Restrict tasks (of job) to only 1 to ensure that the causing Exception asserted for job failure is the deliberately thrown DAGSchedulerSuiteDummyException intended, not an UnsupportedOperationException from any second/subsequent tasks that can propagate from a race condition during code execution.
Author: shellberg <sah@zepler.org>
Closes#9076 from shellberg/shellberg-DAGSchedulerSuite-misbehavedResultHandlerTest-patch-1.
A few more changes:
1. Renamed IDVerifier -> RpcEndpointVerifier
2. Renamed NettyRpcAddress -> RpcEndpointAddress
3. Simplified NettyRpcHandler a bit by removing the connection count tracking. This is OK because I now force spark.shuffle.io.numConnectionsPerPeer to 1
4. Reduced spark.rpc.connect.threads to 64. It would be great to eventually remove this extra thread pool.
5. Minor cleanup & documentation.
Author: Reynold Xin <rxin@databricks.com>
Closes#9112 from rxin/SPARK-11096.
This patch unifies the memory management of the storage and execution regions such that either side can borrow memory from each other. When memory pressure arises, storage will be evicted in favor of execution. To avoid regressions in cases where storage is crucial, we dynamically allocate a fraction of space for storage that execution cannot evict. Several configurations are introduced:
- **spark.memory.fraction (default 0.75)**: fraction of the heap space used for execution and storage. The lower this is, the more frequently spills and cached data eviction occur. The purpose of this config is to set aside memory for internal metadata, user data structures, and imprecise size estimation in the case of sparse, unusually large records.
- **spark.memory.storageFraction (default 0.5)**: size of the storage region within the space set aside by `spark.memory.fraction`. Cached data may only be evicted if total storage exceeds this region.
- **spark.memory.useLegacyMode (default false)**: whether to use the memory management that existed in Spark 1.5 and before. This is mainly for backward compatibility.
For a detailed description of the design, see [SPARK-10000](https://issues.apache.org/jira/browse/SPARK-10000). This patch builds on top of the `MemoryManager` interface introduced in #9000.
Author: Andrew Or <andrew@databricks.com>
Closes#9084 from andrewor14/unified-memory-manager.
I'm going through the implementation right now for post-doc review. Adding more comments and renaming things as I go through them.
I also want to write higher level documentation about how the whole thing works -- but those will come in other pull requests.
Author: Reynold Xin <rxin@databricks.com>
Closes#9091 from rxin/rpc-review.
https://issues.apache.org/jira/browse/SPARK-10858
The issue here is that in resolveURI we default to calling new File(path).getAbsoluteFile().toURI(). But if the path passed in already has a # in it then File(path) will think that is supposed to be part of the actual file path and not a fragment so it changes # to %23. Then when we try to parse that later in Client as a URI it doesn't recognize there is a fragment.
so to fix we just check if there is a fragment, still create the File like we did before and then add the fragment back on.
Author: Tom Graves <tgraves@yahoo-inc.com>
Closes#9035 from tgravescs/SPARK-10858.
This change adds an API that encapsulates information about an app
launched using the library. It also creates a socket-based communication
layer for apps that are launched as child processes; the launching
application listens for connections from launched apps, and once
communication is established, the channel can be used to send updates
to the launching app, or to send commands to the child app.
The change also includes hooks for local, standalone/client and yarn
masters.
Author: Marcelo Vanzin <vanzin@cloudera.com>
Closes#7052 from vanzin/SPARK-8673.
This patch introduces a `MemoryManager` that is the central arbiter of how much memory to grant to storage and execution. This patch is primarily concerned only with refactoring while preserving the existing behavior as much as possible.
This is the first step away from the existing rigid separation of storage and execution memory, which has several major drawbacks discussed on the [issue](https://issues.apache.org/jira/browse/SPARK-10956). It is the precursor of a series of patches that will attempt to address those drawbacks.
