Interesting. That's exactly what we have been doing -- trying to update the same rows in multiple txns. For us to proceed in production, I will take steps to ensure we stop doing that, as it's just an app bug really.
The table in question -- v_messages -- is an empty base table with 76 partitions, with a total of 2.8 billion rows. Let me summarize what I see as the key facts here: (All problems have come from the UPDATE query, all identical except for different "author_id" values.) 1. We did a "link" upgrade Wed night, from 844 to 902 so the upgrade happened in place, no data files were copied. 2. The 1st error was "compressed data is corrupt" at 18:16 3. We got 2 seg fault crashes before turning on cores and getting a 3rd crash with the stack trace. 4. We then got a " invalid memory alloc request size 18446744073449177092" at 23:50. This was an ERROR, not a crash. At this point, is it your suspicion that there is a code bug in 9.0.2, rather than corrupt data? I will post the schema and then work on a test case. -gordon On Fri, Dec 31, 2010 at 8:34 AM, Tom Lane-2 [via PostgreSQL] < ml-node+3323712-1368244686-56...@n5.nabble.com<ml-node%2b3323712-1368244686-56...@n5.nabble.com> > wrote: > > Hmm. This suggests that there's something wrong in the EvalPlanQual > code, which gets invoked when there are concurrent updates to the same > row (ie, the row this UPDATE is trying to change is one that was changed > by some other transaction since the query started). That stuff got > rewritten rather thoroughly for 9.0, so the idea of a new bug there > isn't exactly surprising. But it's going to be hard to find without > a test case. Can you show us the full schema for this table and all > the queries that execute against it up till the point of the failure? > (Turning on log_statement across all sessions would help collect that > info, if you don't have it already.) > > -- View this message in context: http://postgresql.1045698.n5.nabble.com/seg-fault-crashed-the-postmaster-tp3323117p3323796.html Sent from the PostgreSQL - general mailing list archive at Nabble.com.