Ever since its birth, ReorderBufferBuildTupleCidHash() has contained an
assertion that a catalog tuple cannot change Cmax after acquiring one. But
that's wrong: if a subtransaction executes DDL that affects that catalog
tuple, and later aborts and another DDL affects the same tuple, it will
change Cmax. Relax the assertion to merely verify that the Cmax remains
valid and monotonically increasing, instead.
Add a test that tickles the relevant code.
Diagnosed by, and initial patch submitted by: Arseny Sher
Co-authored-by: Arseny Sher
Discussion: https://postgr.es/m/874l9p8hyw.fsf@ars-thinkpad
Repeatedly rewriting a mapped catalog table with VACUUM FULL or
CLUSTER could cause logical decoding to fail with:
ERROR, "could not map filenode \"%s\" to relation OID"
To trigger the problem the rewritten catalog had to have live tuples
with toasted columns.
The problem was triggered as during catalog table rewrites the
heap_insert() check that prevents logical decoding information to be
emitted for system catalogs, failed to treat the new heap's toast table
as a system catalog (because the new heap is not recognized as a
catalog table via RelationIsLogicallyLogged()). The relmapper, in
contrast to the normal catalog contents, does not contain historical
information. After a single rewrite of a mapped table the new relation
is known to the relmapper, but if the table is rewritten twice before
logical decoding occurs, the relfilenode cannot be mapped to a
relation anymore. Which then leads us to error out. This only
happens for toast tables, because the main table contents aren't
re-inserted with heap_insert().
The fix is simple, add a new heap_insert() flag that prevents logical
decoding information from being emitted, and accept during decoding
that there might not be tuple data for toast tables.
Unfortunately that does not fix pre-existing logical decoding
errors. Doing so would require not throwing an error when a filenode
cannot be mapped to a relation during decoding, and that seems too
likely to hide bugs. If it's crucial to fix decoding for an existing
slot, temporarily changing the ERROR in ReorderBufferCommit() to a
WARNING appears to be the best fix.
Author: Andres Freund
Discussion: https://postgr.es/m/20180914021046.oi7dm4ra3ot2g2kt@alap3.anarazel.de
Backpatch: 9.4-, where logical decoding was introduced
Two closely related bugs are fixed. First, xmin of logical slots was
advanced too early. During xl_running_xacts processing, xmin of the
slot was set to the oldest running xid in the record, but that's wrong:
actually, snapshots which will be used for not-yet-replayed transactions
might consider older txns as running too, so we need to keep xmin back
for them. The problem wasn't noticed earlier because DDL which allows
to delete tuple (set xmax) while some another not-yet-committed
transaction looks at it is pretty rare, if not unique: e.g. all forms of
ALTER TABLE which change schema acquire ACCESS EXCLUSIVE lock
conflicting with any inserts. The included test case (test_decoding's
oldest_xmin) uses ALTER of a composite type, which doesn't have such
interlocking.
To deal with this, we must be able to quickly retrieve oldest xmin
(oldest running xid among all assigned snapshots) from ReorderBuffer. To
fix, add another list of ReorderBufferTXNs to the reorderbuffer, where
transactions are sorted by base-snapshot-LSN. This is slightly
different from the existing (sorted by first-LSN) list, because a
transaction can have an earlier LSN but a later Xmin, if its first
record does not obtain an xmin (eg. xl_xact_assignment). Note this new
list doesn't fully replace the existing txn list: we still need that one
to prevent WAL recycling.
The second issue concerns SnapBuilder snapshots and subtransactions.
SnapBuildDistributeNewCatalogSnapshot never assigned a snapshot to a
transaction that is known to be a subtxn, which is good in the common
case that the top-level transaction already has one (no point in doing
so), but a bug otherwise. To fix, arrange to transfer the snapshot from
the subtxn to its top-level txn as soon as the kinship gets known.
test_decoding's snapshot_transfer verifies this.
Also, fix a minor memory leak: refcount of toplevel's old base snapshot
was not decremented when the snapshot is transferred from child.
Liberally sprinkle code comments, and rewrite a few existing ones. This
part is my (Álvaro's) contribution to this commit, as I had to write all
those comments in order to understand the existing code and Arseny's
patch.
