Yes, it is still reproducible. When I try to VACUUM FULL right now, I get the same error.
This is postgres 9.4. I am new to gdb, but based on what you said, I have the following output:Breakpoint 1, 0x0000000000768be0 in errfinish ()
(gdb) bt
#0 0x0000000000768be0 in errfinish ()
#1 0x000000000076998c in elog_finish ()
#2 0x0000000000495960 in ?? ()
#3 0x0000000000496075 in heap_tuple_fetch_attr ()
#4 0x0000000000496572 in toast_insert_or_update ()
#5 0x0000000000492ce1 in ?? ()
#6 0x0000000000493733 in rewrite_heap_tuple ()
#7 0x000000000053ebdf in ?? ()
#8 0x000000000053f68e in cluster_rel ()
#9 0x0000000000590a1b in ?? ()
#10 0x0000000000590f9f in vacuum ()
#11 0x000000000068ee77 in standard_ProcessUtility ()
#12 0x00007f50c6b583d4 in ?? () from /usr/pgsql-9.4/lib/pglogical.so
#13 0x000000000068bb17 in ?? ()
#14 0x000000000068caad in ?? ()
#15 0x000000000068d132 in PortalRun ()
#16 0x000000000068979e in ?? ()
#17 0x000000000068ae18 in PostgresMain ()
#18 0x0000000000635d69 in PostmasterMain ()
#19 0x00000000005cd248 in main ()
(gdb) cont
On Tue, Oct 31, 2017 at 3:59 PM, Tom Lane <tgl@xxxxxxxxxxxxx> wrote:
Arjun Ranade <ranade@xxxxxxxxxxxxxxxxx> writes:
> We had some downtime recently and I thought it would be a good idea to do a
> periodic VACUUM FULL of one of our large Postgres DB's.
> However, when I tried to attempt the VACUUM FULL, I saw the following error:
> INFO: vacuuming "pg_catalog.pg_statistic"
> vacuumdb: vacuuming of database "db1" failed: ERROR: missing chunk number
> 0 for toast value 30382746 in pg_toast_2619
We hear reports like this just often enough to make it seem like there's
some bug that afflicts pg_statistic specifically. Nobody's ever found
a cause though.
> Given that pg_statistic is inessential data (it can be rebuilt by analyzing
> each table), I did a 'DELETE FROM pg_statistic;' which removed all the
> rows.
That would've been my advice ...
> However, when I ran the VACUUM FULL again, I received the same error.
... but that's really interesting. Is this still reproducible? If you
could get a stack trace from the point of the error, that might yield
useful data. (Set a gdb breakpoint at "errfinish", run the VACUUM FULL,
and when it stops, get the stack with "bt". Don't use VERBOSE, or you'll
reach errfinish for each line of verbose output...)
Also, what PG version is this exactly?
regards, tom lane
