On Thu, Apr 23, 2015 at 11:24 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> Having said all that, if we did try to fix it today, I'd imagine changing
> TOAST value identifiers to int64 and inventing a new TOAST pointer format
> for use when 32 bits isn't wide enough for the ID. But I think we're best
> advised to hold off doing that until the need becomes pressing.
Just out of curiosity, has anyone thought about inventing a new TOAST
pointer format on the grounds that our TOAST pointers are unreasonably
large? IIUC, a TOAST pointer right now is 18 bytes: 16 for a
varatt_external, and then that gets embedded in a varattrib_1b_e with
a va_header byte and a va_tag byte. Eliminating one or both of
va_rawsize and va_extsize from the TOAST pointer itself seems like it
could save quite a bit of space on disk. Maybe you could even find a
way to get rid of va_toastrelid; after all, at the point when you
first acquire a pointer to the tuple, you surely know what relation
it's a part of. You'd probably want to force de-TOASTing (or
converting to a more expressive form of TOAST pointer, anyway) when
you extracted the column from the tuple, which might be hard to
arrange.
But the benefits could be pretty significant. Suppose you have a
table where each tuple is 4K untoasted, with all but 100 bytes of that
in a single column. So, as stored, you've got 100 bytes of regular
stuff plus an 18-byte TOAST header. If you could trim 2 of the
above-mentioned 4-byte fields out of the TOAST header, that would
reduce the size of the main relation fork by almost 7%. If you could
trim all 3 of them out, you'd save more than 10%. That's not nothing,
and the benefits could be even larger for rows that contain multiple
TOAST pointers.
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company