been when the update
was originally applied. This would pose a problem if checksums were in
use, and in any case would complicate forensic comparisons between master
and slave servers. So force the "hole" to contain zeroes, both at initial
application of a generically-logged action, and at replay.
Alexander Korotkov, adjusted slightly by me
Since we're requiring pages handled by generic_xlog.c to be standard
format, specify REGBUF_STANDARD when doing a full-page image, so that
xloginsert.c can compress out the "hole" between pd_lower and pd_upper.
Given the current API in which this path will be taken only for a newly
initialized page, the hole is likely to be particularly large in such
cases, so that this oversight could easily be performance-significant.
I don't notice any particular change in the runtime of contrib/bloom's
regression test, though.
Make the inner comparison loops of computeDelta() as tight as possible by
pulling considerations of valid and invalid ranges out of the inner loops,
and extending a match or non-match detection as far as possible before
deciding what to do next. To keep this tractable, give up the possibility
of merging fragments across the pd_lower to pd_upper gap. The fraction of
pages where that could happen (ie, there are 4 or fewer bytes in the gap,
*and* data changes immediately adjacent to it on both sides) is too small
to be worth spending cycles on.
Also, avoid two BLCKSZ-length memcpy()s by computing the delta before
moving data into the target buffer, instead of after. This doesn't save
nearly as many cycles as being tenser about computeDelta(), but it still
seems worth doing.
On my machine, this patch cuts a full 40% off the runtime of
contrib/bloom's regression test.
This routine is unsafe as implemented, because it invalidates the page
image pointers returned by previous GenericXLogRegister() calls.
Rather than complicate the API or the implementation to avoid that,
let's just get rid of it; the use-case for having it seems much
too thin to justify a lot of work here.
While at it, do some wordsmithing on the SGML docs for generic WAL.
Improve commentary, use more specific names for the delta fields,
const-ify pointer arguments where possible, avoid assuming that
initializing only the first element of a local array will guarantee
that the remaining elements end up as we need them. (I think that
code in generic_redo actually worked, but only because InvalidBuffer
is zero; this is a particularly ugly way of depending on that ...)
This feature is controlled by a new old_snapshot_threshold GUC. A
value of -1 disables the feature, and that is the default. The
value of 0 is just intended for testing. Above that it is the
number of minutes a snapshot can reach before pruning and vacuum
are allowed to remove dead tuples which the snapshot would
otherwise protect. The xmin associated with a transaction ID does
still protect dead tuples. A connection which is using an "old"
snapshot does not get an error unless it accesses a page modified
recently enough that it might not be able to produce accurate
results.
This is similar to the Oracle feature, and we use the same SQLSTATE
and error message for compatibility.
This patch is a no-op patch which is intended to reduce the chances
of failures of omission once the functional part of the "snapshot
too old" patch goes in. It adds parameters for snapshot, relation,
and an enum to specify whether the snapshot age check needs to be
done for the page at this point. This initial patch passes NULL
for the first two new parameters and BGP_NO_SNAPSHOT_TEST for the
third. The follow-on patch will change the places where the test
needs to be made.
Now indexes (but only B-tree for now) can contain "extra" column(s) which
doesn't participate in index structure, they are just stored in leaf
tuples. It allows to use index only scan by using single index instead
of two or more indexes.
Author: Anastasia Lubennikova with minor editorializing by me
Reviewers: David Rowley, Peter Geoghegan, Jeff Janes
Benchmarking has shown that the current number of clog buffers limits
scalability. We've previously increased the number in 33aaa139, but
that's not sufficient with a large number of clients.
We've benchmarked the cost of increasing the limit by benchmarking worst
case scenarios; testing showed that 128 buffers don't cause a
regression, even in contrived scenarios, whereas 256 does
There are a number of more complex patches flying around to address
various clog scalability problems, but this is simple enough that we can
get it into 9.6; and is beneficial even after those patches have been
applied.
It is a bit unsatisfactory to increase this in small steps every few
releases, but a better solution seems to require a rewrite of slru.c;
not something done quickly.
