On Mon, Apr 04, 2011 at 10:33:22AM -0500, Kevin Grittner wrote:
> There are patches for all known issues except one. Dan Ports was
> able to replicate the latest issue uncovered by YAMAMOTO Takashi
> using a particular DBT-2 configuration, found the issue, and posted
> a patch:
Well, it would be good to have confirmation from Takashi that it
actually fixed the problem he was seeing. I expect it did. If so, then
yes, we do have a handle on all open SSI issues.
> In investigating the locks which were not being cleaned up properly,
> Dan noticed that the pid wasn't showing on SIReadLock rows in
> pg_locks. He submitted a patch here which would always show the pid
> responsible for the lock:
>
> Jeff Davis questioned whether pid should continue to show after the
> end of the transaction or the closing of the connection (and
> therefore the process which the pid identifies). Not showing it for
> completed transactions would be trivial. Showing it after the
> transaction completes, until the connection closes should be doable,
> but not trivial. Of course, we could just leave it alone, but
> leaving the pid out for these rows looks a little funny and reduces
> useful information a bit.
I see Robert committed that one already. If there's a consensus that
omitting the pid for committed transactions is the right thing to do,
I'm happy to put together a patch. I think that is a better approach
than trying to keep it after commit until the connection closes, but
all of this is sufficiently minor that it's probably not worth worrying
much about.
> The one issue without a reasonable patch is that there are now three
> HTABs in shared memory which can grow until shared memory is
> exhausted, rather than the one in heavyweight locks which we had
> prior to 9.1. I think we're agreed that this is a bad thing, but my
> attempts to address this so far haven't satisfied Heikki. Heikki
> suggested an approach, but didn't respond as to whether I should try
> to code it up. I wasn't sure whether he might be going at it
> himself. I'll happily take a run at it if people want that.
I believe implementing that should just be a matter of making
get_hash_entry bail out before element_alloc if the right flag is set,
because partitioned hash tables already don't split buckets.
Dan
--
Dan R. K. Ports MIT CSAIL http://drkp.net/
