On Sun, Oct 4, 2015 at 1:01 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> Ah, right, sorry. I meant to make its result match what gc_texts would
> get, by not falsely counting entries with dropped texts. That's not
> what you have in your patch but it seems like an easy enough fix.
I'm trying to make mean_query_len representative of *useful* entry
query length. I guess I don't have that within gc_texts in my patch,
but I do have it within entry_dealloc (up to and including considering
dropped texts), which FWIW is far more important.
>> I'd be quite happy if you did everything listed, and also left the
>> extra discrimination against sticky entries within entry_dealloc in --
>> consider what happens when a huge malloc() ends up swapping with an
>> exclusive lock held, and consider that repeated, failed data
>> integration transactions are implicated in this in a big way when a
>> problem appears in the wild. A big part of the problem here was that
>> garbage collection did not run often enough.
>
> Hm. The problem I've got with this is that then mean_query_len means
> something significantly different after entry_dealloc than it does
> after gc_texts.
>
> I'd be okay with changing *both* of those functions to ignore sticky
> entries in the calculation, if that seems reasonable to you.
That seems perfectly reasonable, yes. Should I leave that to you?
>> In other words, I'd be fine with *not* doing the query size filter
>> thing for now, since that is something that seems like an extra
>> defense and not core to the problem. I was kind of ambivalent about
>> doing that part myself, actually.
>
> Agreed on that part.
We're in full agreement on what needs to happen for the next point
release, then. Excellent.
--
Peter Geoghegan