Mike Williams <mike.williams@comodo.com> writes:
> We upgraded from 8.3.5 to 8.3.9 yesterday, after an extended period of
> testing, to 8.3.10 today in hope of fixing it, and we're getting the following
> error on occasion.
> ERROR: could not open segment 1 of relation 1663/743352/743420 (target block
> 2171336): No such file or directory
> CONTEXT: SQL statement "SELECT day FROM min_stats WHERE day<= $1 AND
> unique_ip=-1 ORDER BY day DESC LIMIT 1"
> PL/pgSQL function "update_unique_ips" line 13 at SQL statement
> update_unique_ips() is called once per minute, but we get anywhere from one
> error per minute, to one error per hour.
> Relation 743420 in database 743352 is 441270272 bytes long and growing by a
> few thousand bytes here and there.
> Obviously, as we've gone from 8.3.9 to 8.3.10, we're restarted postgres a few
> times. We have also REINDEXd each table, the database, dropped and re-created
> all the indexes.
> I've even pg_dump'd the database, and loaded it into a fresh database. The
> relation numbers changed, but the 1663 bit remained the same.
Wow. The symptom seems to be what you'd get from a corrupt tuple
pointer in an index, but it's hard to believe that such a thing would
persist over reindexes and even reloads. Is the "target block" number
always the same, or does that change? Can you reproduce the problem on
a different machine? What platform are you running on?
regards, tom lane