[jira] [Commented] (SPARK-2389) globally shared SparkContext / shared Spark application
[ https://issues.apache.org/jira/browse/SPARK-2389?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=14703595#comment-14703595 ] Ashish Rawat commented on SPARK-2389: - Hi [~pwendell], We are facing exactly the problem which you mentioned. And we are looking for exactly the same solution that you mentioned i.e. Driver HA :) I have a few questions/comments on the perspective you shared: 1. If we started to go down this path, we'd need to do things like define a standard serialization format for the RDD data, a global namespace for RDD's, persistence, etc. And then you're building a filesystem. You only need to preserve the RDD metainfo and not the actual RDDs, so there should not be any complexity of serialization format for RDD data. 2. Although the cache data is recoverable, but how to reduce the latency of building back a cache of TBs, for a live application? 3. Can we not just prevent executors from shutting, preserve some important driver info and connect back the driver? Or provide a Hot standby for driver? globally shared SparkContext / shared Spark application - Key: SPARK-2389 URL: https://issues.apache.org/jira/browse/SPARK-2389 Project: Spark Issue Type: Improvement Components: Spark Core Reporter: Robert Stupp The documentation (in Cluster Mode Overview) cites: bq. Each application gets its own executor processes, which *stay up for the duration of the whole application* and run tasks in multiple threads. This has the benefit of isolating applications from each other, on both the scheduling side (each driver schedules its own tasks) and executor side (tasks from different applications run in different JVMs). However, it also means that *data cannot be shared* across different Spark applications (instances of SparkContext) without writing it to an external storage system. IMO this is a limitation that should be lifted to support any number of --driver-- client processes to share executors and to share (persistent / cached) data. This is especially useful if you have a bunch of frontend servers (dump web app servers) that want to use Spark as a _big computing machine_. Most important is the fact that Spark is quite good in caching/persisting data in memory / on disk thus removing load from backend data stores. Means: it would be really great to let different --driver-- client JVMs operate on the same RDDs and benefit from Spark's caching/persistence. It would however introduce some administration mechanisms to * start a shared context * update the executor configuration (# of worker nodes, # of cpus, etc) on the fly * stop a shared context Even conventional batch MR applications would benefit if ran fequently against the same data set. As an implicit requirement, RDD persistence could get a TTL for its materialized state. With such a feature the overall performance of today's web applications could then be increased by adding more web app servers, more spark nodes, more nosql nodes etc -- 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-2389) globally shared SparkContext / shared Spark application
[ https://issues.apache.org/jira/browse/SPARK-2389?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=14291818#comment-14291818 ] Robert Stupp commented on SPARK-2389: - [~srowen] yes, the problem is that drivers cannot share RDDs. IMHO there are a lot of valid scenarios that can benefit from multiple drivers using shared RDDs. globally shared SparkContext / shared Spark application - Key: SPARK-2389 URL: https://issues.apache.org/jira/browse/SPARK-2389 Project: Spark Issue Type: Improvement Components: Spark Core Reporter: Robert Stupp The documentation (in Cluster Mode Overview) cites: bq. Each application gets its own executor processes, which *stay up for the duration of the whole application* and run tasks in multiple threads. This has the benefit of isolating applications from each other, on both the scheduling side (each driver schedules its own tasks) and executor side (tasks from different applications run in different JVMs). However, it also means that *data cannot be shared* across different Spark applications (instances of SparkContext) without writing it to an external storage system. IMO this is a limitation that should be lifted to support any number of --driver-- client processes to share executors and to share (persistent / cached) data. This is especially useful if you have a bunch of frontend servers (dump web app servers) that want to use Spark as a _big computing machine_. Most important is the fact that Spark is quite good in caching/persisting data in memory / on disk thus removing load from backend data stores. Means: it would be really great to let different --driver-- client JVMs operate on the same RDDs and benefit from Spark's caching/persistence. It would however introduce some administration mechanisms to * start a shared context * update the executor configuration (# of worker nodes, # of cpus, etc) on the fly * stop a shared context Even conventional batch MR applications would benefit if ran fequently against the same data set. As an implicit requirement, RDD persistence could get a TTL for its materialized state. With such a feature the overall performance of today's web applications could then be increased by adding more web app servers, more spark nodes, more nosql nodes etc -- 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-2389) globally shared SparkContext / shared Spark application
