During heavy write loads, the CheckpointDestroyerTask has been observed to be taking many seconds to execute:
This results in a large number of Checkpoints being outstanding for deletion, along with large amounts of memory being tied up in them (pending_destruction_bytes):
This causes much lower resident ratio than we would otherwise have, and increases tempOOM backoffs sent to clients.
It also results in the NonIO thread being tied up and otherwise unable to run other NonIO tasks.
|For Gerrit Dashboard: MB-49983|
|167243,3||MB-49983: MemoryTrackingAllocator: support sized deallocation||master||platform||Status: MERGED||+2||+1|