Refactor gc_worker with tikv_util/worker

[Little-Wallace]

Feature Request

Is your feature request related to a problem? Please describe:

Now we use two thread for gc_worker and gc_manager. But in actually, gc manager does not cost much CPU, we can merge them into one thread.

Describe the feature you'd like:

I hope most of threads in TiKV shall be implemented with tikv_util::worker::pool::Worker so that someday we may merge some threads into one thread pool.

Describe alternatives you've considered:

Teachability, Documentation, Adoption, Migration Strategy:

You may need to know something about GC and threadpool.

[Xuanwo]

I am very interested in this project and I will first take the time to understand the existing logic!

[Xuanwo]

Roadmap after offline nigotiation:

• [x] #10649
• [ ] Take down GcManager

[Xuanwo]

For Takedown GcManager, we need more details for the methods.

[Xuanwo]

@Little-Wallace Hi, is there any advice for keeping this task going further?

[Little-Wallace]

I think we shall merge the GcManagerContext,GcManager , GcWorker and GcRunner into two module. A GcManager holder GcManagerContext and Worker<GcTask> . GcRunner will run in worker of GcManager.

[Little-Wallace]

You can see some change in https://github.com/tikv/tikv/pull/8989/files. But not all of it are suitable. I think server.rs shall never hold an object named GcWorker. It is only hold an object Arc<GcManager> and most interface of GcWorker will move to GcManager.

[Little-Wallace]

At first, you can keep GcManager and GcWorker but just cancel the thread in GcManager. You can see that the Worker has supplied a method named on_timeout which means that you can do something in period. So we can just set a period task, every time refresh the safe point just like what TiKV currently does in GcManager::wait_for_next_safe_point.

[Little-Wallace]

Then we can move the method gc_a_round to GcRunner and just replace GcManager::sync_gc with the method GcRunner::gc

[Xuanwo]

Progress Report at 2021-08-15

Hi, I'm busy with my work last two weeks. I'm reading the logic about GcManager and GcWorker for now.

In the next two weeks, I will try to cancel the thread in GcManager first.

[Xuanwo]

At first, you can keep GcManager and GcWorker but just cancel the thread in GcManager. You can see that the Worker has supplied a method named on_timeout which means that you can do something in period. So we can just set a period task, every time refresh the safe point just like what TiKV currently does in GcManager::wait_for_next_safe_point.

I tried following methods:

• Implement a RunnableWithTimer Task called UpdateSafePointTask, but I can't schedule it in GcManager, because our task_scheduler only accept GcTask.
• Add new task type in GcTask, we need to add S: GcSafePointProvider, R: RegionInfoProvider + 'static, E: KvEngine for GcTask, It's too complex.
• Add GcSafePointProvider or AutoGcConfig in GcRunner directly. It looks like the correct way, but there are many changes to do.

Please guide me the correct way before I going further. :smile: