[jira] [Commented] (SPARK-17556) Executor side broadcast for broadcast joins
[ https://issues.apache.org/jira/browse/SPARK-17556?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=17534937#comment-17534937 ] Chris Kimmel commented on SPARK-17556: -- [~jlaskowski] I believe the problem is that "Currently in Spark SQL, in order to perform a broadcast join, the driver must collect the result of an RDD and then broadcast it." So even if you don't collect a result to your driver node, the broadcast join mechanism nevertheless creates a lot of traffic to the driver node. > Executor side broadcast for broadcast joins > --- > > Key: SPARK-17556 > URL: https://issues.apache.org/jira/browse/SPARK-17556 > Project: Spark > Issue Type: New Feature > Components: Spark Core, SQL >Reporter: Reynold Xin >Priority: Major > Attachments: executor broadcast.pdf, executor-side-broadcast.pdf > > > Currently in Spark SQL, in order to perform a broadcast join, the driver must > collect the result of an RDD and then broadcast it. This introduces some > extra latency. It might be possible to broadcast directly from executors. -- This message was sent by Atlassian Jira (v8.20.7#820007) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-17556) Executor side broadcast for broadcast joins
[ https://issues.apache.org/jira/browse/SPARK-17556?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=17533465#comment-17533465 ] Jacek Laskowski commented on SPARK-17556: - Given: # "I'm running a large query with over 100,000 tasks." # "Total size of serialized results ... is bigger than spark.driver.maxResultSize". I think the issue is no a broadcast join but the size of the result (as computed by these 100k tasks). They have to report back to the driver and I can't think of a reason why a broadcast join would make it any worse? I must be missing something obvious (and chimed in to learn a bit about Spark SQL from you today :)) > Executor side broadcast for broadcast joins > --- > > Key: SPARK-17556 > URL: https://issues.apache.org/jira/browse/SPARK-17556 > Project: Spark > Issue Type: New Feature > Components: Spark Core, SQL >Reporter: Reynold Xin >Priority: Major > Attachments: executor broadcast.pdf, executor-side-broadcast.pdf > > > Currently in Spark SQL, in order to perform a broadcast join, the driver must > collect the result of an RDD and then broadcast it. This introduces some > extra latency. It might be possible to broadcast directly from executors. -- This message was sent by Atlassian Jira (v8.20.7#820007) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-17556) Executor side broadcast for broadcast joins
[ https://issues.apache.org/jira/browse/SPARK-17556?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=17533157#comment-17533157 ] Chris Kimmel commented on SPARK-17556: -- Bumping this. I'm running a large query with over 100,000 tasks. Broadcast joins are causing Spark to throw an error that says "Total size of serialized results ... is bigger than spark.driver.maxResultSize". Setting spark.driver.maxResultSize to 0 doesn't work; the traffic just swamps my driver node. I tried disabling broadcast joins entirely, but that creates other problems. Hopefully someone with the time and knowledge will be able to resolve this ticket. > Executor side broadcast for broadcast joins > --- > > Key: SPARK-17556 > URL: https://issues.apache.org/jira/browse/SPARK-17556 > Project: Spark > Issue Type: New Feature > Components: Spark Core, SQL >Reporter: Reynold Xin >Priority: Major > Attachments: executor broadcast.pdf, executor-side-broadcast.pdf > > > Currently in Spark SQL, in order to perform a broadcast join, the driver must > collect the result of an RDD and then broadcast it. This introduces some > extra latency. It might be possible to broadcast directly from executors. -- This message was sent by Atlassian Jira (v8.20.7#820007) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-17556) Executor side broadcast for broadcast joins