[ https://issues.apache.org/jira/browse/SPARK-2389?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=14291797#comment-14291797 ] Robert Stupp commented on SPARK-2389: - bq. That aside, why doesn't it scale? Simply because it's just a single Spark client. If that machine's at its limit for whatever reason (VM memory, OS resources, CPU, network, ...), that's it. Sure, you can run multiple drivers - but each has its own, private set of data. IMO separate preloading is nice for some applications. But data is usually not immutable. By example: * Imagine an application that provides offers for flights worldwide. It's a huge amount of data and a huge amount of processing. It cannot be simply preloaded - prices for tickets vary from minute to minute based on booking status etc etc etc * Overall data set is quite big * Overall load is too big for a single driver to handle - imagine thousands of offer requests per second * Failure of a single driver is an absolute no-go * All clients have to access the same set of data * Preloading is just impossible during runtime (just at initial deployment) So - a suitable approach would be to have: * a Spark cluster holding all the RDDs and doing all offer and booking related operations * a set of Spark clients to abstract Spark from the rest of the application * a huge number of non-uniform frontend clients (could be web app servers, rich clients, SOAP / REST frontends) * everything (except the data) stateless globally shared SparkContext / shared Spark application - Key: SPARK-2389 URL: https://issues.apache.org/jira/browse/SPARK-2389 Project: Spark Issue Type: Improvement Components: Spark Core Reporter: Robert Stupp The documentation (in Cluster Mode Overview) cites: bq. Each application gets its own executor processes, which *stay up for the duration of the whole application* and run tasks in multiple threads. This has the benefit of isolating applications from each other, on both the scheduling side (each driver schedules its own tasks) and executor side (tasks from different applications run in different JVMs). However, it also means that *data cannot be shared* across different Spark applications (instances of SparkContext) without writing it to an external storage system. IMO this is a limitation that should be lifted to support any number of --driver-- client processes to share executors and to share (persistent / cached) data. This is especially useful if you have a bunch of frontend servers (dump web app servers) that want to use Spark as a _big computing machine_. Most important is the fact that Spark is quite good in caching/persisting data in memory / on disk thus removing load from backend data stores. Means: it would be really great to let different --driver-- client JVMs operate on the same RDDs and benefit from Spark's caching/persistence. It would however introduce some administration mechanisms to * start a shared context * update the executor configuration (# of worker nodes, # of cpus, etc) on the fly * stop a shared context Even conventional batch MR applications would benefit if ran fequently against the same data set. As an implicit requirement, RDD persistence could get a TTL for its materialized state. With such a feature the overall performance of today's web applications could then be increased by adding more web app servers, more spark nodes, more nosql nodes etc -- 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-2389) globally shared SparkContext / shared Spark application
[ https://issues.apache.org/jira/browse/SPARK-2389?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=14291795#comment-14291795 ] Murat Eken commented on SPARK-2389: --- [~sowen], I think Robert is talking about fault tolerance when he mentions scalability. Anyway, as I mentioned in my original comment, Tachyon is not an option, at least for us, due to interprocess serialization/deserialization costs. Although we haven't tried HDFS, but I would be surprised if that performed differently. globally shared SparkContext / shared Spark application - Key: SPARK-2389 URL: https://issues.apache.org/jira/browse/SPARK-2389 Project: Spark Issue Type: Improvement Components: Spark Core Reporter: Robert Stupp The documentation (in Cluster Mode Overview) cites: bq. Each application gets its own executor processes, which *stay up for the duration of the whole application* and run tasks in multiple threads. This has the benefit of isolating applications from each other, on both the scheduling side (each driver schedules its own tasks) and executor side (tasks from different applications run in different JVMs). However, it also means that *data cannot be shared* across different Spark applications (instances of SparkContext) without writing it to an external storage system. IMO this is a limitation that should be lifted to support any number of --driver-- client processes to share executors and to share (persistent / cached) data. This is especially useful if you have a bunch of frontend servers (dump web app servers) that want to use Spark as a _big computing machine_. Most important is the fact that Spark is quite good in caching/persisting data in memory / on disk thus removing load from backend data stores. Means: it would be really great to let different --driver-- client JVMs operate on the same RDDs and benefit from Spark's caching/persistence. It would however introduce some administration mechanisms to * start a shared context * update the executor configuration (# of worker nodes, # of cpus, etc) on the fly * stop a shared context Even conventional batch MR applications would benefit if ran fequently against the same data set. As an implicit requirement, RDD persistence could get a TTL for its materialized state. With such a feature the overall performance of today's web applications could then be increased by adding more web app servers, more spark nodes, more nosql nodes etc -- 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-2389) globally shared SparkContext / shared Spark application
