ISTM that heap_compute_xid_horizon_for_tuples() calculates
latestRemovedXid for index deletion callers without sufficient care.
The function only follows line pointer redirects, which is necessary
but not sufficient to visit all relevant heap tuple headers -- it also
needs to traverse HOT chains, but that doesn't happen. AFAICT
heap_compute_xid_horizon_for_tuples() might therefore fail to produce
a sufficiently recent latestRemovedXid value for the index deletion
operation as a whole. This might in turn lead to the REDO routine
(e.g. btree_xlog_delete()) doing conflict processing incorrectly
during hot standby.
Attached is an instrumentation patch. If I run "make check" with the
patch applied, I get test output failures that can be used to get a
general sense of the problem:
$ cat /code/postgresql/patch/build/src/test/regress/regression.diffs |
grep "works okay this time" | wc -l
382
$ cat /code/postgresql/patch/build/src/test/regress/regression.diffs |
grep "hot chain bug"
+WARNING: hot chain bug, latestRemovedXid: 2307,
latestRemovedXidWithHotChain: 2316
+WARNING: hot chain bug, latestRemovedXid: 4468,
latestRemovedXidWithHotChain: 4538
+WARNING: hot chain bug, latestRemovedXid: 4756,
latestRemovedXidWithHotChain: 4809
+WARNING: hot chain bug, latestRemovedXid: 5000,
latestRemovedXidWithHotChain: 5001
+WARNING: hot chain bug, latestRemovedXid: 7683,
latestRemovedXidWithHotChain: 7995
+WARNING: hot chain bug, latestRemovedXid: 13450,
latestRemovedXidWithHotChain: 13453
+WARNING: hot chain bug, latestRemovedXid: 10040,
latestRemovedXidWithHotChain: 10041
So out of 389 calls, we see 7 failures on this occasion, which is
typical. Heap pruning usually saves us in practice (since it is highly
correlated with setting LP_DEAD bits on index pages in the first
place), and even when it doesn't it's not particularly likely that the
issue will make the crucial difference for the deletion operation as a
whole.
The code that is now heap_compute_xid_horizon_for_tuples() ran in REDO
routines directly prior to Postgres 12.
heap_compute_xid_horizon_for_tuples() is a descendant of code added by
Simon’s commit a760893d in 2010 -- pretty close to HOT’s initial
introduction. So this has been around for a long time.
--
Peter Geoghegan
