public inbox for [email protected]
help / color / mirror / Atom feedFrom: Masahiko Sawada <[email protected]>
To: Amit Kapila <[email protected]>
Cc: [email protected] <[email protected]>
Cc: Kyotaro Horiguchi <[email protected]>
Cc: Drouvot, Bertrand <[email protected]>
Cc: PostgreSQL Hackers <[email protected]>
Cc: Oh, Mike <[email protected]>
Subject: Re: [BUG] Logical replication failure "ERROR: could not map filenode "base/13237/442428" to relation OID" with catalog modifying txns
Date: Mon, 25 Jul 2022 10:45:46 +0900
Message-ID: <CAD21AoC_oEy3EvgbJ=asz20Z3ROWtRM01f-F8+F_5VjhCvJX6g@mail.gmail.com> (raw)
In-Reply-To: <CAA4eK1+U1b7P2mMTZHGE0GaT4vYhGsShQx_6070a6xgwj9JQAQ@mail.gmail.com>
References: <[email protected]>
<CAD21AoD00wV4gt-53ze+ZB8n4bqJrdH8J_UnDHddy8S2A+a25g@mail.gmail.com>
<CAA4eK1L_e4r4=10TTzoOiWGQ6hH4vxWAwD=0aofz04czCTV0kQ@mail.gmail.com>
<[email protected]>
<CAA4eK1JoKV2qmp916gFk=9SX=Qo21+sN4n-yjbB2b0Q1xxOKJw@mail.gmail.com>
<CAD21AoAn9T6mS8sMF5=tFV66Wru92s933ihRVunLPEU_+0=Xiw@mail.gmail.com>
<CAA4eK1Lgfj2ww4V_zMxTmduB0BQpEsS-j7RT8BV-cy7Qyp1E3A@mail.gmail.com>
<CAD21AoDNJgYkrO1PQ=k6G9MfvRjhawMN-aZVT2spMr_ymmOhCQ@mail.gmail.com>
<CAD21AoC4x3uOw5rUcSYZkWob5s5ottGt_RPLxCEpHimFRDjrEg@mail.gmail.com>
<CAA4eK1L_Br0wNHwY1PrnusX1H2bvWR+iUnNC=1anKqhPBtnoMg@mail.gmail.com>
<CAD21AoBeMVfY6dvg5J8XmSHGbc3+yTA8k_GVRzwZR24zc4TfSw@mail.gmail.com>
<CAA4eK1Jg+u=ouXbC5EVLVnQ4sUhrR1P+p0LQSjPFmMP4jNYmuw@mail.gmail.com>
<CAD21AoCEF6FAvRRP6LKzq2-2oVvpdQbmZDwzRAdseA9AA2mE-Q@mail.gmail.com>
<OSZPR01MB63106E9CE179D80F46CB3154FD8B9@OSZPR01MB6310.jpnprd01.prod.outlook.com>
<CAD21AoAUx28e-1YQDA6Uhda+=jkidAy+HY8Zv5SYjtXpRcBiVg@mail.gmail.com>
<CAA4eK1LmyPryHUP3kaWRBbz88aCs5J_8-vL2wqwi5wnSwa24tw@mail.gmail.com>
<CAD21AoC1H8tDo=h-xsm5Und-c4FQbNCAOuRoBaMhsxNsJrM8vA@mail.gmail.com>
<CAA4eK1JfcR96cxqb+NYmDueirpvLytfvP-UZPz-E6-U6Lv-+tw@mail.gmail.com>
<CAD21AoCjfmuCRUpsn+nuqRuWW3gQCBdTKuYGYvYiwLVtpTQYSg@mail.gmail.com>
<CAA4eK1LXpPnLEdTttCmMbw3bSnrn=k_58ZUDvx3zbAHLjgod2A@mail.gmail.com>
<CAD21AoBKb_XMdOO0vCAHKieWYqHmJnEN0bGmV5j7PvOGOEkmrQ@mail.gmail.com>
<CAA4eK1LpxNwGKntRJSND8H-DmN8XuZYoaooQmaX4+g-ePqg2=Q@mail.gmail.com>
<CAD21AoCzuV5yhOGw8=nVkymkE0KaHkxMMy5R-HsfBQw9=h0asA@mail.gmail.com>
<CAA4eK1LH54gTKnxhBJRy9GHjoPxp7ggA+B9Cy7VLtXToCVukhw@mail.gmail.com>
<CAD21AoC8NLGJ-Q+u60cGPSnxXqexEFPLo-5-offsgCE_ikVxRw@mail.gmail.com>
<CAA4eK1+U1b7P2mMTZHGE0GaT4vYhGsShQx_6070a6xgwj9JQAQ@mail.gmail.com>
On Sat, Jul 23, 2022 at 8:32 PM Amit Kapila <[email protected]> wrote:
>
> On Fri, Jul 22, 2022 at 11:48 AM Masahiko Sawada <[email protected]> wrote:
> >
> > On Wed, Jul 20, 2022 at 5:50 PM Amit Kapila <[email protected]> wrote:
> > >
> > > On Wed, Jul 20, 2022 at 1:28 PM Masahiko Sawada <[email protected]> wrote:
> > > >
> > >
> > > This is required if we don't want to introduce a new set of functions
> > > as you proposed above. I am not sure which one is better w.r.t back
> > > patching effort later but it seems to me using flag stuff would make
> > > future back patches easier if we make any changes in
> > > SnapBuildCommitTxn.
> >
> > Understood.
> >
> > I've implemented this idea as well for discussion. Both patches have
> > the common change to remember the initial running transactions and to
> > purge them when decoding xl_running_xacts records. The difference is
> > how to mark the transactions as needing to be added to the snapshot.
> >
> > In v7-0001-Fix-catalog-lookup-with-the-wrong-snapshot-during.patch,
> > when the transaction is in the initial running xact list and its
> > commit record has XINFO_HAS_INVAL flag, we mark both the top
> > transaction and its all subtransactions as containing catalog changes
> > (which also means to create ReorderBufferTXN entries for them). These
> > transactions are added to the snapshot in SnapBuildCommitTxn() since
> > ReorderBufferXidHasCatalogChanges () for them returns true.
> >
> > In poc_mark_top_txn_has_inval.patch, when the transaction is in the
> > initial running xacts list and its commit record has XINFO_HAS_INVALS
> > flag, we set a new flag, say RBTXN_COMMIT_HAS_INVALS, only to the top
> > transaction.
> >
>
> It seems that the patch has missed the part to check if the xid is in
> the initial running xacts list?
Oops, right.
>
> > In SnapBuildCommitTxn(), we add all subtransactions to
> > the snapshot without checking ReorderBufferXidHasCatalogChanges() for
> > subtransactions if its top transaction has the RBTXN_COMMIT_HAS_INVALS
> > flag.
> >
> > A difference between the two ideas is the scope of changes: the former
> > changes only snapbuild.c but the latter changes both snapbuild.c and
> > reorderbuffer.c. Moreover, while the former uses the existing flag,
> > the latter adds a new flag to the reorder buffer for dealing with only
> > this case. I think the former idea is simpler in terms of that. But,
> > an advantage of the latter idea is that the latter idea can save to
> > create ReorderBufferTXN entries for subtransactions.
> >
> > Overall I prefer the former for now but I'd like to hear what others think.
> >
>
> I agree that the latter idea can have better performance in extremely
> special scenarios but introducing a new flag for the same sounds a bit
> ugly to me. So, I would also prefer to go with the former idea,
> however, I would also like to hear what Horiguchi-San and others have
> to say.
Agreed.
>
> Few comments on v7-0001-Fix-catalog-lookup-with-the-wrong-snapshot-during:
> 1.
> +void
> +SnapBuildInitialXactSetCatalogChanges(SnapBuild *builder, TransactionId xid,
> + int subxcnt, TransactionId *subxacts,
> + XLogRecPtr lsn)
> +{
>
> I think it is better to name this function as
> SnapBuildXIDSetCatalogChanges as we use this to mark a particular
> transaction as having catalog changes.
>
> 2. Changed/added a few comments in the attached.
Thank you for the comments.
I've attached updated version patches for the master and back branches.
Regards,
--
Masahiko Sawada
EDB: https://www.enterprisedb.com/
Attachments:
[application/octet-stream] REL14-v8-0001-Fix-catalog-lookup-with-the-wrong-snapshot-during.patch (16.3K, ../CAD21AoC_oEy3EvgbJ=asz20Z3ROWtRM01f-F8+F_5VjhCvJX6g@mail.gmail.com/2-REL14-v8-0001-Fix-catalog-lookup-with-the-wrong-snapshot-during.patch)
download | inline diff:
From 09b47b0cd9bfbbb64b1f41b5fae5169a40845e68 Mon Sep 17 00:00:00 2001
From: Masahiko Sawada <[email protected]>
Date: Mon, 11 Jul 2022 21:49:06 +0900
Subject: [PATCH v8] Fix catalog lookup with the wrong snapshot during logical
decoding.
Previously, we relied on HEAP2_NEW_CID records and XACT_INVALIDATION
records to know if the transaction has modified the catalog, and that
information is not serialized to snapshot. Therefore, if the logical
decoding decodes only the commit record of the transaction that
actually has modified a catalog, we missed adding its XID to the
snapshot. We ended up looking at catalogs with the wrong snapshot.
To fix this problem, this changes the snapshot builder so that it
remembers the running transaction written in the xl_running_xacts
record that we decoded first, and mark the transaction as containing
catalog changes if it's in the list of the initial running
transactions and its commit record has XACT_XINFO_HAS_INVALS. To avoid
ABI breakage, we store the array of the initial running transactions
in the static variables InitialRunningXacts and NInitialRunningXacts,
neither SnapBuild nor ReorderBuffer.
This approach has false positive; we could end up adding the
transaction that didn't change catalog to the snapshot since we cannot
distinguish whether the transaction has catalog changes only by
checking the COMMIT record. It doesn't have the information on
which (sub) transaction has catalog changes, and XACT_XINFO_HAS_INVALS
doesn't necessarily indicate that the transaction has catalog
change. But it doesn't become a problem since we use historic snapshot
only for system catalog lookups.
On the master branch, we took a more future-proof approach -- writing
catalog modifying transactions to the serialized snapshot. But we
cannot backpatch it because of change in SnapBuild.
Back-patch to all supported released.
Reported-by: Mike Oh <[email protected]>
Author: Masahiko Sawada <[email protected]>
Reviewed-by: Amit Kapila <[email protected]>
Reviewed-by: Takamichi Osumi <[email protected]>
Reviewed-by: Kyotaro Horiguchi <[email protected]>
Reviewed-by: Bertrand Drouvot <[email protected]>
Reviewed-by: Shi yu <[email protected]>
Reviewed-by: Ahsan Hadi <[email protected]>
Discussion: https://postgr.es/m/81D0D8B0-E7C4-4999-B616-1E5004DBDCD2%40amazon.com
Backpatch-through: 10
---
contrib/test_decoding/Makefile | 2 +-
.../expected/catalog_change_snapshot.out | 44 ++++++
.../specs/catalog_change_snapshot.spec | 39 +++++
src/backend/replication/logical/decode.c | 15 ++
src/backend/replication/logical/snapbuild.c | 143 +++++++++++++++++-
src/include/replication/snapbuild.h | 3 +
6 files changed, 238 insertions(+), 8 deletions(-)
create mode 100644 contrib/test_decoding/expected/catalog_change_snapshot.out
create mode 100644 contrib/test_decoding/specs/catalog_change_snapshot.spec
diff --git a/contrib/test_decoding/Makefile b/contrib/test_decoding/Makefile
index 9a31e0b879..4553252d75 100644
--- a/contrib/test_decoding/Makefile
+++ b/contrib/test_decoding/Makefile
@@ -8,7 +8,7 @@ REGRESS = ddl xact rewrite toast permissions decoding_in_xact \
spill slot truncate stream stats twophase twophase_stream
ISOLATION = mxact delayed_startup ondisk_startup concurrent_ddl_dml \
oldest_xmin snapshot_transfer subxact_without_top concurrent_stream \
- twophase_snapshot
+ twophase_snapshot catalog_change_snapshot
REGRESS_OPTS = --temp-config $(top_srcdir)/contrib/test_decoding/logical.conf
ISOLATION_OPTS = --temp-config $(top_srcdir)/contrib/test_decoding/logical.conf
diff --git a/contrib/test_decoding/expected/catalog_change_snapshot.out b/contrib/test_decoding/expected/catalog_change_snapshot.out
new file mode 100644
index 0000000000..dc4f9b7018
--- /dev/null
+++ b/contrib/test_decoding/expected/catalog_change_snapshot.out
@@ -0,0 +1,44 @@
+Parsed test spec with 2 sessions
+
+starting permutation: s0_init s0_begin s0_savepoint s0_truncate s1_checkpoint s1_get_changes s0_commit s0_begin s0_insert s1_checkpoint s1_get_changes s0_commit s1_get_changes
+step s0_init: SELECT 'init' FROM pg_create_logical_replication_slot('isolation_slot', 'test_decoding');
+?column?
+--------
+init
+(1 row)
+
+step s0_begin: BEGIN;
+step s0_savepoint: SAVEPOINT sp1;
+step s0_truncate: TRUNCATE tbl1;
+step s1_checkpoint: CHECKPOINT;
+step s1_get_changes: SELECT data FROM pg_logical_slot_get_changes('isolation_slot', NULL, NULL, 'skip-empty-xacts', '1', 'include-xids', '0');
+data
+----
+(0 rows)
+
+step s0_commit: COMMIT;
+step s0_begin: BEGIN;
+step s0_insert: INSERT INTO tbl1 VALUES (1);
+step s1_checkpoint: CHECKPOINT;
+step s1_get_changes: SELECT data FROM pg_logical_slot_get_changes('isolation_slot', NULL, NULL, 'skip-empty-xacts', '1', 'include-xids', '0');
+data
+---------------------------------------
+BEGIN
+table public.tbl1: TRUNCATE: (no-flags)
+COMMIT
+(3 rows)
+
+step s0_commit: COMMIT;
+step s1_get_changes: SELECT data FROM pg_logical_slot_get_changes('isolation_slot', NULL, NULL, 'skip-empty-xacts', '1', 'include-xids', '0');
+data
+-------------------------------------------------------------
+BEGIN
+table public.tbl1: INSERT: val1[integer]:1 val2[integer]:null
+COMMIT
+(3 rows)
+
+?column?
+--------
+stop
+(1 row)
+
diff --git a/contrib/test_decoding/specs/catalog_change_snapshot.spec b/contrib/test_decoding/specs/catalog_change_snapshot.spec
new file mode 100644
index 0000000000..662760fbcf
--- /dev/null
+++ b/contrib/test_decoding/specs/catalog_change_snapshot.spec
@@ -0,0 +1,39 @@
+# Test decoding only the commit record of the transaction that have
+# modified catalogs.
+setup
+{
+ DROP TABLE IF EXISTS tbl1;
+ CREATE TABLE tbl1 (val1 integer, val2 integer);
+}
+
+teardown
+{
+ DROP TABLE tbl1;
+ SELECT 'stop' FROM pg_drop_replication_slot('isolation_slot');
+}
+
+session "s0"
+setup { SET synchronous_commit=on; }
+step "s0_init" { SELECT 'init' FROM pg_create_logical_replication_slot('isolation_slot', 'test_decoding'); }
+step "s0_begin" { BEGIN; }
+step "s0_savepoint" { SAVEPOINT sp1; }
+step "s0_truncate" { TRUNCATE tbl1; }
+step "s0_insert" { INSERT INTO tbl1 VALUES (1); }
+step "s0_commit" { COMMIT; }
+
+session "s1"
+setup { SET synchronous_commit=on; }
+step "s1_checkpoint" { CHECKPOINT; }
+step "s1_get_changes" { SELECT data FROM pg_logical_slot_get_changes('isolation_slot', NULL, NULL, 'skip-empty-xacts', '1', 'include-xids', '0'); }
+
+# For the transaction that TRUNCATEd the table tbl1, the last decoding decodes
+# only its COMMIT record, because it starts from the RUNNING_XACT record emitted
+# during the first checkpoint execution. This transaction must be marked as
+# containing catalog changes while decoding the COMMIT record and the decoding
+# of the INSERT record must read the pg_class with the correct historic snapshot.
+#
+# Note that in a case where bgwriter wrote the XACT_RUNNING record between "s0_commit"
+# and "s0_begin", this doesn't happen as the decoding starts from the RUNNING_XACT
+# record written by bgwriter. One might think we can either stop the bgwriter or
+# increase LOG_SNAPSHOT_INTERVAL_MS but it's not practical via tests.