[ https://issues.apache.org/jira/browse/SPARK-17556?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=17198571#comment-17198571 ] L. C. Hsieh commented on SPARK-17556: - We will recently try to pick this up again. > Executor side broadcast for broadcast joins > --- > > Key: SPARK-17556 > URL: https://issues.apache.org/jira/browse/SPARK-17556 > Project: Spark > Issue Type: New Feature > Components: Spark Core, SQL >Reporter: Reynold Xin >Assignee: L. C. Hsieh >Priority: Major > Attachments: executor broadcast.pdf, executor-side-broadcast.pdf > > > Currently in Spark SQL, in order to perform a broadcast join, the driver must > collect the result of an RDD and then broadcast it. This introduces some > extra latency. It might be possible to broadcast directly from executors. -- 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
[jira] [Commented] (SPARK-17556) Executor side broadcast for broadcast joins
[ https://issues.apache.org/jira/browse/SPARK-17556?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=16788206#comment-16788206 ] Eyal Farago commented on SPARK-17556: - why was this abandoned? [~viirya]'s pull request seems promising. I think the last comment by [~LI,Xiao] applies for current implementation as well as executors hold the entire broadcast anyway (assuming they ran task that used it) - so memory footprint on the executors side doesn't change, re. performance regression in case of multiple smaller partitions this also applies for current implementation as the RDD partitions has to be calculated and transferred to the driver. one thing I personally think could be improved in [~viirya]'s PR was the requirement for the RDD to be pre-persisted, I think blocks could be evaluated in the mapPartition operation performed in the newly introduced RDD.broadcast method, this would have solved most comments by [~holdenk_amp] in the PR. > Executor side broadcast for broadcast joins > --- > > Key: SPARK-17556 > URL: https://issues.apache.org/jira/browse/SPARK-17556 > Project: Spark > Issue Type: New Feature > Components: Spark Core, SQL >Reporter: Reynold Xin >Priority: Major > Attachments: executor broadcast.pdf, executor-side-broadcast.pdf > > > Currently in Spark SQL, in order to perform a broadcast join, the driver must > collect the result of an RDD and then broadcast it. This introduces some > extra latency. It might be possible to broadcast directly from executors. -- This message was sent by Atlassian JIRA (v7.6.3#76005) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-17556) Executor side broadcast for broadcast joins
[ https://issues.apache.org/jira/browse/SPARK-17556?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=16770481#comment-16770481 ] t oo commented on SPARK-17556: -- please don't leave us [~scwf] > Executor side broadcast for broadcast joins > --- > > Key: SPARK-17556 > URL: https://issues.apache.org/jira/browse/SPARK-17556 > Project: Spark > Issue Type: New Feature > Components: Spark Core, SQL >Reporter: Reynold Xin >Priority: Major > Attachments: executor broadcast.pdf, executor-side-broadcast.pdf > > > Currently in Spark SQL, in order to perform a broadcast join, the driver must > collect the result of an RDD and then broadcast it. This introduces some > extra latency. It might be possible to broadcast directly from executors. -- This message was sent by Atlassian JIRA (v7.6.3#76005) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-17556) Executor side broadcast for broadcast joins
[ https://issues.apache.org/jira/browse/SPARK-17556?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=15937707#comment-15937707 ] Liang-Chi Hsieh commented on SPARK-17556: - We may need to change the Target Version/s for this. > Executor side broadcast for broadcast joins > --- > > Key: SPARK-17556 > URL: https://issues.apache.org/jira/browse/SPARK-17556 > Project: Spark > Issue Type: New Feature > Components: Spark Core, SQL >Reporter: Reynold Xin > Attachments: executor broadcast.pdf, executor-side-broadcast.pdf > > > Currently in Spark SQL, in order to perform a broadcast join, the driver must > collect the result of an RDD and then broadcast it. This introduces some > extra latency. It might be possible to broadcast directly from executors. -- This message was sent by Atlassian JIRA (v6.3.15#6346) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-17556) Executor side broadcast for broadcast joins