Reported-by: Arseny Sher <a.sher@postgrespro.ru>
Diagnosed-by: Arseny Sher <a.sher@postgrespro.ru>
Co-authored-by: Arseny Sher <a.sher@postgrespro.ru>
Co-authored-by: Álvaro Herrera <alvherre@alvh.no-ip.org>
Reviewed-by: Antonin Houska <ah@cybertec.at>
Discussion: https://postgr.es/m/87lgdyz1wj.fsf@ars-thinkpad
The logical decoding functions do BeginInternalSubTransaction and
RollbackAndReleaseCurrentSubTransaction to clean up after themselves.
It turns out that AtEOSubXact_SPI has an unrecognized assumption that
we always need to cancel the active SPI operation in the SPI context
that surrounds the subtransaction (if there is one). That's true
when the RollbackAndReleaseCurrentSubTransaction call is coming from
the SPI-using function itself, but not when it's happening inside
some unrelated function invoked by a SPI query. In practice the
affected callers are the various PLs.
To fix, record the current subtransaction ID when we begin a SPI
operation, and clean up only if that ID is the subtransaction being
canceled.
Also, remove AtEOSubXact_SPI's assertion that it must have cleaned
up the surrounding SPI context's active tuptable. That's proven
wrong by the same test case.
Also clarify (or, if you prefer, reinterpret) the calling conventions
for _SPI_begin_call and _SPI_end_call. The memory context cleanup
in the latter means that these have always had the flavor of a matched
resource-management pair, but they weren't documented that way before.
Per report from Ben Chobot.
Back-patch to 9.4 where logical decoding came in. In principle,
the SPI changes should go all the way back, since the problem dates
back to commit 7ec1c5a86. But given the lack of field complaints
it seems few people are using internal subtransactions in this way.
So I don't feel a need to take any risks in 9.2/9.3.
Discussion: https://postgr.es/m/73FBA179-C68C-4540-9473-71E865408B15@silentmedia.com
When replica identity full is being used with a wide tuple (above 2^16
bytes after compression) it lead to errors and/or crashes during
decoding because the length field used to store such tuples doesn't
fit into the variable used to store the width in the WAL record.
To fix, discontinue use of xl_heap_header_len.t_len when decoding the
old tuple version, instead compute length of the old tuple by
subtracting the new tuple's length from the record length.
In newer version of postgres this issue is moot because the length is
stored by the new WAL machinery, instead of a xl_heap_header_len
struct. A separate commit will forward-patch the regression test.
Reported-By: "anderson"
Discussion: http://postgr.es/m/20170105144819.f6i5o64vfvy4bn5i@alap3.anarazel.de
The snapshot assembly during the creation of logical slots relied
waiting for transactions in xl_running_xacts to end, by checking for
their commit/abort records. Unfortunately, despite locking, it is
possible to see an xl_running_xact record listing transactions as
ready, that have already WAL-logged an commit/abort record, as the
locking just prevents the ProcArray to be adjusted, and the commit
record has to be logged first.
That lead to either delayed or hanging snapshot creation, because
snapbuild.c would wait "forever" to see commit/abort records for some
transactions. That hang resolved only if a xl_running_xacts record
without any running transactions happened to be logged, far from
certain on a busy server.
It's impractical to prevent that via more heavyweight locking, the
likelihood of deadlocks and significantly increased contention would
be too big.
Instead change the initial snapshot creation to be solely based on
tracking the oldest running transaction via
xl_running_xacts->oldestRunningXid - that actually ends up
significantly simplifying the code. That has two disadvantages:
1) Because we cannot fully "trust" the contents of xl_running_xacts,
we cannot use it to build the initial snapshot. Instead we have to
wait twice for all running transactions to finish.
2) Previously a slot, unless the race occurred, could be created when
the all transaction perceived as running based on commit/abort
records, now we have to wait for the next xl_running_xacts record.
To address that, trigger logging new xl_running_xacts record from
within snapbuild.c exactly when necessary.
Unfortunately snabuild.c's SnapBuild is stored on disk, one of the
stupider ideas of a certain Mr Freund, so we can't change it in a
minor release. As this is going to be backpatched, we have to hack
around a bit to keep on-disk compatibility. A later commit will
rejigger that on master.
