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 f3b5565dd4 was a couple of bricks shy
of a load; specifically, it missed putting pg_trigger_tgrelid_tgname_index
into the relcache init file, because that index is not used by any
syscache. However, we have historically nailed that index into cache for
performance reasons. The upshot was that load_relcache_init_file always
decided that the init file was busted and silently ignored it, resulting
in a significant hit to backend startup speed.
To fix, reinstantiate RelationIdIsInInitFile() as a wrapper around
RelationSupportsSysCache(), which can know about additional relations
that should be in the init file despite being unknown to syscache.c.
Also install some guards against future mistakes of this type: make
write_relcache_init_file Assert that all nailed relations get written to
the init file, and make load_relcache_init_file emit a WARNING if it takes
the "wrong number of nailed relations" exit path. Now that we remove the
init files during postmaster startup, that case should never occur in the
field, even if we are starting a minor-version update that added or removed
rels from the nailed set. So the warning shouldn't ever be seen by end
users, but it will show up in the regression tests if somebody breaks this
logic.
Back-patch to all supported branches, like the previous commit.
When we invalidate the relcache entry for a system catalog or index, we
must also delete the relcache "init file" if the init file contains a copy
of that rel's entry. The old way of doing this relied on a specially
maintained list of the OIDs of relations present in the init file: we made
the list either when reading the file in, or when writing the file out.
The problem is that when writing the file out, we included only rels
present in our local relcache, which might have already suffered some
deletions due to relcache inval events. In such cases we correctly decided
not to overwrite the real init file with incomplete data --- but we still
used the incomplete initFileRelationIds list for the rest of the current
session. This could result in wrong decisions about whether the session's
own actions require deletion of the init file, potentially allowing an init
file created by some other concurrent session to be left around even though
it's been made stale.
Since we don't support changing the schema of a system catalog at runtime,
the only likely scenario in which this would cause a problem in the field
involves a "vacuum full" on a catalog concurrently with other activity, and
even then it's far from easy to provoke. Remarkably, this has been broken
since 2002 (in commit 7863404417), but we had
never seen a reproducible test case until recently. If it did happen in
the field, the symptoms would probably involve unexpected "cache lookup
failed" errors to begin with, then "could not open file" failures after the
next checkpoint, as all accesses to the affected catalog stopped working.
Recovery would require manually removing the stale "pg_internal.init" file.
To fix, get rid of the initFileRelationIds list, and instead consult
syscache.c's list of relations used in catalog caches to decide whether a
relation is included in the init file. This should be a tad more efficient
anyway, since we're replacing linear search of a list with ~100 entries
with a binary search. It's a bit ugly that the init file contents are now
so directly tied to the catalog caches, but in practice that won't make
much difference.
Back-patch to all supported branches.
Instead of initializing a new TransInvalidationInfo for every
transaction or subtransaction, we can just do it for those
transactions or subtransactions that actually need to queue
invalidation messages. That also avoids needing to free those
entries at the end of a transaction or subtransaction that does
not generate any invalidation messages, which is by far the
common case.
Patch by me. Review by Simon Riggs and Andres Freund.
Define padding bytes in SharedInvalidationMessage structs to be
defined. Otherwise the sinvaladt.c ringbuffer, which is accessed by
multiple processes, will cause spurious valgrind warnings about
undefined memory being used. That's because valgrind remembers the
undefined bytes from the last local process's store, not realizing
that another process has written since, filling the previously
uninitialized bytes.
This feature, building on previous commits, allows the write-ahead log
stream to be decoded into a series of logical changes; that is,
inserts, updates, and deletes and the transactions which contain them.
It is capable of handling decoding even across changes to the schema
of the effected tables. The output format is controlled by a
so-called "output plugin"; an example is included. To make use of
this in a real replication system, the output plugin will need to be
modified to produce output in the format appropriate to that system,
and to perform filtering.
Currently, information can be extracted from the logical decoding
system only via SQL; future commits will add the ability to stream
changes via walsender.
Andres Freund, with review and other contributions from many other
people, including Álvaro Herrera, Abhijit Menon-Sen, Peter Gheogegan,
Kevin Grittner, Robert Haas, Heikki Linnakangas, Fujii Masao, Abhijit
Menon-Sen, Michael Paquier, Simon Riggs, Craig Ringer, and Steve
Singer.
