Re: Shared detoast Datum proposal - Mailing list pgsql-hackers

On Tue, May 21, 2024 at 10:02 PM Andy Fan <zhihuifan1213@163.com> wrote:
> One more things I want to highlight it "syscache" is used for metadata
> and *detoast cache* is used for user data. user data is more
> likely bigger than metadata, so cache size control (hence eviction logic
> run more often) is more necessary in this case. The first graph in [1]
> (including the quota text) highlight this idea.

Yes.

> I think that is same idea as design a.

Sounds similar.

> I think the key points includes:
>
> 1. Where to put the *detoast value* so that ExprEval system can find out
> it cheaply. The soluation in A slot->tts_values[*].
>
> 2. How to manage the memory for holding the detoast value.
>
> The soluation in A is:
>
> - using a lazy created MemoryContext in slot.
> - let the detoast system detoast the datum into the designed
> MemoryContext.
>
> 3. How to let Expr evalution run the extra cycles when needed.

I don't understand (3) but I agree that (1) and (2) are key points. In
many - nearly all - circumstances just overwriting slot->tts_values is
unsafe and will cause problems. To make this work, we've got to find
some way of doing this that is compatible with general design of
TupleTableSlot, and I think in that API, just about the only time you
can touch slot->tts_values is when you're trying to populate a virtual
slot. But often we'll have some other kind of slot. And even if we do
have a virtual slot, I think you're only allowed to manipulate
slot->tts_values up until you call ExecStoreVirtualTuple.

That's why I mentioned using something temporary. You could legally
(a) materialize the existing slot, (b) create a new virtual slot, (c)
copy over the tts_values and tts_isnull arrays, (c) detoast any datums
you like from tts_values and store the new datums back into the array,
(d) call ExecStoreVirtualTuple, and then (e) use the new slot in place
of the original one. That would avoid repeated detoasting, but it
would also have a bunch of overhead. Having to fully materialize the
existing slot is a cost that you wouldn't always need to pay if you
didn't do this, but you have to do it to make this work. Creating the
new virtual slot costs something. Copying the arrays costs something.
Detoasting costs quite a lot potentially, if you don't end up using
the values. If you end up needing a heap tuple representation of the
slot, you're going to now have to make a new one rather than just
using the one that came straight from the buffer.

So I don't think we could do something like this unconditionally.
We'd have to do it only when there was a reasonable expectation that
it would work out, which means we'd have to be able to predict fairly
accurately whether it was going to work out. But I do agree with you
that *if* we could make it work, it would probably be better than a
longer-lived cache.

--
Robert Haas
EDB: http://www.enterprisedb.com