Author: Andres Freund, based on a quite different patch from Petr Jelinek
Analyzed-By: Petr Jelinek
Reviewed-By: Petr Jelinek
Discussion: https://postgr.es/m/f37e975c-908f-858e-707f-058d3b1eb214@2ndquadrant.com
Backpatch: 9.4-, where logical decoding has been introduced
A little COLLATE "C" goes a long way.
Pavel Stehule, per suggestion from Craig Ringer
Discussion: <CAFj8pRA8nJZcozgxN=RMSqMmKuHVOkcGAAKPKdFeiMWGDSUDLA@mail.gmail.com>
Before initializing iteration over a subtransaction's changes, the last
few changes were not spilled to disk. That's correct if the transaction
didn't spill to disk, but otherwise... This bug can lead to missed or
misorderd subtransaction contents when they were spilled to disk.
Move spilling of the remaining in-memory changes to
ReorderBufferIterTXNInit(), where it can easily be applied to the top
transaction and, if present, subtransactions.
Since this code had too many bugs already, noticeably increase test
coverage.
Fixes: #14319
Reported-By: Huan Ruan
Discussion: <20160909012610.20024.58169@wrigleys.postgresql.org>
Backport: 9,4-, where logical decoding was added
This oversight could cause logical decoding to fail to decode an outer
transaction containing changes, if a subtransaction had an XID but no
actual changes. Per bug #14279 from Marko Tiikkaja. Patch by Marko
based on analysis by Andrew Gierth.
Discussion: <20160804191757.1430.39011@wrigleys.postgresql.org>
When re-reading an update involving both an old tuple and a new tuple from
disk, reorderbuffer.c was careless about whether the new tuple is suitably
aligned for direct access --- in general, it isn't. We'd missed seeing
this in the buildfarm because the contrib/test_decoding tests exercise this
code path only a few times, and by chance all of those cases have old
tuples with length a multiple of 4, which is usually enough to make the
access to the new tuple's t_len safe. For some still-not-entirely-clear
reason, however, Debian's sparc build gets a bus error, as reported by
Christoph Berg; perhaps it's assuming 8-byte alignment of the pointer?
The lack of previous field reports is probably because you need all of
these conditions to trigger a crash: an alignment-picky platform (not
Intel), a transaction large enough to spill to disk, an update within
that xact that changes a primary-key field and has an odd-length old tuple,
and of course logical decoding tracing the transaction.
Avoid the alignment assumption by using memcpy instead of fetching t_len
directly, and add a test case that exposes the crash on picky platforms.
Back-patch to 9.4 where the bug was introduced.
Discussion: <20160413094117.GC21485@msg.credativ.de>
When decoding the old version of an UPDATE or DELETE change, and if that
tuple was bigger than MaxHeapTupleSize, we either Assert'ed out, or
failed in more subtle ways in non-assert builds. Normally individual
tuples aren't bigger than MaxHeapTupleSize, with big datums toasted.
But that's not the case for the old version of a tuple for logical
decoding; the replica identity is logged as one piece. With the default
replica identity btree limits that to small tuples, but that's not the
case for FULL.
Change the tuple buffer infrastructure to separate allocate over-large
tuples, instead of always going through the slab cache.
This unfortunately requires changing the ReorderBufferTupleBuf
definition, we need to store the allocated size someplace. To avoid
requiring output plugins to recompile, don't store HeapTupleHeaderData
directly after HeapTupleData, but point to it via t_data; that leaves
rooms for the allocated size. As there's no reason for an output plugin
to look at ReorderBufferTupleBuf->t_data.header, remove the field. It
was just a minor convenience having it directly accessible.
Reported-By: Adam Dratwiński
Discussion: CAKg6ypLd7773AOX4DiOGRwQk1TVOQKhNwjYiVjJnpq8Wo+i62Q@mail.gmail.com
Somehow I managed to flip the order of restoring old & new tuples when
de-spooling a change in a large transaction from disk. This happens to
only take effect when a change is spooled to disk which has old/new
versions of the tuple. That only is the case for UPDATEs where he
primary key changed or where replica identity is changed to FULL.