[ https://issues.apache.org/jira/browse/SPARK-17556?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=15537544#comment-15537544 ] Liang-Chi Hsieh commented on SPARK-17556: - Update the design document again to address some review comments. > Executor side broadcast for broadcast joins > --- > > Key: SPARK-17556 > URL: https://issues.apache.org/jira/browse/SPARK-17556 > Project: Spark > Issue Type: New Feature > Components: Spark Core, SQL >Reporter: Reynold Xin > Attachments: executor broadcast.pdf, executor-side-broadcast.pdf > > > Currently in Spark SQL, in order to perform a broadcast join, the driver must > collect the result of an RDD and then broadcast it. This introduces some > extra latency. It might be possible to broadcast directly from executors. -- This message was sent by Atlassian JIRA (v6.3.4#6332) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-17556) Executor side broadcast for broadcast joins
[ https://issues.apache.org/jira/browse/SPARK-17556?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=15534774#comment-15534774 ] Liang-Chi Hsieh commented on SPARK-17556: - Update the design document to add more description for using this feature and new config for it. > Executor side broadcast for broadcast joins > --- > > Key: SPARK-17556 > URL: https://issues.apache.org/jira/browse/SPARK-17556 > Project: Spark > Issue Type: New Feature > Components: Spark Core, SQL >Reporter: Reynold Xin > Attachments: executor broadcast.pdf, executor-side-broadcast.pdf > > > Currently in Spark SQL, in order to perform a broadcast join, the driver must > collect the result of an RDD and then broadcast it. This introduces some > extra latency. It might be possible to broadcast directly from executors. -- This message was sent by Atlassian JIRA (v6.3.4#6332) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-17556) Executor side broadcast for broadcast joins
[ https://issues.apache.org/jira/browse/SPARK-17556?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=15529476#comment-15529476 ] Liang-Chi Hsieh commented on SPARK-17556: - For example, assume we have 3 executors and the broadcasting object is split to 3 pieces too. You think executor side broadcast will end to 3 executors connecting to 3 executors. Let us see. t1: E1 E2 E3 p1 p2 p3 t2: E1 going to fetch p2 from E2, E2 going to fetch p3 from E3, E3 going to fetch p2 from E2 E1E2 E3 p1, p2p2, p3 p2, p3 t3: E1 going to fetch p3 from E2, E2 going to fetch p1 from E1, E3 going to fetch p1 from E1 E1E2 E3 p1, p2, p3p1, p2, p3 p1, p2, p3 Now all executors get all pieces of data. During the broadcast, E1 connected to E2, E2 connected to E1, E3, E3 connected to E1, E2. In above, E1 doesn't connect to E3. The simple analysis is based on the assumption that the operations is synchronize. But it already shows that the BitTorrent approach can relieve the all-to-all transferring problem. > Executor side broadcast for broadcast joins > --- > > Key: SPARK-17556 > URL: https://issues.apache.org/jira/browse/SPARK-17556 > Project: Spark > Issue Type: New Feature > Components: Spark Core, SQL >Reporter: Reynold Xin > Attachments: executor broadcast.pdf, executor-side-broadcast.pdf > > > Currently in Spark SQL, in order to perform a broadcast join, the driver must > collect the result of an RDD and then broadcast it. This introduces some > extra latency. It might be possible to broadcast directly from executors. -- This message was sent by Atlassian JIRA (v6.3.4#6332) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-17556) Executor side broadcast for broadcast joins
[ https://issues.apache.org/jira/browse/SPARK-17556?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=15529439#comment-15529439 ] Liang-Chi Hsieh commented on SPARK-17556: - Actually I am hesitant to add the feature to handle the case you said. One reason is, I don't think it is highly possible to have the case. It only occurs when the executor exists before its piece of object data is fetched by other executors. I prefer to add a config to enable the executor side broadcast and add the detailed document. > Executor side broadcast for broadcast joins > --- > > Key: SPARK-17556 > URL: https://issues.apache.org/jira/browse/SPARK-17556 > Project: Spark > Issue Type: New Feature > Components: Spark Core, SQL >Reporter: Reynold Xin > Attachments: executor broadcast.pdf, executor-side-broadcast.pdf > > > Currently in Spark SQL, in order to perform a broadcast join, the driver must > collect the result of an RDD and then broadcast it. This introduces some > extra latency. It might be possible to broadcast directly from executors. -- This message was sent by Atlassian JIRA (v6.3.4#6332) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-17556) Executor side broadcast for broadcast joins
