Details
-
New Feature
-
Status: Resolved
-
Minor
-
Resolution: Fixed
-
None
-
None
-
None
-
None
Description
This relates to allowing a resource manager (e.g., hadoop on demand) to grow and (rarely) shrink jobs on the fly.
Growing is already supported. Shrinking could be done in 2 ways - (1) consider the machine dead and allow speculative execution to take care of it or (2) moving the existing map outputs from that machine somewhere else (another machine, dfs) - "task tracker checkpointing"
In the case of IO only intensive jobs, checkpointing the tasktracker doesn't do much for you. But, in the case of CPU or other scarce resource (e.g., a DB or Webpage cache...), the checkpointing could be very useful. The question is how often is this the case and how useful?
Attachments
Issue Links
- is related to
-
MAPREDUCE-443 snapshot a map-reduce to DFS ... and restore
- Resolved