Full rebuild after killing daemon

[zjturner]

Full rebuild after killing daemon

[JakobDegen]

Yes, this is expected behavior. The only caching that is available across daemon instances is whatever is provided by the RE action cache.

We've considered improvements here, typically under names like "dice rehydration" or something like that, which would preserve daemon state across restarts. Nothing that's on any near future roadmap though, unfortunately

[zjturner]

to be clear, is there no concept of a local action cache?

[JakobDegen]

There is not, no, although this is an interesting alternative to dice rehydration that I hadn't considered. I think @scottcao knows some things about this, maybe he can weigh in

[zjturner]

I was under the impression that hash_all_commands basically enabled the action cache and that it would have benefits even in local build scenarios. Did I misunderstand?

[JakobDegen]

I was under the impression that hash_all_commands basically enabled the action cache and that it would have benefits even in local build scenarios.

It does, but not across daemons.

Buck2 always has at least two strategies by which it avoids unnecessarily rerunning actions. The first is a remote action cache, which isn't relevant here. The second is DICE. DICE backs the implementation of most of buck2 core and is the reason that running the same command twice doesn't cause it to do all the work twice - it prevents not just unnecessarily rerunning actions, but also unnecessarily reevaluating buck files and things like that.

Dice, however, is not a cache, but rather an incrementality engine. It's primarily concerned with detecting what did or didn't change since the last build. As a result, if an action was already run five commands ago but invalidated since then, dice won't detect that it can skip recomputing that. Similarly, if two actions "swap places" in the build graph, dice won't detect to not rerun them either.

My understanding of hash_all_commands is that it activates a local action cache that can additionally avoid redoing work because of the above restrictions. However, that cache is still limited to the lifetime of a single daemon

[zjturner]

Just curious, what does DICE stand for? I keep seeing it and wondering.

In any case, I know buck1 had a dir cache, and the topic of dir cache has come up in the context of buck2 several times in the past. What's the current status of that? It seems like, if dir cache were implemented in buck1, would that also include action cache? And have there been any updates about implementing it on Meta's side?

Hypothetically speaking, let's say someone outside of Meta wanted to implement dir cache and upstream the work. Is it feasible without having access to internal Meta systems?

[JakobDegen]

Just curious, what does DICE stand for? I keep seeing it and wondering.

Distributed incremental computation engine. The "distributed" is ehhh... "future work"

What's the current status of that? It seems like, if dir cache were implemented in buck1, would that also include action cache?

I don't actually know what the dir cache in buck1 did, so I may have to get back to you on that.

Internally, I don't think we currently have any concrete plans for additional local caches - remote caches are working well for us.

Hypothetically speaking, let's say someone outside of Meta wanted to implement dir cache and upstream the work. Is it feasible without having access to internal Meta systems?

Yes, certainly; most of the internal-only code is around RE and this should basically work separately. We'd also be very happy to generally review design proposals or anything like that.

But even if that weren't the case - if there is some change that generally makes sense for OSS but we have no use for internally, feel free to send a PR anyway, I have no problem #[cfg]ing it out in our builds