[ https://issues.apache.org/jira/browse/SPARK-17556?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=15522817#comment-15522817 ] Apache Spark commented on SPARK-17556: -- User 'scwf' has created a pull request for this issue: https://github.com/apache/spark/pull/15240 > Executor side broadcast for broadcast joins > --- > > Key: SPARK-17556 > URL: https://issues.apache.org/jira/browse/SPARK-17556 > Project: Spark > Issue Type: New Feature > Components: Spark Core, SQL >Reporter: Reynold Xin > Attachments: executor broadcast.pdf, executor-side-broadcast.pdf > > > Currently in Spark SQL, in order to perform a broadcast join, the driver must > collect the result of an RDD and then broadcast it. This introduces some > extra latency. It might be possible to broadcast directly from executors. -- This message was sent by Atlassian JIRA (v6.3.4#6332) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-17556) Executor side broadcast for broadcast joins
[ https://issues.apache.org/jira/browse/SPARK-17556?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=15519497#comment-15519497 ] Yan commented on SPARK-17556: - For 2), I think BitTorrent won't help in the case of all-to-all transfers, unlike the one-to-all such as the driver-to-cluster broadcast, or few-to-all, transfers. Thanks. > Executor side broadcast for broadcast joins > --- > > Key: SPARK-17556 > URL: https://issues.apache.org/jira/browse/SPARK-17556 > Project: Spark > Issue Type: New Feature > Components: Spark Core, SQL >Reporter: Reynold Xin > Attachments: executor broadcast.pdf, executor-side-broadcast.pdf > > > Currently in Spark SQL, in order to perform a broadcast join, the driver must > collect the result of an RDD and then broadcast it. This introduces some > extra latency. It might be possible to broadcast directly from executors. -- This message was sent by Atlassian JIRA (v6.3.4#6332) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-17556) Executor side broadcast for broadcast joins
[ https://issues.apache.org/jira/browse/SPARK-17556?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=15518790#comment-15518790 ] Liang-Chi Hsieh commented on SPARK-17556: - For 1). It is true only if your driver is outside of the cluster. So you can avoid uploading data from the driver to the cluster. If it is in cluster mode, then I think it is no obvious difference between uploading data from the driver and any executor. For 2). I think it is not exactly correct. Basically we perform a BitTorrent-like approach to fetch block, the slaves do need to connect to all others by the end. > Executor side broadcast for broadcast joins > --- > > Key: SPARK-17556 > URL: https://issues.apache.org/jira/browse/SPARK-17556 > Project: Spark > Issue Type: New Feature > Components: Spark Core, SQL >Reporter: Reynold Xin > Attachments: executor broadcast.pdf, executor-side-broadcast.pdf > > > Currently in Spark SQL, in order to perform a broadcast join, the driver must > collect the result of an RDD and then broadcast it. This introduces some > extra latency. It might be possible to broadcast directly from executors. -- This message was sent by Atlassian JIRA (v6.3.4#6332) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-17556) Executor side broadcast for broadcast joins
[ https://issues.apache.org/jira/browse/SPARK-17556?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=15518598#comment-15518598 ] Yan commented on SPARK-17556: - A few comments of mine are as follows: 1) The "one-executor collection" approach is different from the driver-side collection and broadcasting, in that it avoids uploading data from the driver back to cluster. The primary concern of the "one-executor collection" approach, as pointed out, is that the sole executor could get bottlenecked similar to the latency issue with the "driver-side collection" approach, to a large degree; 2) The "all-executor collection" approach is more balanced and scalable, but it might suffer from the network storming since all slaves needs to connect to all others. 3) the real issue is the repeated, and thus wasted, work of collection of pieces of the broadcast data by multiple collectors/broadcasters, against the extended latency if the collection/broadcasting is performed once and for all. This is actually not quite different from the scenario of multiple- vs single-reducer in a map/reduce execution. Final output from a single reducer is ready to use; while those from multiple-reducers require final assemblies by the end users, particularly if the final result is to be organized, e.g., totally ordered. But using multiple-reducers is more scalable, balanced and likely faster. 