Re: rethinking dense_alloc (HashJoin) as a memory context

On 07/13/2016 07:37 PM, Tom Lane wrote:
> Peter Geoghegan <pg@heroku.com> writes:
>> On Wed, Jul 13, 2016 at 7:53 AM, Tomas Vondra
>> <tomas.vondra@2ndquadrant.com> wrote:
>>> In the thread [1] dealing with hashjoin bug introduced in 9.5, Tom voiced
>>> his dislike of dense_alloc. I kinda agree with him that introducing "local
>>> allocators" may not be the best idea, and as dense_alloc was introduced by
>>> me I was playing with the idea to wrap this into a regular memory context,
>>> perhaps with some restrictions (e.g. no pfree). But I'm having trouble with
>>> that approach ...
>
>> I think that the "no pfree()" restriction would be necessary to get
>> the same benefit. But, doesn't that undermine the whole idea of making
>> it a memory context?
>
> The other thing that doesn't seem to square at all with a general-purpose
> memory context is the desire to walk through the stored tuples directly,
> knowing that they are adjacent.  That means nothing else can be allocated
> via the same mechanism.  So I tend to agree that if we accept Tomas' three
> requirements as non-negotiable, then trying to make the allocator match
> the MemoryContext API is probably impractical.
>
> My feeling at this point is that we should leave it alone until/unless
> we see similar requirements elsewhere, and then look to see if we can
> derive a common abstraction.  I always find that it's easier to design
> APIs based on concrete use-cases than on guesses about what will be
> needed.

I agree with both points.

I think the MemoryContext API may not be right abstraction for this. 
Given a hammer big enough it would probably work in the end, but it'd 
probably require changes to the public MemoryContext API (e.g. relaxing 
the StandardChunkHeader requirement). And that seems a bit too risky.

So we probably need a new independent abstraction for this, but doing 
that based on a single use case is a bit silly.

>
> I wonder though if we don't already have another similar use-case in
> the ad-hoc "slab allocators" in reorderbuffer.c.  We already know that
> that code has performance issues, cf bug #14231, so I suspect there's
> a redesign in its future anyway.
>

I'm not sure - I'm not familiar with reorderbuffer.c, but it seems to do 
a fair number of pfrees and such. Also, pfrees seem to be the root of 
the performance issue. I suspect the slab allocator (or rather the 
allocation strategy in general) may need rethinking, but let's discuss 
that in that thread.

regards

-- 
Tomas Vondra                  http://www.2ndQuadrant.com
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services