[
https://issues.apache.org/jira/browse/HADOOP-19986?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Jose Luis López updated HADOOP-19986:
-------------------------------------
Description:
{{TestBalancerWithHANameNodes#testBalancerWithObserverWithFailedNode}}
intermittently
times out after 180s. The hang is in teardown, not in the balancer.
h3. Cause
{{MiniDFSCluster.shutdownDataNodes()}} tears DataNodes down strictly serially –
stop one, join it, move to the next. The DataNodes not yet reached keep retrying
the NameNode the test has already killed.
DataNodes in a single JVM share an {{ipc.Client}} through {{{}ClientCache{}}},
and so
share its per-address {{Connection}} objects. A surviving DataNode's
{{BPServiceActor}} holds a {{Connection}} monitor across its connect-retry
sleeps
in {{{}handleConnectionFailure{}}}, while an actor of the DataNode being joined
sits
BLOCKED on that same monitor in {{{}Client.addCall{}}}. {{Thread.interrupt()}}
cannot
dislodge a BLOCKED thread, so the {{stop()}} issued by
{{BlockPoolManager.shutDownAll}} is ineffective and the join waits on scheduling
luck: the holder releases and re-acquires roughly every 2s, and unfair monitors
starve the blocked thread.
Measured on a 2-core runner: DataNode 2 took 165s to shut down, after which DN1
and
DN0 finished in ~10ms.
h3. Fix
Signal every DataNode before joining any of them, so the monitor holder aborts
its
sleep and releases. Adds {{BlockPoolManager#signalShutDownAll}} (the
stop-without-join
half of the existing {{{}shutDownAll{}}}) and
{{{}DataNode#signalBlockPoolShutdown{}}}.
{{stop()}} is idempotent, so the existing per-DataNode shutdown path is
unchanged.
h3. Verification
several (at least 3) 20 runs of the test on a 2-core GitHub runner, before and
after:
|| ||before||after||
|failures|4/20|0/20|
|wall-clock spread|53s - 182s|44s - 48s|
|test method time|up to 181s|14.4s - 15.0s|
The two sub-timeout outliers before the fix (116.7s, 113.9s) are the same
starvation landing under the deadline, so the test was silently burning ~2
minutes
on runs that reported as passing.
was:pending
> MiniDFSCluster.shutdownDataNodes() should signal all DataNodes before joining
> any
> ---------------------------------------------------------------------------------
>
> Key: HADOOP-19986
> URL: https://issues.apache.org/jira/browse/HADOOP-19986
> Project: Hadoop Common
> Issue Type: Sub-task
> Components: hdfs, test
> Reporter: Jose Luis López
> Priority: Minor
>
> {{TestBalancerWithHANameNodes#testBalancerWithObserverWithFailedNode}}
> intermittently
> times out after 180s. The hang is in teardown, not in the balancer.
> h3. Cause
> {{MiniDFSCluster.shutdownDataNodes()}} tears DataNodes down strictly serially
> –
> stop one, join it, move to the next. The DataNodes not yet reached keep
> retrying
> the NameNode the test has already killed.
> DataNodes in a single JVM share an {{ipc.Client}} through
> {{{}ClientCache{}}}, and so
> share its per-address {{Connection}} objects. A surviving DataNode's
> {{BPServiceActor}} holds a {{Connection}} monitor across its connect-retry
> sleeps
> in {{{}handleConnectionFailure{}}}, while an actor of the DataNode being
> joined sits
> BLOCKED on that same monitor in {{{}Client.addCall{}}}.
> {{Thread.interrupt()}} cannot
> dislodge a BLOCKED thread, so the {{stop()}} issued by
> {{BlockPoolManager.shutDownAll}} is ineffective and the join waits on
> scheduling
> luck: the holder releases and re-acquires roughly every 2s, and unfair
> monitors
> starve the blocked thread.
> Measured on a 2-core runner: DataNode 2 took 165s to shut down, after which
> DN1 and
> DN0 finished in ~10ms.
> h3. Fix
> Signal every DataNode before joining any of them, so the monitor holder
> aborts its
> sleep and releases. Adds {{BlockPoolManager#signalShutDownAll}} (the
> stop-without-join
> half of the existing {{{}shutDownAll{}}}) and
> {{{}DataNode#signalBlockPoolShutdown{}}}.
> {{stop()}} is idempotent, so the existing per-DataNode shutdown path is
> unchanged.
> h3. Verification
> several (at least 3) 20 runs of the test on a 2-core GitHub runner, before
> and after:
> || ||before||after||
> |failures|4/20|0/20|
> |wall-clock spread|53s - 182s|44s - 48s|
> |test method time|up to 181s|14.4s - 15.0s|
> The two sub-timeout outliers before the fix (116.7s, 113.9s) are the same
> starvation landing under the deadline, so the test was silently burning ~2
> minutes
> on runs that reported as passing.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