Although user-defined relations can't be directly created in
pg_catalog, it's possible for them to end up there, because you can
create them in some other schema and then use ALTER TABLE .. SET SCHEMA
to move them there. Previously, such relations couldn't afterwards
be manipulated, because IsSystemRelation()/IsSystemClass() rejected
all attempts to modify objects in the pg_catalog schema, regardless
of their origin. With this patch, they now reject only those
objects in pg_catalog which were created at initdb-time, allowing
most operations on user-created tables in pg_catalog to proceed
normally.
This patch also adds new functions IsCatalogRelation() and
IsCatalogClass(), which is similar to IsSystemRelation() and
IsSystemClass() but with a slightly narrower definition: only TOAST
tables of system catalogs are included, rather than *all* TOAST tables.
This is currently used only for making decisions about when
invalidation messages need to be sent, but upcoming logical decoding
patches will find other uses for this information.
Andres Freund, with some modifications by me.
Future patches are expected to introduce logical replication that
works by decoding WAL. WAL contains relfilenodes rather than relation
OIDs, so this infrastructure will be needed to find the relation OID
based on WAL contents.
If logical replication does not make it into this release, we probably
should consider reverting this, since it will add some overhead to DDL
operations that create new relations. One additional index insert per
pg_class row is not a large overhead, but it's more than zero.
Another way of meeting the needs of logical replication would be to
the relation OID to WAL, but that would burden DML operations, not
only DDL.
Andres Freund, with some changes by me. Design review, in earlier
versions, by Álvaro Herrera.
SnapshotNow scans have the undesirable property that, in the face of
concurrent updates, the scan can fail to see either the old or the new
versions of the row. In many cases, we work around this by requiring
DDL operations to hold AccessExclusiveLock on the object being
modified; in some cases, the existing locking is inadequate and random
failures occur as a result. This commit doesn't change anything
related to locking, but will hopefully pave the way to allowing lock
strength reductions in the future.
The major issue has held us back from making this change in the past
is that taking an MVCC snapshot is significantly more expensive than
using a static special snapshot such as SnapshotNow. However, testing
of various worst-case scenarios reveals that this problem is not
severe except under fairly extreme workloads. To mitigate those
problems, we avoid retaking the MVCC snapshot for each new scan;
instead, we take a new snapshot only when invalidation messages have
been processed. The catcache machinery already requires that
invalidation messages be sent before releasing the related heavyweight
lock; else other backends might rely on locally-cached data rather
than scanning the catalog at all. Thus, making snapshot reuse
dependent on the same guarantees shouldn't break anything that wasn't
already subtly broken.
Patch by me. Review by Michael Paquier and Andres Freund.
This reduces unnecessary exposure of other headers through htup.h, which
is very widely included by many files.
I have chosen to move the function prototypes to the new file as well,
because that means htup.h no longer needs to include tupdesc.h. In
itself this doesn't have much effect in indirect inclusion of tupdesc.h
throughout the tree, because it's also required by execnodes.h; but it's
something to explore in the future, and it seemed best to do the htup.h
change now while I'm busy with it.
By my count there are 18 callers of CacheRegisterSyscacheCallback in the
core code in HEAD, so we are potentially leaving as few as 2 slots for any
add-on code to use (though possibly not all these callers would actually
activate in any particular session). That doesn't seem like a lot of
headroom, so let's pump it up a little.
This requires adjusting the API for syscache callback functions: they now
get a hash value, not a TID, to identify the target tuple. Most of them
weren't paying any attention to that argument anyway, but plancache did
require a small amount of fixing.
Also, improve performance a trifle by avoiding sending duplicate inval
messages when a heap_update isn't changing the catcache lookup columns.
The previous code tried to synchronize by unlinking the init file twice,
but that doesn't actually work: it leaves a window wherein a third process
could read the already-stale init file but miss the SI messages that would
tell it the data is stale. The result would be bizarre failures in catalog
accesses, typically "could not read block 0 in file ..." later during
startup.
Instead, hold RelCacheInitLock across both the unlink and the sending of
the SI messages. This is more straightforward, and might even be a bit
faster since only one unlink call is needed.
This has been wrong since it was put in (in 2002!), so back-patch to all
supported releases.
