You are viewing a plain text version of this content. The canonical link for it is here.
Posted to yarn-issues@hadoop.apache.org by "Suma Shivaprasad (JIRA)" <ji...@apache.org> on 2019/01/24 18:42:00 UTC
[jira] [Commented] (YARN-8901) Restart "NEVER" policy does not work
with component dependency
[ https://issues.apache.org/jira/browse/YARN-8901?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16751421#comment-16751421 ]
Suma Shivaprasad commented on YARN-8901:
----------------------------------------
Currently downstream components that depend on components with restartPolicy=NEVER/ON_FAILURE are not started until they finish. But this breaks the notion/assumption that when downstream components can be started when the upstream component reaches READY state. Reverting the behaviour for restartPolicy = NEVER/ON_FAILURE to be the same as ALWAYS restart policy in the attached patch.
If downstream components need to start up only after a certain condition is met, then that should be supported as a separate feature in the downstream component and can be addressed as part of another jira.
> Restart "NEVER" policy does not work with component dependency
> --------------------------------------------------------------
>
> Key: YARN-8901
> URL: https://issues.apache.org/jira/browse/YARN-8901
> Project: Hadoop YARN
> Issue Type: Bug
> Affects Versions: 3.1.1
> Reporter: Yesha Vora
> Assignee: Suma Shivaprasad
> Priority: Critical
>
> Scenario:
> 1) Launch an application with two components. master and worker. Here, worker is dependent on master. ( Worker should be launched only after master is launched )
> 2) Set restart_policy = NEVER for both master and worker.
> {code:title=sample launch.json}
> {
> "name": "mawo-hadoop-ut",
> "artifact": {
> "type": "DOCKER",
> "id": "xxx"
> },
> "configuration": {
> "env": {
> "YARN_CONTAINER_RUNTIME_DOCKER_CONTAINER_NETWORK": "hadoop"
> },
> "properties": {
> "docker.network": "hadoop"
> }
> },
> "components": [{
> "dependencies": [],
> "resource": {
> "memory": "2048",
> "cpus": "1"
> },
> "name": "master",
> "run_privileged_container": true,
> "number_of_containers": 1,
> "launch_command": "start master",
> "restart_policy": "NEVER",
> }, {
> "dependencies": ["master"],
> "resource": {
> "memory": "8072",
> "cpus": "1"
> },
> "name": "worker",
> "run_privileged_container": true,
> "number_of_containers": 10,
> "launch_command": "start worker",
> "restart_policy": "NEVER",
> }],
> "lifetime": -1,
> "version": 1.0
> }{code}
> When restart policy is selected to NEVER, AM never launches Worker component. It get stuck with below message.
> {code}
> 2018-10-17 15:11:58,560 [Component dispatcher] INFO component.Component - [COMPONENT master] Transitioned from FLEXING to STABLE on CHECK_STABLE event.
> 2018-10-17 15:11:58,560 [pool-7-thread-1] INFO instance.ComponentInstance - [COMPINSTANCE master-0 : container_e41_1539027682947_0020_01_000002] Transitioned from STARTED to READY on BECOME_READY event
> 2018-10-17 15:11:58,560 [pool-7-thread-1] INFO component.Component - [COMPONENT worker]: Dependency master not satisfied, only 1 of 1 instances are ready or the dependent component has not completed
> 2018-10-17 15:12:28,556 [pool-7-thread-1] INFO component.Component - [COMPONENT worker]: Dependency master not satisfied, only 1 of 1 instances are ready or the dependent component has not completed
> 2018-10-17 15:12:58,556 [pool-7-thread-1] INFO component.Component - [COMPONENT worker]: Dependency master not satisfied, only 1 of 1 instances are ready or the dependent component has not completed
> 2018-10-17 15:13:28,556 [pool-7-thread-1] INFO component.Component - [COMPONENT worker]: Dependency master not satisfied, only 1 of 1 instances are ready or the dependent component has not completed
> 2018-10-17 15:13:58,556 [pool-7-thread-1] INFO component.Component - [COMPONENT worker]: Dependency master not satisfied, only 1 of 1 instances are ready or the dependent component has not completed
> 2018-10-17 15:14:28,556 [pool-7-thread-1] INFO component.Component - [COMPONENT worker]: Dependency master not satisfied, only 1 of 1 instances are ready or the dependent component has not completed {code}
> 'NEVER' restart policy expects master component to be finished before starting workers. Master component can not finish the job without workers. Thus, it create a deadlock.
> The logic for 'NEVER' restart policy should be fixed to allow worker components to be launched as soon as master component is in READY state.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: yarn-issues-unsubscribe@hadoop.apache.org
For additional commands, e-mail: yarn-issues-help@hadoop.apache.org