Details
Major Description
On remote filesystems like Azure, GCP and S3, the splits contain a filler location of "localhost".
This is worse than having no location information at all - on large clusters yarn waits upto 200[1] seconds for heartbeat from "localhost" before allocating a container.
To speed up this process, the split affinity provider should scrub the bogus "localhost" from the locations and allow for the allocation of "*" containers instead on each heartbeat.
[1] - yarn.scheduler.capacity.node-locality-delay=40 x heartbeat of 5s