Tom, may be I didn't make my point clear.
There are two issues :
1) Why this value was left as odd
2) Why backend entry is still pending in beentry for backend process even after it was killed/terminated.
I am talking about 2nd issue. My understanding is query on pg_stat_activity goes via all backend entries via beentry and it finds this wrong/corrupted entry. When a process terminates, ideally this backend entry into beentery should have also been cleaned. But why this still there? Whose responsibility it is to remove entry from beentry when process terminates ? May be that part is broken?
Ideally this issue should have stayed till this process was active. If entry from beentry would have been removed after process was killed, system should have auto recovered.
neeraj kumar <neeru.cse@gmail.com> writes:
> We got more information about this issue. There is one backend process
> still present into beentry which has changecount as odd value. However this
> process is long gone/terminated. It means when this process was
> killed/terminated its entry was not cleaned from beentry list. There seems
> to be some shutdown hook which cleans beentry if process is
> killed/terminated that somehow was not kicked off ?
No, that's the point: there isn't any such cleanup code. The design
idea here was that we'd only have short straight-line code segments
with no possible way to throw an error while changecount is odd.
Then somebody abused the heck out of that ...
regards, tom lane
--
-------------------------------------
Thanks
Neeraj Kumar,
+1 (206) 427-7267