The tests didn't catch this because either spooled updates, or updates
that changed primary keys, were tested; not both at the same time.
Found while adding tests for the following commit.
Backpatch: 9.4, where logical decoding was added
Logical decoding's reorderbuffer keeps transactions in an LSN ordered
list for efficiency. To make that's efficiently possible upper-level
xids are forced to be logged before nested subtransaction xids. That
only works though if these records are all looked at: Unfortunately we
didn't do so for e.g. row level locks, which are otherwise uninteresting
for logical decoding.
This could lead to errors like:
"ERROR: subxact logged without previous toplevel record".
It's not sufficient to just look at row locking records, the xid could
appear first due to a lot of other types of records (which will trigger
the transaction to be marked logged with MarkCurrentTransactionIdLoggedIfAny).
So invent infrastructure to tell reorderbuffer about xids seen, when
they'd otherwise not pass through reorderbuffer.c.
Reported-By: Jarred Ward
Bug: #13844
Discussion: 20160105033249.1087.66040@wrigleys.postgresql.org
Backpatch: 9.4, where logical decoding was added
Otherwise running installcheck-force on a server with
synchronous_commit=off will result in the tests failing. All the other
tests already do so...
Backpatch: 9.4, where logical decoding was added
The existing hint talked about "may only contain letters", but the
actual requirement is more strict: only lower case letters are allowed.
Reported-By: Rushabh Lathia
Author: Rushabh Lathia
Discussion: AGPqQf2x50qcwbYOBKzb4x75sO_V3g81ZsA8+Ji9iN5t_khFhQ@mail.gmail.com
Backpatch: 9.4-, where replication slots were added
A bunch of tests missed specifying that empty transactions shouldn't be
displayed. That causes problems when e.g. autovacuum runs in an
unfortunate moment. The tests in question only run for a very short
time, making this quite unlikely.
Reported-By: Buildfarm member axolotl
Backpatch: 9.4, where logical decoding was introduced
test_decoding used fastgetattr() to extract column values. That's wrong
when decoding updates and deletes if a table's replica identity is set
to FULL and new columns have been added since the old version of the
tuple was created. Due to the lack of a crosscheck with the datum's
natts values an invalid value will be output, leading to errors or
worse.
Bug: #13470
Reported-By: Krzysztof Kotlarski
Discussion: 20150626100333.3874.90852@wrigleys.postgresql.org
Backpatch to 9.4, where the feature, including the bug, was added.
The logical decoding patchset introduced PROC_IN_LOGICAL_DECODING flag
PGXACT flag, that allows such backends to be skipped when computing
the xmin horizon/snapshots. That's fine and sensible for walsenders
streaming out logical changes, but not at all fine for SQL backends
doing logical decoding. If the latter set that flag any change they
have performed outside of logical decoding will not be regarded as
visible - which e.g. can lead to that change being vacuumed away.
Note that not setting the flag for SQL backends isn't particularly
bothersome - the SQL backend doesn't do streaming, so it only runs for
a limited amount of time.
Per buildfarm member 'tick' and Alvaro.
Backpatch to 9.4, where logical decoding was introduced.
There are basically three situations in which logical decoding needs
to perform cache invalidation. During/After replaying a transaction
with catalog changes, when skipping a uninteresting transaction that
performed catalog changes and when erroring out while replaying a
transaction. Unfortunately these three cases were all done slightly
differently - partially because 8de3e410fa, which greatly simplifies
matters, got committed in the midst of the development of logical
decoding.
The actually problematic case was when logical decoding skipped
transaction commits (and thus processed invalidations). When used via
the SQL interface cache invalidation could access the catalog - bad,
because we didn't set up enough state to allow that correctly. It'd
not be hard to setup sufficient state, but the simpler solution is to
always perform cache invalidation outside a valid transaction.
Also make the different cache invalidation cases look as similar as
possible, to ease code review.
This fixes the assertion failure reported by Antonin Houska in
53EE02D9.7040702@gmail.com. The presented testcase has been expanded
into a regression test.
Backpatch to 9.4, where logical decoding was introduced.
Heikki noticed in 544E23C0.8090605@vmware.com that slot.c and
snapbuild.c were missing the FIN_CRC32 call when computing/checking
checksums of on disk files. That doesn't lower the the error detection
capabilities of the checksum, but is inconsistent with other usages.
