holdenk commented on pull request #29855: URL: https://github.com/apache/spark/pull/29855#issuecomment-702227373
For a bit more context: I am ok with us using external shuffle service on K8s, I just think that we should keep in mind that for many configurations it isn’t a good fit due to limitations around volumes so we should be sure to document it clearly and when possible support alternatives. On Tue, Sep 29, 2020 at 2:57 PM Min Shen <[email protected]> wrote: > > > *@Victsm* commented on this pull request. > > > > > ------------------------------ > > > > > In > common/network-shuffle/src/main/java/org/apache/spark/network/shuffle/MergedShuffleFileManager.java > <https://github.com/apache/spark/pull/29855#discussion_r497084532>: > > > > + > > +import java.io.IOException; > > + > > +import org.apache.spark.network.buffer.ManagedBuffer; > > +import org.apache.spark.network.client.StreamCallbackWithID; > > +import org.apache.spark.network.shuffle.protocol.FinalizeShuffleMerge; > > +import org.apache.spark.network.shuffle.protocol.MergeStatuses; > > +import org.apache.spark.network.shuffle.protocol.PushBlockStream; > > + > > + > > +/** > > + * The MergedShuffleFileManager is used to process push based shuffle when enabled. It works > > + * along side {@link ExternalBlockHandler} and serves as an RPCHandler for > > + * {@link org.apache.spark.network.server.RpcHandler#receiveStream}, where it processes the > > + * remotely pushed streams of shuffle blocks to merge them into merged shuffle files. Right > > + * now, push based shuffle can only be enabled when external shuffle service in YARN mode > > > > That comment might not be clear enough. > > > It's rather that right now we only implemented support for push-based > shuffle when external shuffle service in YARN mode is used. > > > Extending to standalone mode is definitely doable, considering the > similarity of the usage of ESS in that mode compared with YARN. > > > For K8S, we discussed briefly with @dbtsai <https://github.com/dbtsai> > and @holdenk <https://github.com/holdenk> before. > > > They prefer to not use external shuffle service in K8S due to the > limitation of mounting persistent volume in that environment, i.e. the > volume can only be mounted in one pod for I/O. > > > So some work might be needed there to extend support into K8S. > > > There's also the need of supporting disaggregated shuffle storage, which > we are also evaluating. > > > I have created SPARK-32925 > <https://issues.apache.org/jira/browse/SPARK-32925> for discussions on > this. > > > > > — > You are receiving this because you were mentioned. > Reply to this email directly, view it on GitHub > <https://github.com/apache/spark/pull/29855#discussion_r497084532>, or > unsubscribe > <https://github.com/notifications/unsubscribe-auth/AAAOT5LN25SS2CXYIAKPAPLSIJJ6HANCNFSM4RXN5ZFQ> > . > > > -- Cell : 425-233-8271 ---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: [email protected] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
