Shashikant Banerjee created HDDS-3227:
-----------------------------------------
Summary: Ensure eviction of stateMachineData from cache only when
both followers catch up
Key: HDDS-3227
URL: https://issues.apache.org/jira/browse/HDDS-3227
Project: Hadoop Distributed Data Store
Issue Type: Bug
Components: Ozone Datanode
Reporter: Shashikant Banerjee
Assignee: Shashikant Banerjee
Fix For: 0.6.0
Currently, the data in the StateMachineCache is evicted as soon as the
applyTransaction call is issued for a transaction in Ratis. In our testing with
keys in few kbs of size, it was figured that the data is evicted from the cache
before append requests can be processed in a slightly slow follower thereby
making leader read the chunk data from underlying fs/disk very frequently. This
leads to slowing down the leader as well as well as overall throughput of the
pipeline.
The idea here is to ensure the data is evicted from the cache only when both
followers have caught up with the match index. If a follower is really slow, it
will eventually be marked slow after nodeFailureTimeout and pipeline will be
destroyed.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
