Craig Ringer <craig@postnewspapers.com.au> writes:
> After turning autovacuum off completely, though, it does crash when
> ANALYZE is run.
>> postgres.exe!pfree(void * pointer=0x68f08610) Line 591 + 0x3 bytes C
>> postgres.exe!examine_attribute(RelationData * onerel=0x00000000, int attnum=5, Node * index_expr=0x00000000) Line
877 C
>> postgres.exe!do_analyze_rel(RelationData * onerel=0x01747b48, VacuumStmt * vacstmt=0x01690580, char
update_reltuples='',char inh=0) Line 357 + 0xa bytes C
Hmm. That is suspiciously close to the location of some last-minute
changes in Postgres 9.0. I wonder whether Andrea is using a version of
PostGIS that was compiled against pre-9.0RC1 Postgres sources. If they
weren't accounting for this patch:
http://git.postgresql.org/gitweb?p=postgresql.git;a=commitdiff;h=67becf8d41a082eaaf6db6e0860d49409b79e32b
then we could easily have a crash right about here --- in fact it looks
like this is exactly what you'd get, because the extension would think
that the compute_stats field is where attrtype now is, so the
"pfree(stats->attrtype)" would be trying to pfree a function address.
In short, what we've got here is a version skew problem. That doubtless
explains why Craig couldn't duplicate it on his Linux machine.
regards, tom lane