+permutation "s0_init" "s0_begin" "s0_savepoint" "s0_truncate" "s1_checkpoint" "s1_get_changes" "s0_commit" "s0_begin" "s0_insert" "s1_checkpoint" "s1_get_changes" "s0_commit" "s1_get_changes"
diff --git a/src/backend/replication/logical/decode.c b/src/backend/replication/logical/decode.c
index 92dfafc632..6fefe9e964 100644
--- a/src/backend/replication/logical/decode.c
+++ b/src/backend/replication/logical/decode.c
@@ -691,6 +691,21 @@ DecodeCommit(LogicalDecodingContext *ctx, XLogRecordBuffer *buf,
commit_time = parsed->origin_timestamp;
}
+ /*
+ * If the COMMIT record has invalidation messages, it could have catalog
+ * changes. It is possible that we didn't mark this transaction as
+ * containing catalog changes when the decoding starts from a commit record
+ * without decoding the transaction's other changes. So, we ensure to mark
+ * such transactions as containing catalog change.
+ *
+ * This must be done before SnapBuildCommitTxn() so that we can include
+ * these transactions in the historic snapshot.
+ */
+ if (parsed->xinfo & XACT_XINFO_HAS_INVALS)
+ SnapBuildXidSetCatalogChanges(ctx->snapshot_builder, xid,
+ parsed->nsubxacts, parsed->subxacts,
+ buf->origptr);
+
SnapBuildCommitTxn(ctx->snapshot_builder, buf->origptr, xid,
parsed->nsubxacts, parsed->subxacts);
diff --git a/src/backend/replication/logical/snapbuild.c b/src/backend/replication/logical/snapbuild.c
index 6df602485b..17e93aac67 100644
--- a/src/backend/replication/logical/snapbuild.c
+++ b/src/backend/replication/logical/snapbuild.c
@@ -250,8 +250,38 @@ struct SnapBuild
static ResourceOwner SavedResourceOwnerDuringExport = NULL;
static bool ExportInProgress = false;
-/* ->committed manipulation */
-static void SnapBuildPurgeCommittedTxn(SnapBuild *builder);
+/*
+ * Array of transactions and subtransactions that were running when
+ * the xl_running_xacts record that we decoded first was written.
+ * The array is sorted in xidComparator order. We remove Xids from
+ * this array when they become old enough to matter. This array is
+ * allocated in builder->context so its lifetime is the same as the
+ * snapshot builder.
+ *
+ * We rely on HEAP2_NEW_CID records and XACT_INVALIDATIONS to know
+ * if the transaction has changed the catalog, and that information
+ * is not serialized to SnapBuilder. Therefore, if the logical
+ * decoding decodes the commit record of the transaction that actually
+ * has done catalog changes without these records, we miss to add
+ * the xid to the snapshot, and end up looking at catalogs with the
+ * wrong snapshot. To avoid this problem, if the COMMIT record of
+ * the xid listed in InitialRunningXacts has XACT_XINFO_HAS_INVALS
+ * flag, we mark both the top transaction and its substransactions
+ * as containing catalog changes.
+ *
+ * We could end up adding the transaction that didn't change catalog
+ * to the snapshot since we cannot distinguish whether the transaction
+ * has catalog changes only by checking the COMMIT record. It doesn't
+ * have the information on which (sub) transaction has catalog changes,
+ * and XACT_XINFO_HAS_INVALS doesn't necessarily indicate that the
+ * transaction has catalog change. But it doesn't become a problem since
+ * we use historic snapshot only for reading system catalogs.
+ */
+static TransactionId *InitialRunningXacts = NULL;
+static int NInitialRunningXacts = 0;
+
+/* ->committed and InitailRunningXacts manipulation */
+static void SnapBuildPurgeOlderTxn(SnapBuild *builder);
/* snapshot building/manipulation/distribution functions */
static Snapshot SnapBuildBuildSnapshot(SnapBuild *builder);
@@ -879,12 +909,17 @@ SnapBuildAddCommittedTxn(SnapBuild *builder, TransactionId xid)
}
/*
- * Remove knowledge about transactions we treat as committed that are smaller
- * than ->xmin. Those won't ever get checked via the ->committed array but via
- * the clog machinery, so we don't need to waste memory on them.
+ * Remove knowledge about transactions we treat as committed and the initial
+ * running transactions that are smaller than ->xmin. Those won't ever get
+ * checked via the ->committed or InitialRunningXacts array, respectively.
+ * The committed xids will get checked via the clog machinery.
+ *
+ * We can ideally remove the transaction from InitialRunningXacts array
+ * once it is finished (committed/aborted) but that could be costly as we need
+ * to maintain the xids order in the array.
*/
static void
-SnapBuildPurgeCommittedTxn(SnapBuild *builder)
+SnapBuildPurgeOlderTxn(SnapBuild *builder)
{
int off;
TransactionId *workspace;
@@ -919,6 +954,49 @@ SnapBuildPurgeCommittedTxn(SnapBuild *builder)
builder->committed.xcnt = surviving_xids;
pfree(workspace);
+
+ /* Quick exit if there is no initial running transactions */
+ if (likely(NInitialRunningXacts == 0))
+ return;
+
+ /* bound check if there is at least one transaction to remove */
+ if (!NormalTransactionIdPrecedes(InitialRunningXacts[0],
+ builder->xmin))
+ return;
+
+ /*
+ * purge xids in InitialRunningXacts as well. The purged array must also be
+ * sorted in xidComparator order.
+ */
+ workspace =
+ MemoryContextAlloc(builder->context,
+ NInitialRunningXacts * sizeof(TransactionId));
+ surviving_xids = 0;
+ for (off = 0; off < NInitialRunningXacts; off++)
+ {
+ if (NormalTransactionIdPrecedes(InitialRunningXacts[off],
+ builder->xmin))
+ ; /* remove */
+ else
+ workspace[surviving_xids++] = InitialRunningXacts[off];
+ }
+
+ if (surviving_xids > 0)
+ memcpy(InitialRunningXacts, workspace,
+ sizeof(TransactionId) * surviving_xids);
+ else
+ {
+ pfree(InitialRunningXacts);
+ InitialRunningXacts = NULL;
+ }
+
+ elog(DEBUG3, "purged initial running transactions from %u to %u, oldest running xid %u",
+ (uint32) NInitialRunningXacts,
+ (uint32) surviving_xids,
+ builder->xmin);
+
+ NInitialRunningXacts = surviving_xids;
+ pfree(workspace);
}
/*
@@ -1104,6 +1182,21 @@ SnapBuildProcessRunningXacts(SnapBuild *builder, XLogRecPtr lsn, xl_running_xact
*/
if (builder->state < SNAPBUILD_CONSISTENT)
{
+ /*
+ * Remember the transactions and subtransactions that were running
+ * when xl_running_xacts record that we decoded first was written.
+ */
+ if (builder->state == SNAPBUILD_START)
+ {
+ int nxacts = running->subxcnt + running->xcnt;
+ Size sz = sizeof(TransactionId) * nxacts;
+
+ NInitialRunningXacts = nxacts;
+ InitialRunningXacts = MemoryContextAlloc(builder->context, sz);
+ memcpy(InitialRunningXacts, running->xids, sz);
+ qsort(InitialRunningXacts, nxacts, sizeof(TransactionId), xidComparator);
+ }
+
/* returns false if there's no point in performing cleanup just yet */
if (!SnapBuildFindSnapshot(builder, lsn, running))
return;
@@ -1126,7 +1219,7 @@ SnapBuildProcessRunningXacts(SnapBuild *builder, XLogRecPtr lsn, xl_running_xact
builder->xmin = running->oldestRunningXid;
/* Remove transactions we don't need to keep track off anymore */
- SnapBuildPurgeCommittedTxn(builder);
+ SnapBuildPurgeOlderTxn(builder);
/*
* Advance the xmin limit for the current replication slot, to allow
@@ -1993,3 +2086,39 @@ CheckPointSnapBuild(void)
}
FreeDir(snap_dir);
}
+
+/*
+ * If the given xid is in the list of the initial running xacts, we mark the
+ * transaction and its subtransactions as containing catalog changes. See
+ * comments
+ */
+void
+SnapBuildXidSetCatalogChanges(SnapBuild *builder, TransactionId xid, int subxcnt,
+ TransactionId *subxacts, XLogRecPtr lsn)
+{
+ /*
+ * Skip if there is no initial running xacts information or the
+ * transaction is already marked as containing catalog changes.
+ */
+ if (likely((NInitialRunningXacts == 0) ||
+ ReorderBufferXidHasCatalogChanges(builder->reorder, xid)))
+ return;
+
+ /*
+ * If this committed transaction is the one that was running at the time
+ * when decoding the first RUNNING_XACTS record and have done catalog
+ * changes, we can mark both the top transaction and its subtransactions
+ * as containing catalog changes.
+ */
+ if (bsearch(&xid, InitialRunningXacts, NInitialRunningXacts,
+ sizeof(TransactionId), xidComparator) != NULL)
+ {
+ ReorderBufferXidSetCatalogChanges(builder->reorder, xid, lsn);
+
+ for (int i = 0; i < subxcnt; i++)
+ {
+ ReorderBufferAssignChild(builder->reorder, xid, subxacts[i], lsn);
+ ReorderBufferXidSetCatalogChanges(builder->reorder, subxacts[i], lsn);
+ }
+ }
+}
diff --git a/src/include/replication/snapbuild.h b/src/include/replication/snapbuild.h
index 3604621e88..a19b59e100 100644
--- a/src/include/replication/snapbuild.h
+++ b/src/include/replication/snapbuild.h
@@ -90,4 +90,7 @@ extern void SnapBuildProcessRunningXacts(SnapBuild *builder, XLogRecPtr lsn,
struct xl_running_xacts *running);
extern void SnapBuildSerializationPoint(SnapBuild *builder, XLogRecPtr lsn);
+extern void SnapBuildXidSetCatalogChanges(SnapBuild *builder, TransactionId xid,
+ int subxcnt, TransactionId *subxacts,
+ XLogRecPtr lsn);
#endif /* SNAPBUILD_H */
--
2.24.3 (Apple Git-128)
[application/octet-stream] REL10-v8-0001-Fix-catalog-lookup-with-the-wrong-snapshot-during.patch (16.1K, ../CAD21AoC_oEy3EvgbJ=asz20Z3ROWtRM01f-F8+F_5VjhCvJX6g@mail.gmail.com/3-REL10-v8-0001-Fix-catalog-lookup-with-the-wrong-snapshot-during.patch)
download | inline diff:
From 83341fb2d72e30bb8ec75783e58990d63bd5d1ca Mon Sep 17 00:00:00 2001
From: Masahiko Sawada <[email protected]>
Date: Sun, 17 Jul 2022 21:28:35 +0900
Subject: [PATCH v8] Fix catalog lookup with the wrong snapshot during logical
decoding.
Previously, we relied on HEAP2_NEW_CID records and XACT_INVALIDATION
records to know if the transaction has modified the catalog, and that
information is not serialized to snapshot. Therefore, if the logical
decoding decodes only the commit record of the transaction that
actually has modified a catalog, we missed adding its XID to the
snapshot. We ended up looking at catalogs with the wrong snapshot.
To fix this problem, this changes the snapshot builder so that it
remembers the running transaction written in the xl_running_xacts
record that we decoded first, and mark the transaction as containing
catalog changes if it's in the list of the initial running
transactions and its commit record has XACT_XINFO_HAS_INVALS. To avoid
ABI breakage, we store the array of the initial running transactions
in the static variables InitialRunningXacts and NInitialRunningXacts,
neither SnapBuild nor ReorderBuffer.
This approach has false positive; we could end up adding the
transaction that didn't change catalog to the snapshot since we cannot
distinguish whether the transaction has catalog changes only by
checking the COMMIT record. It doesn't have the information on
which (sub) transaction has catalog changes, and XACT_XINFO_HAS_INVALS
doesn't necessarily indicate that the transaction has catalog
change. But it doesn't become a problem since we use historic snapshot
only for system catalog lookups.
On the master branch, we took a more future-proof approach -- writing
catalog modifying transactions to the serialized snapshot. But we
cannot backpatch it because of change in SnapBuild.
Back-patch to all supported released.
Reported-by: Mike Oh <[email protected]>
Author: Masahiko Sawada <[email protected]>
Reviewed-by: Amit Kapila <[email protected]>
Reviewed-by: Takamichi Osumi <[email protected]>
Reviewed-by: Kyotaro Horiguchi <[email protected]>
Reviewed-by: Bertrand Drouvot <[email protected]>
Reviewed-by: Shi yu <[email protected]>
Reviewed-by: Ahsan Hadi <[email protected]>
Discussion: https://postgr.es/m/81D0D8B0-E7C4-4999-B616-1E5004DBDCD2%40amazon.com
Backpatch-through: 10
---
contrib/test_decoding/Makefile | 2 +-
.../expected/catalog_change_snapshot.out | 41 +++++
.../specs/catalog_change_snapshot.spec | 39 +++++
src/backend/replication/logical/decode.c | 15 ++
src/backend/replication/logical/snapbuild.c | 143 +++++++++++++++++-
src/include/replication/snapbuild.h | 3 +
6 files changed, 235 insertions(+), 8 deletions(-)
create mode 100644 contrib/test_decoding/expected/catalog_change_snapshot.out
create mode 100644 contrib/test_decoding/specs/catalog_change_snapshot.spec
diff --git a/contrib/test_decoding/Makefile b/contrib/test_decoding/Makefile
index 2db2b2774b..73bc0fe1fe 100644
--- a/contrib/test_decoding/Makefile
+++ b/contrib/test_decoding/Makefile
@@ -51,7 +51,7 @@ regresscheck-install-force: | submake-regress submake-test_decoding temp-install
$(REGRESSCHECKS)
ISOLATIONCHECKS=mxact delayed_startup ondisk_startup concurrent_ddl_dml \
- oldest_xmin snapshot_transfer subxact_without_top
+ oldest_xmin snapshot_transfer subxact_without_top catalog_change_snapshot
isolationcheck: | submake-isolation submake-test_decoding temp-install
$(pg_isolation_regress_check) \
diff --git a/contrib/test_decoding/expected/catalog_change_snapshot.out b/contrib/test_decoding/expected/catalog_change_snapshot.out
new file mode 100644
index 0000000000..15f9540b3f
--- /dev/null
+++ b/contrib/test_decoding/expected/catalog_change_snapshot.out
@@ -0,0 +1,41 @@
+Parsed test spec with 2 sessions
+
+starting permutation: s0_init s0_begin s0_savepoint s0_truncate s1_checkpoint s1_get_changes s0_commit s0_begin s0_insert s1_checkpoint s1_get_changes s0_commit s1_get_changes
+step s0_init: SELECT 'init' FROM pg_create_logical_replication_slot('isolation_slot', 'test_decoding');
+?column?
+--------
+init
+(1 row)
+
+step s0_begin: BEGIN;
+step s0_savepoint: SAVEPOINT sp1;
+step s0_truncate: TRUNCATE tbl1;
+step s1_checkpoint: CHECKPOINT;
+step s1_get_changes: SELECT data FROM pg_logical_slot_get_changes('isolation_slot', NULL, NULL, 'skip-empty-xacts', '1', 'include-xids', '0');
+data
+----
+(0 rows)
+
+step s0_commit: COMMIT;
+step s0_begin: BEGIN;
+step s0_insert: INSERT INTO tbl1 VALUES (1);
+step s1_checkpoint: CHECKPOINT;
+step s1_get_changes: SELECT data FROM pg_logical_slot_get_changes('isolation_slot', NULL, NULL, 'skip-empty-xacts', '1', 'include-xids', '0');
+data
+----
+(0 rows)
+
+step s0_commit: COMMIT;
+step s1_get_changes: SELECT data FROM pg_logical_slot_get_changes('isolation_slot', NULL, NULL, 'skip-empty-xacts', '1', 'include-xids', '0');
+data
+-------------------------------------------------------------
+BEGIN
+table public.tbl1: INSERT: val1[integer]:1 val2[integer]:null
+COMMIT
+(3 rows)
+
+?column?
+--------
+stop
+(1 row)
+
diff --git a/contrib/test_decoding/specs/catalog_change_snapshot.spec b/contrib/test_decoding/specs/catalog_change_snapshot.spec
new file mode 100644
index 0000000000..662760fbcf
--- /dev/null
+++ b/contrib/test_decoding/specs/catalog_change_snapshot.spec
@@ -0,0 +1,39 @@
+# Test decoding only the commit record of the transaction that have
+# modified catalogs.
