WIP patch: reducing overhead for repeat de-TOASTing

From: Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us>
To: pgsql-hackers(at)postgreSQL(dot)org
Subject: WIP patch: reducing overhead for repeat de-TOASTing
Date: 2008-06-29 20:57:10
Message-ID: 5184.1214773030@sss.pgh.pa.us
Views: Raw Message | Whole Thread | Download mbox | Resend email
Lists: pgsql-hackers

Attached is a worked-out patch for the approach proposed here:
namely, that cache management for de-TOASTed datums is handled
by TupleTableSlots.

To avoid premature detoasting of values that we might never need, the
patch introduces a concept of an "indirect TOAST pointer", which has
the same 0x80 or 0x01 header as an external TOAST pointer, but can
be told apart by having a different length byte. Within that we have
* pointer to original toasted field within the Slot's tuple
* pointer to the owning Slot
* pointer to decompressed copy, or NULL if not decompressed yet
Some fairly straightforward extensions to the TupleTableSlot code,
heaptuple.c, and tuptoaster.c make it all go.

My original thoughts had included turning FREE_IF_COPY() into a no-op,
but on investigation that seems impractical. One case that still
depends on that pfree is where we have palloc'd a 4-byte-header copy
of a short-header datum to support code that needs properly aligned
datum content. The solution adopted in the patch is to arrange for
pfree() applied to a cacheable detoasted object to be a no-op, whereas
it still works normally for non-cached detoasted objects. We do this
by inserting a dummy chunk header that points to a dummy memory context
whose pfree support method does nothing. I think this part of the patch
would be required for any toast caching method, not just this one.

What I like about this patch is that it's a fairly small-footprint
change, it doesn't add much overhead, and it covers caching of
decompression for in-line-compressed datums as well as the out-of-line

One thing I really *don't* like about it is that it requires everyplace
that copies Datums to know about indirect pointers: in general, the copy
must be a copy of the original toasted Datum, not of the indirect
pointer, else we have indirect pointers that can outlive their owning
TupleTableSlot (or at least outlive its current tuple cycle). There
only seem to be about half a dozen such places in the current code,
but still it seems a rather fragile coding rule.

After playing with it for a little bit, I'm not convinced that it buys
enough performance win to be worth applying --- the restriction of cache
lifespan to one tuple cycle of a TupleTableSlot is awfully restrictive.
(For example, sorts that involve toasted sort keys continue to suck,
because the tuples being sorted aren't in Slots.) It would probably
fix the specific case that the PostGIS hackers were complaining of,
but I think we need something more.

Still, I wanted to get it into the archives because the idea of indirect
toast pointers might be useful for something else.

regards, tom lane

Attachment Content-Type Size
toast-indirect-1.patch.gz application/octet-stream 11.3 KB


Browse pgsql-hackers by date

  From Date Subject
Next Message Robert Treat 2008-06-30 02:47:03 Re: Does anything dump per-database config settings? (was Re: ALTER DATABASE vs pg_dump)
Previous Message Karl Schnaitter 2008-06-29 19:33:24 Re: TODO item: Allow data to be pulled directly from indexes