[ https://issues.apache.org/jira/browse/SPARK-2389?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=14291788#comment-14291788 ] Sean Owen commented on SPARK-2389: -- Yes, the SPOF problem makes sense. It doesn't seem to be what this JIRA was about though, which seems to be what the jobserver-style approach addresses. That aside, why doesn't it scale? because of work that needs to be done on the driver? You can of course still run a bunch of drivers, just not one per client. The preloading cache issue is what off-heap caching in Tachyon or HDFS is supposed to ameliorate. globally shared SparkContext / shared Spark application - Key: SPARK-2389 URL: https://issues.apache.org/jira/browse/SPARK-2389 Project: Spark Issue Type: Improvement Components: Spark Core Reporter: Robert Stupp The documentation (in Cluster Mode Overview) cites: bq. Each application gets its own executor processes, which *stay up for the duration of the whole application* and run tasks in multiple threads. This has the benefit of isolating applications from each other, on both the scheduling side (each driver schedules its own tasks) and executor side (tasks from different applications run in different JVMs). However, it also means that *data cannot be shared* across different Spark applications (instances of SparkContext) without writing it to an external storage system. IMO this is a limitation that should be lifted to support any number of --driver-- client processes to share executors and to share (persistent / cached) data. This is especially useful if you have a bunch of frontend servers (dump web app servers) that want to use Spark as a _big computing machine_. Most important is the fact that Spark is quite good in caching/persisting data in memory / on disk thus removing load from backend data stores. Means: it would be really great to let different --driver-- client JVMs operate on the same RDDs and benefit from Spark's caching/persistence. It would however introduce some administration mechanisms to * start a shared context * update the executor configuration (# of worker nodes, # of cpus, etc) on the fly * stop a shared context Even conventional batch MR applications would benefit if ran fequently against the same data set. As an implicit requirement, RDD persistence could get a TTL for its materialized state. With such a feature the overall performance of today's web applications could then be increased by adding more web app servers, more spark nodes, more nosql nodes etc -- 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-2389) globally shared SparkContext / shared Spark application
[ https://issues.apache.org/jira/browse/SPARK-2389?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=14291798#comment-14291798 ] Robert Stupp commented on SPARK-2389: - bq. fault tolerance when he mentions scalability both play well together in a stateless application ;) globally shared SparkContext / shared Spark application - Key: SPARK-2389 URL: https://issues.apache.org/jira/browse/SPARK-2389 Project: Spark Issue Type: Improvement Components: Spark Core Reporter: Robert Stupp The documentation (in Cluster Mode Overview) cites: bq. Each application gets its own executor processes, which *stay up for the duration of the whole application* and run tasks in multiple threads. This has the benefit of isolating applications from each other, on both the scheduling side (each driver schedules its own tasks) and executor side (tasks from different applications run in different JVMs). However, it also means that *data cannot be shared* across different Spark applications (instances of SparkContext) without writing it to an external storage system. IMO this is a limitation that should be lifted to support any number of --driver-- client processes to share executors and to share (persistent / cached) data. This is especially useful if you have a bunch of frontend servers (dump web app servers) that want to use Spark as a _big computing machine_. Most important is the fact that Spark is quite good in caching/persisting data in memory / on disk thus removing load from backend data stores. Means: it would be really great to let different --driver-- client JVMs operate on the same RDDs and benefit from Spark's caching/persistence. It would however introduce some administration mechanisms to * start a shared context * update the executor configuration (# of worker nodes, # of cpus, etc) on the fly * stop a shared context Even conventional batch MR applications would benefit if ran fequently against the same data set. As an implicit requirement, RDD persistence could get a TTL for its materialized state. With such a feature the overall performance of today's web applications could then be increased by adding more web app servers, more spark nodes, more nosql nodes etc -- 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-2389) globally shared SparkContext / shared Spark application