Author: Amit Kapila and Andres Freund
Discussion: CAA4eK1+-=18HOrdqtLXqOMwZDbC_15WTyHiFruz7BvVArZPaAw@mail.gmail.com
The code that estimates what parallel degree should be uesd for the
scan of a relation is currently rather stupid, so add a parallel_degree
reloption that can be used to override the planner's rather limited
judgement.
Julien Rouhaud, reviewed by David Rowley, James Sewell, Amit Kapila,
and me. Some further hacking by me.
Contention on the relation extension lock can become quite fierce when
multiple processes are inserting data into the same relation at the same
time at a high rate. Experimentation shows the extending the relation
multiple blocks at a time improves scalability.
Dilip Kumar, reviewed by Petr Jelinek, Amit Kapila, and me.
While prior to this patch the user-visible effect on the database
of any set of successfully committed serializable transactions was
always consistent with some one-at-a-time order of execution of
those transactions, the presence of declarative constraints could
allow errors to occur which were not possible in any such ordering,
and developers had no good workarounds to prevent user-facing
errors where they were not necessary or desired. This patch adds
a check for serialization failure ahead of duplicate key checking
so that if a developer explicitly (redundantly) checks for the
pre-existing value they will get the desired serialization failure
where the problem is caused by a concurrent serializable
transaction; otherwise they will get a duplicate key error.
While it would be better if the reads performed by the constraints
could count as part of the work of the transaction for
serialization failure checking, and we will hopefully get there
some day, this patch allows a clean and reliable way for developers
to work around the issue. In many cases existing code will already
be doing the right thing for this to "just work".
Author: Thomas Munro, with minor editing of docs by me
Reviewed-by: Marko Tiikkaja, Kevin Grittner
Now that pg_dump will properly dump out any ACL changes made to
functions which exist in pg_catalog, switch to using the GRANT system
to manage access to those functions.
This means removing 'if (!superuser()) ereport()' checks from the
functions themselves and then REVOKEing EXECUTE right from 'public' for
these functions in system_views.sql.
Reviews by Alexander Korotkov, Jose Luis Tallon
API and mechanism to allow generic messages to be inserted into WAL that are
intended to be read by logical decoding plugins. This commit adds an optional
new callback to the logical decoding API.
Messages are either text or bytea. Messages can be transactional, or not, and
are identified by a prefix to allow multiple concurrent decoding plugins.
(Not to be confused with Generic WAL records, which are intended to allow crash
recovery of extensible objects.)
Author: Petr Jelinek and Andres Freund
Reviewers: Artur Zakirov, Tomas Vondra, Simon Riggs
Discussion: 5685F999.6010202@2ndquadrant.com
Previously non-exclusive backups had to be done using the replication protocol
and pg_basebackup. With this commit it's now possible to make them using
pg_start_backup/pg_stop_backup as well, as long as the backup program can
maintain a persistent connection to the database.
Doing this, backup_label and tablespace_map are returned as results from
pg_stop_backup() instead of being written to the data directory. This makes
the server safe from a crash during an ongoing backup, which can be a problem
with exclusive backups.
The old syntax of the functions remain and work exactly as before, but since the
new syntax is safer this should eventually be deprecated and removed.
Only reference documentation is included. The main section on backup still needs
to be rewritten to cover this, but since that is already scheduled for a separate
large rewrite, it's not included in this patch.
Reviewed by David Steele and Amit Kapila
Replay of XLOG_BTREE_VACUUM during Hot Standby was previously thought to require
complex interlocking that matched the requirements on the master. This required
an O(N) operation that became a significant problem with large indexes, causing
replication delays of seconds or in some cases minutes while the
XLOG_BTREE_VACUUM was replayed.
This commit skips the pin scan that was previously required, by observing in
detail when and how it is safe to do so, with full documentation. The pin
scan is skipped only in replay; the VACUUM code path on master is not
touched here and WAL is identical.
The current commit applies in all cases, effectively replacing commit
687f2cd7a0.
This interface is designed to give an access to WAL for extensions which
could implement new access method, for example. Previously it was
impossible because restoring from custom WAL would need to access system
catalog to find a redo custom function. This patch suggests generic way
to describe changes on page with standart layout.
Bump XLOG_PAGE_MAGIC because of new record type.
