On 11/7/2005 3:40 PM, Christopher Browne wrote:
> Hannu Krosing <hannu at skype.net> writes:
>> On E, 2005-11-07 at 07:29 -0500, Jan Wieck wrote:
>>> On 11/7/2005 4:11 AM, Hannu Krosing wrote:
>>> > On L, 2005-11-05 at 12:48 +0000, Dave Page wrote:
>>> >> It doesn't :-(
>>> > 
>>> > What are the issues ?
>>> > 
>>> > Is anybody working on it ?
>>> 
>>> The issue is a change in a backend internal interface function to 
>>> LISTEN/NOTIFY, that we use to cleanup stale pg_listener entries after a 
>>> postmaster crash. I do have a plan and am working on it. Fixes will be 
>>> applied to HEAD and 1_1_STABLE soon.
>>
>> Would it be worthwhile to change postgresql (too late for 8.1 but
>> maybe for 8.2 if LISTEN/NOTIFY is not completely rewritten there) by
>> adding "TRUNACTE PG_LISTENER;" to postgresql's startup code and also
>> in cleanup after kill -9 initiated restarts ?
> 
> That seems overly aggressive, and I'm not sure it would work anyways.
> 
> The upcoming problem is that pg_listener won't be tracking PIDs the
> same way, and we use the PIDs at present to detect whether a
> particular node is "claimed" by a slon process yet.

This isn't entirely accurate.

The current problem is just that the call to Async_Unlisten() lost its 
pid argument, so one backend cannot remove stale entries from another 
one any more via this call. As a quick workaround we could have done a 
DELETE FROM pg_listener ... instead, but since we can already foresee 
that we're going to get into trouble with that again as soon as the 8.2 
rewrite of LISTEN/NOTIFY removes that relation eventually, I think it is 
better to reimplement our interlock together with the underlying 
information for terminating all node connections on failover completely 
right now.


Jan

> 
> Truncating pg_listener is beside the point; the problem is that, in
> 8.1, we lose the "interlock" where pg_listener is used to have each
> node uniquely claim its node.
> 
> By 8.2, I don't think we'll be using LISTEN/NOTIFY the way we are at
> present...


-- 
#======================================================================#
# It's easier to get forgiveness for being wrong than for being right. #
# Let's break this rule - forgive me.                                  #
#================================================== JanWieck at Yahoo.com #