"Jim C. Nasby" <jnasby@pervasive.com> writes:
> Doesn't clog use the same code?
Yeah, but all three of your examples were referencing pg_subtrans,
as proven by the stack traces and the contents of the shared control
block.
Even though the bug seems completely clog.c's fault, this is not a
coincidence: if subtransactions are being used heavily, then pg_subtrans
would have far greater I/O volume than any of the other clog-managed
logs, and hence have more exposure to the race condition.
We really ought to fix that code so that pg_subtrans can have more
buffers than pg_clog...
regards, tom lane