+setup
+{
+ DROP TABLE IF EXISTS tbl1;
+ CREATE TABLE tbl1 (val1 integer, val2 integer);
+}
+
+teardown
+{
+ DROP TABLE tbl1;
+ SELECT 'stop' FROM pg_drop_replication_slot('isolation_slot');
+}
+
+session "s0"
+setup { SET synchronous_commit=on; }
+step "s0_init" { SELECT 'init' FROM pg_create_logical_replication_slot('isolation_slot', 'test_decoding'); }
+step "s0_begin" { BEGIN; }
+step "s0_savepoint" { SAVEPOINT sp1; }
+step "s0_truncate" { TRUNCATE tbl1; }
+step "s0_insert" { INSERT INTO tbl1 VALUES (1); }
+step "s0_commit" { COMMIT; }
+
+session "s1"
+setup { SET synchronous_commit=on; }
+step "s1_checkpoint" { CHECKPOINT; }
+step "s1_get_changes" { SELECT data FROM pg_logical_slot_get_changes('isolation_slot', NULL, NULL, 'skip-empty-xacts', '1', 'include-xids', '0'); }
+
+# For the transaction that TRUNCATEd the table tbl1, the last decoding decodes
+# only its COMMIT record, because it starts from the RUNNING_XACT record emitted
+# during the first checkpoint execution. This transaction must be marked as
+# containing catalog changes while decoding the COMMIT record and the decoding
+# of the INSERT record must read the pg_class with the correct historic snapshot.
+#
+# Note that in a case where bgwriter wrote the XACT_RUNNING record between "s0_commit"
+# and "s0_begin", this doesn't happen as the decoding starts from the RUNNING_XACT
+# record written by bgwriter. One might think we can either stop the bgwriter or
+# increase LOG_SNAPSHOT_INTERVAL_MS but it's not practical via tests.
+permutation "s0_init" "s0_begin" "s0_savepoint" "s0_truncate" "s1_checkpoint" "s1_get_changes" "s0_commit" "s0_begin" "s0_insert" "s1_checkpoint" "s1_get_changes" "s0_commit" "s1_get_changes"
diff --git a/src/backend/replication/logical/decode.c b/src/backend/replication/logical/decode.c
index 6f8920f52c..fd4d457e64 100644
--- a/src/backend/replication/logical/decode.c
+++ b/src/backend/replication/logical/decode.c
@@ -552,6 +552,21 @@ DecodeCommit(LogicalDecodingContext *ctx, XLogRecordBuffer *buf,
commit_time = parsed->origin_timestamp;
}
+ /*
+ * If the COMMIT record has invalidation messages, it could have catalog
+ * changes. It is possible that we didn't mark this transaction as
+ * containing catalog changes when the decoding starts from a commit record
+ * without decoding the transaction's other changes. So, we ensure to mark
+ * such transactions as containing catalog change.
+ *
+ * This must be done before SnapBuildCommitTxn() so that we can include
+ * these transactions in the historic snapshot.
+ */
+ if (parsed->xinfo & XACT_XINFO_HAS_INVALS)
+ SnapBuildXidSetCatalogChanges(ctx->snapshot_builder, xid,
+ parsed->nsubxacts, parsed->subxacts,
+ buf->origptr);
+
/*
* Process invalidation messages, even if we're not interested in the
* transaction's contents, since the various caches need to always be
diff --git a/src/backend/replication/logical/snapbuild.c b/src/backend/replication/logical/snapbuild.c
index 1010a2e869..9519b953d2 100644
--- a/src/backend/replication/logical/snapbuild.c
+++ b/src/backend/replication/logical/snapbuild.c
@@ -258,8 +258,38 @@ struct SnapBuild
static ResourceOwner SavedResourceOwnerDuringExport = NULL;
static bool ExportInProgress = false;
-/* ->committed manipulation */
-static void SnapBuildPurgeCommittedTxn(SnapBuild *builder);
+/*
+ * Array of transactions and subtransactions that were running when
+ * the xl_running_xacts record that we decoded first was written.
+ * The array is sorted in xidComparator order. We remove Xids from
+ * this array when they become old enough to matter. This array is
+ * allocated in builder->context so its lifetime is the same as the
+ * snapshot builder.
+ *
+ * We rely on HEAP2_NEW_CID records and XACT_INVALIDATIONS to know
+ * if the transaction has changed the catalog, and that information
+ * is not serialized to SnapBuilder. Therefore, if the logical
+ * decoding decodes the commit record of the transaction that actually
+ * has done catalog changes without these records, we miss to add
+ * the xid to the snapshot, and end up looking at catalogs with the
+ * wrong snapshot. To avoid this problem, if the COMMIT record of
+ * the xid listed in InitialRunningXacts has XACT_XINFO_HAS_INVALS
+ * flag, we mark both the top transaction and its substransactions
+ * as containing catalog changes.
+ *
+ * We could end up adding the transaction that didn't change catalog
+ * to the snapshot since we cannot distinguish whether the transaction
+ * has catalog changes only by checking the COMMIT record. It doesn't
+ * have the information on which (sub) transaction has catalog changes,
+ * and XACT_XINFO_HAS_INVALS doesn't necessarily indicate that the
+ * transaction has catalog change. But it doesn't become a problem since
+ * we use historic snapshot only for reading system catalogs.
+ */
+static TransactionId *InitialRunningXacts = NULL;
+static int NInitialRunningXacts = 0;
+
+/* ->committed and InitailRunningXacts manipulation */
+static void SnapBuildPurgeOlderTxn(SnapBuild *builder);
/* snapshot building/manipulation/distribution functions */
static Snapshot SnapBuildBuildSnapshot(SnapBuild *builder);
@@ -896,12 +926,17 @@ SnapBuildAddCommittedTxn(SnapBuild *builder, TransactionId xid)
}
/*
- * Remove knowledge about transactions we treat as committed that are smaller
- * than ->xmin. Those won't ever get checked via the ->committed array but via
- * the clog machinery, so we don't need to waste memory on them.
+ * Remove knowledge about transactions we treat as committed and the initial
+ * running transactions that are smaller than ->xmin. Those won't ever get
+ * checked via the ->committed or InitialRunningXacts array, respectively.
+ * The committed xids will get checked via the clog machinery.
+ *
+ * We can ideally remove the transaction from InitialRunningXacts array
+ * once it is finished (committed/aborted) but that could be costly as we need
+ * to maintain the xids order in the array.
*/
static void
-SnapBuildPurgeCommittedTxn(SnapBuild *builder)
+SnapBuildPurgeOlderTxn(SnapBuild *builder)
{
int off;
TransactionId *workspace;
@@ -936,6 +971,49 @@ SnapBuildPurgeCommittedTxn(SnapBuild *builder)
builder->committed.xcnt = surviving_xids;
pfree(workspace);
+
+ /* Quick exit if there is no initial running transactions */
+ if (likely(NInitialRunningXacts == 0))
+ return;
+
+ /* bound check if there is at least one transaction to remove */
+ if (!NormalTransactionIdPrecedes(InitialRunningXacts[0],
+ builder->xmin))
+ return;
+
+ /*
+ * purge xids in InitialRunningXacts as well. The purged array must also be
+ * sorted in xidComparator order.
+ */
+ workspace =
+ MemoryContextAlloc(builder->context,
+ NInitialRunningXacts * sizeof(TransactionId));
+ surviving_xids = 0;
+ for (off = 0; off < NInitialRunningXacts; off++)
+ {
+ if (NormalTransactionIdPrecedes(InitialRunningXacts[off],
+ builder->xmin))
+ ; /* remove */
+ else
+ workspace[surviving_xids++] = InitialRunningXacts[off];
+ }
+
+ if (surviving_xids > 0)
+ memcpy(InitialRunningXacts, workspace,
+ sizeof(TransactionId) * surviving_xids);
+ else
+ {
+ pfree(InitialRunningXacts);
+ InitialRunningXacts = NULL;
+ }
+
+ elog(DEBUG3, "purged initial running transactions from %u to %u, oldest running xid %u",
+ (uint32) NInitialRunningXacts,
+ (uint32) surviving_xids,
+ builder->xmin);
+
+ NInitialRunningXacts = surviving_xids;
+ pfree(workspace);
}
/*
@@ -1121,6 +1199,21 @@ SnapBuildProcessRunningXacts(SnapBuild *builder, XLogRecPtr lsn, xl_running_xact
*/
if (builder->state < SNAPBUILD_CONSISTENT)
{
+ /*
+ * Remember the transactions and subtransactions that were running
+ * when xl_running_xacts record that we decoded first was written.
+ */
+ if (builder->state == SNAPBUILD_START)
+ {
+ int nxacts = running->subxcnt + running->xcnt;
+ Size sz = sizeof(TransactionId) * nxacts;
+
+ NInitialRunningXacts = nxacts;
+ InitialRunningXacts = MemoryContextAlloc(builder->context, sz);
+ memcpy(InitialRunningXacts, running->xids, sz);
+ qsort(InitialRunningXacts, nxacts, sizeof(TransactionId), xidComparator);
+ }
+
/* returns false if there's no point in performing cleanup just yet */
if (!SnapBuildFindSnapshot(builder, lsn, running))
return;
@@ -1143,7 +1236,7 @@ SnapBuildProcessRunningXacts(SnapBuild *builder, XLogRecPtr lsn, xl_running_xact
builder->xmin = running->oldestRunningXid;
/* Remove transactions we don't need to keep track off anymore */
- SnapBuildPurgeCommittedTxn(builder);
+ SnapBuildPurgeOlderTxn(builder);
/*
* Advance the xmin limit for the current replication slot, to allow
@@ -1997,3 +2090,39 @@ CheckPointSnapBuild(void)
}
FreeDir(snap_dir);
}
+
+/*
+ * If the given xid is in the list of the initial running xacts, we mark the
+ * transaction and its subtransactions as containing catalog changes. See
+ * comments
+ */
+void
+SnapBuildXidSetCatalogChanges(SnapBuild *builder, TransactionId xid, int subxcnt,
+ TransactionId *subxacts, XLogRecPtr lsn)
+{
+ /*
+ * Skip if there is no initial running xacts information or the
+ * transaction is already marked as containing catalog changes.
+ */
+ if (likely((NInitialRunningXacts == 0) ||
+ ReorderBufferXidHasCatalogChanges(builder->reorder, xid)))
+ return;
+
+ /*
+ * If this committed transaction is the one that was running at the time
+ * when decoding the first RUNNING_XACTS record and have done catalog
+ * changes, we can mark both the top transaction and its subtransactions
+ * as containing catalog changes.
+ */
+ if (bsearch(&xid, InitialRunningXacts, NInitialRunningXacts,
+ sizeof(TransactionId), xidComparator) != NULL)
+ {
+ ReorderBufferXidSetCatalogChanges(builder->reorder, xid, lsn);
+
+ for (int i = 0; i < subxcnt; i++)
+ {
+ ReorderBufferAssignChild(builder->reorder, xid, subxacts[i], lsn);
+ ReorderBufferXidSetCatalogChanges(builder->reorder, subxacts[i], lsn);
+ }
+ }
+}
diff --git a/src/include/replication/snapbuild.h b/src/include/replication/snapbuild.h
index b95f56eec3..7a796ce136 100644
--- a/src/include/replication/snapbuild.h
+++ b/src/include/replication/snapbuild.h
@@ -88,4 +88,7 @@ extern void SnapBuildProcessRunningXacts(SnapBuild *builder, XLogRecPtr lsn,
struct xl_running_xacts *running);
extern void SnapBuildSerializationPoint(SnapBuild *builder, XLogRecPtr lsn);
+extern void SnapBuildXidSetCatalogChanges(SnapBuild *builder, TransactionId xid,
+ int subxcnt, TransactionId *subxacts,
+ XLogRecPtr lsn);
#endif /* SNAPBUILD_H */
--
2.24.3 (Apple Git-128)
[application/octet-stream] REL13-v8-0001-Fix-catalog-lookup-with-the-wrong-snapshot-during.patch (16.3K, ../CAD21AoC_oEy3EvgbJ=asz20Z3ROWtRM01f-F8+F_5VjhCvJX6g@mail.gmail.com/4-REL13-v8-0001-Fix-catalog-lookup-with-the-wrong-snapshot-during.patch)
download | inline diff:
From a01f57fec21e03bb09563f66c3e26816538dd78a Mon Sep 17 00:00:00 2001
From: Masahiko Sawada <[email protected]>
Date: Mon, 11 Jul 2022 21:49:06 +0900
Subject: [PATCH v8] Fix catalog lookup with the wrong snapshot during logical
decoding.
Previously, we relied on HEAP2_NEW_CID records and XACT_INVALIDATION
records to know if the transaction has modified the catalog, and that
information is not serialized to snapshot. Therefore, if the logical
decoding decodes only the commit record of the transaction that
actually has modified a catalog, we missed adding its XID to the
snapshot. We ended up looking at catalogs with the wrong snapshot.
To fix this problem, this changes the snapshot builder so that it
remembers the running transaction written in the xl_running_xacts
record that we decoded first, and mark the transaction as containing
catalog changes if it's in the list of the initial running
transactions and its commit record has XACT_XINFO_HAS_INVALS. To avoid
ABI breakage, we store the array of the initial running transactions
in the static variables InitialRunningXacts and NInitialRunningXacts,
neither SnapBuild nor ReorderBuffer.
This approach has false positive; we could end up adding the
transaction that didn't change catalog to the snapshot since we cannot
distinguish whether the transaction has catalog changes only by
checking the COMMIT record. It doesn't have the information on
which (sub) transaction has catalog changes, and XACT_XINFO_HAS_INVALS
doesn't necessarily indicate that the transaction has catalog
change. But it doesn't become a problem since we use historic snapshot
only for system catalog lookups.
On the master branch, we took a more future-proof approach -- writing
catalog modifying transactions to the serialized snapshot. But we
cannot backpatch it because of change in SnapBuild.
Back-patch to all supported released.
Reported-by: Mike Oh <[email protected]>
Author: Masahiko Sawada <[email protected]>
Reviewed-by: Amit Kapila <[email protected]>
Reviewed-by: Takamichi Osumi <[email protected]>
Reviewed-by: Kyotaro Horiguchi <[email protected]>
Reviewed-by: Bertrand Drouvot <[email protected]>
Reviewed-by: Shi yu <[email protected]>
Reviewed-by: Ahsan Hadi <[email protected]>
Discussion: https://postgr.es/m/81D0D8B0-E7C4-4999-B616-1E5004DBDCD2%40amazon.com
Backpatch-through: 10
---
contrib/test_decoding/Makefile | 2 +-
.../expected/catalog_change_snapshot.out | 44 ++++++
.../specs/catalog_change_snapshot.spec | 39 +++++
src/backend/replication/logical/decode.c | 15 ++
src/backend/replication/logical/snapbuild.c | 143 +++++++++++++++++-
src/include/replication/snapbuild.h | 3 +
6 files changed, 238 insertions(+), 8 deletions(-)
create mode 100644 contrib/test_decoding/expected/catalog_change_snapshot.out
create mode 100644 contrib/test_decoding/specs/catalog_change_snapshot.spec
diff --git a/contrib/test_decoding/Makefile b/contrib/test_decoding/Makefile
index f439c582a5..6ec09ab192 100644
--- a/contrib/test_decoding/Makefile
+++ b/contrib/test_decoding/Makefile
@@ -7,7 +7,7 @@ REGRESS = ddl xact rewrite toast permissions decoding_in_xact \
decoding_into_rel binary prepared replorigin time messages \
spill slot truncate
ISOLATION = mxact delayed_startup ondisk_startup concurrent_ddl_dml \
- oldest_xmin snapshot_transfer subxact_without_top
+ oldest_xmin snapshot_transfer subxact_without_top catalog_change_snapshot
REGRESS_OPTS = --temp-config $(top_srcdir)/contrib/test_decoding/logical.conf
ISOLATION_OPTS = --temp-config $(top_srcdir)/contrib/test_decoding/logical.conf
diff --git a/contrib/test_decoding/expected/catalog_change_snapshot.out b/contrib/test_decoding/expected/catalog_change_snapshot.out
new file mode 100644
index 0000000000..dc4f9b7018
--- /dev/null
+++ b/contrib/test_decoding/expected/catalog_change_snapshot.out
@@ -0,0 +1,44 @@
+Parsed test spec with 2 sessions
+
+starting permutation: s0_init s0_begin s0_savepoint s0_truncate s1_checkpoint s1_get_changes s0_commit s0_begin s0_insert s1_checkpoint s1_get_changes s0_commit s1_get_changes
+step s0_init: SELECT 'init' FROM pg_create_logical_replication_slot('isolation_slot', 'test_decoding');
+?column?
