Claudio Natoli <claudio.natoli@memetrics.com> writes:
> Actually, if I was going to argue anything, I'd say that if a backend goes
> nuts and zeroes the whole shmem segment you've probably some bigger things
> to worry about (Aside: Would postgres actually recover from such an
> occurence? BTW, I'd be pretty impressed if it did, but not all that
> surprised ;-).
It should, although there are limits (for instance, if someone is
actively writing out a page of WAL at the same time the bogus backend
comes by and zeroes that buffer, you might lose WAL entries for
already-committed transactions, which would be unhappy-making).
As a developer, though, I crash backends all the time, and I can say
that this mechanism is indeed pretty robust. The postmaster never goes
down (what, never? well, hardly ever) and it's *extremely* seldom that
a crash results in on-disk corruption, because the postmaster generally
manages to kill the other backends before any corruption in shared
memory gets propagated to disk.
regards, tom lane