Manu Zhang created SPARK-31219:
----------------------------------
Summary: YarnShuffleService doesn't close idle netty channel
Key: SPARK-31219
URL: https://issues.apache.org/jira/browse/SPARK-31219
Project: Spark
Issue Type: Improvement
Components: Shuffle
Affects Versions: 2.4.5, 3.0.0
Reporter: Manu Zhang
Recently, we find our YarnShuffleService has a lot of [half-open
connections|https://blog.stephencleary.com/2009/05/detection-of-half-open-dropped.html]
where shuffle servers' connections are active while clients have already
closed.
For example, from server's `ss -nt sport = :7337` output we have
{code:java}
ESTAB 0 0 server:7337 client:port
{code}
However, on client `ss -nt dport =: 7337 | grep server` would return nothing.
Looking at the code, `YarnShuffleService` creates a `TransportContext` with
`closeIdleConnections` set to false.
{code:java}
public class YarnShuffleService extends AuxiliaryService {
...
@Override protected void serviceInit(Configuration conf) throws Exception {
...
transportContext = new TransportContext(transportConf, blockHandler);
...
}
...
}
public class TransportContext implements Closeable {
...
public TransportContext(TransportConf conf, RpcHandler rpcHandler) {
this(conf, rpcHandler, false, false);
}
public TransportContext(TransportConf conf, RpcHandler rpcHandler, boolean
closeIdleConnections) {
this(conf, rpcHandler, closeIdleConnections, false);
}
...
}{code}
Hence, it's possible the channel may never get closed at server side if the
server misses the event that the client has closed it.
I find that parameter is true for `ExternalShuffleService`.
Is there any reason for the difference here ? Will it be valuable to add a
configuration to allow enabling closeIdleConnections ?
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
