By Hiding the shuffleblockmanager behind Shufflemanager, we decouple the shuffle data's block mapping management work from Diskblockmananger. This give a more clear interface and more easy for other shuffle manager to implement their own block management logic. the jira ticket have more details.
Author: Raymond Liu <raymond.liu@intel.com>
Closes#1241 from colorant/shuffle and squashes the following commits:
0e01ae3 [Raymond Liu] Move ShuffleBlockmanager behind shuffleManager
This is a simplified fix for SPARK-1912.
Author: Reynold Xin <rxin@apache.org>
Closes#2179 from rxin/SPARK-1912 and squashes the following commits:
b2f0e9e [Reynold Xin] Fix unit tests.
a8eddfe [Reynold Xin] [SPARK-1912] Lazily initialize buffers for local shuffle blocks.
**Summary of the changes**
The bulk of this PR is comprised of tests and documentation; the actual fix is really just adding 1 line of code (see `BlockObjectWriter.scala`). We currently do not run the `External*` test suites with different compression codecs, and this would have caught the bug reported in [SPARK-3277](https://issues.apache.org/jira/browse/SPARK-3277). This PR extends the existing code to test spilling using all compression codecs known to Spark, including `LZ4`.
**The bug itself**
In `DiskBlockObjectWriter`, we only report the shuffle bytes written before we close the streams. With `LZ4`, all the bytes written reported by our metrics were 0 because `flush()` was not taking effect for some reason. In general, compression codecs may write additional bytes to the file after we call `close()`, and so we must also capture those bytes in our shuffle write metrics.
Thanks mridulm and pwendell for help with debugging.
Author: Andrew Or <andrewor14@gmail.com>
Author: Patrick Wendell <pwendell@gmail.com>
Closes#2187 from andrewor14/fix-lz4-spilling and squashes the following commits:
1b54bdc [Andrew Or] Speed up tests by not compressing everything
1c4624e [Andrew Or] Merge branch 'master' of github.com:apache/spark into fix-lz4-spilling
6b2e7d1 [Andrew Or] Fix compilation error
92e251b [Patrick Wendell] Better documentation for BlockObjectWriter.
a1ad536 [Andrew Or] Fix tests
089593f [Andrew Or] Actually fix SPARK-3277 (tests still fail)
4bbcf68 [Andrew Or] Update tests to actually test all compression codecs
b264a84 [Andrew Or] ExternalAppendOnlyMapSuite code style fixes (minor)
1bfa743 [Andrew Or] Add more information to assert for better debugging
Netty functionality will be added back in subsequent PRs by using the BlockTransferService interface.
Author: Reynold Xin <rxin@apache.org>
Closes#2181 from rxin/SPARK-3281 and squashes the following commits:
5494b0e [Reynold Xin] Fix extra port.
ff6d1e1 [Reynold Xin] [SPARK-3281] Remove Netty specific code in BlockManager.
compeleted stage only need to remove its own partitions that are no longer cached. However, "StorageTab" may lost some rdds which are cached actually. Not only in "StorageTab", "ExectutorTab" may also lose some rdd info which have been overwritten by last rdd in a same task.
1. "StorageTab": when multiple stages run simultaneously, completed stage will remove rdd info which belong to other stages that are still running.
2. "ExectutorTab": taskcontext may lose some "updatedBlocks" info of rdds in a dependency chain. Like the following example:
val r1 = sc.paralize(..).cache()
val r2 = r1.map(...).cache()
val n = r2.count()
When count the r2, r1 and r2 will be cached finally. So in CacheManager.getOrCompute, the taskcontext should contain "updatedBlocks" of r1 and r2. Currently, the "updatedBlocks" only contain the info of r2.
Author: uncleGen <hustyugm@gmail.com>
Closes#2131 from uncleGen/master_ui_fix and squashes the following commits:
a6a8a0b [uncleGen] fix some coding style
3a1bc15 [uncleGen] fix some error in unit test
56ea488 [uncleGen] there's some line too long
c82ba82 [uncleGen] Bug Fix: RDD info loss in "StorageTab" and "ExecutorTab"
As a workaround for SPARK-3015, the ContextCleaner was made "blocking", that is, it cleaned items one-by-one. But shuffles can take a long time to be deleted. Given that the RC for 1.1 is imminent, this PR makes a narrow change in the context cleaner - not wait for shuffle cleanups to complete. Also it changes the error messages on failure to delete to be milder warnings, as exceptions in the delete code path for one item does not really stop the actual functioning of the system.
Author: Tathagata Das <tathagata.das1565@gmail.com>
Closes#2143 from tdas/cleaner-shuffle-fix and squashes the following commits:
9c84202 [Tathagata Das] Restoring default blocking behavior in ContextCleanerSuite, and added docs to identify that spark.cleaner.referenceTracking.blocking does not control shuffle.
2181329 [Tathagata Das] Mark shuffle cleanup as non-blocking.
e337cc2 [Tathagata Das] Changed semantics based on PR comments.
387b578 [Tathagata Das] Made ContextCleaner to not block on shuffles
This is a HOTFIX for 1.1.
Author: Reynold Xin <rxin@apache.org>
Author: Kay Ousterhout <kayousterhout@gmail.com>
Closes#2127 from rxin/SPARK-3224 and squashes the following commits:
effb1ce [Reynold Xin] Move log message.
49282b3 [Reynold Xin] Kay's feedback.
3f01847 [Reynold Xin] Merge pull request #2 from kayousterhout/SPARK-3224
796d282 [Kay Ousterhout] Added unit test for SPARK-3224
3d3d356 [Reynold Xin] Remove map output loc even for repeated FetchFaileds.
1dd3eb5 [Reynold Xin] [SPARK-3224] FetchFailed reduce stages should only show up once in the failed stages UI.
Simple way to reproduce this in the UI:
```scala
val f = new java.io.File("/tmp/test")
f.delete()
sc.parallelize(1 to 2, 2).map(x => (x,x )).repartition(3).mapPartitionsWithContext { case (context, iter) =>
if (context.partitionId == 0) {
val f = new java.io.File("/tmp/test")
if (!f.exists) {
f.mkdir()
System.exit(0);
}
}
iter
}.count()
```
Author: Reynold Xin <rxin@apache.org>
Closes#1545 from rxin/stage-attempt and squashes the following commits:
3ee1d2a [Reynold Xin] - Rename attempt to retry in UI. - Properly report stage failure in FetchFailed.
40a6bd5 [Reynold Xin] Updated test suites.
c414c36 [Reynold Xin] Fixed the hanging in JobCancellationSuite.
b3e2eed [Reynold Xin] Oops previous code didn't compile.
0f36075 [Reynold Xin] Mark unknown stage attempt with id -1 and drop that in JobProgressListener.
6c08b07 [Reynold Xin] Addressed code review feedback.
4e5faa2 [Reynold Xin] [SPARK-2298] Encode stage attempt in SparkListener & UI.
This PR fixes two bugs related to `spark.local.dirs` and `SPARK_LOCAL_DIRS`, one where `Utils.getLocalDir()` might return an invalid directory (SPARK-2974) and another where the `SPARK_LOCAL_DIRS` override didn't affect the driver, which could cause problems when running tasks in local mode (SPARK-2975).
This patch fixes both issues: the new `Utils.getOrCreateLocalRootDirs(conf: SparkConf)` utility method manages the creation of local directories and handles the precedence among the different configuration options, so we should see the same behavior whether we're running in local mode or on a worker.
It's kind of a pain to mock out environment variables in tests (no easy way to mock System.getenv), so I added a `private[spark]` method to SparkConf for accessing environment variables (by default, it just delegates to System.getenv). By subclassing SparkConf and overriding this method, we can mock out SPARK_LOCAL_DIRS in tests.
I also fixed a typo in PySpark where we used `SPARK_LOCAL_DIR` instead of `SPARK_LOCAL_DIRS` (I think this was technically innocuous, but it seemed worth fixing).
Author: Josh Rosen <joshrosen@apache.org>
Closes#2002 from JoshRosen/local-dirs and squashes the following commits:
efad8c6 [Josh Rosen] Address review comments:
1dec709 [Josh Rosen] Minor updates to Javadocs.
7f36999 [Josh Rosen] Use env vars to detect if running in YARN container.