This allows us to reliably remove all leftover temporary relation
files on cluster startup without reference to system catalogs or WAL;
therefore, we no longer include temporary relations in XLOG_XACT_COMMIT
and XLOG_XACT_ABORT WAL records.
Since these changes require including a backend ID in each
SharedInvalSmgrMsg, the size of the SharedInvalidationMessage.id
field has been reduced from two bytes to one, and the maximum number
of connections has been reduced from INT_MAX / 4 to 2^23-1. It would
be possible to remove these restrictions by increasing the size of
SharedInvalidationMessage by 4 bytes, but right now that doesn't seem
like a good trade-off.
Review by Jaime Casanova and Tom Lane.
The purpose of this change is to eliminate the need for every caller
of SearchSysCache, SearchSysCacheCopy, SearchSysCacheExists,
GetSysCacheOid, and SearchSysCacheList to know the maximum number
of allowable keys for a syscache entry (currently 4). This will
make it far easier to increase the maximum number of keys in a
future release should we choose to do so, and it makes the code
shorter, too.
Design and review by Tom Lane.
where a database has a non-default tablespaceid. Pass thru MyDatabaseId
and MyDatabaseTableSpace to allow file path to be re-created in
standby and correct invalidation to take place in all cases.
Update and rework xact_commit_desc() debug messages.
Bug report from Tom by code inspection. Fix by me.
VACUUM FULL INPLACE), along with a boatload of subsidiary code and complexity.
Per discussion, the use case for this method of vacuuming is no longer large
enough to justify maintaining it; not to mention that we don't wish to invest
the work that would be needed to make it play nicely with Hot Standby.
Aside from the code directly related to old-style VACUUM FULL, this commit
removes support for certain WAL record types that could only be generated
within VACUUM FULL, redirect-pointer removal in heap_page_prune, and
nontransactional generation of cache invalidation sinval messages (the last
being the sticking point for Hot Standby).
We still have to retain all code that copes with finding HEAP_MOVED_OFF and
HEAP_MOVED_IN flag bits on existing tuples. This can't be removed as long
as we want to support in-place update from pre-9.0 databases.
of shared or nailed system catalogs. This has two key benefits:
* The new CLUSTER-based VACUUM FULL can be applied safely to all catalogs.
* We no longer have to use an unsafe reindex-in-place approach for reindexing
shared catalogs.
CLUSTER on nailed catalogs now works too, although I left it disabled on
shared catalogs because the resulting pg_index.indisclustered update would
only be visible in one database.
Since reindexing shared system catalogs is now fully transactional and
crash-safe, the former special cases in REINDEX behavior have been removed;
shared catalogs are treated the same as non-shared.
This commit does not do anything about the recently-discussed problem of
deadlocks between VACUUM FULL/CLUSTER on a system catalog and other
concurrent queries; will address that in a separate patch. As a stopgap,
parallel_schedule has been tweaked to run vacuum.sql by itself, to avoid
such failures during the regression tests.
the relfilenode of currently-not-relocatable system catalogs.
1. Get rid of inval.c's dependency on relfilenode, by not having it emit
smgr invalidations as a result of relcache flushes. Instead, smgr sinval
messages are sent directly from smgr.c when an actual relation delete or
truncate is done. This makes considerably more structural sense and allows
elimination of a large number of useless smgr inval messages that were
formerly sent even in cases where nothing was changing at the
physical-relation level. Note that this reintroduces the concept of
nontransactional inval messages, but that's okay --- because the messages
are sent by smgr.c, they will be sent in Hot Standby slaves, just from a
lower logical level than before.
2. Move setNewRelfilenode out of catalog/index.c, where it never logically
belonged, into relcache.c; which is a somewhat debatable choice as well but
better than before. (I considered catalog/storage.c, but that seemed too
low level.) Rename to RelationSetNewRelfilenode.
3. Cosmetic cleanups of some other relfilenode manipulations.
deletion, so that we attempt to unlink the correct filepath. unlink()
errors are ignorable there, so lack of a DatabasePath initialization step
did not cause visible problems until a related bug showed up on Solaris.
Code refactored from xact_redo_commit() to
ProcessCommittedInvalidationMessages() in inval.c. Recovery may replay
shared invalidation messages for many databases, so we cannot
SetDatabasePath() once as we do in normal backends. Read the databaseid
from the shared invalidation messages, then set DatabasePath
temporarily before calling RelationCacheInitFileInvalidate().
