On Saturday, September 15, 2012 06:29:25 PM Tom Lane wrote:
> Robert Haas <robertmhaas@gmail.com> writes:
> > Definitions aside, I think it's a pretty scary issue. It basically means
> > that if you have a recovery (crash or archive) during which you read a
> > buffer into memory, the buffer won't be checkpointed. So if, before the
> > buffer is next evicted, you have a crash, and if at least one checkpoint
> > has intervened between the most recent WAL-logged operation on the
> > buffer and the crash, you're hosed. That's not a terribly unlikely
> > scenario.
>
> This is only an issue on standby slaves or when doing a PITR recovery, no?
> As far as I can tell from the discussion, it would not affect crash
> recovery, because we don't do restartpoints during crash recovery.
I think unfortunately it does. At the end of recovery we perform a
END_OF_RECOVERY checkpoint that seems to suffer from these issues. While
CreateCheckPoint() itself treats that kind of checkpoint similarly to a
shutdown checkpoint it doesn't pass that similarity to BufferSync (via
CheckPointGuts->CheckPointBuffers).
I hope I missed something ...
Greetings,
Andres
-- Andres Freund http://www.2ndQuadrant.com/PostgreSQL Development, 24x7 Support, Training &
Services