[ https://issues.apache.org/jira/browse/SPARK-2389?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=14291804#comment-14291804 ] Sean Owen commented on SPARK-2389: -- Yes, makes sense. Maxing out one driver isn't an issue since you can have many drivers (or push work into the cluster). The issue is really that each driver then has its own RDDs, and if you need 100s of drivers to keep up, that just won't work. (Although then I'd question how so much work is being done on the Spark driver?) In theory the redundancy of all those RDDs is what HDFS caching and Tachyon could in theory help with, although those help share outside Spark. Whether that works for a particular use case right now is a different question, although I suspect it makes more sense to make those work than start yet another solution. What you are describing -- mutating lots shared in-memory state -- doesn't sound like a problem Spark helps solve per se. That is, it doesn't sound like work that has to live in a Spark driver program, even if it needs to ask a Spark driver-based service for some results. Naturally you know your problem better than I, but I am wondering if the answer here isn't just using Spark differently, for what it's for. globally shared SparkContext / shared Spark application - Key: SPARK-2389 URL: https://issues.apache.org/jira/browse/SPARK-2389 Project: Spark Issue Type: Improvement Components: Spark Core Reporter: Robert Stupp The documentation (in Cluster Mode Overview) cites: bq. Each application gets its own executor processes, which *stay up for the duration of the whole application* and run tasks in multiple threads. This has the benefit of isolating applications from each other, on both the scheduling side (each driver schedules its own tasks) and executor side (tasks from different applications run in different JVMs). However, it also means that *data cannot be shared* across different Spark applications (instances of SparkContext) without writing it to an external storage system. IMO this is a limitation that should be lifted to support any number of --driver-- client processes to share executors and to share (persistent / cached) data. This is especially useful if you have a bunch of frontend servers (dump web app servers) that want to use Spark as a _big computing machine_. Most important is the fact that Spark is quite good in caching/persisting data in memory / on disk thus removing load from backend data stores. Means: it would be really great to let different --driver-- client JVMs operate on the same RDDs and benefit from Spark's caching/persistence. It would however introduce some administration mechanisms to * start a shared context * update the executor configuration (# of worker nodes, # of cpus, etc) on the fly * stop a shared context Even conventional batch MR applications would benefit if ran fequently against the same data set. As an implicit requirement, RDD persistence could get a TTL for its materialized state. With such a feature the overall performance of today's web applications could then be increased by adding more web app servers, more spark nodes, more nosql nodes etc -- 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-2389) globally shared SparkContext / shared Spark application
[ https://issues.apache.org/jira/browse/SPARK-2389?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=14291724#comment-14291724 ] Murat Eken commented on SPARK-2389: --- +1. We're using a Spark cluster as a real-time query engine, and unfortunately we're running into the same issues as Robert mentions. Although Spark provides a plethora of solutions when it comes to making its cluster fault-tolerant and resilient, we need the same resilience for the front layer, from where the Spark cluster is accessed; meaning multiple instances of Spark clients, hence multiple SparkContexts from those clients connecting to the same cluster with the same computing power. Performance is crucial for us, hence our choice for caching the data in memory and utilizing the full hardware resources in the executors. Alternative solutions, such as using Tachyon for the data, and restarting executors for each query just don't give the same performance. We're looking into using https://github.com/spark-jobserver/spark-jobserver but that's not a proper solution as we still would have the jobserver as a single point of failure in our setup, which is a problem for us. I'd appreciate it if a Spark developer could give some information about the feasibility of this change request; if this turns out to be difficult or even impossible due to the choices made in the architecture, it would be good to know that so that we can consider our alternatives. globally shared SparkContext / shared Spark application - Key: SPARK-2389 URL: https://issues.apache.org/jira/browse/SPARK-2389 Project: Spark Issue Type: Improvement Components: Spark Core Reporter: Robert Stupp The documentation (in Cluster Mode Overview) cites: bq. Each application gets its own executor processes, which *stay up for the duration of the whole application* and run tasks in multiple threads. This has the benefit of isolating applications from each other, on both the scheduling side (each driver schedules its own tasks) and executor side (tasks from different applications run in different JVMs). However, it also means that *data cannot be shared* across different Spark applications (instances of SparkContext) without writing it to an external storage system. IMO this is a limitation that should be lifted to support any number of --driver-- client processes to share executors and to share (persistent / cached) data. This is especially useful if you have a bunch of frontend servers (dump web app servers) that want to use Spark as a _big computing machine_. Most important is the fact that Spark is quite good in caching/persisting data in memory / on disk thus removing load from backend data stores. Means: it would be really great to let different --driver-- client JVMs operate on the same RDDs and benefit from Spark's caching/persistence. It would however introduce some administration mechanisms to * start a shared context * update the executor configuration (# of worker nodes, # of cpus, etc) on the fly * stop a shared context Even conventional batch MR applications would benefit if ran fequently against the same data set. As an implicit requirement, RDD persistence could get a TTL for its materialized state. With such a feature the overall performance of today's web applications could then be increased by adding more web app servers, more spark nodes, more nosql nodes etc -- 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-2389) globally shared SparkContext / shared Spark application