399ac25 [Josh Rosen] Update getLocalDir() documentation.
bb3ad89 [Josh Rosen] Remove duplicated YARN getLocalDirs() code.
3e92d44 [Josh Rosen] Move local dirs override logic into Utils; fix bugs:
b2c4736 [Josh Rosen] Add failing tests for SPARK-2974 and SPARK-2975.
007298b [Josh Rosen] Allow environment variables to be mocked in tests.
6d9259b [Josh Rosen] Fix typo in PySpark: SPARK_LOCAL_DIR should be SPARK_LOCAL_DIRS
This is a re-implementation of TorrentBroadcast, with the following changes:
1. Removes most of the mutable, transient state from TorrentBroadcast (e.g. totalBytes, num of blocks fetched).
2. Removes TorrentInfo and TorrentBlock
3. Replaces the BlockManager.getSingle call in readObject with a getLocal, resuling in one less RPC call to the BlockManagerMasterActor to find the location of the block.
4. Removes the metadata block, resulting in one less block to fetch.
5. Removes an extra memory copy for deserialization (by using Java's SequenceInputStream).
Basically for a regular broadcasted object with only one block, the number of RPC calls goes from 5+1 to 2+1).
Old TorrentBroadcast for object of a single block:
1 RPC to ask for location of the broadcast variable
1 RPC to ask for location of the metadata block
1 RPC to fetch the metadata block
1 RPC to ask for location of the first data block
1 RPC to fetch the first data block
1 RPC to tell the driver we put the first data block in
i.e. 5 + 1
New TorrentBroadcast for object of a single block:
1 RPC to ask for location of the first data block
1 RPC to get the first data block
1 RPC to tell the driver we put the first data block in
i.e. 2 + 1
Author: Reynold Xin <rxin@apache.org>
Closes#2030 from rxin/torrentBroadcast and squashes the following commits:
5bacb9d [Reynold Xin] Always add the object to driver's block manager.
0d8ed5b [Reynold Xin] Added getBytes to BlockManager and uses that in TorrentBroadcast.
2d6a5fb [Reynold Xin] Use putBytes/getRemoteBytes throughout.
3670f00 [Reynold Xin] Code review feedback.
c1185cd [Reynold Xin] [SPARK-3119] Re-implementation of TorrentBroadcast.
Previous pull request (#1907) was reverted. This brings it back. Still looking into the hang.
Author: Reynold Xin <rxin@apache.org>
Closes#1971 from rxin/netty1 and squashes the following commits:
b0be96f [Reynold Xin] Added test to make sure outstandingRequests are cleaned after firing the events.
4c6d0ee [Reynold Xin] Pass callbacks cleanly.
603dce7 [Reynold Xin] Upgrade Netty to 4.0.23 to fix the DefaultFileRegion bug.
88be1d4 [Reynold Xin] Downgrade to 4.0.21 to work around a bug in writing DefaultFileRegion.
002626a [Reynold Xin] Remove netty-test-file.txt.
db6e6e0 [Reynold Xin] Revert "Revert "[SPARK-2468] Netty based block server / client module""
Author: Chandan Kumar <chandan.kumar@imaginea.com>
Closes#1787 from nrchandan/spark-2862 and squashes the following commits:
a76bbf6 [Chandan Kumar] [SPARK-2862] Fix for a broken test case and add new test cases
4211eea [Chandan Kumar] [SPARK-2862] Add Scala bug id
13854f1 [Chandan Kumar] [SPARK-2862] Use shorthand range notation to avoid Scala bug
This is intended to fix SPARK-2977. Before, there was an implicit ordering dependency where we needed to know the ShuffleManager implementation before creating the ShuffleBlockManager. This patch makes that dependency explicit by adding ShuffleManager to a bunch of constructors.
I think it's a little odd for BlockManager to take a ShuffleManager only to pass it to ShuffleBlockManager without using it itself; there's an opportunity to clean this up later if we sever the circular dependencies between BlockManager and other components and pass those components to BlockManager's constructor.
Author: Josh Rosen <joshrosen@apache.org>
Closes#1976 from JoshRosen/SPARK-2977 and squashes the following commits:
a9cd1e1 [Josh Rosen] [SPARK-2977] Ensure ShuffleManager is created before ShuffleBlockManager.