Author: Andrew Or <andrew@databricks.com>
Author: Josh Rosen <joshrosen@databricks.com>
Author: andrewor14 <andrew@databricks.com>
Closes#9000 from andrewor14/memory-manager.
This PR just reverted 02144d6745 to remerge #6457 and also included the commits in #8905.
Author: zsxwing <zsxwing@gmail.com>
Closes#8944 from zsxwing/SPARK-6028.
Compatibility between history server script and functionality
The history server has its argument parsing class in HistoryServerArguments. However, this doesn't get involved in the start-history-server.sh codepath where the $0 arg is assigned to spark.history.fs.logDirectory and all other arguments discarded (e.g --property-file.)
This stops the other options being usable from this script
Author: Joshi <rekhajoshm@gmail.com>
Author: Rekha Joshi <rekhajoshm@gmail.com>
Closes#8758 from rekhajoshm/SPARK-10317.
Fix the following issues in StandaloneDynamicAllocationSuite:
1. It should not assume master and workers start in order
2. It should not assume master and workers get ready at once
3. It should not assume the application is already registered with master after creating SparkContext
4. It should not access Master.app and idToApp which are not thread safe
The changes includes:
* Use `eventually` to wait until master and workers are ready to fix 1 and 2
* Use `eventually` to wait until the application is registered with master to fix 3
* Use `askWithRetry[MasterStateResponse](RequestMasterState)` to get the application info to fix 4
Author: zsxwing <zsxwing@gmail.com>
Closes#8914 from zsxwing/fix-StandaloneDynamicAllocationSuite.
This makes two changes:
- Allow reduce tasks to fetch multiple map output partitions -- this is a pretty small change to HashShuffleFetcher
- Move shuffle locality computation out of DAGScheduler and into ShuffledRDD / MapOutputTracker; this was needed because the code in DAGScheduler wouldn't work for RDDs that fetch multiple map output partitions from each reduce task
I also added an AdaptiveSchedulingSuite that creates RDDs depending on multiple map output partitions.
Author: Matei Zaharia <matei@databricks.com>
Closes#8844 from mateiz/spark-9852.
The DiskBlockObjectWriter constructor took a BlockId parameter but never used it. As part of some general cleanup in these interfaces, this patch refactors its constructor to eliminate this parameter.
Author: Josh Rosen <joshrosen@databricks.com>
Closes#8871 from JoshRosen/disk-block-object-writer-blockid-cleanup.
The current shuffle code has an interface named ShuffleReader with only one implementation, HashShuffleReader. This naming is confusing, since the same read path code is used for both sort- and hash-based shuffle. This patch addresses this by renaming HashShuffleReader to BlockStoreShuffleReader.
Author: Josh Rosen <joshrosen@databricks.com>
Closes#8825 from JoshRosen/shuffle-reader-cleanup.
The job group, and job descriptions information is passed through thread local properties, and get inherited by child threads. In case of spark streaming, the streaming jobs inherit these properties from the thread that called streamingContext.start(). This may not make sense.
1. Job group: This is mainly used for cancelling a group of jobs together. It does not make sense to cancel streaming jobs like this, as the effect will be unpredictable. And its not a valid usecase any way, to cancel a streaming context, call streamingContext.stop()
2. Job description: This is used to pass on nice text descriptions for jobs to show up in the UI. The job description of the thread that calls streamingContext.start() is not useful for all the streaming jobs, as it does not make sense for all of the streaming jobs to have the same description, and the description may or may not be related to streaming.
The solution in this PR is meant for the Spark master branch, where local properties are inherited by cloning the properties. The job group and job description in the thread that starts the streaming scheduler are explicitly removed, so that all the subsequent child threads does not inherit them. Also, the starting is done in a new child thread, so that setting the job group and description for streaming, does not change those properties in the thread that called streamingContext.start().
Author: Tathagata Das <tathagata.das1565@gmail.com>
Closes#8781 from tdas/SPARK-10649.
Track pending tasks by partition ID instead of Task objects.