+--------
+init
+(1 row)
+
+step s0_begin: BEGIN;
+step s0_savepoint: SAVEPOINT sp1;
+step s0_truncate: TRUNCATE tbl1;
+step s1_checkpoint: CHECKPOINT;
+step s1_get_changes: SELECT data FROM pg_logical_slot_get_changes('isolation_slot', NULL, NULL, 'skip-empty-xacts', '1', 'include-xids', '0');
+data
+----
+(0 rows)
+
+step s0_commit: COMMIT;
+step s0_begin: BEGIN;
+step s0_insert: INSERT INTO tbl1 VALUES (1);
+step s1_checkpoint: CHECKPOINT;
+step s1_get_changes: SELECT data FROM pg_logical_slot_get_changes('isolation_slot', NULL, NULL, 'skip-empty-xacts', '1', 'include-xids', '0');
+data
+---------------------------------------
+BEGIN
+table public.tbl1: TRUNCATE: (no-flags)
+COMMIT
+(3 rows)
+
+step s0_commit: COMMIT;
+step s1_get_changes: SELECT data FROM pg_logical_slot_get_changes('isolation_slot', NULL, NULL, 'skip-empty-xacts', '1', 'include-xids', '0');
+data
+-------------------------------------------------------------
+BEGIN
+table public.tbl1: INSERT: val1[integer]:1 val2[integer]:null
+COMMIT
+(3 rows)
+
+?column?
+--------
+stop
+(1 row)
+
diff --git a/contrib/test_decoding/specs/catalog_change_snapshot.spec b/contrib/test_decoding/specs/catalog_change_snapshot.spec
new file mode 100644
index 0000000000..662760fbcf
--- /dev/null
+++ b/contrib/test_decoding/specs/catalog_change_snapshot.spec
@@ -0,0 +1,39 @@
+# Test decoding only the commit record of the transaction that have
+# modified catalogs.
+setup
+{
+ DROP TABLE IF EXISTS tbl1;
+ CREATE TABLE tbl1 (val1 integer, val2 integer);
+}
+
+teardown
+{
+ DROP TABLE tbl1;
+ SELECT 'stop' FROM pg_drop_replication_slot('isolation_slot');
+}
+
+session "s0"
+setup { SET synchronous_commit=on; }
+step "s0_init" { SELECT 'init' FROM pg_create_logical_replication_slot('isolation_slot', 'test_decoding'); }
+step "s0_begin" { BEGIN; }
+step "s0_savepoint" { SAVEPOINT sp1; }
+step "s0_truncate" { TRUNCATE tbl1; }
+step "s0_insert" { INSERT INTO tbl1 VALUES (1); }
+step "s0_commit" { COMMIT; }
+
+session "s1"
+setup { SET synchronous_commit=on; }
+step "s1_checkpoint" { CHECKPOINT; }
+step "s1_get_changes" { SELECT data FROM pg_logical_slot_get_changes('isolation_slot', NULL, NULL, 'skip-empty-xacts', '1', 'include-xids', '0'); }
+
+# For the transaction that TRUNCATEd the table tbl1, the last decoding decodes
+# only its COMMIT record, because it starts from the RUNNING_XACT record emitted
+# during the first checkpoint execution. This transaction must be marked as
+# containing catalog changes while decoding the COMMIT record and the decoding
+# of the INSERT record must read the pg_class with the correct historic snapshot.
+#
+# Note that in a case where bgwriter wrote the XACT_RUNNING record between "s0_commit"
+# and "s0_begin", this doesn't happen as the decoding starts from the RUNNING_XACT
+# record written by bgwriter. One might think we can either stop the bgwriter or
+# increase LOG_SNAPSHOT_INTERVAL_MS but it's not practical via tests.
+permutation "s0_init" "s0_begin" "s0_savepoint" "s0_truncate" "s1_checkpoint" "s1_get_changes" "s0_commit" "s0_begin" "s0_insert" "s1_checkpoint" "s1_get_changes" "s0_commit" "s1_get_changes"
diff --git a/src/backend/replication/logical/decode.c b/src/backend/replication/logical/decode.c
index 5a2b828aa3..33a2fd2ee8 100644
--- a/src/backend/replication/logical/decode.c
+++ b/src/backend/replication/logical/decode.c
@@ -572,6 +572,21 @@ DecodeCommit(LogicalDecodingContext *ctx, XLogRecordBuffer *buf,
commit_time = parsed->origin_timestamp;
}
+ /*
+ * If the COMMIT record has invalidation messages, it could have catalog
+ * changes. It is possible that we didn't mark this transaction as
+ * containing catalog changes when the decoding starts from a commit record
+ * without decoding the transaction's other changes. So, we ensure to mark
+ * such transactions as containing catalog change.
+ *
+ * This must be done before SnapBuildCommitTxn() so that we can include
+ * these transactions in the historic snapshot.
+ */
+ if (parsed->xinfo & XACT_XINFO_HAS_INVALS)
+ SnapBuildXidSetCatalogChanges(ctx->snapshot_builder, xid,
+ parsed->nsubxacts, parsed->subxacts,
+ buf->origptr);
+
/*
* Process invalidation messages, even if we're not interested in the
* transaction's contents, since the various caches need to always be
diff --git a/src/backend/replication/logical/snapbuild.c b/src/backend/replication/logical/snapbuild.c
index be46bf0363..b5a9b1e5fb 100644
--- a/src/backend/replication/logical/snapbuild.c
+++ b/src/backend/replication/logical/snapbuild.c
@@ -252,8 +252,38 @@ struct SnapBuild
static ResourceOwner SavedResourceOwnerDuringExport = NULL;
static bool ExportInProgress = false;
-/* ->committed manipulation */
-static void SnapBuildPurgeCommittedTxn(SnapBuild *builder);
+/*
+ * Array of transactions and subtransactions that were running when
+ * the xl_running_xacts record that we decoded first was written.
+ * The array is sorted in xidComparator order. We remove Xids from
+ * this array when they become old enough to matter. This array is
+ * allocated in builder->context so its lifetime is the same as the
+ * snapshot builder.
+ *
+ * We rely on HEAP2_NEW_CID records and XACT_INVALIDATIONS to know
+ * if the transaction has changed the catalog, and that information
+ * is not serialized to SnapBuilder. Therefore, if the logical
+ * decoding decodes the commit record of the transaction that actually
+ * has done catalog changes without these records, we miss to add
+ * the xid to the snapshot, and end up looking at catalogs with the
+ * wrong snapshot. To avoid this problem, if the COMMIT record of
+ * the xid listed in InitialRunningXacts has XACT_XINFO_HAS_INVALS
+ * flag, we mark both the top transaction and its substransactions
+ * as containing catalog changes.
+ *
+ * We could end up adding the transaction that didn't change catalog
+ * to the snapshot since we cannot distinguish whether the transaction
+ * has catalog changes only by checking the COMMIT record. It doesn't
+ * have the information on which (sub) transaction has catalog changes,
+ * and XACT_XINFO_HAS_INVALS doesn't necessarily indicate that the
+ * transaction has catalog change. But it doesn't become a problem since
+ * we use historic snapshot only for reading system catalogs.
+ */
+static TransactionId *InitialRunningXacts = NULL;
+static int NInitialRunningXacts = 0;
+
+/* ->committed and InitailRunningXacts manipulation */
+static void SnapBuildPurgeOlderTxn(SnapBuild *builder);
/* snapshot building/manipulation/distribution functions */
static Snapshot SnapBuildBuildSnapshot(SnapBuild *builder);
@@ -890,12 +920,17 @@ SnapBuildAddCommittedTxn(SnapBuild *builder, TransactionId xid)
}
/*
- * Remove knowledge about transactions we treat as committed that are smaller
- * than ->xmin. Those won't ever get checked via the ->committed array but via
- * the clog machinery, so we don't need to waste memory on them.
+ * Remove knowledge about transactions we treat as committed and the initial
+ * running transactions that are smaller than ->xmin. Those won't ever get
+ * checked via the ->committed or InitialRunningXacts array, respectively.
+ * The committed xids will get checked via the clog machinery.
+ *
+ * We can ideally remove the transaction from InitialRunningXacts array
+ * once it is finished (committed/aborted) but that could be costly as we need
+ * to maintain the xids order in the array.
*/
static void
-SnapBuildPurgeCommittedTxn(SnapBuild *builder)
+SnapBuildPurgeOlderTxn(SnapBuild *builder)
{
int off;
TransactionId *workspace;
@@ -930,6 +965,49 @@ SnapBuildPurgeCommittedTxn(SnapBuild *builder)
builder->committed.xcnt = surviving_xids;
pfree(workspace);
+
+ /* Quick exit if there is no initial running transactions */
+ if (likely(NInitialRunningXacts == 0))
+ return;
+
+ /* bound check if there is at least one transaction to remove */
+ if (!NormalTransactionIdPrecedes(InitialRunningXacts[0],
+ builder->xmin))
+ return;
+
+ /*
+ * purge xids in InitialRunningXacts as well. The purged array must also be
+ * sorted in xidComparator order.
+ */
+ workspace =
+ MemoryContextAlloc(builder->context,
+ NInitialRunningXacts * sizeof(TransactionId));
+ surviving_xids = 0;
+ for (off = 0; off < NInitialRunningXacts; off++)
+ {
+ if (NormalTransactionIdPrecedes(InitialRunningXacts[off],
+ builder->xmin))
+ ; /* remove */
+ else
+ workspace[surviving_xids++] = InitialRunningXacts[off];
+ }
+
+ if (surviving_xids > 0)
+ memcpy(InitialRunningXacts, workspace,
+ sizeof(TransactionId) * surviving_xids);
+ else
+ {
+ pfree(InitialRunningXacts);
+ InitialRunningXacts = NULL;
+ }
+
+ elog(DEBUG3, "purged initial running transactions from %u to %u, oldest running xid %u",
+ (uint32) NInitialRunningXacts,
+ (uint32) surviving_xids,
+ builder->xmin);
+
+ NInitialRunningXacts = surviving_xids;
+ pfree(workspace);
}
/*
@@ -1115,6 +1193,21 @@ SnapBuildProcessRunningXacts(SnapBuild *builder, XLogRecPtr lsn, xl_running_xact
*/
if (builder->state < SNAPBUILD_CONSISTENT)
{
+ /*
+ * Remember the transactions and subtransactions that were running
+ * when xl_running_xacts record that we decoded first was written.
+ */
+ if (builder->state == SNAPBUILD_START)
+ {
+ int nxacts = running->subxcnt + running->xcnt;
+ Size sz = sizeof(TransactionId) * nxacts;
+
+ NInitialRunningXacts = nxacts;
+ InitialRunningXacts = MemoryContextAlloc(builder->context, sz);
+ memcpy(InitialRunningXacts, running->xids, sz);
+ qsort(InitialRunningXacts, nxacts, sizeof(TransactionId), xidComparator);
+ }
+
/* returns false if there's no point in performing cleanup just yet */
if (!SnapBuildFindSnapshot(builder, lsn, running))
return;
@@ -1137,7 +1230,7 @@ SnapBuildProcessRunningXacts(SnapBuild *builder, XLogRecPtr lsn, xl_running_xact
builder->xmin = running->oldestRunningXid;
/* Remove transactions we don't need to keep track off anymore */
- SnapBuildPurgeCommittedTxn(builder);
+ SnapBuildPurgeOlderTxn(builder);
/*
* Advance the xmin limit for the current replication slot, to allow
@@ -2030,3 +2123,39 @@ CheckPointSnapBuild(void)
}
FreeDir(snap_dir);
}
+
+/*
+ * If the given xid is in the list of the initial running xacts, we mark the
+ * transaction and its subtransactions as containing catalog changes. See
+ * comments
+ */
+void
+SnapBuildXidSetCatalogChanges(SnapBuild *builder, TransactionId xid, int subxcnt,
+ TransactionId *subxacts, XLogRecPtr lsn)
+{
+ /*
+ * Skip if there is no initial running xacts information or the
+ * transaction is already marked as containing catalog changes.
+ */
+ if (likely((NInitialRunningXacts == 0) ||
+ ReorderBufferXidHasCatalogChanges(builder->reorder, xid)))
+ return;
+
+ /*
+ * If this committed transaction is the one that was running at the time
+ * when decoding the first RUNNING_XACTS record and have done catalog
+ * changes, we can mark both the top transaction and its subtransactions
+ * as containing catalog changes.
+ */
+ if (bsearch(&xid, InitialRunningXacts, NInitialRunningXacts,
+ sizeof(TransactionId), xidComparator) != NULL)
+ {
+ ReorderBufferXidSetCatalogChanges(builder->reorder, xid, lsn);
+
+ for (int i = 0; i < subxcnt; i++)
+ {
+ ReorderBufferAssignChild(builder->reorder, xid, subxacts[i], lsn);
+ ReorderBufferXidSetCatalogChanges(builder->reorder, subxacts[i], lsn);
+ }
+ }
+}
diff --git a/src/include/replication/snapbuild.h b/src/include/replication/snapbuild.h
index b048dc7484..17d2f93300 100644
--- a/src/include/replication/snapbuild.h
+++ b/src/include/replication/snapbuild.h
@@ -88,4 +88,7 @@ extern void SnapBuildProcessRunningXacts(SnapBuild *builder, XLogRecPtr lsn,
struct xl_running_xacts *running);
extern void SnapBuildSerializationPoint(SnapBuild *builder, XLogRecPtr lsn);
+extern void SnapBuildXidSetCatalogChanges(SnapBuild *builder, TransactionId xid,
+ int subxcnt, TransactionId *subxacts,
+ XLogRecPtr lsn);
#endif /* SNAPBUILD_H */
--
2.24.3 (Apple Git-128)
[application/octet-stream] REl11-v8-0001-Fix-catalog-lookup-with-the-wrong-snapshot-during.patch (16.2K, ../CAD21AoC_oEy3EvgbJ=asz20Z3ROWtRM01f-F8+F_5VjhCvJX6g@mail.gmail.com/5-REl11-v8-0001-Fix-catalog-lookup-with-the-wrong-snapshot-during.patch)
download | inline diff:
From 950f16af07d4ddf448f2fe0d66a52b856aa825c9 Mon Sep 17 00:00:00 2001
From: Masahiko Sawada <[email protected]>
Date: Sun, 17 Jul 2022 07:30:23 +0900
Subject: [PATCH v8] Fix catalog lookup with the wrong snapshot during logical
decoding.
Previously, we relied on HEAP2_NEW_CID records and XACT_INVALIDATION
records to know if the transaction has modified the catalog, and that
information is not serialized to snapshot. Therefore, if the logical
decoding decodes only the commit record of the transaction that
actually has modified a catalog, we missed adding its XID to the
snapshot. We ended up looking at catalogs with the wrong snapshot.
To fix this problem, this changes the snapshot builder so that it
remembers the running transaction written in the xl_running_xacts
record that we decoded first, and mark the transaction as containing
catalog changes if it's in the list of the initial running
transactions and its commit record has XACT_XINFO_HAS_INVALS. To avoid
ABI breakage, we store the array of the initial running transactions
in the static variables InitialRunningXacts and NInitialRunningXacts,
neither SnapBuild nor ReorderBuffer.
This approach has false positive; we could end up adding the
transaction that didn't change catalog to the snapshot since we cannot
distinguish whether the transaction has catalog changes only by
checking the COMMIT record. It doesn't have the information on
which (sub) transaction has catalog changes, and XACT_XINFO_HAS_INVALS
doesn't necessarily indicate that the transaction has catalog
change. But it doesn't become a problem since we use historic snapshot
only for system catalog lookups.
On the master branch, we took a more future-proof approach -- writing
catalog modifying transactions to the serialized snapshot. But we
cannot backpatch it because of change in SnapBuild.
Back-patch to all supported released.
Reported-by: Mike Oh <[email protected]>
Author: Masahiko Sawada <[email protected]>
Reviewed-by: Amit Kapila <[email protected]>
Reviewed-by: Takamichi Osumi <[email protected]>
Reviewed-by: Kyotaro Horiguchi <[email protected]>
Reviewed-by: Bertrand Drouvot <[email protected]>
Reviewed-by: Shi yu <[email protected]>
Reviewed-by: Ahsan Hadi <[email protected]>
Discussion: https://postgr.es/m/81D0D8B0-E7C4-4999-B616-1E5004DBDCD2%40amazon.com
Backpatch-through: 10
---
contrib/test_decoding/Makefile | 2 +-
.../expected/catalog_change_snapshot.out | 44 ++++++
.../specs/catalog_change_snapshot.spec | 39 +++++
src/backend/replication/logical/decode.c | 15 ++
src/backend/replication/logical/snapbuild.c | 143 +++++++++++++++++-
src/include/replication/snapbuild.h | 3 +
6 files changed, 238 insertions(+), 8 deletions(-)
create mode 100644 contrib/test_decoding/expected/catalog_change_snapshot.out
create mode 100644 contrib/test_decoding/specs/catalog_change_snapshot.spec
diff --git a/contrib/test_decoding/Makefile b/contrib/test_decoding/Makefile
index 65a91a8014..973b94738a 100644
--- a/contrib/test_decoding/Makefile
+++ b/contrib/test_decoding/Makefile
@@ -51,7 +51,7 @@ regresscheck-install-force: | submake-regress submake-test_decoding temp-install
$(REGRESSCHECKS)
ISOLATIONCHECKS=mxact delayed_startup ondisk_startup concurrent_ddl_dml \
- oldest_xmin snapshot_transfer subxact_without_top
+ oldest_xmin snapshot_transfer subxact_without_top catalog_change_snapshot
isolationcheck: | submake-isolation submake-test_decoding temp-install
$(pg_isolation_regress_check) \
diff --git a/contrib/test_decoding/expected/catalog_change_snapshot.out b/contrib/test_decoding/expected/catalog_change_snapshot.out
new file mode 100644
index 0000000000..dc4f9b7018
--- /dev/null
+++ b/contrib/test_decoding/expected/catalog_change_snapshot.out
@@ -0,0 +1,44 @@
+Parsed test spec with 2 sessions
+
+starting permutation: s0_init s0_begin s0_savepoint s0_truncate s1_checkpoint s1_get_changes s0_commit s0_begin s0_insert s1_checkpoint s1_get_changes s0_commit s1_get_changes
+step s0_init: SELECT 'init' FROM pg_create_logical_replication_slot('isolation_slot', 'test_decoding');
+?column?