Problem report by Robert Treat, analysis and fix by me.
Enabled by recovery_connections = on (default) and forcing archive recovery using a recovery.conf. Recovery processing now emulates the original transactions as they are replayed, providing full locking and MVCC behaviour for read only queries. Recovery must enter consistent state before connections are allowed, so there is a delay, typically short, before connections succeed. Replay of recovering transactions can conflict and in some cases deadlock with queries during recovery; these result in query cancellation after max_standby_delay seconds have expired. Infrastructure changes have minor effects on normal running, though introduce four new types of WAL record.
New test mode "make standbycheck" allows regression tests of static command behaviour on a standby server while in recovery. Typical and extreme dynamic behaviours have been checked via code inspection and manual testing. Few port specific behaviours have been utilised, though primary testing has been on Linux only so far.
This commit is the basic patch. Additional changes will follow in this release to enhance some aspects of behaviour, notably improved handling of conflicts, deadlock detection and query cancellation. Changes to VACUUM FULL are also required.
Simon Riggs, with significant and lengthy review by Heikki Linnakangas, including streamlined redesign of snapshot creation and two-phase commit.
Important contributions from Florian Pflug, Mark Kirkwood, Merlin Moncure, Greg Stark, Gianni Ciolli, Gabriele Bartolini, Hannu Krosing, Robert Haas, Tatsuo Ishii, Hiroyuki Yamada plus support and feedback from many other community members.
when user-defined functions used in a plan are modified. Also invalidate
plans when schemas, operators, or operator classes are modified; but for these
cases we just invalidate everything rather than tracking exact dependencies,
since these types of objects seldom change in a production database.
Tom Lane; loosely based on a patch by Martin Pihlak.
unnecessary cache resets. The major changes are:
* When the queue overflows, we only issue a cache reset to the specific
backend or backends that still haven't read the oldest message, rather
than resetting everyone as in the original coding.
* When we observe backend(s) falling well behind, we signal SIGUSR1
to only one backend, the one that is furthest behind and doesn't already
have a signal outstanding for it. When it finishes catching up, it will
in turn signal SIGUSR1 to the next-furthest-back guy, if there is one that
is far enough behind to justify a signal. The PMSIGNAL_WAKEN_CHILDREN
mechanism is removed.
* We don't attempt to clean out dead messages after every message-receipt
operation; rather, we do it on the insertion side, and only when the queue
fullness passes certain thresholds.
* Split SInvalLock into SInvalReadLock and SInvalWriteLock so that readers
don't block writers nor vice versa (except during the infrequent queue
cleanout operations).
* Transfer multiple sinval messages for each acquisition of a read or
write lock.
corresponding struct definitions. This allows other headers to avoid including
certain highly-loaded headers such as rel.h and relscan.h, instead using just
relcache.h, heapam.h or genam.h, which are more lightweight and thus cause less
unnecessary dependencies.
messages if the calling transaction aborts later on. Collapsing out line
pointer redirects is a done deal as soon as we complete the page update,
so syscache *must* be notified even if the VACUUM FULL as a whole doesn't
complete. To fix, add some functionality to inval.c to allow the pending
inval messages to be sent immediately while heap_page_prune is still
running. The implementation is a bit chintzy: it will only work in the
context of VACUUM FULL. But that's all we need now, and it can always be
extended later if needed. Per my trouble report of a week ago.
but no database changes have been made since the last CommandCounterIncrement.
This should result in a significant improvement in the number of "commands"
that can typically be performed within a transaction before hitting the 2^32
CommandId size limit. In particular this buys back (and more) the possible
adverse consequences of my previous patch to fix plan caching behavior.
The implementation requires tracking whether the current CommandCounter
value has been "used" to mark any tuples. CommandCounter values stored into
snapshots are presumed not to be used for this purpose. This requires some
small executor changes, since the executor used to conflate the curcid of
the snapshot it was using with the command ID to mark output tuples with.
Separating these concepts allows some small simplifications in executor APIs.
Something for the TODO list: look into having CommandCounterIncrement not do
AcceptInvalidationMessages. It seems fairly bogus to be doing it there,
but exactly where to do it instead isn't clear, and I'm disinclined to mess
with asynchronous behavior during late beta.