Before this change, failure & retry could result in a case where a stage got submitted before the map output from its dependencies get registered. This was due to an error in the condition for registering map outputs.
Author: hushan[胡珊] <hushan@xiaomi.com>
Author: Imran Rashid <irashid@cloudera.com>
Closes#7699 from squito/SPARK-5259.
I noticed only one block manager registered with master in an unsuccessful build (https://amplab.cs.berkeley.edu/jenkins/job/Spark-Master-SBT/AMPLAB_JENKINS_BUILD_PROFILE=hadoop2.2,label=spark-test/3534/)
```
15/09/16 13:02:30.981 pool-1-thread-1-ScalaTest-running-BroadcastSuite INFO SparkContext: Running Spark version 1.6.0-SNAPSHOT
...
15/09/16 13:02:38.133 sparkDriver-akka.actor.default-dispatcher-19 INFO BlockManagerMasterEndpoint: Registering block manager localhost:48196 with 530.3 MB RAM, BlockManagerId(0, localhost, 48196)
```
In addition, the first block manager needed 7+ seconds to start. But the test expected 2 block managers so it failed.
However, there was no exception in this log file. So I checked a successful build (https://amplab.cs.berkeley.edu/jenkins/job/Spark-Master-SBT/3536/AMPLAB_JENKINS_BUILD_PROFILE=hadoop2.2,label=spark-test/) and it needed 4-5 seconds to set up the local cluster:
```
15/09/16 18:11:27.738 sparkWorker1-akka.actor.default-dispatcher-5 INFO Worker: Running Spark version 1.6.0-SNAPSHOT
...
15/09/16 18:11:30.838 sparkDriver-akka.actor.default-dispatcher-20 INFO BlockManagerMasterEndpoint: Registering block manager localhost:54202 with 530.3 MB RAM, BlockManagerId(1, localhost, 54202)
15/09/16 18:11:32.112 sparkDriver-akka.actor.default-dispatcher-20 INFO BlockManagerMasterEndpoint: Registering block manager localhost:32955 with 530.3 MB RAM, BlockManagerId(0, localhost, 32955)
```
In this build, the first block manager needed only 3+ seconds to start.
Comparing these two builds, I guess it's possible that the local cluster in `BroadcastSuite` cannot be ready in 10 seconds if the Jenkins worker is busy. So I just increased the timeout to 60 seconds to see if this can fix the issue.
Author: zsxwing <zsxwing@gmail.com>
Closes#8813 from zsxwing/fix-BroadcastSuite.
It does not make much sense to set `spark.shuffle.spill` or `spark.sql.planner.externalSort` to false: I believe that these configurations were initially added as "escape hatches" to guard against bugs in the external operators, but these operators are now mature and well-tested. In addition, these configurations are not handled in a consistent way anymore: SQL's Tungsten codepath ignores these configurations and will continue to use spilling operators. Similarly, Spark Core's `tungsten-sort` shuffle manager does not respect `spark.shuffle.spill=false`.
This pull request removes these configurations, adds warnings at the appropriate places, and deletes a large amount of code which was only used in code paths that did not support spilling.
Author: Josh Rosen <joshrosen@databricks.com>
Closes#8831 from JoshRosen/remove-ability-to-disable-spilling.
When speculative execution is enabled, consider a scenario where the authorized committer of a particular output partition fails during the OutputCommitter.commitTask() call. In this case, the OutputCommitCoordinator is supposed to release that committer's exclusive lock on committing once that task fails. However, due to a unit mismatch (we used task attempt number in one place and task attempt id in another) the lock will not be released, causing Spark to go into an infinite retry loop.
This bug was masked by the fact that the OutputCommitCoordinator does not have enough end-to-end tests (the current tests use many mocks). Other factors contributing to this bug are the fact that we have many similarly-named identifiers that have different semantics but the same data types (e.g. attemptNumber and taskAttemptId, with inconsistent variable naming which makes them difficult to distinguish).
This patch adds a regression test and fixes this bug by always using task attempt numbers throughout this code.
