Details
-
Bug
-
Status: Resolved
-
Major
-
Resolution: Duplicate
-
None
-
None
-
None
-
None
Description
The namenode's handling of block reports should be done from a batch queue instead of while the rpc call from the client is waiting. I'd propose:
void blockReport(DataNodeRegistration, Block[]) throws IOException;
instead of passing back obsolete blocks they can be handed back on later heartbeat.
The blockReport calls should just queue the registration and list of blocks and there should be a thread that processes the block reports when possible.
The goal is to prevent rpc timeouts and therefore retransmissions of the block lists.
Attachments
Issue Links
- duplicates
-
HDFS-2379 0.20: Allow block reports to proceed without holding FSDataset lock
- Closed