+--------
+init
+(1 row)
+
+step s0_begin: BEGIN;
+step s0_savepoint: SAVEPOINT sp1;
+step s0_truncate: TRUNCATE tbl1;
+step s1_checkpoint: CHECKPOINT;
+step s1_get_changes: SELECT data FROM pg_logical_slot_get_changes('isolation_slot', NULL, NULL, 'skip-empty-xacts', '1', 'include-xids', '0');
+data
+----
+(0 rows)
+
+step s0_commit: COMMIT;
+step s0_begin: BEGIN;
+step s0_insert: INSERT INTO tbl1 VALUES (1);
+step s1_checkpoint: CHECKPOINT;
+step s1_get_changes: SELECT data FROM pg_logical_slot_get_changes('isolation_slot', NULL, NULL, 'skip-empty-xacts', '1', 'include-xids', '0');
+data
+---------------------------------------
+BEGIN
+table public.tbl1: TRUNCATE: (no-flags)
+COMMIT
+(3 rows)
+
+step s0_commit: COMMIT;
+step s1_get_changes: SELECT data FROM pg_logical_slot_get_changes('isolation_slot', NULL, NULL, 'skip-empty-xacts', '1', 'include-xids', '0');
+data
+-------------------------------------------------------------
+BEGIN
+table public.tbl1: INSERT: val1[integer]:1 val2[integer]:null
+COMMIT
+(3 rows)
+
+?column?
+--------
+stop
+(1 row)
+
diff --git a/contrib/test_decoding/specs/catalog_change_snapshot.spec b/contrib/test_decoding/specs/catalog_change_snapshot.spec
new file mode 100644
index 0000000000..662760fbcf
--- /dev/null
+++ b/contrib/test_decoding/specs/catalog_change_snapshot.spec
@@ -0,0 +1,39 @@
+# Test decoding only the commit record of the transaction that have
+# modified catalogs.
+setup
+{
+ DROP TABLE IF EXISTS tbl1;
+ CREATE TABLE tbl1 (val1 integer, val2 integer);
+}
+
+teardown
+{
+ DROP TABLE tbl1;
+ SELECT 'stop' FROM pg_drop_replication_slot('isolation_slot');
+}
+
+session "s0"
+setup { SET synchronous_commit=on; }
+step "s0_init" { SELECT 'init' FROM pg_create_logical_replication_slot('isolation_slot', 'test_decoding'); }
+step "s0_begin" { BEGIN; }
+step "s0_savepoint" { SAVEPOINT sp1; }
+step "s0_truncate" { TRUNCATE tbl1; }
+step "s0_insert" { INSERT INTO tbl1 VALUES (1); }
+step "s0_commit" { COMMIT; }
+
+session "s1"
+setup { SET synchronous_commit=on; }
+step "s1_checkpoint" { CHECKPOINT; }
+step "s1_get_changes" { SELECT data FROM pg_logical_slot_get_changes('isolation_slot', NULL, NULL, 'skip-empty-xacts', '1', 'include-xids', '0'); }
+
+# For the transaction that TRUNCATEd the table tbl1, the last decoding decodes
+# only its COMMIT record, because it starts from the RUNNING_XACT record emitted
+# during the first checkpoint execution. This transaction must be marked as
+# containing catalog changes while decoding the COMMIT record and the decoding
+# of the INSERT record must read the pg_class with the correct historic snapshot.
+#
+# Note that in a case where bgwriter wrote the XACT_RUNNING record between "s0_commit"
+# and "s0_begin", this doesn't happen as the decoding starts from the RUNNING_XACT
+# record written by bgwriter. One might think we can either stop the bgwriter or
+# increase LOG_SNAPSHOT_INTERVAL_MS but it's not practical via tests.
+permutation "s0_init" "s0_begin" "s0_savepoint" "s0_truncate" "s1_checkpoint" "s1_get_changes" "s0_commit" "s0_begin" "s0_insert" "s1_checkpoint" "s1_get_changes" "s0_commit" "s1_get_changes"
diff --git a/src/backend/replication/logical/decode.c b/src/backend/replication/logical/decode.c
index c085f7b0f3..beadabb804 100644
--- a/src/backend/replication/logical/decode.c
+++ b/src/backend/replication/logical/decode.c
@@ -576,6 +576,21 @@ DecodeCommit(LogicalDecodingContext *ctx, XLogRecordBuffer *buf,
commit_time = parsed->origin_timestamp;
}
+ /*
+ * If the COMMIT record has invalidation messages, it could have catalog
+ * changes. It is possible that we didn't mark this transaction as
+ * containing catalog changes when the decoding starts from a commit record
+ * without decoding the transaction's other changes. So, we ensure to mark
+ * such transactions as containing catalog change.
+ *
+ * This must be done before SnapBuildCommitTxn() so that we can include
+ * these transactions in the historic snapshot.
+ */
+ if (parsed->xinfo & XACT_XINFO_HAS_INVALS)
+ SnapBuildXidSetCatalogChanges(ctx->snapshot_builder, xid,
+ parsed->nsubxacts, parsed->subxacts,
+ buf->origptr);
+
/*
* Process invalidation messages, even if we're not interested in the
* transaction's contents, since the various caches need to always be
diff --git a/src/backend/replication/logical/snapbuild.c b/src/backend/replication/logical/snapbuild.c
index 1c52bc64e3..8190e7a911 100644
--- a/src/backend/replication/logical/snapbuild.c
+++ b/src/backend/replication/logical/snapbuild.c
@@ -258,8 +258,38 @@ struct SnapBuild
static ResourceOwner SavedResourceOwnerDuringExport = NULL;
static bool ExportInProgress = false;
-/* ->committed manipulation */
-static void SnapBuildPurgeCommittedTxn(SnapBuild *builder);
+/*
+ * Array of transactions and subtransactions that were running when
+ * the xl_running_xacts record that we decoded first was written.
+ * The array is sorted in xidComparator order. We remove Xids from
+ * this array when they become old enough to matter. This array is
+ * allocated in builder->context so its lifetime is the same as the
+ * snapshot builder.
+ *
+ * We rely on HEAP2_NEW_CID records and XACT_INVALIDATIONS to know
+ * if the transaction has changed the catalog, and that information
+ * is not serialized to SnapBuilder. Therefore, if the logical
+ * decoding decodes the commit record of the transaction that actually
+ * has done catalog changes without these records, we miss to add
+ * the xid to the snapshot, and end up looking at catalogs with the
+ * wrong snapshot. To avoid this problem, if the COMMIT record of
+ * the xid listed in InitialRunningXacts has XACT_XINFO_HAS_INVALS
+ * flag, we mark both the top transaction and its substransactions
+ * as containing catalog changes.
+ *
+ * We could end up adding the transaction that didn't change catalog
+ * to the snapshot since we cannot distinguish whether the transaction
+ * has catalog changes only by checking the COMMIT record. It doesn't
+ * have the information on which (sub) transaction has catalog changes,
+ * and XACT_XINFO_HAS_INVALS doesn't necessarily indicate that the
+ * transaction has catalog change. But it doesn't become a problem since
+ * we use historic snapshot only for reading system catalogs.
+ */
+static TransactionId *InitialRunningXacts = NULL;
+static int NInitialRunningXacts = 0;
+
+/* ->committed and InitailRunningXacts manipulation */
+static void SnapBuildPurgeOlderTxn(SnapBuild *builder);
/* snapshot building/manipulation/distribution functions */
static Snapshot SnapBuildBuildSnapshot(SnapBuild *builder);
@@ -896,12 +926,17 @@ SnapBuildAddCommittedTxn(SnapBuild *builder, TransactionId xid)
}
/*
- * Remove knowledge about transactions we treat as committed that are smaller
- * than ->xmin. Those won't ever get checked via the ->committed array but via
- * the clog machinery, so we don't need to waste memory on them.
+ * Remove knowledge about transactions we treat as committed and the initial
+ * running transactions that are smaller than ->xmin. Those won't ever get
+ * checked via the ->committed or InitialRunningXacts array, respectively.
+ * The committed xids will get checked via the clog machinery.
+ *
+ * We can ideally remove the transaction from InitialRunningXacts array
+ * once it is finished (committed/aborted) but that could be costly as we need
+ * to maintain the xids order in the array.
*/
static void
-SnapBuildPurgeCommittedTxn(SnapBuild *builder)
+SnapBuildPurgeOlderTxn(SnapBuild *builder)
{
int off;
TransactionId *workspace;
@@ -936,6 +971,49 @@ SnapBuildPurgeCommittedTxn(SnapBuild *builder)
builder->committed.xcnt = surviving_xids;
pfree(workspace);
+
+ /* Quick exit if there is no initial running transactions */
+ if (likely(NInitialRunningXacts == 0))
+ return;
+
+ /* bound check if there is at least one transaction to remove */
+ if (!NormalTransactionIdPrecedes(InitialRunningXacts[0],
+ builder->xmin))
+ return;
+
+ /*
+ * purge xids in InitialRunningXacts as well. The purged array must also be
+ * sorted in xidComparator order.
+ */
+ workspace =
+ MemoryContextAlloc(builder->context,
+ NInitialRunningXacts * sizeof(TransactionId));
+ surviving_xids = 0;
+ for (off = 0; off < NInitialRunningXacts; off++)
+ {
+ if (NormalTransactionIdPrecedes(InitialRunningXacts[off],
+ builder->xmin))
+ ; /* remove */
+ else
+ workspace[surviving_xids++] = InitialRunningXacts[off];
+ }
+
+ if (surviving_xids > 0)
+ memcpy(InitialRunningXacts, workspace,
+ sizeof(TransactionId) * surviving_xids);
+ else
+ {
+ pfree(InitialRunningXacts);
+ InitialRunningXacts = NULL;
+ }
+
+ elog(DEBUG3, "purged initial running transactions from %u to %u, oldest running xid %u",
+ (uint32) NInitialRunningXacts,
+ (uint32) surviving_xids,
+ builder->xmin);
+
+ NInitialRunningXacts = surviving_xids;
+ pfree(workspace);
}
/*
@@ -1121,6 +1199,21 @@ SnapBuildProcessRunningXacts(SnapBuild *builder, XLogRecPtr lsn, xl_running_xact
*/
if (builder->state < SNAPBUILD_CONSISTENT)
{
+ /*
+ * Remember the transactions and subtransactions that were running
+ * when xl_running_xacts record that we decoded first was written.
+ */
+ if (builder->state == SNAPBUILD_START)
+ {
+ int nxacts = running->subxcnt + running->xcnt;
+ Size sz = sizeof(TransactionId) * nxacts;
+
+ NInitialRunningXacts = nxacts;
+ InitialRunningXacts = MemoryContextAlloc(builder->context, sz);
+ memcpy(InitialRunningXacts, running->xids, sz);
+ qsort(InitialRunningXacts, nxacts, sizeof(TransactionId), xidComparator);
+ }
+
/* returns false if there's no point in performing cleanup just yet */
if (!SnapBuildFindSnapshot(builder, lsn, running))
return;
@@ -1143,7 +1236,7 @@ SnapBuildProcessRunningXacts(SnapBuild *builder, XLogRecPtr lsn, xl_running_xact
builder->xmin = running->oldestRunningXid;
/* Remove transactions we don't need to keep track off anymore */
- SnapBuildPurgeCommittedTxn(builder);
+ SnapBuildPurgeOlderTxn(builder);
/*
* Advance the xmin limit for the current replication slot, to allow
@@ -1996,3 +2089,39 @@ CheckPointSnapBuild(void)
}
FreeDir(snap_dir);
}
+
+/*
+ * If the given xid is in the list of the initial running xacts, we mark the
+ * transaction and its subtransactions as containing catalog changes. See
+ * comments
+ */
+void
+SnapBuildXidSetCatalogChanges(SnapBuild *builder, TransactionId xid, int subxcnt,
+ TransactionId *subxacts, XLogRecPtr lsn)
+{
+ /*
+ * Skip if there is no initial running xacts information or the
+ * transaction is already marked as containing catalog changes.
+ */
+ if (likely((NInitialRunningXacts == 0) ||
+ ReorderBufferXidHasCatalogChanges(builder->reorder, xid)))
+ return;
+
+ /*
+ * If this committed transaction is the one that was running at the time
+ * when decoding the first RUNNING_XACTS record and have done catalog
+ * changes, we can mark both the top transaction and its subtransactions
+ * as containing catalog changes.
+ */
+ if (bsearch(&xid, InitialRunningXacts, NInitialRunningXacts,
+ sizeof(TransactionId), xidComparator) != NULL)
+ {
+ ReorderBufferXidSetCatalogChanges(builder->reorder, xid, lsn);
+
+ for (int i = 0; i < subxcnt; i++)
+ {
+ ReorderBufferAssignChild(builder->reorder, xid, subxacts[i], lsn);
+ ReorderBufferXidSetCatalogChanges(builder->reorder, subxacts[i], lsn);
+ }
+ }
+}
diff --git a/src/include/replication/snapbuild.h b/src/include/replication/snapbuild.h
index 1df66a3c75..4df3c3f2f7 100644
--- a/src/include/replication/snapbuild.h
+++ b/src/include/replication/snapbuild.h
@@ -88,4 +88,7 @@ extern void SnapBuildProcessRunningXacts(SnapBuild *builder, XLogRecPtr lsn,
struct xl_running_xacts *running);
extern void SnapBuildSerializationPoint(SnapBuild *builder, XLogRecPtr lsn);
+extern void SnapBuildXidSetCatalogChanges(SnapBuild *builder, TransactionId xid,
+ int subxcnt, TransactionId *subxacts,
+ XLogRecPtr lsn);
#endif /* SNAPBUILD_H */
--
2.24.3 (Apple Git-128)
[application/octet-stream] REl12-v8-0001-Fix-catalog-lookup-with-the-wrong-snapshot-during.patch (16.3K, ../CAD21AoC_oEy3EvgbJ=asz20Z3ROWtRM01f-F8+F_5VjhCvJX6g@mail.gmail.com/6-REl12-v8-0001-Fix-catalog-lookup-with-the-wrong-snapshot-during.patch)
download | inline diff:
From 6eb109bf6429c2da07b96d26b65571c3172ec568 Mon Sep 17 00:00:00 2001
From: Masahiko Sawada <[email protected]>
Date: Sun, 17 Jul 2022 07:19:00 +0900
Subject: [PATCH v8] Fix catalog lookup with the wrong snapshot during logical
decoding.
Previously, we relied on HEAP2_NEW_CID records and XACT_INVALIDATION
records to know if the transaction has modified the catalog, and that
information is not serialized to snapshot. Therefore, if the logical
decoding decodes only the commit record of the transaction that
actually has modified a catalog, we missed adding its XID to the
snapshot. We ended up looking at catalogs with the wrong snapshot.
To fix this problem, this changes the snapshot builder so that it
remembers the running transaction written in the xl_running_xacts
record that we decoded first, and mark the transaction as containing
catalog changes if it's in the list of the initial running
transactions and its commit record has XACT_XINFO_HAS_INVALS. To avoid
ABI breakage, we store the array of the initial running transactions
in the static variables InitialRunningXacts and NInitialRunningXacts,
neither SnapBuild nor ReorderBuffer.
This approach has false positive; we could end up adding the
transaction that didn't change catalog to the snapshot since we cannot
distinguish whether the transaction has catalog changes only by
checking the COMMIT record. It doesn't have the information on
which (sub) transaction has catalog changes, and XACT_XINFO_HAS_INVALS
doesn't necessarily indicate that the transaction has catalog
change. But it doesn't become a problem since we use historic snapshot
only for system catalog lookups.
On the master branch, we took a more future-proof approach -- writing
catalog modifying transactions to the serialized snapshot. But we
cannot backpatch it because of change in SnapBuild.
Back-patch to all supported released.
