Slashq

• Message processing demo inspired by below conference
• Specific implementation can be different
• slash 23 (https://toss.im/slash-23/session-detail/B1-7)

Feature

• Message Listening will be handled by netty io thread (default cpu)
• Message processing will be handled by dedicated single thread

Not implemented

• Data Store with async -> can be implemente through reactive redis template

Performance

• 비동기는 단일 요청의 입장에서는 어떤것도 빠르게 만들지는 않는다. Async never makes something faster...
• 다수의 request에서 평균 wait시간이 감속하는 효과
• Kafka가 성능이 느린건 I/O가 ... 아무리 kafka read/write이 효율이 좋다고 하지만 file i/o의 한계인듯

Why

• why single thread executor
  • sequence matters
• why policy is drop last
  • urgency matters

Try?

• Sequence matters -> always single thread?
  • parallel and sort ?
• parallel 처리 후 sort 한다면 .. nlgn
• sort can be O(1) if it has range, and we have (timestamp)
  • Map<TimeStamp, Record>
• caveat
  • timestamp can have multiple records
  • But many threads will access in parallel to update record
  • ConcurrentHashMap is disaster in performance

Tips

• TheadPoolExecutor -> core 수 설정하더라도 executor task가 들어오지 않으면 spawn 하지 않음
• Spring에서 제공하는 TaskExecutor가 아닌, jdk Executor를 사용할 경우 아래와 같이 warm up 코드가 존재하는 것이 좋음

this.threadPoolExecutor.execute(()->log.info("Thread ready :{}",Thread.currentThread().getName()));