Re: TOAST code ignores freespace (was Tweak TOAST code)

On 5/2/2010 10:34 AM, Tom Lane wrote:
> Simon Riggs <simon@2ndQuadrant.com> writes:
>> Not commenting further on that patch, but I notice that when we UPDATE
>> the toasting algorithm takes no account of the available freespace on
>> the current block. If we are updating and the space available would make
>> a difference to the row length chosen, it seems like it would be more
>> beneficial to trim the row and encourage HOT updates.
> 
> That doesn't strike me as a terribly good idea: it would make the
> behavior of TOAST significantly more difficult to predict.  Also, what
> happens if we force a row to a smaller size and then it doesn't fit
> anyway (eg because someone else inserted another row on the page while
> we were busy doing this)?  Spend even more cycles to un-toast back to
> the normal size, to be consistent with ordinary cross-page updates?
> 
> Pretty much every previous discussion of tweaking the TOAST behavior
> has focused on giving the user more control (indeed, the patch you
> mention could be seen as doing that).  What you're suggesting here
> would give the user less control, as well as less predictability.

Correct. And on top of that, the cost/benefit of the proposed change 
will be extremely hard to evaluate since freespace and the value of HOT 
depend very much on access patterns.

If we want to substantially do better, we need to use a bigger hammer.

TOAST's largest performance benefit lies in the fact that it reduces the 
size of the main tuple, which is the data that travels in intermediate 
result sets throughout the executor. Reducing that size results in 
smaller sort sets, more in memory operations, fewer blocks seqscanned 
for keys and all that.

Suppose we had something similar to the NULL value bitmap, specifying 
plain or compressed values (not TOAST references), that are moved to a 
shadow tuple inside the toast table. Suppose further we had some 
statistics about how often attributes appear in a qualification (i.e. 
end up in a scan key or scan filter or other parts of the qual 
expression list). Not sure, maybe we even want to know how often or 
seldom an attribute is heap_getattr()'d at all. Those don't need to be 
accurate counts. Small random samples will probably do. ANALYZE could 
evaluate those statistics and adjust the "shadow" storage settings per 
attribute accordingly.

I can imagine many applications, where this would shrink the main tuples 
to almost nothing at all.

There are for sure a lot of "if's" and "suppose" in the above and the 
impact of a fundamental on disk storage format change needs to be 
justified by a really big gain. And yes, Simon, this also depends a lot 
on access patterns. But if you try to gain more from TOAST, I'd look for 
something like this instead of making the target tuple size dynamic.


Jan

-- 
Anyone who trades liberty for security deserves neither
liberty nor security. -- Benjamin Franklin