Reported-by: Mike Oh <[email protected]>
Author: Masahiko Sawada <[email protected]>
Reviewed-by: Amit Kapila <[email protected]>
Reviewed-by: Takamichi Osumi <[email protected]>
Reviewed-by: Kyotaro Horiguchi <[email protected]>
Reviewed-by: Bertrand Drouvot <[email protected]>
Reviewed-by: Shi yu <[email protected]>
Reviewed-by: Ahsan Hadi <[email protected]>
Discussion: https://postgr.es/m/81D0D8B0-E7C4-4999-B616-1E5004DBDCD2%40amazon.com
Backpatch-through: 10
---
contrib/test_decoding/Makefile | 2 +-
.../expected/catalog_change_snapshot.out | 44 ++++++
.../specs/catalog_change_snapshot.spec | 39 +++++
src/backend/replication/logical/decode.c | 15 ++
src/backend/replication/logical/snapbuild.c | 143 +++++++++++++++++-
src/include/replication/snapbuild.h | 3 +
6 files changed, 238 insertions(+), 8 deletions(-)
create mode 100644 contrib/test_decoding/expected/catalog_change_snapshot.out
create mode 100644 contrib/test_decoding/specs/catalog_change_snapshot.spec
diff --git a/contrib/test_decoding/Makefile b/contrib/test_decoding/Makefile
index f439c582a5..6ec09ab192 100644
--- a/contrib/test_decoding/Makefile
+++ b/contrib/test_decoding/Makefile
@@ -7,7 +7,7 @@ REGRESS = ddl xact rewrite toast permissions decoding_in_xact \
decoding_into_rel binary prepared replorigin time messages \
spill slot truncate
ISOLATION = mxact delayed_startup ondisk_startup concurrent_ddl_dml \
- oldest_xmin snapshot_transfer subxact_without_top
+ oldest_xmin snapshot_transfer subxact_without_top catalog_change_snapshot
REGRESS_OPTS = --temp-config $(top_srcdir)/contrib/test_decoding/logical.conf
ISOLATION_OPTS = --temp-config $(top_srcdir)/contrib/test_decoding/logical.conf
diff --git a/contrib/test_decoding/expected/catalog_change_snapshot.out b/contrib/test_decoding/expected/catalog_change_snapshot.out
new file mode 100644
index 0000000000..dc4f9b7018
--- /dev/null
+++ b/contrib/test_decoding/expected/catalog_change_snapshot.out
@@ -0,0 +1,44 @@
+Parsed test spec with 2 sessions
+
+starting permutation: s0_init s0_begin s0_savepoint s0_truncate s1_checkpoint s1_get_changes s0_commit s0_begin s0_insert s1_checkpoint s1_get_changes s0_commit s1_get_changes
+step s0_init: SELECT 'init' FROM pg_create_logical_replication_slot('isolation_slot', 'test_decoding');
+?column?
+--------
+init
+(1 row)
+
+step s0_begin: BEGIN;
+step s0_savepoint: SAVEPOINT sp1;
+step s0_truncate: TRUNCATE tbl1;
+step s1_checkpoint: CHECKPOINT;
+step s1_get_changes: SELECT data FROM pg_logical_slot_get_changes('isolation_slot', NULL, NULL, 'skip-empty-xacts', '1', 'include-xids', '0');
+data
+----
+(0 rows)
+
+step s0_commit: COMMIT;
+step s0_begin: BEGIN;
+step s0_insert: INSERT INTO tbl1 VALUES (1);
+step s1_checkpoint: CHECKPOINT;
+step s1_get_changes: SELECT data FROM pg_logical_slot_get_changes('isolation_slot', NULL, NULL, 'skip-empty-xacts', '1', 'include-xids', '0');
+data
+---------------------------------------
+BEGIN
+table public.tbl1: TRUNCATE: (no-flags)
+COMMIT
+(3 rows)
+
+step s0_commit: COMMIT;
+step s1_get_changes: SELECT data FROM pg_logical_slot_get_changes('isolation_slot', NULL, NULL, 'skip-empty-xacts', '1', 'include-xids', '0');
+data
+-------------------------------------------------------------
+BEGIN
+table public.tbl1: INSERT: val1[integer]:1 val2[integer]:null
+COMMIT
+(3 rows)
+
+?column?
+--------
+stop
+(1 row)
+
diff --git a/contrib/test_decoding/specs/catalog_change_snapshot.spec b/contrib/test_decoding/specs/catalog_change_snapshot.spec
new file mode 100644
index 0000000000..662760fbcf
--- /dev/null
+++ b/contrib/test_decoding/specs/catalog_change_snapshot.spec
@@ -0,0 +1,39 @@
+# Test decoding only the commit record of the transaction that have
+# modified catalogs.
+setup
+{
+ DROP TABLE IF EXISTS tbl1;
+ CREATE TABLE tbl1 (val1 integer, val2 integer);
+}
+
+teardown
+{
+ DROP TABLE tbl1;
+ SELECT 'stop' FROM pg_drop_replication_slot('isolation_slot');
+}
+
+session "s0"
+setup { SET synchronous_commit=on; }
+step "s0_init" { SELECT 'init' FROM pg_create_logical_replication_slot('isolation_slot', 'test_decoding'); }
+step "s0_begin" { BEGIN; }
+step "s0_savepoint" { SAVEPOINT sp1; }
+step "s0_truncate" { TRUNCATE tbl1; }
+step "s0_insert" { INSERT INTO tbl1 VALUES (1); }
+step "s0_commit" { COMMIT; }
+
+session "s1"
+setup { SET synchronous_commit=on; }
+step "s1_checkpoint" { CHECKPOINT; }
+step "s1_get_changes" { SELECT data FROM pg_logical_slot_get_changes('isolation_slot', NULL, NULL, 'skip-empty-xacts', '1', 'include-xids', '0'); }
+
+# For the transaction that TRUNCATEd the table tbl1, the last decoding decodes
+# only its COMMIT record, because it starts from the RUNNING_XACT record emitted
+# during the first checkpoint execution. This transaction must be marked as
+# containing catalog changes while decoding the COMMIT record and the decoding
+# of the INSERT record must read the pg_class with the correct historic snapshot.
+#
+# Note that in a case where bgwriter wrote the XACT_RUNNING record between "s0_commit"
+# and "s0_begin", this doesn't happen as the decoding starts from the RUNNING_XACT
+# record written by bgwriter. One might think we can either stop the bgwriter or
+# increase LOG_SNAPSHOT_INTERVAL_MS but it's not practical via tests.
+permutation "s0_init" "s0_begin" "s0_savepoint" "s0_truncate" "s1_checkpoint" "s1_get_changes" "s0_commit" "s0_begin" "s0_insert" "s1_checkpoint" "s1_get_changes" "s0_commit" "s1_get_changes"
diff --git a/src/backend/replication/logical/decode.c b/src/backend/replication/logical/decode.c
index 60d07ce4eb..56a0e3e255 100644
--- a/src/backend/replication/logical/decode.c
+++ b/src/backend/replication/logical/decode.c
@@ -575,6 +575,21 @@ DecodeCommit(LogicalDecodingContext *ctx, XLogRecordBuffer *buf,
commit_time = parsed->origin_timestamp;
}
+ /*
+ * If the COMMIT record has invalidation messages, it could have catalog
+ * changes. It is possible that we didn't mark this transaction as
+ * containing catalog changes when the decoding starts from a commit record
+ * without decoding the transaction's other changes. So, we ensure to mark
+ * such transactions as containing catalog change.
+ *
+ * This must be done before SnapBuildCommitTxn() so that we can include
+ * these transactions in the historic snapshot.
+ */
+ if (parsed->xinfo & XACT_XINFO_HAS_INVALS)
+ SnapBuildXidSetCatalogChanges(ctx->snapshot_builder, xid,
+ parsed->nsubxacts, parsed->subxacts,
+ buf->origptr);
+
/*
* Process invalidation messages, even if we're not interested in the
* transaction's contents, since the various caches need to always be
diff --git a/src/backend/replication/logical/snapbuild.c b/src/backend/replication/logical/snapbuild.c
index 5a1bce5acc..9d1039f7df 100644
--- a/src/backend/replication/logical/snapbuild.c
+++ b/src/backend/replication/logical/snapbuild.c
@@ -257,8 +257,38 @@ struct SnapBuild
static ResourceOwner SavedResourceOwnerDuringExport = NULL;
static bool ExportInProgress = false;
-/* ->committed manipulation */
-static void SnapBuildPurgeCommittedTxn(SnapBuild *builder);
+/*
+ * Array of transactions and subtransactions that were running when
+ * the xl_running_xacts record that we decoded first was written.
+ * The array is sorted in xidComparator order. We remove Xids from
+ * this array when they become old enough to matter. This array is
+ * allocated in builder->context so its lifetime is the same as the
+ * snapshot builder.
+ *
+ * We rely on HEAP2_NEW_CID records and XACT_INVALIDATIONS to know
+ * if the transaction has changed the catalog, and that information
+ * is not serialized to SnapBuilder. Therefore, if the logical
+ * decoding decodes the commit record of the transaction that actually
+ * has done catalog changes without these records, we miss to add
+ * the xid to the snapshot, and end up looking at catalogs with the
+ * wrong snapshot. To avoid this problem, if the COMMIT record of
+ * the xid listed in InitialRunningXacts has XACT_XINFO_HAS_INVALS
+ * flag, we mark both the top transaction and its substransactions
+ * as containing catalog changes.
+ *
+ * We could end up adding the transaction that didn't change catalog
+ * to the snapshot since we cannot distinguish whether the transaction
+ * has catalog changes only by checking the COMMIT record. It doesn't
+ * have the information on which (sub) transaction has catalog changes,
+ * and XACT_XINFO_HAS_INVALS doesn't necessarily indicate that the
+ * transaction has catalog change. But it doesn't become a problem since
+ * we use historic snapshot only for reading system catalogs.
+ */
+static TransactionId *InitialRunningXacts = NULL;
+static int NInitialRunningXacts = 0;
+
+/* ->committed and InitailRunningXacts manipulation */
+static void SnapBuildPurgeOlderTxn(SnapBuild *builder);
/* snapshot building/manipulation/distribution functions */
static Snapshot SnapBuildBuildSnapshot(SnapBuild *builder);
@@ -895,12 +925,17 @@ SnapBuildAddCommittedTxn(SnapBuild *builder, TransactionId xid)
}
/*
- * Remove knowledge about transactions we treat as committed that are smaller
- * than ->xmin. Those won't ever get checked via the ->committed array but via
- * the clog machinery, so we don't need to waste memory on them.
+ * Remove knowledge about transactions we treat as committed and the initial
+ * running transactions that are smaller than ->xmin. Those won't ever get
+ * checked via the ->committed or InitialRunningXacts array, respectively.
+ * The committed xids will get checked via the clog machinery.
+ *
+ * We can ideally remove the transaction from InitialRunningXacts array
+ * once it is finished (committed/aborted) but that could be costly as we need
+ * to maintain the xids order in the array.
*/
static void
-SnapBuildPurgeCommittedTxn(SnapBuild *builder)
+SnapBuildPurgeOlderTxn(SnapBuild *builder)
{
int off;
TransactionId *workspace;
@@ -935,6 +970,49 @@ SnapBuildPurgeCommittedTxn(SnapBuild *builder)
builder->committed.xcnt = surviving_xids;
pfree(workspace);
+
+ /* Quick exit if there is no initial running transactions */
+ if (likely(NInitialRunningXacts == 0))
+ return;
+
+ /* bound check if there is at least one transaction to remove */
+ if (!NormalTransactionIdPrecedes(InitialRunningXacts[0],
+ builder->xmin))
+ return;
+
+ /*
+ * purge xids in InitialRunningXacts as well. The purged array must also be
+ * sorted in xidComparator order.
+ */
+ workspace =
+ MemoryContextAlloc(builder->context,
+ NInitialRunningXacts * sizeof(TransactionId));
+ surviving_xids = 0;
+ for (off = 0; off < NInitialRunningXacts; off++)
+ {
+ if (NormalTransactionIdPrecedes(InitialRunningXacts[off],
+ builder->xmin))
+ ; /* remove */
+ else
+ workspace[surviving_xids++] = InitialRunningXacts[off];
+ }
+
+ if (surviving_xids > 0)
+ memcpy(InitialRunningXacts, workspace,
+ sizeof(TransactionId) * surviving_xids);
+ else
+ {
+ pfree(InitialRunningXacts);
+ InitialRunningXacts = NULL;
+ }
+
+ elog(DEBUG3, "purged initial running transactions from %u to %u, oldest running xid %u",
+ (uint32) NInitialRunningXacts,
+ (uint32) surviving_xids,
+ builder->xmin);
+
+ NInitialRunningXacts = surviving_xids;
+ pfree(workspace);
}
/*
@@ -1120,6 +1198,21 @@ SnapBuildProcessRunningXacts(SnapBuild *builder, XLogRecPtr lsn, xl_running_xact
*/
if (builder->state < SNAPBUILD_CONSISTENT)
{
+ /*
+ * Remember the transactions and subtransactions that were running
+ * when xl_running_xacts record that we decoded first was written.
+ */
+ if (builder->state == SNAPBUILD_START)
+ {
+ int nxacts = running->subxcnt + running->xcnt;
+ Size sz = sizeof(TransactionId) * nxacts;
+
+ NInitialRunningXacts = nxacts;
+ InitialRunningXacts = MemoryContextAlloc(builder->context, sz);
+ memcpy(InitialRunningXacts, running->xids, sz);
+ qsort(InitialRunningXacts, nxacts, sizeof(TransactionId), xidComparator);
+ }
+
/* returns false if there's no point in performing cleanup just yet */
if (!SnapBuildFindSnapshot(builder, lsn, running))
return;
@@ -1142,7 +1235,7 @@ SnapBuildProcessRunningXacts(SnapBuild *builder, XLogRecPtr lsn, xl_running_xact
builder->xmin = running->oldestRunningXid;
/* Remove transactions we don't need to keep track off anymore */
- SnapBuildPurgeCommittedTxn(builder);
+ SnapBuildPurgeOlderTxn(builder);
/*
* Advance the xmin limit for the current replication slot, to allow
@@ -2035,3 +2128,39 @@ CheckPointSnapBuild(void)
}
FreeDir(snap_dir);
}
+
+/*
+ * If the given xid is in the list of the initial running xacts, we mark the
+ * transaction and its subtransactions as containing catalog changes. See
+ * comments
+ */
+void
+SnapBuildXidSetCatalogChanges(SnapBuild *builder, TransactionId xid, int subxcnt,
+ TransactionId *subxacts, XLogRecPtr lsn)
+{
+ /*
+ * Skip if there is no initial running xacts information or the
+ * transaction is already marked as containing catalog changes.
+ */
+ if (likely((NInitialRunningXacts == 0) ||
+ ReorderBufferXidHasCatalogChanges(builder->reorder, xid)))
+ return;
+
+ /*
+ * If this committed transaction is the one that was running at the time
+ * when decoding the first RUNNING_XACTS record and have done catalog
+ * changes, we can mark both the top transaction and its subtransactions
+ * as containing catalog changes.
+ */
+ if (bsearch(&xid, InitialRunningXacts, NInitialRunningXacts,
+ sizeof(TransactionId), xidComparator) != NULL)
+ {
+ ReorderBufferXidSetCatalogChanges(builder->reorder, xid, lsn);
+
+ for (int i = 0; i < subxcnt; i++)
+ {
+ ReorderBufferAssignChild(builder->reorder, xid, subxacts[i], lsn);
+ ReorderBufferXidSetCatalogChanges(builder->reorder, subxacts[i], lsn);
+ }
+ }
+}
diff --git a/src/include/replication/snapbuild.h b/src/include/replication/snapbuild.h
index 3acf68f5bd..2eb9532a1b 100644
--- a/src/include/replication/snapbuild.h
+++ b/src/include/replication/snapbuild.h
@@ -88,4 +88,7 @@ extern void SnapBuildProcessRunningXacts(SnapBuild *builder, XLogRecPtr lsn,
struct xl_running_xacts *running);
extern void SnapBuildSerializationPoint(SnapBuild *builder, XLogRecPtr lsn);
+extern void SnapBuildXidSetCatalogChanges(SnapBuild *builder, TransactionId xid,
+ int subxcnt, TransactionId *subxacts,
+ XLogRecPtr lsn);
#endif /* SNAPBUILD_H */
--
2.24.3 (Apple Git-128)
[application/octet-stream] master-v8-0001-Add-catalog-modifying-transactions-to-logical-dec.patch (27.3K, ../CAD21AoC_oEy3EvgbJ=asz20Z3ROWtRM01f-F8+F_5VjhCvJX6g@mail.gmail.com/7-master-v8-0001-Add-catalog-modifying-transactions-to-logical-dec.patch)
download | inline diff:
From 27c6daab9a4787683c60445fc6cc5dba550bdeb2 Mon Sep 17 00:00:00 2001
From: Masahiko Sawada <[email protected]>
Date: Wed, 6 Jul 2022 12:53:36 +0900
Subject: [PATCH v8] Add catalog modifying transactions to logical decoding
serialized snapshot.
