Re: PG10.1 autovac killed building extended stats

On Fri, Nov 17, 2017 at 03:43:52PM -0600, Justin Pryzby wrote:
> On Fri, Nov 17, 2017 at 01:36:00PM -0300, Alvaro Herrera wrote:
> > Justin Pryzby wrote:
> >
> > > Core was generated by `postgres: autovacuum worker process gtt '.
> > > Program terminated with signal 11, Segmentation fault.
> > > #0 statext_ndistinct_build (totalrows=300, numrows=300, rows=0x21df3e8, attrs=<value optimized out>, stats=0x0) at mvdistinct.c:103
> > > 103 item->attrs = bms_add_member(item->attrs,
> > >
> > > (gdb) p stats
> > > $5 = (VacAttrStats **) 0x0
> > > => This is an error, no ??
> >
> > Not necessarily, but then I think this previous code is busted:
>
> > If I recall things correctly, the "continue" should be executed
> > regardless of IsAutoVacuumWorkerProcess() (only the log should be
> > conditional on that). I'm not sure how I arrived at the current coding,
> > which I added in bf2a691e02d7 -- probably just fuzzy thinking.
>
> Is it useful to run with that change ?

FYI: the affected server ran without crashing since Nov 18 with this change.

--- postgresql-10.1.orig/src/backend/statistics/extended_stats.c 2017-11-06 20:46:52.000000000 -0400
+++ postgresql-10.1/src/backend/statistics/extended_stats.c 2017-11-18 18:23:44.057988419 -0400
@@ -98,20 +98,21 @@
if (!stats && !IsAutoVacuumWorkerProcess())
{
ereport(WARNING,
(errcode(ERRCODE_INVALID_OBJECT_DEFINITION),
errmsg("statistics object \"%s.%s\" could not be computed for relation \"%s.%s\"",
stat->schema, stat->name,
get_namespace_name(onerel->rd_rel->relnamespace),
RelationGetRelationName(onerel)),
errtable(onerel)));
- continue;
}
+ if (!stats)
+ continue;

/* check allowed number of dimensions */
Assert(bms_num_members(stat->columns) >= 2 &&
bms_num_members(stat->columns) <= STATS_MAX_DIMENSIONS);

/* compute statistic of each requested type */
foreach(lc2, stat->types)
{
char t = (char) lfirst_int(lc2);

Justin