Author: Josh Rosen <joshrosen@databricks.com>
Closes#8544 from JoshRosen/SPARK-10381.
*Note: this is for master branch only.* The fix for branch-1.5 is at #8721.
The query execution ID is currently passed from a thread to its children, which is not the intended behavior. This led to `IllegalArgumentException: spark.sql.execution.id is already set` when running queries in parallel, e.g.:
```
(1 to 100).par.foreach { _ =>
sc.parallelize(1 to 5).map { i => (i, i) }.toDF("a", "b").count()
}
```
The cause is `SparkContext`'s local properties are inherited by default. This patch adds a way to exclude keys we don't want to be inherited, and makes SQL go through that code path.
Author: Andrew Or <andrew@databricks.com>
Closes#8710 from andrewor14/concurrent-sql-executions.
This patch adds support for submitting map stages in a DAG individually so that we can make downstream decisions after seeing statistics about their output, as part of SPARK-9850. I also added more comments to many of the key classes in DAGScheduler. By itself, the patch is not super useful except maybe to switch between a shuffle and broadcast join, but with the other subtasks of SPARK-9850 we'll be able to do more interesting decisions.
The main entry point is SparkContext.submitMapStage, which lets you run a map stage and see stats about the map output sizes. Other stats could also be collected through accumulators. See AdaptiveSchedulingSuite for a short example.
Author: Matei Zaharia <matei@databricks.com>
Closes#8180 from mateiz/spark-9851.
This is a follow-up patch to #8723. I missed one case there.
Author: Andrew Or <andrew@databricks.com>
Closes#8727 from andrewor14/fix-threading-suite.
This is a followup to #8499 which adds a Scalastyle rule to mandate the use of SparkHadoopUtil's JobContext accessor methods and fixes the existing violations.
Author: Josh Rosen <joshrosen@databricks.com>
Closes#8521 from JoshRosen/SPARK-10330-part2.
Fix a few Java API test style issues: unused generic types, exceptions, wrong assert argument order
Author: Sean Owen <sowen@cloudera.com>
Closes#8706 from srowen/SPARK-10547.
This commit ensures if an assertion fails within a thread, it will ultimately fail the test. Otherwise we end up potentially masking real bugs by not propagating assertion failures properly.
Author: Andrew Or <andrew@databricks.com>
Closes#8723 from andrewor14/fix-threading-suite.
ShuffleManager implementations are currently not given type information for
the key, value and combiner classes. Serialization of shuffle objects relies
on objects being JavaSerializable, with methods defined for reading/writing
the object or, alternatively, serialization via Kryo which uses reflection.
Serialization systems like Avro, Thrift and Protobuf generate classes with
zero argument constructors and explicit schema information
(e.g. IndexedRecords in Avro have get, put and getSchema methods).
By serializing the key, value and combiner class names in ShuffleDependency,
shuffle implementations will have access to schema information when
registerShuffle() is called.
Author: Matt Massie <massie@cs.berkeley.edu>
Closes#7403 from massie/shuffle-classtags.
This is a regression introduced in #4960, this commit fixes it and adds a test.
tnachen andrewor14 please review, this should be an easy one.
Author: Iulian Dragos <jaguarul@gmail.com>
Closes#8653 from dragos/issue/mesos/fine-grained-maxExecutorCores.
The architecture is that, in YARN mode, if the driver detects that an executor has disconnected, it asks the ApplicationMaster why the executor died. If the ApplicationMaster is aware that the executor died because of preemption, all tasks associated with that executor are not marked as failed. The executor
is still removed from the driver's list of available executors, however.
There's a few open questions:
1. Should standalone mode have a similar "get executor loss reason" as well? I localized this change as much as possible to affect only YARN, but there could be a valid case to differentiate executor losses in standalone mode as well.