Previously, we relied on HEAP2_NEW_CID records and XACT_INVALIDATION
records to know if the transaction has modified the catalog, and that
information is not serialized to snapshot. Therefore, if the logical
decoding decodes only the commit record of the transaction that
actually has modified a catalog, we missed adding its XID to the
snapshot. We ended up looking at catalogs with the wrong snapshot.
To fix this problem, this change adds the list of transaction IDs and
sub-transaction IDs, that have modified catalogs and are running when
snapshot serialization, to the serialized snapshot. When decoding a
COMMIT record, we check both the list and the ReorderBuffer to see if
the transaction has modified catalogs.
Since this adds additional information to the serialized snapshot, we
cannot backpatch it. For back branches, we take another approach;
remember the last-running-xacts list of the first decoded
RUNNING_XACTS record, check if the transaction whose commit record
has XACT_XINFO_HAS_INVALS and whose XID is in the list, and mark the
top-level transaction and sub transactions as containing catalog
changes. This doesn't require any file format changes but the
transaction will end up being added to the snapshot even if it has
only relcache invalidations. It doesn't become a problem as we use the
historical snapshot for only catalog lookups.
This commit bumps SNAPBUILD_VERSION because of change in SnapBuild.
Back-patch to all supported released.
Reported-by: Mike Oh <[email protected]>
Author: Masahiko Sawada <[email protected]>
Reviewed-by: Amit Kapila <[email protected]>
Reviewed-by: Takamichi Osumi <[email protected]>
Reviewed-by: Kyotaro Horiguchi <[email protected]>
Reviewed-by: Bertrand Drouvot <[email protected]>
Reviewed-by: Shi yu <[email protected]>
Reviewed-by: Ahsan Hadi <[email protected]>
Discussion: https://postgr.es/m/81D0D8B0-E7C4-4999-B616-1E5004DBDCD2%40amazon.com
Backpatch-through: 10
---
contrib/test_decoding/Makefile | 2 +-
.../expected/catalog_change_snapshot.out | 44 +++
.../specs/catalog_change_snapshot.spec | 39 +++
.../replication/logical/reorderbuffer.c | 71 ++++-
src/backend/replication/logical/snapbuild.c | 269 ++++++++++++------
src/include/replication/reorderbuffer.h | 12 +
6 files changed, 347 insertions(+), 90 deletions(-)
create mode 100644 contrib/test_decoding/expected/catalog_change_snapshot.out
create mode 100644 contrib/test_decoding/specs/catalog_change_snapshot.spec
diff --git a/contrib/test_decoding/Makefile b/contrib/test_decoding/Makefile
index b220906479..c7ce603706 100644
--- a/contrib/test_decoding/Makefile
+++ b/contrib/test_decoding/Makefile
@@ -8,7 +8,7 @@ REGRESS = ddl xact rewrite toast permissions decoding_in_xact \
spill slot truncate stream stats twophase twophase_stream
ISOLATION = mxact delayed_startup ondisk_startup concurrent_ddl_dml \
oldest_xmin snapshot_transfer subxact_without_top concurrent_stream \
- twophase_snapshot slot_creation_error
+ twophase_snapshot slot_creation_error catalog_change_snapshot
REGRESS_OPTS = --temp-config $(top_srcdir)/contrib/test_decoding/logical.conf
ISOLATION_OPTS = --temp-config $(top_srcdir)/contrib/test_decoding/logical.conf
diff --git a/contrib/test_decoding/expected/catalog_change_snapshot.out b/contrib/test_decoding/expected/catalog_change_snapshot.out
new file mode 100644
index 0000000000..dc4f9b7018
--- /dev/null
+++ b/contrib/test_decoding/expected/catalog_change_snapshot.out
@@ -0,0 +1,44 @@
+Parsed test spec with 2 sessions
+
+starting permutation: s0_init s0_begin s0_savepoint s0_truncate s1_checkpoint s1_get_changes s0_commit s0_begin s0_insert s1_checkpoint s1_get_changes s0_commit s1_get_changes
+step s0_init: SELECT 'init' FROM pg_create_logical_replication_slot('isolation_slot', 'test_decoding');
+?column?
+--------
+init
+(1 row)
+
+step s0_begin: BEGIN;
+step s0_savepoint: SAVEPOINT sp1;
+step s0_truncate: TRUNCATE tbl1;
+step s1_checkpoint: CHECKPOINT;
+step s1_get_changes: SELECT data FROM pg_logical_slot_get_changes('isolation_slot', NULL, NULL, 'skip-empty-xacts', '1', 'include-xids', '0');
+data
+----
+(0 rows)
+
+step s0_commit: COMMIT;
+step s0_begin: BEGIN;
+step s0_insert: INSERT INTO tbl1 VALUES (1);
+step s1_checkpoint: CHECKPOINT;
+step s1_get_changes: SELECT data FROM pg_logical_slot_get_changes('isolation_slot', NULL, NULL, 'skip-empty-xacts', '1', 'include-xids', '0');
+data
+---------------------------------------
+BEGIN
+table public.tbl1: TRUNCATE: (no-flags)
+COMMIT
+(3 rows)
+
+step s0_commit: COMMIT;
+step s1_get_changes: SELECT data FROM pg_logical_slot_get_changes('isolation_slot', NULL, NULL, 'skip-empty-xacts', '1', 'include-xids', '0');
+data
+-------------------------------------------------------------
+BEGIN
+table public.tbl1: INSERT: val1[integer]:1 val2[integer]:null
+COMMIT
+(3 rows)
+
+?column?
+--------
+stop
+(1 row)
+
diff --git a/contrib/test_decoding/specs/catalog_change_snapshot.spec b/contrib/test_decoding/specs/catalog_change_snapshot.spec
new file mode 100644
index 0000000000..2971ddc69c
--- /dev/null
+++ b/contrib/test_decoding/specs/catalog_change_snapshot.spec
@@ -0,0 +1,39 @@
+# Test decoding only the commit record of the transaction that have
+# modified catalogs.
+setup
+{
+ DROP TABLE IF EXISTS tbl1;
+ CREATE TABLE tbl1 (val1 integer, val2 integer);
+}
+
+teardown
+{
+ DROP TABLE tbl1;
+ SELECT 'stop' FROM pg_drop_replication_slot('isolation_slot');
+}
+
+session "s0"
+setup { SET synchronous_commit=on; }
+step "s0_init" { SELECT 'init' FROM pg_create_logical_replication_slot('isolation_slot', 'test_decoding'); }
+step "s0_begin" { BEGIN; }
+step "s0_savepoint" { SAVEPOINT sp1; }
+step "s0_truncate" { TRUNCATE tbl1; }
+step "s0_insert" { INSERT INTO tbl1 VALUES (1); }
+step "s0_commit" { COMMIT; }
+
+session "s1"
+setup { SET synchronous_commit=on; }
+step "s1_checkpoint" { CHECKPOINT; }
+step "s1_get_changes" { SELECT data FROM pg_logical_slot_get_changes('isolation_slot', NULL, NULL, 'skip-empty-xacts', '1', 'include-xids', '0'); }
+
+# For the transaction that TRUNCATEd the table tbl1, the last decoding decodes
+# only its COMMIT record, because it starts from the RUNNING_XACTS record emitted
+# during the first checkpoint execution. This transaction must be marked as
+# containing catalog changes while decoding the COMMIT record and the decoding
+# of the INSERT record must read the pg_class with the correct historic snapshot.
+#
+# Note that in a case where bgwriter wrote the RUNNING_XACTS record between "s0_commit"
+# and "s0_begin", this doesn't happen as the decoding starts from the RUNNING_XACTS
+# record written by bgwriter. One might think we can either stop the bgwriter or
+# increase LOG_SNAPSHOT_INTERVAL_MS but it's not practical via tests.
+permutation "s0_init" "s0_begin" "s0_savepoint" "s0_truncate" "s1_checkpoint" "s1_get_changes" "s0_commit" "s0_begin" "s0_insert" "s1_checkpoint" "s1_get_changes" "s0_commit" "s1_get_changes"
diff --git a/src/backend/replication/logical/reorderbuffer.c b/src/backend/replication/logical/reorderbuffer.c
index 88a37fde72..0b2d9b7930 100644
--- a/src/backend/replication/logical/reorderbuffer.c
+++ b/src/backend/replication/logical/reorderbuffer.c
@@ -349,6 +349,8 @@ ReorderBufferAllocate(void)
buffer->by_txn_last_xid = InvalidTransactionId;
buffer->by_txn_last_txn = NULL;
+ buffer->catchange_ntxns = 0;
+
buffer->outbuf = NULL;
buffer->outbufsize = 0;
buffer->size = 0;
@@ -366,6 +368,7 @@ ReorderBufferAllocate(void)
dlist_init(&buffer->toplevel_by_lsn);
dlist_init(&buffer->txns_by_base_snapshot_lsn);
+ dlist_init(&buffer->catchange_txns);
/*
* Ensure there's no stale data from prior uses of this slot, in case some
@@ -1526,14 +1529,22 @@ ReorderBufferCleanupTXN(ReorderBuffer *rb, ReorderBufferTXN *txn)
}
/*
- * Remove TXN from its containing list.
+ * Remove TXN from its containing lists.
*
* Note: if txn is known as subxact, we are deleting the TXN from its
* parent's list of known subxacts; this leaves the parent's nsubxacts
* count too high, but we don't care. Otherwise, we are deleting the TXN
- * from the LSN-ordered list of toplevel TXNs.
+ * from the LSN-ordered list of toplevel TXNs. We remove TXN from
+ * the list of catalog modifying transactions as well.
*/
dlist_delete(&txn->node);
+ if (rbtxn_has_catalog_changes(txn))
+ {
+ dlist_delete(&txn->catchange_node);
+ rb->catchange_ntxns--;
+
+ Assert(rb->catchange_ntxns >= 0);
+ }
/* now remove reference from buffer */
hash_search(rb->by_txn,
@@ -3275,10 +3286,16 @@ ReorderBufferXidSetCatalogChanges(ReorderBuffer *rb, TransactionId xid,
XLogRecPtr lsn)
{
ReorderBufferTXN *txn;
+ ReorderBufferTXN *toptxn;
txn = ReorderBufferTXNByXid(rb, xid, true, NULL, lsn, true);
- txn->txn_flags |= RBTXN_HAS_CATALOG_CHANGES;
+ if (!rbtxn_has_catalog_changes(txn))
+ {
+ txn->txn_flags |= RBTXN_HAS_CATALOG_CHANGES;
+ dlist_push_tail(&rb->catchange_txns, &txn->catchange_node);
+ rb->catchange_ntxns++;
+ }
/*
* Mark top-level transaction as having catalog changes too if one of its
@@ -3286,8 +3303,52 @@ ReorderBufferXidSetCatalogChanges(ReorderBuffer *rb, TransactionId xid,
* conveniently check just top-level transaction and decide whether to
* build the hash table or not.
*/
- if (txn->toptxn != NULL)
- txn->toptxn->txn_flags |= RBTXN_HAS_CATALOG_CHANGES;
+ toptxn = txn->toptxn;
+ if (toptxn != NULL && !rbtxn_has_catalog_changes(toptxn))
+ {
+ toptxn->txn_flags |= RBTXN_HAS_CATALOG_CHANGES;
+ dlist_push_tail(&rb->catchange_txns, &toptxn->catchange_node);
+ rb->catchange_ntxns++;
+ }
+}
+
+/*
+ * Return palloc'ed array of the transactions that have changed catalogs.
+ * The returned array is sorted in xidComparator order.
+ *
+ * The caller must free the returned array when done with it.
+ */
+TransactionId *
+ReorderBufferGetCatalogChangesXacts(ReorderBuffer *rb)
+{
+ dlist_iter iter;
+ TransactionId *xids = NULL;
+ size_t xcnt = 0;
+
+ /* Quick return if the list is empty */
+ if (dlist_is_empty(&rb->catchange_txns))
+ {
+ Assert(rb->catchange_ntxns == 0);
+ return NULL;
+ }
+
+ /* Initialize XID array */
+ xids = (TransactionId *) palloc(sizeof(TransactionId) * rb->catchange_ntxns);
+ dlist_foreach(iter, &rb->catchange_txns)
+ {
+ ReorderBufferTXN *txn = dlist_container(ReorderBufferTXN,
+ catchange_node,
+ iter.cur);
+
+ Assert(rbtxn_has_catalog_changes(txn));
+
+ xids[xcnt++] = txn->xid;
+ }
+
+ qsort(xids, xcnt, sizeof(TransactionId), xidComparator);
+
+ Assert(xcnt == rb->catchange_ntxns);
+ return xids;
}
/*
diff --git a/src/backend/replication/logical/snapbuild.c b/src/backend/replication/logical/snapbuild.c
index 73c0f15214..4f766866b2 100644
--- a/src/backend/replication/logical/snapbuild.c
+++ b/src/backend/replication/logical/snapbuild.c
@@ -241,6 +241,30 @@ struct SnapBuild
*/
TransactionId *xip;
} committed;
+
+ /*
+ * Array of transactions and subtransactions that had modified catalogs
+ * and were running when the snapshot was serialized.
+ *
+ * We normally rely on HEAP2_NEW_CID and XLOG_XACT_INVALIDATIONS records to
+ * know if the transaction has changed the catalog. But it could happen that
+ * the logical decoding decodes only the commit record of the transaction.
+ * This array stores the transactions that have modified catalogs and were
+ * running when serializing a snapshot, and this array is used to add such
+ * transactions to the snapshot.
+ *
+ * This array is set once when restoring the snapshot, we remove xids from
+ * this array when they become old enough to matter, and then it eventually
+ * becomes empty.
+ */
+ struct
+ {
+ /* number of transactions */
+ size_t xcnt;
+
+ /* This array must be sorted in xidComparator order */
+ TransactionId *xip;
+ } catchange;
};
/*
@@ -250,8 +274,8 @@ struct SnapBuild
static ResourceOwner SavedResourceOwnerDuringExport = NULL;
static bool ExportInProgress = false;
-/* ->committed manipulation */
-static void SnapBuildPurgeCommittedTxn(SnapBuild *builder);
+/* ->committed and ->catchange manipulation */
+static void SnapBuildPurgeOlderTxn(SnapBuild *builder);
/* snapshot building/manipulation/distribution functions */
static Snapshot SnapBuildBuildSnapshot(SnapBuild *builder);
@@ -262,6 +286,8 @@ static void SnapBuildSnapIncRefcount(Snapshot snap);
static void SnapBuildDistributeNewCatalogSnapshot(SnapBuild *builder, XLogRecPtr lsn);
+static inline bool SnapBuildXidHasCatalogChanges(SnapBuild *builder, TransactionId xid);
+
/* xlog reading helper functions for SnapBuildProcessRunningXacts */
static bool SnapBuildFindSnapshot(SnapBuild *builder, XLogRecPtr lsn, xl_running_xacts *running);
static void SnapBuildWaitSnapshot(xl_running_xacts *running, TransactionId cutoff);
@@ -269,6 +295,7 @@ static void SnapBuildWaitSnapshot(xl_running_xacts *running, TransactionId cutof
/* serialization functions */
static void SnapBuildSerialize(SnapBuild *builder, XLogRecPtr lsn);
static bool SnapBuildRestore(SnapBuild *builder, XLogRecPtr lsn);
+static void SnapBuildRestoreContents(int fd, char *dest, Size size, const char *path);
/*
* Allocate a new snapshot builder.
@@ -306,6 +333,9 @@ AllocateSnapshotBuilder(ReorderBuffer *reorder,
palloc0(builder->committed.xcnt_space * sizeof(TransactionId));
builder->committed.includes_all_transactions = true;
+ builder->catchange.xcnt = 0;
+ builder->catchange.xip = NULL;
+
builder->initial_xmin_horizon = xmin_horizon;
builder->start_decoding_at = start_lsn;
builder->building_full_snapshot = need_full_snapshot;
@@ -888,12 +918,17 @@ SnapBuildAddCommittedTxn(SnapBuild *builder, TransactionId xid)
}
/*
- * Remove knowledge about transactions we treat as committed that are smaller
- * than ->xmin. Those won't ever get checked via the ->committed array but via
- * the clog machinery, so we don't need to waste memory on them.
+ * Remove knowledge about transactions we treat as committed or containing catalog
+ * changes that are smaller than ->xmin. Those won't ever get checked via
+ * the ->committed or ->catchange array, respectively. The committed xids will
+ * get checked via the clog machinery.
