We have support for multiprocess execution, where we can execute ops concurrently (this is the default executor). Issue
#4041 is tracking the async executor which executes the ops in a single process using threads. This avoids the process spin-up cost for each op. It’s still very much on our roadmap but we have some conflicts with the current implementation of the compute log manager, which we’ll need to rewrite. I know some users have used the
NoOpComputeLogManager
(to basically disable the compute log manager) and written their own async executor successfully to achieve the same effect.