I looked at the thread dump in more detail, and the situation is now less clear.

(1) I see no obvious ManifoldCF-mediated deadlock at all.  Everything
looks as it should.
(2) There are a lot of threads waiting for one particular thread to finish.
(3) That thread is waiting to complete a query that *should* complete,
according to everything else in ManifoldCF that I can see.  If it
doesn't complete it should detect database deadlock, in any case.

The possibility I am worried about now is that postgresql processes
are dying due to some issue, like memory consumption.  We therefore
need to correlate postgresql processes with threads, so if you can get
a ps -elf | grep postgres from the postgres machine and also include
that, it would be great.
Karl

On Mon, Feb 11, 2013 at 11:14 AM, Karl Wright <[email protected]> wrote:
> I've confirmed that there's a deadlock of some kind; it's occurring as
> a result of the hopcount depth tracking.  The issue seems to be nested
> transactions; PostgreSQL doesn't appear to be dealing with these
> properly in all cases, and winds up getting a deadlock that it doesn't
> detect.
>
> I have to look carefully at the code to see if it can be restructured.
>
> Karl
>
> On Mon, Feb 11, 2013 at 10:56 AM, Karl Wright <[email protected]> wrote:
>> That is not the reason for the stop.  I suspect a deadlock.  Trying to
>> verify that now.
>>
>> Karl
>>
>>
>> On Mon, Feb 11, 2013 at 10:44 AM, Erlend Garåsen
>> <[email protected]> wrote:
>>> On 11.02.13 16.26, Karl Wright wrote:
>>>>
>>>> trunk has a different schema than 1.1.1.  So yes, you'd have to blow
>>>> away the old database to go back.
>>>
>>>
>>> OK, who knows. Maybe this was the reason why it just stopped. I'll clean up
>>> by deleting all tables and related db resources and try again.
>>>
>>> Erlend
>>>
>>>
>>> --
>>> Erlend Garåsen
>>> Center for Information Technology Services
>>> University of Oslo
>>> P.O. Box 1086 Blindern, N-0317 OSLO, Norway
>>> Ph: (+47) 22840193, Fax: (+47) 22852970, Mobile: (+47) 91380968, VIP: 31050

Reply via email to