Research: thread per core model

[tomershafir]

I wonder, did you ever consider/research a thread-per-core/shard-per-core model for Clickhouse, like ScyllaDB seastar?

[nikitamikhaylov]

I would say yes, but we didn't proceed further with this initiative as it will require us to rewrite the whole codebase.
What we did so far is:

• Implemented custom background jobs (merges, mutations) execution mechanism which allows us to suspend and resume their execution, but only in pre-defined places e.g. after executing a block.
• Implemented a custom pipeline executor and introduced a task abstraction called IProcessor which allow us to better control the concurrency and the overall execution process.
• We do use boost coroutines in some places (for distributed query execution).

[alexey-milovidov]

Thread per core model requires userspace context switching and scheduling, which is painful to use and support.
OS threads are good enough.