Re: postgres crash SOS

Felde Norbert <fenor77(at)gmail(dot)com> writes:
> The message is the same for the original pg_clog/0003, the 0003
> containing binary 0 and after pg_resetxlog:
> pg_dump: Error message from server: ERROR: could not access status of
> transaction 3974799
> DETAIL: Could not read from file "pg_clog/0003" at offset 204800: No error.
> pg_dump: The command was: COPY public.active_sessions_split (ct_sid,
> ct_name, ct_pos, ct_val, ct_changed) TO stdout;
> pg_dump: *** aborted because of error

> If create the bigger 0003 containing 0 than I get that:
> pg_dump: Error message from server: ERROR: xlog flush request
> 0/A19F5BF8 is not satisfied --- flushed only to 0/A02A1AC8
> CONTEXT: writing block 1149 of relation 1663/4192208/4192508
> pg_dump: The command was: COPY public.history (historyid, adatkod,
> elemid, userid, ido, actionid, targyid, szuloid, opvalue, longfield,
> longtext) TO stdout;
> pg_dump: *** aborted because of error

I'm afraid this means you're screwed :-(. Both of those symptoms imply
that one part of the database storage is out of sync with another part:
the first error says there are transaction IDs in the
active_sessions_split table that don't exist in pg_clog, and the second
error says that there are pages in the history table that were last
updated by WAL records that don't exist in pg_xlog. If there are two
such errors, there are probably more.

You weren't too specific about how you got into this state, but I
suppose that it must have been a system crash or power failure. Even
then, you would not have gotten burnt if the filesystem and hardware
did what they're supposed to do. I suspect you have a setup wherein
fsync() calls aren't being honored properly. You may need to disable
write caching on your disks, and/or switch to another filesystem or OS.
(Personally I'd never run a database I cared about on Windows.)

regards, tom lane