Author: Alexander Korotkov with a help of Petr Jelinek, Markus Nullmeier and
minor editorization by my
Reviewers: Petr Jelinek, Alvaro Herrera, Teodor Sigaev, Jim Nasby,
Michael Paquier
When decoding from a logical slot, it's necessary for xlog reading to be
able to read xlog from historical (i.e. not current) timelines;
otherwise, decoding fails after failover, because the archives are in
the historical timeline. This is required to make "failover logical
slots" possible; it currently has no other use, although theoretically
it could be used by an extension that creates a slot on a standby and
continues to replay from the slot when the standby is promoted.
This commit includes a module in src/test/modules with functions to
manipulate the slots (which is not otherwise possible in SQL code) in
order to enable testing, and a new test in src/test/recovery to ensure
that the behavior is as expected.
Author: Craig Ringer
Reviewed-By: Oleksii Kliukin, Andres Freund, Petr Jelínek
Some minor tweaks and comment additions, for cleanliness sake and to
avoid having the upcoming timeline-following patch be polluted with
unrelated cleanup.
Extracted from a larger patch by Craig Ringer, reviewed by Andres
Freund, with some additions by myself.
During scan sometimes it would be very helpful to know some information about
parent node or all ancestor nodes. Right now reconstructedValue could be used
but it's not a right usage of it (range opclass uses that).
traversalValue is arbitrary piece of memory in separate MemoryContext while
reconstructedVale should have the same type as indexed column.
Subsequent patches for range opclass and quad4d tree will use it.
Author: Alexander Lebedev, Teodor Sigaev
In this mode, the master waits for the transaction to be applied on
the remote side, not just written to disk. That means that you can
count on a transaction started on the standby to see all commits
previously acknowledged by the master.
To make this work, the standby sends a reply after replaying each
commit record generated with synchronous_commit >= 'remote_apply'.
This introduces a small inefficiency: the extra replies will be sent
even by standbys that aren't the current synchronous standby. But
previously-existing synchronous_commit levels make no attempt at all
to optimize which replies are sent based on what the primary cares
about, so this is no worse, and at least avoids any extra replies for
people not using the feature at all.
Thomas Munro, reviewed by Michael Paquier and by me. Some additional
tweaks by me.
This enables external code to create access methods. This is useful so
that extensions can add their own access methods which can be formally
tracked for dependencies, so that DROP operates correctly. Also, having
explicit support makes pg_dump work correctly.
Currently only index AMs are supported, but we expect different types to
be added in the future.
Authors: Alexander Korotkov, Petr Jelínek
Reviewed-By: Teodor Sigaev, Petr Jelínek, Jim Nasby
Commitfest-URL: https://commitfest.postgresql.org/9/353/
Discussion: https://www.postgresql.org/message-id/CAPpHfdsXwZmojm6Dx+TJnpYk27kT4o7Ri6X_4OSWcByu1Rm+VA@mail.gmail.com
The distinction between "archive" and "hot_standby" existed only because
at the time "hot_standby" was added, there was some uncertainty about
stability. This is now a long time ago. We would like to move forward
with simplifying the replication configuration, but this distinction is
in the way, because a primary server cannot tell (without asking a
standby or predicting the future) which one of these would be the
appropriate level.
Pick a new name for the combined setting to make it clearer that it
covers all (non-logical) backup and replication uses. The old values
are still accepted but are converted internally.
Reviewed-by: Michael Paquier <michael.paquier@gmail.com>
Reviewed-by: David Steele <david@pgmasters.net>
Commit d88976cfa1 removed this code from ginFreeScanKeys():
- if (entry->list)
- pfree(entry->list);
evidently in the belief that that ItemPointer array is allocated in the
keyCtx and so would be reclaimed by the following MemoryContextReset.
Unfortunately, it isn't and it won't. It'd likely be a good idea for
that to become so, but as a simple and back-patchable fix in the
meantime, restore this code to ginFreeScanKeys().
Also, add a similar pfree to where startScanEntry() is about to zero out
entry->list. I am not sure if there are any code paths where this
change prevents a leak today, but it seems like cheap future-proofing.
