feat: Custom Balancer

[metcoder95]

• Add histograms to workers - feat: add histogram to workers #619
• Expose workers lifecycle events - feat: add state properties to worker #620
• Expose pool lifecycle events - feat: pool events for workers #624
• Decouple handling of when new worker available and when a worker has finished a task
• Abstract event handling through AsyncResourcePool
• Decouple LB from scheduler (LB should return either worker or null)
• Implement LB branching
  • LB on new worker
  • LB once worker finished task
  • LB on new task

[jerome-benoit]

How is the worker histogram initialized in that case?
If a balancer query the runtime and that is initialized to zero, the balancer will wrongly pickup only new workers.
The same goes for worker recreation in the error case.

[metcoder95]

For each new worker, the histogram is instantiated to zero, but the whole WorkerPool has a global histogram that can be queried.

Nevertheless, to query it, you need to have a reference to the Worker Pool (Piscina instance). Maybe worth it to pass a reference to the global Piscina instance so it can query this data

[jerome-benoit]

Maybe. Or you can initialize to the pool one if it's not a cumulative histogram.
On poolifier, each entries init to the min of worker matching entries.

[metcoder95]

That would be useful, I'll follow that path; thanks!

[jerome-benoit]

What is the difference between closingUp and destroying flag? I think the two can be merged.

[metcoder95]

Thanks for the feedback @jerome-benoit, I'll kick-off the second round along the week

[metcoder95]

@ronag I've updated the docs, feel free to take a look now 👍

[jerome-benoit]

@ronag I've updated the docs, feel free to take a look now 👍

If you are sure that no regressions are introduced by that PR, you should merge it and flag load balancer feature as experimental. Then do an alpha or beta or ... release and communicate about it to allow people to play with it and do feedbacks.

[metcoder95]

Agree, that's the plan. After this is merged, I'll start issuing alpha of the new version of the library to start gathering feedback and allow people use it.

I'll reflect this in the documentation of this PR

[metcoder95]

I'm merging now, if there's more feedback please write it here, I'll be iterating over it in upcoming PRs

Let's land and release some alpha release of the new version to get some feedback; once the Atomics implementation is also merged, I'll release the version.