FlashCacheComponent using RegionManager's reclaim workers

[sumo-orange]

First of all, thanks for open-sourcing cachelib.

I was curious what the longer-term plan was for cachelib/interface/CacheComponent.h was. I was looking into writing a folly::coro wrapper for the navy block cache, but if CacheComponent and its impls are likely to remain in the repo, we'll probably try to use those.

I also have a specific question about how FlashCacheComponent is scheduling work: it looks like it's using the region manager's flush/reclaim worker NavyThread(s) to do the call into insert/find on the block cache (see FlashCacheComponent::onWorkerThread) -- is the long term plan to continue to do that?

Thanks!

[rlyerly]

Hey, thanks for reaching out!

The goal is to provide a single cache interface that can be implemented by individual "Cache Components" (RAM cache, flash cache, etc.). Then, we want to build a generic inter-component framework on top of these that abstracts away all the inter-cache movement, including synchronization, optimized data copy, etc. At first this is going to look similar to the existing hybrid cache implementation, but the goal is to be able to support more exotic architectures and new components (e.g., CXL caching). FlashCacheComponent is part of the first step of breaking out individual components for composition.

To answer your questions:

1. These are staying in the repo and are under active development
2. No, we're going to move FlashCacheComponent's async work onto a separate dedicated pool of workers. Ideally we'd launch the work on the executing coroutine (and then the caller would only need to configure the sizes once), but we need a translation layer because BlockCache uses fibers for async work. We may revisit the fibers abstraction in the future, but for now we're using a coro <-> fiber interop layer.