Hitesh Sharma created TEZ-4011:
----------------------------------
Summary: Don't consider task attempt failed if container fails to
launch and thus times out
Key: TEZ-4011
URL: https://issues.apache.org/jira/browse/TEZ-4011
Project: Apache Tez
Issue Type: Improvement
Reporter: Hitesh Sharma
If a container fails to start (never heartbeats back to the AM during launch)
then the container is considered timed out and the task attempt assigned to the
container is failed. This is counted towards the failure count for the task. In
some environments this may not be desirable (due to high probability of these
events) as the task itself never got the chance to run, but since it counts
towards the max task attempts it could lead to a failure. If we configure the
timeout value for container heartbeat to be bigger then we have job slowness as
the job just waits for the container to launch. An alternative here is to
instead kill the task attempt if the container times out during launch. This is
because the killed containers are not counted towards task attempt failures and
allows one to have a bit more aggressive launch timeout. This behavior would be
off by default and users could opt into it if it makes more sense to do so in
their environments.
Thoughts?
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
