Thanks for working on this, Corey.
On Fri, Mar 14, 2025 at 04:03:16PM -0400, Corey Huinker wrote:
> 0003 -
>
> Storing the restore function calls in the archive entry hogged a lot of
> memory and made people nervous. This introduces a new function pointer that
> generates those restore SQL calls right before they're written to disk,
> thus reducing the memory load from "stats for every object to be dumped" to
> just one object. Thanks to Nathan for diagnosing some weird quirks with
> various formats.
>
> 0004 -
>
> This replaces the query in the prepared statement with one that batches
> them 100 relations at a time, and then maintains that result set until it
> is consumed. It seems to have obvious speedups.
I've been doing a variety of tests with my toy database of 100K relations
[0], and I'm seeing around 20% less memory usage. That's still 20% more
than without stats, but that's still a pretty nice improvement.
I'd propose two small changes to the design:
* I tested a variety of batch sizes, and to my suprise, I saw the best
results with around 64 relations per batch. I imagine the absolute best
batch size will vary greatly depending on the workload. It might also
depend on work_mem and friends.
* The custom format actually does two WriteToc() calls, and since these
patches move the queries to this part of pg_dump, it means we'll run all
the queries twice. The comments around this code suggest that the second
pass isn't strictly necessary and that it is really only useful for
data/parallel restore, so we could probably skip it for no-data dumps.
With those two changes, a pg_upgrade-style dump of my test database goes
from ~21.6 seconds without these patches to ~11.2 seconds with them. For
reference, the same dump without stats takes ~7 seconds on HEAD.
[0] https://postgr.es/m/Z9R9-mFbxukqKmg4%40nathan
--
nathan