In a followup mail Heikki also noticed that, contrary to a comment,
the 'version' and 'length' struct fields of replication slot's on disk
data where not covered by the checksum. That's not likely to lead to
actually missed corruption as those fields are cross checked with the
expected version and the actual file length. But it's wrong
nonetheless.
As fixing these issues makes existing on disk files unreadable, bump
the expected versions of on disk files for both slots and logical
decoding historic catalog snapshots. This means that loading old
files will fail with
ERROR: "replication slot file ... has unsupported version 1"
and
ERROR: "snapbuild state file ... has unsupported version 1 instead of
2" respectively. Given the low likelihood of anybody already using
these new features in a production setup that seems acceptable.
Fixing these issues made me notice that there's no regression test
covering the loading of historic snapshot from disk - so add one.
Backpatch to 9.4 where these features were introduced.
The regression tests for contrib/test_decoding regularly failed on
postgres instances that were very slow. Either because the hardware
itself was slow or because very expensive debugging options like
CLOBBER_CACHE_ALWAYS were used.
The reason they failed was just that some additional transactions were
decoded. Analyze and vacuum, triggered by autovac.
To fix just add a option to test_decoding to only display transactions
in which a change was actually displayed. That's not pretty because it
removes information from the tests; but better than constantly failing
tests in very likely harmless ways.
Backpatch to 9.4 where logical decoding was introduced.
Discussion: 20140629142511.GA26930@awork2.anarazel.de
Commit 1b86c81d2d fixed the decoding of toasted columns for the rows
contained in one xl_heap_multi_insert record. But that's not actually
enough, because heap_multi_insert() will actually first toast all
passed in rows and then emit several *_multi_insert records; one for
each page it fills with tuples.
Add a XLOG_HEAP_LAST_MULTI_INSERT flag which is set in
xl_heap_multi_insert->flag denoting that this multi_insert record is
the last emitted by one heap_multi_insert() call. Then use that flag
in decode.c to only set clear_toast_afterwards in the right situation.
Expand the number of rows inserted via COPY in the corresponding
regression test to make sure that more than one heap page is filled
with tuples by one heap_multi_insert() call.
Backpatch to 9.4 like the previous commit.
When decoding the results of a HEAP2_MULTI_INSERT (currently only
generated by COPY FROM) toast columns for all but the last tuple
weren't replaced by their actual contents before being handed to the
output plugin. The reassembled toast datums where disregarded after
every REORDER_BUFFER_CHANGE_(INSERT|UPDATE|DELETE) which is correct
for plain inserts, updates, deletes, but not multi inserts - there we
generate several REORDER_BUFFER_CHANGE_INSERTs for a single
xl_heap_multi_insert record.
To solve the problem add a clear_toast_afterwards boolean to
ReorderBufferChange's union member that's used by modifications. All
row changes but multi_inserts always set that to true, but
multi_insert sets it only for the last change generated.
Add a regression test covering decoding of multi_inserts - there was
none at all before.
Backpatch to 9.4 where logical decoding was introduced.
Bug found by Petr Jelinek.
pg_stat_replication shows connected replication clients. The ddl test case
never has any replication clients connected, so querying pg_stat_replication
is pointless. To check that a slot has been dropped correctly, query
pg_replication_slots instead.
Andres Freund
contrib/test_decoding's "make check" runs two sets of tests. Unless we
specify separate output directories for each set the isolation tests
will overwrite the output from the normal regression set. Doing this
will help the buildfarm collect complete logs.
In b89e151054a I had assumed it was ok to use anonymous unions as
struct members, but while a longstanding extension in many compilers,
it's only been standardized in C11.
To fix, remove one of the anonymous unions which tried to hide some
implementation specific enum values and give the other a name. The
latter unfortunately requires changes in output plugins, but since the
feature has only been added a few days ago...
Andres Freund
Commit 6f37c08057685ee3c6c63222dba0dac012760dde removed whitespace
from the SQL file but not the expected-output file, and commit
7e8db2dc420099df3fa73987cf2d2d6d1a609d86 changed the error message
without updating the expected outputs.
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.