4) It's probably good to have a configurable # of executors acting as collectors/broadcasters, each of which just collects and broadcasts a portion of the broadcast table for the final join executions. > Executor side broadcast for broadcast joins > --- > > Key: SPARK-17556 > URL: https://issues.apache.org/jira/browse/SPARK-17556 > Project: Spark > Issue Type: New Feature > Components: Spark Core, SQL >Reporter: Reynold Xin > Attachments: executor broadcast.pdf > > > Currently in Spark SQL, in order to perform a broadcast join, the driver must > collect the result of an RDD and then broadcast it. This introduces some > extra latency. It might be possible to broadcast directly from executors. -- This message was sent by Atlassian JIRA (v6.3.4#6332) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-17556) Executor side broadcast for broadcast joins
[ https://issues.apache.org/jira/browse/SPARK-17556?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=15516772#comment-15516772 ] Fei Wang commented on SPARK-17556: -- [~viirya] in this case how about notify driver to re-persist the rdd? > Executor side broadcast for broadcast joins > --- > > Key: SPARK-17556 > URL: https://issues.apache.org/jira/browse/SPARK-17556 > Project: Spark > Issue Type: New Feature > Components: Spark Core, SQL >Reporter: Reynold Xin > Attachments: executor broadcast.pdf > > > Currently in Spark SQL, in order to perform a broadcast join, the driver must > collect the result of an RDD and then broadcast it. This introduces some > extra latency. It might be possible to broadcast directly from executors. -- This message was sent by Atlassian JIRA (v6.3.4#6332) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-17556) Executor side broadcast for broadcast joins
[ https://issues.apache.org/jira/browse/SPARK-17556?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=15516732#comment-15516732 ] Fei Wang commented on SPARK-17556: -- That's a good point! In your solution, the broadcast rdd must persist first right? How you handle the case for executors lost (all the replication of a piece lost)? > Executor side broadcast for broadcast joins > --- > > Key: SPARK-17556 > URL: https://issues.apache.org/jira/browse/SPARK-17556 > Project: Spark > Issue Type: New Feature > Components: Spark Core, SQL >Reporter: Reynold Xin > Attachments: executor broadcast.pdf > > > Currently in Spark SQL, in order to perform a broadcast join, the driver must > collect the result of an RDD and then broadcast it. This introduces some > extra latency. It might be possible to broadcast directly from executors. -- This message was sent by Atlassian JIRA (v6.3.4#6332) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-17556) Executor side broadcast for broadcast joins
[ https://issues.apache.org/jira/browse/SPARK-17556?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=15516668#comment-15516668 ] Liang-Chi Hsieh commented on SPARK-17556: - No. It doesn't. I think the point is not only the overhead to the driver, but also the extra latency mentioned in the jira description. With the solution in my PR, all executors are going to fetch RDD content from other executors. It doesn't do "collect the data first and then broadcast it". > Executor side broadcast for broadcast joins > --- > > Key: SPARK-17556 > URL: https://issues.apache.org/jira/browse/SPARK-17556 > Project: Spark > Issue Type: New Feature > Components: Spark Core, SQL >Reporter: Reynold Xin > Attachments: executor broadcast.pdf > > > Currently in Spark SQL, in order to perform a broadcast join, the driver must > collect the result of an RDD and then broadcast it. This introduces some > extra latency. It might be possible to broadcast directly from executors. -- This message was sent by Atlassian JIRA (v6.3.4#6332) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-17556) Executor side broadcast for broadcast joins