Author: Reynold Xin <rxin@apache.org>
Closes#1948 from rxin/kryo and squashes the following commits:
a3a80d8 [Reynold Xin] [SPARK-3046] use executor's class loader as the default serializer classloader
3d13277 [Reynold Xin] Reverted that in TestJavaSerializerImpl too.
196f3dc [Reynold Xin] Ok one more commit to revert the classloader change.
c49b50c [Reynold Xin] Removed JavaSerializer change.
afbf37d [Reynold Xin] Moved the test case also.
a2e693e [Reynold Xin] Removed the Kryo bug fix from this pull request.
c81bd6c [Reynold Xin] Use defaultClassLoader when executing user specified custom registrator.
68f261e [Reynold Xin] Added license check excludes.
0c28179 [Reynold Xin] [SPARK-3045] Make Serializer interface Java friendly [SPARK-3046] Set executor's class loader as the default serializer class loader
The serializer is not always used in an executor thread (e.g. connection manager, broadcast), in which case the classloader might not have the user jar set, leading to corruption in deserialization.
https://issues.apache.org/jira/browse/SPARK-3046https://issues.apache.org/jira/browse/SPARK-2878
Author: Reynold Xin <rxin@apache.org>
Closes#1972 from rxin/kryoBug and squashes the following commits:
c1c7bf0 [Reynold Xin] Made change to JavaSerializer.
7204c33 [Reynold Xin] Added imports back.
d879e67 [Reynold Xin] [SPARK-3046] use executor's class loader as the default serializer class loader.
This is a rewrite of the original Netty module that was added about 1.5 years ago. The old code was turned off by default and didn't really work because it lacked a frame decoder (only worked with very very small blocks).
For this pull request, I tried to make the changes non-instrusive to the rest of Spark. I only added an init and shutdown to BlockManager/DiskBlockManager, and a bunch of comments to help me understand the existing code base.
Compared with the old Netty module, this one features:
- It appears to work :)
- SPARK-2941: option to specicy nio vs oio vs epoll for channel/transport. By default nio is used. (Not using Epoll yet because I have found some bugs with its implementation)
- SPARK-2943: options to specify send buf and receive buf for users who want to do hyper tuning
- SPARK-2942: io errors are reported from server to client (the protocol uses negative length to indicate error)
- SPARK-2940: fetching multiple blocks in a single request to reduce syscalls
- SPARK-2959: clients share a single thread pool
- SPARK-2990: use PooledByteBufAllocator to reduce GC (basically a Netty managed pool of buffers with jmalloc)
- SPARK-2625: added fetchWaitTime metric and fixed thread-safety issue in metrics update.
- SPARK-2367: bump Netty version to 4.0.21.Final to address an Epoll bug (https://groups.google.com/forum/#!topic/netty/O7m-HxCJpCA)
Compared with the existing communication manager, this one features:
- IMO it is substantially easier to understand
- zero-copy send for the server for on-disk blocks
- one-copy receive (due to a frame decoder)
- don't quote me on this, but I think a lot less sys calls
- SPARK-2990: use PooledByteBufAllocator to reduce GC (basically a Netty managed pool of buffers with jmalloc)
- SPARK-2941: option to specicy nio vs oio vs epoll for channel/transport. By default nio is used. (Not using Epoll yet because I have found some bugs with its implementation)
- SPARK-2943: options to specify send buf and receive buf for users who want to do hyper tuning
TODOs before it can fully replace the existing ConnectionManager, if that ever happens (most of them should probably be done in separate PRs since this needs to be turned on explicitly)
- [x] Basic test cases
- [ ] More unit/integration tests for failures
- [ ] Performance analysis
- [ ] Support client connection reuse so we don't need to keep opening new connections (not sure how useful this would be)
- [ ] Support putting blocks in addition to fetching blocks (i.e. two way transfer)
- [x] Support serving non-disk blocks
- [ ] Support SASL authentication
For a more comprehensive list, see https://issues.apache.org/jira/browse/SPARK-2468
Thanks to @coderplay for peer coding with me on a Sunday.
Author: Reynold Xin <rxin@apache.org>
Closes#1907 from rxin/netty and squashes the following commits:
f921421 [Reynold Xin] Upgrade Netty to 4.0.22.Final to fix another Epoll bug.
