cubicweb #1910817 switch from the non-scalable thread/request handling to something using processes and queues [rejected]


  • have a top-level process dispatching requests to worker processes
  • process/job control belongs to the top-level and should be 'transparent'
  • the number of actual process should be 'auto' (one process per available/perceived core) or configuration-fixed
  • use queues and queue size to have some sort of internal automatic load-balancing

This would replace the need to configure separate 'worker instances' on one machine to benefit from multiple cores. It would also avoid the current strangulation/performance drop that arises when too many requests are treated.

Memory consumption could get higher.

done in<not specified>
load left0.000
closed by<not specified>