During our benchmark tests, we found that the memory usage of Fluentd increases as num_threads increases (we keep the other configurations exactly the same). It's almost linear.

We are testing with pretty low QPS, which should not require that many threads as specified in num_threads. It makes sense to pre-allocate memory for each thread. But seems Fluentd could have saved a lot of resource if it uses some dynamic thread pool / kill idle threads and add new ones only when necessary. That way the baseline of memory usage should be lower.

Is this something that's possible to enable with the current Fluentd (we haven't tried v1.0.X yet)?

Thanks!

• Fluentd version
  v0.14.25
• OS
  ubuntu-1604-xenial
• Configuration

<source>
  @type forward
  port 24224
  bind 127.0.0.1
</source>

<source>
  @type prometheus
  port 24231
</source>

<source>
  @type prometheus_monitor
</source>

<match **>
  @type google_cloud
  buffer_chunk_limit 1M
  flush_interval 5s
  max_retry_wait 300
  disable_retry_limit
  num_threads 8
  use_grpc true
  enable_monitoring true
  monitoring_type prometheus
</match>