[ https://issues.apache.org/jira/browse/SPARK-17556?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=15516639#comment-15516639 ] Fei Wang commented on SPARK-17556: -- Yes, the main different is is does not introduce overhead to driver, for broadcast the executor do need all the result of an RDD, i task a look of your PR, i think you also collect all the result of that rdd to executor, right? > Executor side broadcast for broadcast joins > --- > > Key: SPARK-17556 > URL: https://issues.apache.org/jira/browse/SPARK-17556 > Project: Spark > Issue Type: New Feature > Components: Spark Core, SQL >Reporter: Reynold Xin > Attachments: executor broadcast.pdf > > > Currently in Spark SQL, in order to perform a broadcast join, the driver must > collect the result of an RDD and then broadcast it. This introduces some > extra latency. It might be possible to broadcast directly from executors. -- This message was sent by Atlassian JIRA (v6.3.4#6332) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-17556) Executor side broadcast for broadcast joins
[ https://issues.apache.org/jira/browse/SPARK-17556?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=15516577#comment-15516577 ] Liang-Chi Hsieh commented on SPARK-17556: - In other words, from the jira description we say "the driver must collect the result of an RDD and then broadcast it." You solution is just "the executor must collect the result of an RDD and then broadcast it." > Executor side broadcast for broadcast joins > --- > > Key: SPARK-17556 > URL: https://issues.apache.org/jira/browse/SPARK-17556 > Project: Spark > Issue Type: New Feature > Components: Spark Core, SQL >Reporter: Reynold Xin > Attachments: executor broadcast.pdf > > > Currently in Spark SQL, in order to perform a broadcast join, the driver must > collect the result of an RDD and then broadcast it. This introduces some > extra latency. It might be possible to broadcast directly from executors. -- This message was sent by Atlassian JIRA (v6.3.4#6332) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-17556) Executor side broadcast for broadcast joins
[ https://issues.apache.org/jira/browse/SPARK-17556?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=15516568#comment-15516568 ] Liang-Chi Hsieh commented on SPARK-17556: - OK. You create the broadcast object on one executor. So, is it any different than collecting data to the driver? You just replace the driver with one executor as the data collector... > Executor side broadcast for broadcast joins > --- > > Key: SPARK-17556 > URL: https://issues.apache.org/jira/browse/SPARK-17556 > Project: Spark > Issue Type: New Feature > Components: Spark Core, SQL >Reporter: Reynold Xin > Attachments: executor broadcast.pdf > > > Currently in Spark SQL, in order to perform a broadcast join, the driver must > collect the result of an RDD and then broadcast it. This introduces some > extra latency. It might be possible to broadcast directly from executors. -- This message was sent by Atlassian JIRA (v6.3.4#6332) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-17556) Executor side broadcast for broadcast joins
[ https://issues.apache.org/jira/browse/SPARK-17556?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=15516545#comment-15516545 ] Fei Wang commented on SPARK-17556: -- Not correct, I just collect the broadcast ref to the driver but not the data:) . > Executor side broadcast for broadcast joins > --- > > Key: SPARK-17556 > URL: https://issues.apache.org/jira/browse/SPARK-17556 > Project: Spark > Issue Type: New Feature > Components: Spark Core, SQL >Reporter: Reynold Xin > Attachments: executor broadcast.pdf > > > Currently in Spark SQL, in order to perform a broadcast join, the driver must > collect the result of an RDD and then broadcast it. This introduces some > extra latency. It might be possible to broadcast directly from executors. -- This message was sent by Atlassian JIRA (v6.3.4#6332) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-17556) Executor side broadcast for broadcast joins
[ https://issues.apache.org/jira/browse/SPARK-17556?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=15516546#comment-15516546 ] Fei Wang commented on SPARK-17556: -- Not correct, I just collect the broadcast ref to the driver but not the data:) . > Executor side broadcast for broadcast joins > --- > > Key: SPARK-17556 > URL: https://issues.apache.org/jira/browse/SPARK-17556 > Project: Spark > Issue Type: New Feature > Components: Spark Core, SQL >Reporter: Reynold Xin > Attachments: executor broadcast.pdf > > > Currently in Spark SQL, in order to perform a broadcast join, the driver must > collect the result of an RDD and then broadcast it. This introduces some > extra latency. It might be possible to broadcast directly from executors. -- This message was sent by Atlassian JIRA (v6.3.4#6332) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-17556) Executor side broadcast for broadcast joins