In passing, make the initial allocation of so->entries[] use palloc
not palloc0. The code doesn't depend on unused entries being zero;
if it did, the array-enlargement code in ginFillScanEntry() would be
wrong. So using palloc0 initially can only serve to confuse readers
about what the invariant is.
Per report from Felipe de Jesús Molina Bravo, via Jaime Casanova in
<CAJGNTeMR1ndMU2Thpr8GPDUfiHTV7idELJRFusA5UXUGY1y-eA@mail.gmail.com>
When a process is waiting for a heavyweight lock, we will now indicate
the type of heavyweight lock for which it is waiting. Also, you can
now see when a process is waiting for a lightweight lock - in which
case we will indicate the individual lock name or the tranche, as
appropriate - or for a buffer pin.
Amit Kapila, Ildus Kurbangaliev, reviewed by me. Lots of helpful
discussion and suggestions by many others, including Alexander
Korotkov, Vladimir Borodin, and many others.
Previously 2PC header was fixed at 200 bytes, which in most cases wasted
WAL space for a workload using 2PC heavily.
Pavan Deolasee, reviewed by Petr Jelinek
Renaming a file using rename(2) is not guaranteed to be durable in face
of crashes. Use the previously added durable_rename()/durable_link_or_rename()
in various places where we previously just renamed files.
Most of the changed call sites are arguably not critical, but it seems
better to err on the side of too much durability. The most prominent
known case where the previously missing fsyncs could cause data loss is
crashes at the end of a checkpoint. After the actual checkpoint has been
performed, old WAL files are recycled. When they're filled, their
contents are fdatasynced, but we did not fsync the containing
directory. An OS/hardware crash in an unfortunate moment could then end
up leaving that file with its old name, but new content; WAL replay
would thus not replay it.
Reported-By: Tomas Vondra
Author: Michael Paquier, Tomas Vondra, Andres Freund
Discussion: 56583BDD.9060302@2ndquadrant.com
Backpatch: All supported branches
An index search using a row comparison such as ROW(a, b) > ROW('x', 'y')
would stop upon reaching a NULL entry in the "b" column, ignoring the
fact that there might be non-NULL "b" values associated with later values
of "a". This happens because _bt_mark_scankey_required() marks the
subsidiary scankey for "b" as required, which is just wrong: it's for
a column after the one with the first inequality key (namely "a"), and
thus can't be considered a required match.
This bit of brain fade dates back to the very beginnings of our support
for indexed ROW() comparisons, in 2006. Kind of astonishing that no one
came across it before Glen Takahashi, in bug #14010.
Back-patch to all supported versions.
Note: the given test case doesn't actually fail in unpatched 9.1, evidently
because the fix for bug #6278 (i.e., stopping at nulls in either scan
direction) is required to make it fail. I'm sure I could devise a case
that fails in 9.1 as well, perhaps with something involving making a cursor
back up; but it doesn't seem worth the trouble.
Using this facility, any utility command can report the target relation
upon which it is operating, if there is one, and up to 10 64-bit
counters; the intent of this is that users should be able to figure out
what a utility command is doing without having to resort to ugly hacks
like attaching strace to a backend.
As a demonstration, this adds very crude reporting to lazy vacuum; we
just report the target relation and nothing else. A forthcoming patch
will make VACUUM report a bunch of additional data that will make this
much more interesting. But this gets the basic framework in place.
Vinayak Pokale, Rahila Syed, Amit Langote, Robert Haas, reviewed by
Kyotaro Horiguchi, Jim Nasby, Thom Brown, Masahiko Sawada, Fujii Masao,
and Masanori Oyama.
Commit a892234f83 added a second bit per
page to the visibility map, which still seems like a good idea, but it
also added a second page-level bit alongside PD_ALL_VISIBLE to track
whether the visibility map bit was set. That no longer seems like a
clever plan, because we don't really need that bit for anything. We
always clear both bits when the page is modified anyway.
Patch by me, reviewed by Kyotaro Horiguchi and Masahiko Sawada.
Previously recovery_min_apply_delay was applied even before recovery
had reached consistency. This could cause us to wait a long time
unexpectedly for read-only connections to be allowed. It's problematic
because the standby was useless during that wait time.
