On Thu, Dec 15, 2022 at 01:56:30PM -0500, Tom Lane wrote:
> * To fix vacuumdb properly, it might be enough to get it to
> batch VACUUMs, say by naming up to 1000 tables per command
> instead of just one. I'm not sure how that would interact
> with its parallelization logic, though. It's not really
> solving the O(N^2) issue either, just pushing it further out.
I have been thinking about this part, and using a hardcoded rule for
the batches would be tricky. The list of relations returned by the
scan of pg_class are ordered by relpages, so depending on the
distribution of the sizes (few tables with a large size and a lot of
table with small sizes, exponential distribution of table sizes), we
may finish with more downsides than upsides in some cases, even if we
use a linear rule based on the number of relations, or even if we
distribute the relations across the slots in a round robin fashion for
example.
In order to control all that, rather than a hardcoded rule, could it
be as simple as introducing an option like vacuumdb --batch=N
defaulting to 1 to let users control the number of relations grouped
in a single command with a round robin distribution for each slot?
--
Michael