[ https://issues.apache.org/jira/browse/SPARK-17556?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=15516490#comment-15516490 ] Liang-Chi Hsieh commented on SPARK-17556: - [~Fei Wang] I quickly go through your design doc. Looks like you still need to collect the content of RDD to the driver. I don't think it is not executor side broadcast means in this jira's description. You can refer to the PR I submitted with which we don't need to collect the RDD back to the driver. > Executor side broadcast for broadcast joins > --- > > Key: SPARK-17556 > URL: https://issues.apache.org/jira/browse/SPARK-17556 > Project: Spark > Issue Type: New Feature > Components: Spark Core, SQL >Reporter: Reynold Xin > Attachments: executor broadcast.pdf > > > Currently in Spark SQL, in order to perform a broadcast join, the driver must > collect the result of an RDD and then broadcast it. This introduces some > extra latency. It might be possible to broadcast directly from executors. -- This message was sent by Atlassian JIRA (v6.3.4#6332) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-17556) Executor side broadcast for broadcast joins
[ https://issues.apache.org/jira/browse/SPARK-17556?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=15516478#comment-15516478 ] Liang-Chi Hsieh commented on SPARK-17556: - [~scwf]I already submitted a PR for this. Can you also help to review it? Thanks. > Executor side broadcast for broadcast joins > --- > > Key: SPARK-17556 > URL: https://issues.apache.org/jira/browse/SPARK-17556 > Project: Spark > Issue Type: New Feature > Components: Spark Core, SQL >Reporter: Reynold Xin > Attachments: executor broadcast.pdf > > > Currently in Spark SQL, in order to perform a broadcast join, the driver must > collect the result of an RDD and then broadcast it. This introduces some > extra latency. It might be possible to broadcast directly from executors. -- This message was sent by Atlassian JIRA (v6.3.4#6332) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-17556) Executor side broadcast for broadcast joins
[ https://issues.apache.org/jira/browse/SPARK-17556?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=15516386#comment-15516386 ] Fei Wang commented on SPARK-17556: -- [~rxin] attached a design doc for the executor based broadcast. Will soon file a PR for this. [~viirya] We have a executor based broadcast implementation in our inner produce system which is based on the design doc i attached. Can you help to review this, thanks. > Executor side broadcast for broadcast joins > --- > > Key: SPARK-17556 > URL: https://issues.apache.org/jira/browse/SPARK-17556 > Project: Spark > Issue Type: New Feature > Components: Spark Core, SQL >Reporter: Reynold Xin > Attachments: executor broadcast.pdf > > > Currently in Spark SQL, in order to perform a broadcast join, the driver must > collect the result of an RDD and then broadcast it. This introduces some > extra latency. It might be possible to broadcast directly from executors. -- This message was sent by Atlassian JIRA (v6.3.4#6332) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-17556) Executor side broadcast for broadcast joins
[ https://issues.apache.org/jira/browse/SPARK-17556?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=15509228#comment-15509228 ] Apache Spark commented on SPARK-17556: -- User 'viirya' has created a pull request for this issue: https://github.com/apache/spark/pull/15178 > Executor side broadcast for broadcast joins > --- > > Key: SPARK-17556 > URL: https://issues.apache.org/jira/browse/SPARK-17556 > Project: Spark > Issue Type: New Feature > Components: Spark Core, SQL >Reporter: Reynold Xin > > Currently in Spark SQL, in order to perform a broadcast join, the driver must > collect the result of an RDD and then broadcast it. This introduces some > extra latency. It might be possible to broadcast directly from executors. -- This message was sent by Atlassian JIRA (v6.3.4#6332) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
