And another question: does the controller need to have a lot of cpu and memory, or just brokers and servers (i.e. servers have relatively small jvm heap size but a lot of extra memory for mmap'ed segments).

Not necessarily

That's good to know:) We use gcs to back up the segments, but in general it can be on a smaller node, no heavy network, cpu requirements?

if you are using uri based push, then technically the only load is for zk read/write and some small operational related tasks like retention, validation, segment assignment etc

Since we use gcs as the data directory that would mean we're using uri based push, right?

for your segment push job , is job type segmetpush or segmentUriPush

sounds good, thanks!