In #3602 we noticed that during a VACUUM container counts would climb up and max out the workers.

This seems to imply that GC was held up somewhere. If containers are being created, queries are running to create those, so it's not like the whole table is locked up.

We currently run all of our collectors one after another:

Maybe one of the collectors was blocked waiting on the VACUUM? Should we run these in parallel on independent loops instead?

Can we confirm the slowdown by observing logs on the algorithm environment? (This is time-sensitive, since they don't currently aggregate anywhere.)