2. I make a pretty strong assumption in YarnAllocator that getExecutorLossReason(executorId) will only be called once per executor id; I do this so that I can remove the metadata from the in-memory map to avoid object accumulation. It's not clear if I'm being overly zealous to save space, however.
cc vanzin specifically for review because it collided with some earlier YARN scheduling work.
cc JoshRosen because it's similar to output commit coordination we did in the past
cc andrewor14 for our discussion on how to get executor exit codes and loss reasons
Author: mcheah <mcheah@palantir.com>
Closes#8007 from mccheah/feature/preemption-handling.
We introduced the Netty network module for shuffle in Spark 1.2, and has turned it on by default for 3 releases. The old ConnectionManager is difficult to maintain. If we merge the patch now, by the time it is released, it would be 1 yr for which ConnectionManager is off by default. It's time to remove it.
Author: Reynold Xin <rxin@databricks.com>
Closes#8161 from rxin/SPARK-9767.
[SPARK-9591](https://issues.apache.org/jira/browse/SPARK-9591)
When we getting the broadcast variable, we can fetch the block form several location,but now when connecting the lost blockmanager(idle for enough time removed by driver when using dynamic resource allocate and so on) will cause task fail,and the worse case will cause the job fail.
Author: jeanlyn <jeanlyn92@gmail.com>
Closes#7927 from jeanlyn/catch_exception.
This contribution is my original work and I license the work to the project under the project's open source license.
Author: Pat Shields <yeoldefortran@gmail.com>
Closes#7979 from pashields/env-loading-on-driver.
This is pretty minor, just trying to improve the readability of `DAGSchedulerSuite`, I figure every bit helps. Before whenever I read this test, I never knew what "should work" and "should be ignored" really meant -- this adds some asserts & updates comments to make it more clear. Also some reformatting per a suggestion from markhamstra on https://github.com/apache/spark/pull/7699
Author: Imran Rashid <irashid@cloudera.com>
Closes#8434 from squito/SPARK-10247.
The ```Stage``` class now tracks whether there were a sufficient number of consecutive failures of that stage to trigger an abort.
To avoid an infinite loop of stage retries, we abort the job completely after 4 consecutive stage failures for one stage. We still allow more than 4 consecutive stage failures if there is an intervening successful attempt for the stage, so that in very long-lived applications, where a stage may get reused many times, we don't abort the job after failures that have been recovered from successfully.
I've added test cases to exercise the most obvious scenarios.
Author: Ilya Ganelin <ilya.ganelin@capitalone.com>
Closes#5636 from ilganeli/SPARK-5945.
SPARK-4223.
Currently we support setting view and modify acls but you have to specify a list of users. It would be nice to support * meaning all users have access.
Manual tests to verify that: "*" works for any user in:
a. Spark ui: view and kill stage. Done.
b. Spark history server. Done.
c. Yarn application killing. Done.
Author: zhuol <zhuol@yahoo-inc.com>
Closes#8398 from zhuoliu/4223.
In SMJ, the first ExternalSorter could consume all the memory before spilling, then the second can not even acquire the first page.
Before we have a better memory allocator, SMJ should call prepare() before call any compute() of it's children.
cc rxin JoshRosen
Author: Davies Liu <davies@databricks.com>
Closes#8511 from davies/smj_memory.
This change aims at speeding up the dev cycle a little bit, by making
sure that all tests behave the same w.r.t. where the code to be tested
is loaded from. Namely, that means that tests don't rely on the assembly
anymore, rather loading all needed classes from the build directories.
The main change is to make sure all build directories (classes and test-classes)
are added to the classpath of child processes when running tests.
YarnClusterSuite required some custom code since the executors are run
differently (i.e. not through the launcher library, like standalone and
Mesos do).
I also found a couple of tests that could leak a SparkContext on failure,
and added code to handle those.
With this patch, it's possible to run the following command from a clean
source directory and have all tests pass:
mvn -Pyarn -Phadoop-2.4 -Phive-thriftserver install
Author: Marcelo Vanzin <vanzin@cloudera.com>
Closes#7629 from vanzin/SPARK-9284.
