(encoding of confusion: unreferenced state group deletion either doesn't make sense, or can explode size of state_groups_state)

[reivilibre]

This is intended as a demonstrative example for #18217 and is not a PR to merge.

---

The docstring

Deletes no longer referenced state groups and de-deltas any state
groups that reference them.

does not make sense; or rather, it's contradictory — if a state group is no longer referenced, then it follows that no state groups reference it and thus there should be no de-delta-ing to be done.

BUT if we edit it slightly, maybe the docstring starts to make sense?

Deletes state groups that are no longer referenced by events and de-deltas any state
groups that reference them.

frankly: I don't actually know if this is the intended meaning. But it sounds plausible.

If we go by that assumption, then we can write a test to demonstrate a situation in which state group deletion causes more rows to be added to state_groups_state than deleted.

[erikjohnston]

Ah, I think the thinko here is that it assumes that all state groups before it have already been deleted. This would make sense given we generally purge all history before a certain time, however that doesn't take into account the fact we do naturally have unreferenced-by-events state groups between two referenced state groups.

Two options, either don't delete unreferenced-by-events state groups that:

1. are referenced by other state groups; or
2. reference other state groups.

The motivating reason why we delete unreferenced-by-events state groups is the case where you have long chains of state groups S_1, S_2, …, S_n where all but S_n are unreferenced-by-events. The number of state events in these chains can get large (especially when using the state compressor), so deleting everything and putting the full state in S_n can remove a lot of rows.