4b174ca [Reynold Xin] Shivaram's code review comment.
4a3dfe7 [Reynold Xin] Switched to nio for default (instead of epoll on Linux).
56bfb9d [Reynold Xin] Bump Netty version to 4.0.21.Final for some bug fixes.
b443a4b [Reynold Xin] Added debug message to help debug Jenkins failures.
57fc4d7 [Reynold Xin] Added test cases for BlockHeaderEncoder and BlockFetchingClientHandlerSuite.
22623e9 [Reynold Xin] Added exception handling and test case for BlockServerHandler and BlockFetchingClientHandler.
6550dd7 [Reynold Xin] Fixed block mgr init bug.
60c2edf [Reynold Xin] Beefed up server/client integration tests.
38d88d5 [Reynold Xin] Added missing test files.
6ce3f3c [Reynold Xin] Added some basic test cases.
47f7ce0 [Reynold Xin] Created server and client packages and moved files there.
b16f412 [Reynold Xin] Added commit count.
f13022d [Reynold Xin] Remove unused clone() in BlockFetcherIterator.
c57d68c [Reynold Xin] Added back missing files.
842dfa7 [Reynold Xin] Made everything work with proper reference counting.
3fae001 [Reynold Xin] Connected the new netty network module with rest of Spark.
1a8f6d4 [Reynold Xin] Completed protocol documentation.
2951478 [Reynold Xin] New Netty implementation.
cc7843d [Reynold Xin] Basic skeleton.
Note this also passes the TaskContext itself to the TaskCompletionListener. In the future we can mark TaskContext with the exception object if exception occurs during task execution.
Author: Reynold Xin <rxin@apache.org>
Closes#1938 from rxin/TaskContext and squashes the following commits:
145de43 [Reynold Xin] Added JavaTaskCompletionListenerImpl for Java API friendly guarantee.
f435ea5 [Reynold Xin] Added license header for TaskCompletionListener.
dc4ed27 [Reynold Xin] [SPARK-3027] TaskContext: tighten the visibility and provide Java friendly callback API
The previous behaviour of swallowing ClassNotFound exceptions when running a custom Kryo registrator could lead to difficult to debug problems later on at serialisation / deserialisation time, see SPARK-2878. Instead it is better to fail fast.
Added test case.
Author: Graham Dennis <graham.dennis@gmail.com>
Closes#1827 from GrahamDennis/feature/spark-2893 and squashes the following commits:
fbe4cb6 [Graham Dennis] [SPARK-2878]: Update the test case to match the updated exception message
65e53c5 [Graham Dennis] [SPARK-2893]: Improve message when a spark.kryo.registrator fails.
f480d85 [Graham Dennis] [SPARK-2893] Fix typo.
b59d2c2 [Graham Dennis] SPARK-2893: Do not swallow Exceptions when running a custom spark.kryo.registrator
Currently, local execution of Spark jobs is only used by take(), and it can be problematic as it can load a significant amount of data onto the driver. The worst case scenarios occur if the RDD is cached (guaranteed to load whole partition), has very large elements, or the partition is just large and we apply a filter with high selectivity or computational overhead.
Additionally, jobs that run locally in this manner do not show up in the web UI, and are thus harder to track or understand what is occurring.
This PR adds a flag to disable local execution, which is turned OFF by default, with the intention of perhaps eventually removing this functionality altogether. Removing it now is a tougher proposition since it is part of the public runJob API. An alternative solution would be to limit the flag to take()/first() to avoid impacting any external users of this API, but such usage (or, at least, reliance upon the feature) is hopefully minimal.
Author: Aaron Davidson <aaron@databricks.com>
Closes#1321 from aarondav/allowlocal and squashes the following commits:
136b253 [Aaron Davidson] Fix DAGSchedulerSuite
5599d55 [Aaron Davidson] [RFC] Disable local execution of Spark jobs by default
Instead of requiring "org.apache.spark.io.LZ4CompressionCodec", it is easier for users if Spark just accepts "lz4", "lzf", "snappy".
Author: Reynold Xin <rxin@apache.org>
Closes#1873 from rxin/compressionCodecShortForm and squashes the following commits:
9f50962 [Reynold Xin] Specify short-form compression codec names first.