+ *
+ * We can ideally remove the transaction
+ * from catchange array once it is finished (committed/aborted) but that could
+ * be costly as we need to maintain the xids order in the array.
*/
static void
-SnapBuildPurgeCommittedTxn(SnapBuild *builder)
+SnapBuildPurgeOlderTxn(SnapBuild *builder)
{
int off;
TransactionId *workspace;
@@ -928,6 +963,40 @@ SnapBuildPurgeCommittedTxn(SnapBuild *builder)
builder->committed.xcnt = surviving_xids;
pfree(workspace);
+
+ /*
+ * purge xids in ->catchange as well. The purged array must also be
+ * sorted in xidComparator order.
+ */
+ if (builder->catchange.xcnt > 0)
+ {
+ /*
+ * Since catchange.xip is sorted, we find the lower bound of
+ * xids that still are interesting.
+ */
+ for (off = 0; off < builder->catchange.xcnt; off++)
+ {
+ if (TransactionIdFollowsOrEquals(builder->catchange.xip[off],
+ builder->xmin))
+ break;
+ }
+
+ surviving_xids = builder->catchange.xcnt - off;
+ if (surviving_xids > 0)
+ memmove(builder->catchange.xip, &(builder->catchange.xip[off]),
+ surviving_xids * sizeof(TransactionId));
+ else
+ {
+ /* catchange list becomes empty */
+ pfree(builder->catchange.xip);
+ builder->catchange.xip = NULL;
+ }
+
+ elog(DEBUG3, "purged catalog modifying transactions from %u to %u, xmin %u",
+ (uint32) builder->catchange.xcnt, (uint32) surviving_xids,
+ builder->xmin);
+ builder->catchange.xcnt = surviving_xids;
+ }
}
/*
@@ -983,7 +1052,7 @@ SnapBuildCommitTxn(SnapBuild *builder, XLogRecPtr lsn, TransactionId xid,
* Add subtransaction to base snapshot if catalog modifying, we don't
* distinguish to toplevel transactions there.
*/
- if (ReorderBufferXidHasCatalogChanges(builder->reorder, subxid))
+ if (SnapBuildXidHasCatalogChanges(builder, subxid))
{
sub_needs_timetravel = true;
needs_snapshot = true;
@@ -1012,7 +1081,7 @@ SnapBuildCommitTxn(SnapBuild *builder, XLogRecPtr lsn, TransactionId xid,
}
/* if top-level modified catalog, it'll need a snapshot */
- if (ReorderBufferXidHasCatalogChanges(builder->reorder, xid))
+ if (SnapBuildXidHasCatalogChanges(builder, xid))
{
elog(DEBUG2, "found top level transaction %u, with catalog changes",
xid);
@@ -1089,6 +1158,21 @@ SnapBuildCommitTxn(SnapBuild *builder, XLogRecPtr lsn, TransactionId xid,
}
}
+/*
+ * Check the reorder buffer and the snapshot to see if the given transaction has
+ * modified catalogs.
+ */
+static inline bool
+SnapBuildXidHasCatalogChanges(SnapBuild *builder, TransactionId xid)
+{
+ if (ReorderBufferXidHasCatalogChanges(builder->reorder, xid))
+ return true;
+
+ /* Check the catchange XID array */
+ return ((builder->catchange.xcnt > 0) &&
+ (bsearch(&xid, builder->catchange.xip, builder->catchange.xcnt,
+ sizeof(TransactionId), xidComparator) != NULL));
+}
/* -----------------------------------
* Snapshot building functions dealing with xlog records
@@ -1135,7 +1219,7 @@ SnapBuildProcessRunningXacts(SnapBuild *builder, XLogRecPtr lsn, xl_running_xact
builder->xmin = running->oldestRunningXid;
/* Remove transactions we don't need to keep track off anymore */
- SnapBuildPurgeCommittedTxn(builder);
+ SnapBuildPurgeOlderTxn(builder);
/*
* Advance the xmin limit for the current replication slot, to allow
@@ -1438,6 +1522,7 @@ SnapBuildWaitSnapshot(xl_running_xacts *running, TransactionId cutoff)
*
* struct SnapBuildOnDisk;
* TransactionId * committed.xcnt; (*not xcnt_space*)
+ * TransactionId * catchange.xcnt;
*
*/
typedef struct SnapBuildOnDisk
@@ -1467,7 +1552,7 @@ typedef struct SnapBuildOnDisk
offsetof(SnapBuildOnDisk, version)
#define SNAPBUILD_MAGIC 0x51A1E001
-#define SNAPBUILD_VERSION 4
+#define SNAPBUILD_VERSION 5
/*
* Store/Load a snapshot from disk, depending on the snapshot builder's state.
@@ -1493,6 +1578,9 @@ SnapBuildSerialize(SnapBuild *builder, XLogRecPtr lsn)
{
Size needed_length;
SnapBuildOnDisk *ondisk = NULL;
+ TransactionId *catchange_xip = NULL;
+ MemoryContext old_ctx;
+ size_t catchange_xcnt;
char *ondisk_c;
int fd;
char tmppath[MAXPGPATH];
@@ -1578,10 +1666,16 @@ SnapBuildSerialize(SnapBuild *builder, XLogRecPtr lsn)
(errcode_for_file_access(),
errmsg("could not remove file \"%s\": %m", tmppath)));
+ old_ctx = MemoryContextSwitchTo(builder->context);
+
+ /* Get the catalog modifying transactions that are yet not committed */
+ catchange_xip = ReorderBufferGetCatalogChangesXacts(builder->reorder);
+ catchange_xcnt = builder->reorder->catchange_ntxns;
+
needed_length = sizeof(SnapBuildOnDisk) +
- sizeof(TransactionId) * builder->committed.xcnt;
+ sizeof(TransactionId) * (builder->committed.xcnt + catchange_xcnt);
- ondisk_c = MemoryContextAllocZero(builder->context, needed_length);
+ ondisk_c = palloc0(needed_length);
ondisk = (SnapBuildOnDisk *) ondisk_c;
ondisk->magic = SNAPBUILD_MAGIC;
ondisk->version = SNAPBUILD_VERSION;
@@ -1598,16 +1692,31 @@ SnapBuildSerialize(SnapBuild *builder, XLogRecPtr lsn)
ondisk->builder.snapshot = NULL;
ondisk->builder.reorder = NULL;
ondisk->builder.committed.xip = NULL;
+ ondisk->builder.catchange.xip = NULL;
+ /* update catchange only on disk data */
+ ondisk->builder.catchange.xcnt = catchange_xcnt;
COMP_CRC32C(ondisk->checksum,
&ondisk->builder,
sizeof(SnapBuild));
/* copy committed xacts */
- sz = sizeof(TransactionId) * builder->committed.xcnt;
- memcpy(ondisk_c, builder->committed.xip, sz);
- COMP_CRC32C(ondisk->checksum, ondisk_c, sz);
- ondisk_c += sz;
+ if (builder->committed.xcnt > 0)
+ {
+ sz = sizeof(TransactionId) * builder->committed.xcnt;
+ memcpy(ondisk_c, builder->committed.xip, sz);
+ COMP_CRC32C(ondisk->checksum, ondisk_c, sz);
+ ondisk_c += sz;
+ }
+
+ /* copy catalog modifying xacts */
+ if (catchange_xcnt > 0)
+ {
+ sz = sizeof(TransactionId) * catchange_xcnt;
+ memcpy(ondisk_c, catchange_xip, sz);
+ COMP_CRC32C(ondisk->checksum, ondisk_c, sz);
+ ondisk_c += sz;
+ }
FIN_CRC32C(ondisk->checksum);
@@ -1688,12 +1797,16 @@ SnapBuildSerialize(SnapBuild *builder, XLogRecPtr lsn)
*/
builder->last_serialized_snapshot = lsn;
+ MemoryContextSwitchTo(old_ctx);
+
out:
ReorderBufferSetRestartPoint(builder->reorder,
builder->last_serialized_snapshot);
/* be tidy */
if (ondisk)
pfree(ondisk);
+ if (catchange_xip)
+ pfree(catchange_xip);
}
/*
@@ -1707,7 +1820,6 @@ SnapBuildRestore(SnapBuild *builder, XLogRecPtr lsn)
int fd;
char path[MAXPGPATH];
Size sz;
- int readBytes;
pg_crc32c checksum;
/* no point in loading a snapshot if we're already there */
@@ -1739,29 +1851,7 @@ SnapBuildRestore(SnapBuild *builder, XLogRecPtr lsn)
/* read statically sized portion of snapshot */
- pgstat_report_wait_start(WAIT_EVENT_SNAPBUILD_READ);
- readBytes = read(fd, &ondisk, SnapBuildOnDiskConstantSize);
- pgstat_report_wait_end();
- if (readBytes != SnapBuildOnDiskConstantSize)
- {
- int save_errno = errno;
-
- CloseTransientFile(fd);
-
- if (readBytes < 0)
- {
- errno = save_errno;
- ereport(ERROR,
- (errcode_for_file_access(),
- errmsg("could not read file \"%s\": %m", path)));
- }
- else
- ereport(ERROR,
- (errcode(ERRCODE_DATA_CORRUPTED),
- errmsg("could not read file \"%s\": read %d of %zu",
- path, readBytes,
- (Size) SnapBuildOnDiskConstantSize)));
- }
+ SnapBuildRestoreContents(fd, (char *) &ondisk, SnapBuildOnDiskConstantSize, path);
if (ondisk.magic != SNAPBUILD_MAGIC)
ereport(ERROR,
@@ -1781,56 +1871,26 @@ SnapBuildRestore(SnapBuild *builder, XLogRecPtr lsn)
SnapBuildOnDiskConstantSize - SnapBuildOnDiskNotChecksummedSize);
/* read SnapBuild */
- pgstat_report_wait_start(WAIT_EVENT_SNAPBUILD_READ);
- readBytes = read(fd, &ondisk.builder, sizeof(SnapBuild));
- pgstat_report_wait_end();
- if (readBytes != sizeof(SnapBuild))
- {
- int save_errno = errno;
-
- CloseTransientFile(fd);
-
- if (readBytes < 0)
- {
- errno = save_errno;
- ereport(ERROR,
- (errcode_for_file_access(),
- errmsg("could not read file \"%s\": %m", path)));
- }
- else
- ereport(ERROR,
- (errcode(ERRCODE_DATA_CORRUPTED),
- errmsg("could not read file \"%s\": read %d of %zu",
- path, readBytes, sizeof(SnapBuild))));
- }
+ SnapBuildRestoreContents(fd, (char *) &ondisk.builder, sizeof(SnapBuild), path);
COMP_CRC32C(checksum, &ondisk.builder, sizeof(SnapBuild));
/* restore committed xacts information */
- sz = sizeof(TransactionId) * ondisk.builder.committed.xcnt;
- ondisk.builder.committed.xip = MemoryContextAllocZero(builder->context, sz);
- pgstat_report_wait_start(WAIT_EVENT_SNAPBUILD_READ);
- readBytes = read(fd, ondisk.builder.committed.xip, sz);
- pgstat_report_wait_end();
- if (readBytes != sz)
+ if (ondisk.builder.committed.xcnt > 0)
{
- int save_errno = errno;
-
- CloseTransientFile(fd);
+ sz = sizeof(TransactionId) * ondisk.builder.committed.xcnt;
+ ondisk.builder.committed.xip = MemoryContextAllocZero(builder->context, sz);
+ SnapBuildRestoreContents(fd, (char *) ondisk.builder.committed.xip, sz, path);
+ COMP_CRC32C(checksum, ondisk.builder.committed.xip, sz);
+ }
- if (readBytes < 0)
- {
- errno = save_errno;
- ereport(ERROR,
- (errcode_for_file_access(),
- errmsg("could not read file \"%s\": %m", path)));
- }
- else
- ereport(ERROR,
- (errcode(ERRCODE_DATA_CORRUPTED),
- errmsg("could not read file \"%s\": read %d of %zu",
- path, readBytes, sz)));
+ /* restore catalog modifying xacts information */
+ if (ondisk.builder.catchange.xcnt > 0)
+ {
+ sz = sizeof(TransactionId) * ondisk.builder.catchange.xcnt;
+ ondisk.builder.catchange.xip = MemoryContextAllocZero(builder->context, sz);
+ SnapBuildRestoreContents(fd, (char *) ondisk.builder.catchange.xip, sz, path);
+ COMP_CRC32C(checksum, ondisk.builder.catchange.xip, sz);
}
- COMP_CRC32C(checksum, ondisk.builder.committed.xip, sz);
if (CloseTransientFile(fd) != 0)
ereport(ERROR,
@@ -1885,6 +1945,13 @@ SnapBuildRestore(SnapBuild *builder, XLogRecPtr lsn)
}
ondisk.builder.committed.xip = NULL;
+ /* set catalog modifying transactions */
+ if (builder->catchange.xip)
+ pfree(builder->catchange.xip);
+ builder->catchange.xcnt = ondisk.builder.catchange.xcnt;
+ builder->catchange.xip = ondisk.builder.catchange.xip;
+ ondisk.builder.catchange.xip = NULL;
+
/* our snapshot is not interesting anymore, build a new one */
if (builder->snapshot != NULL)
{
@@ -1906,9 +1973,43 @@ SnapBuildRestore(SnapBuild *builder, XLogRecPtr lsn)
snapshot_not_interesting:
if (ondisk.builder.committed.xip != NULL)
pfree(ondisk.builder.committed.xip);
+ if (ondisk.builder.catchange.xip != NULL)
+ pfree(ondisk.builder.catchange.xip);
return false;
}
+/*
+ * Read the contents of the serialized snapshot to the dest.
+ */
+static void
+SnapBuildRestoreContents(int fd, char *dest, Size size, const char *path)
+{
+ int readBytes;
+
+ pgstat_report_wait_start(WAIT_EVENT_SNAPBUILD_READ);
+ readBytes = read(fd, dest, size);
+ pgstat_report_wait_end();
+ if (readBytes != size)
+ {
+ int save_errno = errno;
+
+ CloseTransientFile(fd);
+
+ if (readBytes < 0)
+ {
+ errno = save_errno;
+ ereport(ERROR,
+ (errcode_for_file_access(),
+ errmsg("could not read file \"%s\": %m", path)));
+ }
+ else
+ ereport(ERROR,
+ (errcode(ERRCODE_DATA_CORRUPTED),
+ errmsg("could not read file \"%s\": read %d of %zu",
+ path, readBytes, sizeof(SnapBuild))));
+ }
+}
+
/*
* Remove all serialized snapshots that are not required anymore because no
* slot can need them. This doesn't actually have to run during a checkpoint,
diff --git a/src/include/replication/reorderbuffer.h b/src/include/replication/reorderbuffer.h
index d109d0baed..fd84f175c0 100644
--- a/src/include/replication/reorderbuffer.h
+++ b/src/include/replication/reorderbuffer.h
@@ -380,6 +380,11 @@ typedef struct ReorderBufferTXN
*/
dlist_node node;
+ /*
+ * A node in the list of catalog modifying transactions
+ */
+ dlist_node catchange_node;
+
/*
* Size of this transaction (changes currently in memory, in bytes).
*/
@@ -526,6 +531,12 @@ struct ReorderBuffer
*/
dlist_head txns_by_base_snapshot_lsn;
+ /*
+ * Transactions and subtransactions that have modified system catalogs.
+ */
+ dlist_head catchange_txns;
+ int catchange_ntxns;
+
/*
* one-entry sized cache for by_txn. Very frequently the same txn gets
* looked up over and over again.
@@ -677,6 +688,7 @@ extern void ReorderBufferSkipPrepare(ReorderBuffer *rb, TransactionId xid);
extern void ReorderBufferPrepare(ReorderBuffer *rb, TransactionId xid, char *gid);
extern ReorderBufferTXN *ReorderBufferGetOldestTXN(ReorderBuffer *);
extern TransactionId ReorderBufferGetOldestXmin(ReorderBuffer *rb);
+extern TransactionId *ReorderBufferGetCatalogChangesXacts(ReorderBuffer *rb);
extern void ReorderBufferSetRestartPoint(ReorderBuffer *, XLogRecPtr ptr);
--
2.24.3 (Apple Git-128)
view thread (136+ messages) latest in thread
reply
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Reply to all the recipients using the --to and --cc options:
reply via email
To: [email protected]
Cc: [email protected], [email protected], [email protected], [email protected], [email protected], [email protected], [email protected]
Subject: Re: [BUG] Logical replication failure "ERROR: could not map filenode "base/13237/442428" to relation OID" with catalog modifying txns
In-Reply-To: <CAD21AoC_oEy3EvgbJ=asz20Z3ROWtRM01f-F8+F_5VjhCvJX6g@mail.gmail.com>
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
This inbox is served by agora; see mirroring instructions
for how to clone and mirror all data and code used for this inbox