Mike Percy created KUDU-1991:
--------------------------------
Summary: Master does not retry TS maintenance tasks if TS not
registered
Key: KUDU-1991
URL: https://issues.apache.org/jira/browse/KUDU-1991
Project: Kudu
Issue Type: Bug
Components: master
Affects Versions: 1.3.1
Reporter: Mike Percy
If a master had just started up and gets a heartbeat from a tablet server that
triggers some action on another tablet server (for example, tombstoning an
evicted replica or adding a new replica to an under-replicated tablet config)
then if the target tablet server (for example, the replica that the master
thought was the leader) has not yet registered with the master (since its
restart) the action will fail to be sent and will not be retried.
This is because:
# There is a logic error in the catalog manager task management code that
assumes all tablet servers have registered with the master at the time a task
is started; and
# These kinds of tasks are edge-triggered (based on a response to a tablet
report) instead of level-triggered (based on periodic state polling) on the
master side.
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