63f78ee [Reynold Xin] Updated configuration documentation.
47b3848 [Reynold Xin] [SPARK-2953] Allow using short names for io compression codecs
This addresses SPARK-2931, a bug where getAllowedLocalityLevel() could throw ArrayIndexOutOfBoundsException. The fix here is to reset currentLocalityIndex after recomputing the locality levels.
Thanks to kayousterhout, mridulm, and lirui-intel for helping me to debug this.
Author: Josh Rosen <joshrosen@apache.org>
Closes#1896 from JoshRosen/SPARK-2931 and squashes the following commits:
48b60b5 [Josh Rosen] Move FakeRackUtil.cleanUp() info beforeEach().
6fec474 [Josh Rosen] Set currentLocalityIndex after recomputing locality levels.
9384897 [Josh Rosen] Update SPARK-2931 test to reflect changes in 63bdb1f41b.
9ecd455 [Josh Rosen] Apply @mridulm's patch for reproducing SPARK-2931.
To enable Python consistency and `Experimental` label of the `sampleByKeyExact` API.
Author: Doris Xin <doris.s.xin@gmail.com>
Author: Xiangrui Meng <meng@databricks.com>
Closes#1866 from dorx/stratified and squashes the following commits:
0ad97b2 [Doris Xin] reviewer comments.
2948aae [Doris Xin] remove unrelated changes
e990325 [Doris Xin] Merge branch 'master' into stratified
555a3f9 [Doris Xin] separate out sampleByKeyExact as its own API
616e55c [Doris Xin] merge master
245439e [Doris Xin] moved minSamplingRate to getUpperBound
eaf5771 [Doris Xin] bug fixes.
17a381b [Doris Xin] fixed a merge issue and a failed unit
ea7d27f [Doris Xin] merge master
b223529 [Xiangrui Meng] use approx bounds for poisson fix poisson mean for waitlisting add unit tests for Java
b3013a4 [Xiangrui Meng] move math3 back to test scope
eecee5f [Doris Xin] Merge branch 'master' into stratified
f4c21f3 [Doris Xin] Reviewer comments
a10e68d [Doris Xin] style fix
a2bf756 [Doris Xin] Merge branch 'master' into stratified
680b677 [Doris Xin] use mapPartitionWithIndex instead
9884a9f [Doris Xin] style fix
bbfb8c9 [Doris Xin] Merge branch 'master' into stratified
ee9d260 [Doris Xin] addressed reviewer comments
6b5b10b [Doris Xin] Merge branch 'master' into stratified
254e03c [Doris Xin] minor fixes and Java API.
4ad516b [Doris Xin] remove unused imports from PairRDDFunctions
bd9dc6e [Doris Xin] unit bug and style violation fixed
1fe1cff [Doris Xin] Changed fractionByKey to a map to enable arg check
944a10c [Doris Xin] [SPARK-2145] Add lower bound on sampling rate
0214a76 [Doris Xin] cleanUp
90d94c0 [Doris Xin] merge master
9e74ab5 [Doris Xin] Separated out most of the logic in sampleByKey
7327611 [Doris Xin] merge master
50581fc [Doris Xin] added a TODO for logging in python
46f6c8c [Doris Xin] fixed the NPE caused by closures being cleaned before being passed into the aggregate function
7e1a481 [Doris Xin] changed the permission on SamplingUtil
1d413ce [Doris Xin] fixed checkstyle issues
9ee94ee [Doris Xin] [SPARK-2082] stratified sampling in PairRDDFunctions that guarantees exact sample size
e3fd6a6 [Doris Xin] Merge branch 'master' into takeSample
7cab53a [Doris Xin] fixed import bug in rdd.py
ffea61a [Doris Xin] SPARK-1939: Refactor takeSample method in RDD
1441977 [Doris Xin] SPARK-1939 Refactor takeSample method in RDD to use ScaSRS
Author: GuoQiang Li <witgo@qq.com>
Closes#1836 from witgo/SPARK-2897 and squashes the following commits:
23cdc5b [GuoQiang Li] review commit
ada4fba [GuoQiang Li] TorrentBroadcast does not support broadcast compression
fb91792 [GuoQiang Li] org.apache.spark.broadcast.TorrentBroadcast does use the serializer class specified in the spark option "spark.serializer"
Author: Erik Erlandson <eerlands@redhat.com>
Closes#1841 from erikerlandson/spark-2911-pr and squashes the following commits:
4699e2f [Erik Erlandson] [SPARK-2911]: provide rdd.parent[T](j) to obtain jth parent RDD
Author: Sandy Ryza <sandy@cloudera.com>
Closes#1507 from sryza/sandy-spark-2565 and squashes the following commits:
74dad41 [Sandy Ryza] SPARK-2565. Update ShuffleReadMetrics as blocks are fetched
As described in https://issues.apache.org/jira/browse/SPARK-2787, right now sort-based shuffle is more expensive than hash-based for map operations that do no partial aggregation or sorting, such as groupByKey. This is because it has to serialize each data item twice (once when spilling to intermediate files, and then again when merging these files object-by-object). This patch adds a code path to just write separate files directly if the # of output partitions is small, and concatenate them at the end to produce a sorted file.
