Details
-
Bug
-
Status: Closed
-
Minor
-
Resolution: Invalid
-
0.20.5, 0.89.20100621
-
None
-
None
Description
Now that the region servers can heartbeat back to the master as soon as they have a message (region open, close, etc), they also recompute their metrics. I see two problems:
- The metrics are recomputed way too often, and by the looks of that code it seems quite intense.
- Our UIs are expecting that the value returned by the metrics computation is per second, but when regions are moving a RS can ping back a few times within a single second. This gives a very wrong view of the load.
I think it's time we move metrics computation in it's own separate thread.