Peter Eisentraut <peter.eisentraut@2ndquadrant.com> writes:
> In a non-data_only dump, the order of the tables doesn't matter, because
> the foreign keys are added at the very end. In parallel dumps, the
> tables are in addition sorted by size, so the resultant order is
> different from a single-threaded dump. This can be seen by comparing
> the dumped TOCs of the defaults_dir_format and defaults_parallel cases.
> But it all happens to pass the tests right now.
I noticed that business about sorting the TOC by size yesterday.
I think that's a completely bletcherous hack, and we ought to get
rid of it in favor of keeping the TOC order the same between parallel
and non-parallel cases, and instead doing size comparisons during
parallel worker dispatch.
The primary reason why it's a bletcherous hack is that it's designed
around the assumption that parallel dumps will be restored in parallel
and non-parallel dumps will be restored non-parallel. That assumption
is wrong on its face. But it explains why the code sorts both tables
and indexes, even though pg_dump doesn't have any need to parallelize
index dumps: it's expecting that a subsequent parallel restore will have
use for building indexes largest-first. Of course, if you made the dump
non-parallel, you're out of luck on that.
Another reason why the code is bogus is that it sorts only a consecutive
sequence of DO_INDEX items. This fails to work at all for indexes that
are constraints, and even for ones that aren't, there's no very good
reason to assume that they aren't interspersed with other sorts of
objects due to dependencies. So I suspect an awful lot of parallelism
is being left on the table so far as indexes are concerned.
So if we made this less of a quick-n-dirty kluge, we could remove the
hazard of different-looking dump results in parallel and serial cases,
and probably improve restore performance as well.
A small problem with postponing sorting to the restore side is that
I don't think we record relpages info anywhere in the archive format.
However, at least for the directory-format case (which I think is the
only one supported for parallel restore), we could make it compare the
file sizes of the TABLE DATA items. That'd work pretty well as a proxy
for both the amount of effort needed for table restore, and the amount
of effort needed to build indexes on the tables afterwards.
regards, tom lane
