Re: pg_listener entries deleted under heavy NOTIFY load only on Windows

Any thoughts on whether or not this patch should be included in a future
PostgreSQL release? If so, any thoughts on additional testing, etc,
that needs to be done? Should the patch be added to a minor release of
8.3, or only included forward in 8.4? In my opinion, this is a pretty
serious error causing some major problems and workarounds in systems I
am working on, should I would vote for inclusion in minor releases.

The post below seems to have become disassociated from the rest of the
thread on this bug, perhaps because it was issued in February, and the
rest of the thread was posted in January. Is there any way to make the
mailing list archives link these messages back to the rest of the
thread?

-----Original Message-----
I've completed my test for changes to the Windows implementation of
pgkill. I ran two clients constantly doing a pair of notifications,
with one client subscribing to those events. I ran the test for 90
hours. I found 760 cases where the CallNamedPipe function initially
failed, but succeeded after recalling. In all but one of those cases,
only a single retry was needed; the other case required two retries.

As Tom has pointed out, there is no principled way to determine the
correct number of times to retry. Therefore, I have set the maximum
number of retries to 2, which is the worst case found in a fairly long
stress test. This strikes me as a reasonable balance between failing
quickly in truly erroneous cases, while still avoiding transient
pseudo-failures.

A patch with this change is attached. It should be applied to the
8.3.5 version of src/port/kill.c

Steve