Description

Ray Data’s read / map / write APIs already let you specify concurrency with a tuple (min, max) when the argument fn is a Python class.
However, in my long-tail workload the last few tasks are extremely expensive, so I set min = 1 to keep at least one actor alive. This forces Ray to start with a single actor and scale up to max through autoscaling, which adds noticeable startup latency.

I believe this requirement is reasonable, as both Kubernetes and Spark already provide similar knobs: in Kubernetes we have replicas, maxReplicas, and minReplicas, while Spark offers spark.dynamicAllocation.initialExecutors, spark.dynamicAllocation.maxExecutors, and spark.dynamicAllocation.minExecutors.

Use case

Perhaps we could extend the concurrency parameter to concurrency = (min, max, init).