✅ Deploy Preview for salsa-rs canceled.

CodSpeed Performance Report

Merging #839 will not alter performance

_{Comparing ibraheemdev:interned-gc (02a5bc6) with master (14318b7)}

Summary

✅ 12 untouched benchmarks

I'm seeing a lot of panics in red-knot, mostly Data was not interned in the latest revision for its durability. I'm not sure exactly why this is happening, but I suspect it's related to durabilities.

Interesting. I suspect this is new but I'm investigating a similar issue related to interned in cycles

That makes me wonder if your PR might help to come up with simpler repros for the fixpoint issues that we're seeing. I should give this a try tomorrow.

On the perf side. You could try using the knot_benchmark script in the ruff repository

I quickly tried the branch in Red Knot but I also got new panics that we didn't see before. That makes me suspect that some revision tracking might be off in this PR?

What I like to do to debug those issues is to compare the tracing output and see where things go different

I think the issue is unrelated to durabilities but about interned query arguments. If a query takes an interned struct as an argument then an older memo may be incorrectly verified after the slot has been reused, but the IDs still match. I think we have to check that the memo and first_interned_at revisions match to avoid this, but I'm not exactly sure where that would be.

I think I understand. What you're saying is that query(interned) may be checked for changes in a later revision where interned has been collected.

I think will need something similar to what we have in tracked structs where salsa informs other ingredients that the memo was collected so that they can free resources as well if necessary:

https://github.com/MichaReiser/salsa/blob/953dc34ce5cfd7fca6ad040241c3cbebc11830e7/src/tracked_struct.rs#L597-L655

The second part in tracked struct that protects against this is that we separately track created_at and updated_at. maybe_changed_after can then compare created_at to see if this struct was recrated since rev_before.

I updated the PR to add read dependencies and clear memos on reuse (I believe we will need to defer reclamation of the memos), but there are a still some ty tests failing that I'm unsure about.

Hmm I'm not sure we can remove the read dependency on creation with generational IDs, we still need last_interned_at to be updated for any dependent queries. Maybe it would be cheaper to update that field directly when an interned value is read?

Hmm I'm not sure we can remove the read dependency on creation with generational IDs, we still need last_interned_at to be updated for any dependent queries. Maybe it would be cheaper to update that field directly when an interned value is read?

No, I don't think we can remove it from the writing query. We need the dependency to ensure the interned value gets re-created when it was claimed previously

I updated the PR description with the new changes, this should be ready for review.

We also cannot collect low durability values, which would require &mut db. This could be supported with the above.

I believe you mean high durability here? It would be helpful to add some details to why

Slots that have not been accessed in the last 3 active revisions are always reused, and nothing else. Active revisions are those in which interned values were actually read to avoid counting empty revisions.

I believe we originally discussed a GC where interned values would move from green, to yellow, to red generations. I assume what you implemented now isn't that. Can you give some context to why?

My main concern is that this change adds 3 word-sized fields to interned values: an Id and two pointers for the intrusive LRU list. This seems unavoidable to implement garbage collection, but still something to note.

Is this after removing the created_at field or in addition to it?

I believe we originally discussed a GC where interned values would move from green, to yellow, to red generations. I assume what you implemented now isn't that. Can you give some context to why?

The reason is that it doesn't really make sense to have a specific capacity for the LRU. Instead, we care more about reusing values that are unlikely to be accessed again. Bounding the LRU means we might end up evicting values that were read in a recent revision. It also makes it a lot more expensive to grow the LRU and shuffle around the values based on their color, especially for a cold run.

The LRU implemented in this PR does have some idea of coloring based on revision, i.e. values that have already been read in the current revision are not moved in the LRU, which as I understand is the main benefit of the coloring idea.

I believe you mean high durability here? It would be helpful to add some details to why

Yes, sorry. I believe issue was that we need to make sure that maybe_changed_after (and the update to last_interned_at) is not short-circuited for higher durability values that are reused , which would require a synthetic write to invalidate the revision for the given durability.

I removed the first_interned_at field, so the memory usage impact is only two words per-value for the LRU list.

Would it be possible to release med/high durability revisions if we detect a med/high durability change? I guess the problem remains that med/high durability in most cases would need to be skipped, which could outweight the benefits of reusing them in the less common case of a medium/high durability change.