This patch changes recovery_min_apply_delay so that it's applied once
the database has reached the consistent state. That is, even if the delay
is set, the standby tries to replay WAL records as fast as possible until
it has reached consistency.
Author: Michael Paquier
Reviewed-By: Julien Rouhaud
Reported-By: Greg Clough
Backpatch: 9.4, where recovery_min_apply_delay was added
Bug: #13770
Discussion: http://www.postgresql.org/message-id/20151111155006.2644.84564@wrigleys.postgresql.org
A simple SELECT is handled by PortalRunSelect, not ProcessQuery. Also,
the previous indentation was unclear: change it so that a deeper level
of indentation indicates that the outer function calls the inner one.
Stas Kelvich
Originally, we didn't have nworkers_launched, so code that used parallel
contexts had to be preprared for the possibility that not all of the
workers requested actually got launched. But now we can count on knowing
the number of workers that were successfully launched, which can shave
off a few cycles and simplify some code slightly.
Amit Kapila, reviewed by Haribabu Kommi, per a suggestion from Peter
Geoghegan.
606c0123d6 attempted to reduce cost of index scans using > and <
strategies, though got that completely wrong in a few complex cases.
Revert whole patch until we find a safe optimization.
The new bit indicates whether every tuple on the page is already frozen.
It is cleared only when the all-visible bit is cleared, and it can be
set only when we vacuum a page and find that every tuple on that page is
both visible to every transaction and in no need of any future
vacuuming.
A future commit will use this new bit to optimize away full-table scans
that would otherwise be triggered by XID wraparound considerations. A
page which is merely all-visible must still be scanned in that case, but
a page which is all-frozen need not be. This commit does not attempt
that optimization, although that optimization is the goal here. It
seems better to get the basic infrastructure in place first.
Per discussion, it's very desirable for pg_upgrade to automatically
migrate existing VM forks from the old format to the new format. That,
too, will be handled in a follow-on patch.
Masahiko Sawada, reviewed by Kyotaro Horiguchi, Fujii Masao, Amit
Kapila, Simon Riggs, Andres Freund, and others, and substantially
revised by me.
StartupSUBTRANS() incorrectly handled cases near the max pageid in the subtrans
data structure, which in some cases could lead to errors in startup for Hot
Standby.
This patch wraps the pageids correctly, avoiding any such errors.
Identified by exhaustive crash testing by Jeff Janes.
Jeff Janes
Commit 4de82f7d7 increased the WAL flush rate, mainly to increase the
likelihood that hint bits can be set quickly. More quickly set hint bits
can reduce contention around the clog et al. But unfortunately the
increased flush rate can have a significant negative performance impact,
I have measured up to a factor of ~4. The reason for this slowdown is
that if there are independent writes to the underlying devices, for
example because shared buffers is a lot smaller than the hot data set,
or because a checkpoint is ongoing, the fdatasync() calls force cache
flushes to be emitted to the storage.
This is achieved by flushing WAL only if the last flush was longer than
wal_writer_delay ago, or if more than wal_writer_flush_after (new GUC)
unflushed blocks are pending. Based on some tests the default for
wal_writer_delay is 1MB, which seems to work well both on SSD and
rotational media.
To avoid negative performance impact due to 4de82f7d7 an earlier
commit (db76b1e) made SetHintBits() more likely to succeed; preventing
performance regressions in the pgbench tests I performed.
Discussion: 20160118163908.GW10941@awork2.anarazel.de
NextXID has been rendered in the form of a pg_lsn even though it
really is not. This can cause confusion, so change the format from
%u/%u to %u:%u, per discussion on hackers.
Complaint by me, patch by me and Bruce, reviewed by Michael Paquier
and Alvaro. Applied to HEAD only.
Author: Joe Conway, Bruce Momjian
Reviewed-by: Michael Paquier, Alvaro Herrera
Backpatch-through: master
Robert Haas
Tom Lane
Alvaro Herrera
Heikki Linnakangas
Peter Eisentraut
Simon Riggs
Fujii Masao
Andres Freund
Magnus Hagander
Greg Stark
Kevin Grittner
Teodor Sigaev
Stephen Frost
Joe Conway