[ https://issues.apache.org/jira/browse/SPARK-2389?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=14291737#comment-14291737 ] Sean Owen commented on SPARK-2389: -- Why can't N front-ends talk to a process built around one long-running Spark app? I think that's what the OP is talking about, and can be done right now. One Spark app having many contexts doesn't quite make sense as an app is a SparkContext. But [~meken] are you really talking about HA for the driver? globally shared SparkContext / shared Spark application - Key: SPARK-2389 URL: https://issues.apache.org/jira/browse/SPARK-2389 Project: Spark Issue Type: Improvement Components: Spark Core Reporter: Robert Stupp The documentation (in Cluster Mode Overview) cites: bq. Each application gets its own executor processes, which *stay up for the duration of the whole application* and run tasks in multiple threads. This has the benefit of isolating applications from each other, on both the scheduling side (each driver schedules its own tasks) and executor side (tasks from different applications run in different JVMs). However, it also means that *data cannot be shared* across different Spark applications (instances of SparkContext) without writing it to an external storage system. IMO this is a limitation that should be lifted to support any number of --driver-- client processes to share executors and to share (persistent / cached) data. This is especially useful if you have a bunch of frontend servers (dump web app servers) that want to use Spark as a _big computing machine_. Most important is the fact that Spark is quite good in caching/persisting data in memory / on disk thus removing load from backend data stores. Means: it would be really great to let different --driver-- client JVMs operate on the same RDDs and benefit from Spark's caching/persistence. It would however introduce some administration mechanisms to * start a shared context * update the executor configuration (# of worker nodes, # of cpus, etc) on the fly * stop a shared context Even conventional batch MR applications would benefit if ran fequently against the same data set. As an implicit requirement, RDD persistence could get a TTL for its materialized state. With such a feature the overall performance of today's web applications could then be increased by adding more web app servers, more spark nodes, more nosql nodes etc -- 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-2389) globally shared SparkContext / shared Spark application
[ https://issues.apache.org/jira/browse/SPARK-2389?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=14291755#comment-14291755 ] Murat Eken commented on SPARK-2389: --- Yes [~sowen], it's about HA for the driver. Our approach is to have a single app that's responsible for initializing the cache at start up (quite expensive) and then serve queries on that cached data (very fast). When you mention N front-ends talking to a process built around one long running Spark app that can be done right now, are you referring to something like the spark-jobserver (or any alternative) I mentioned? If yes, the problem with that is the single point of failure, as we're moving that from the driver to the jobserver instance. Or is there something else we've missed? globally shared SparkContext / shared Spark application - Key: SPARK-2389 URL: https://issues.apache.org/jira/browse/SPARK-2389 Project: Spark Issue Type: Improvement Components: Spark Core Reporter: Robert Stupp The documentation (in Cluster Mode Overview) cites: bq. Each application gets its own executor processes, which *stay up for the duration of the whole application* and run tasks in multiple threads. This has the benefit of isolating applications from each other, on both the scheduling side (each driver schedules its own tasks) and executor side (tasks from different applications run in different JVMs). However, it also means that *data cannot be shared* across different Spark applications (instances of SparkContext) without writing it to an external storage system. IMO this is a limitation that should be lifted to support any number of --driver-- client processes to share executors and to share (persistent / cached) data. This is especially useful if you have a bunch of frontend servers (dump web app servers) that want to use Spark as a _big computing machine_. Most important is the fact that Spark is quite good in caching/persisting data in memory / on disk thus removing load from backend data stores. Means: it would be really great to let different --driver-- client JVMs operate on the same RDDs and benefit from Spark's caching/persistence. It would however introduce some administration mechanisms to * start a shared context * update the executor configuration (# of worker nodes, # of cpus, etc) on the fly * stop a shared context Even conventional batch MR applications would benefit if ran fequently against the same data set. As an implicit requirement, RDD persistence could get a TTL for its materialized state. With such a feature the overall performance of today's web applications could then be increased by adding more web app servers, more spark nodes, more nosql nodes etc -- 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