On the unit test side, I added some tests that force or don't force this bypass path to be used, and checked that our tests for other features (e.g. all the operations) cover both cases.
Author: Matei Zaharia <matei@databricks.com>
Closes#1799 from mateiz/SPARK-2787 and squashes the following commits:
88cf26a [Matei Zaharia] Fix rebase
10233af [Matei Zaharia] Review comments
398cb95 [Matei Zaharia] Fix looking up shuffle manager in conf
ca3efd9 [Matei Zaharia] Add docs for shuffle manager properties, and allow short names for them
d0ae3c5 [Matei Zaharia] Fix some comments
90d084f [Matei Zaharia] Add code path to bypass merge-sort in ExternalSorter, and tests
31e5d7c [Matei Zaharia] Move existing logic for writing partitioned files into ExternalSorter
fix bug of countApproxDistinct() when have more than one partition
Author: Davies Liu <davies.liu@gmail.com>
Closes#1812 from davies/approx and squashes the following commits:
bf757ce [Davies Liu] fix bug of countApproxDistinct() when have more than one partition
This patch modifies the ConnectionManager so that error messages are sent in reply when uncaught exceptions occur during message processing. This prevents message senders from hanging while waiting for an acknowledgment if the remote message processing failed.
This is an updated version of sarutak's PR, #1490. The main change is to use Futures / Promises to signal errors.
Author: Kousuke Saruta <sarutak@oss.nttdata.co.jp>
Author: Josh Rosen <joshrosen@apache.org>
Closes#1758 from JoshRosen/connection-manager-fixes and squashes the following commits:
68620cb [Josh Rosen] Fix test in BlockFetcherIteratorSuite:
83673de [Josh Rosen] Error ACKs should trigger IOExceptions, so catch only those exceptions in the test.
b8bb4d4 [Josh Rosen] Fix manager.id vs managerServer.id typo that broke security tests.
659521f [Josh Rosen] Include previous exception when throwing new one
a2f745c [Josh Rosen] Remove sendMessageReliablySync; callers can wait themselves.
c01c450 [Josh Rosen] Return Try[Message] from sendMessageReliablySync.
f1cd1bb [Josh Rosen] Clean up @sarutak's PR #1490 for [SPARK-2583]: ConnectionManager error reporting
7399c6b [Josh Rosen] Merge remote-tracking branch 'origin/pr/1490' into connection-manager-fixes
ee91bb7 [Kousuke Saruta] Modified BufferMessage.scala to keep the spark code style
9dfd0d8 [Kousuke Saruta] Merge branch 'master' of git://git.apache.org/spark into SPARK-2583
e7d9aa6 [Kousuke Saruta] rebase to master
326a17f [Kousuke Saruta] Add test cases to ConnectionManagerSuite.scala for SPARK-2583
2a18d6b [Kousuke Saruta] Merge branch 'master' of git://git.apache.org/spark into SPARK-2583
22d7ebd [Kousuke Saruta] Add test cases to BlockManagerSuite for SPARK-2583
e579302 [Kousuke Saruta] Merge branch 'master' of git://git.apache.org/spark into SPARK-2583
281589c [Kousuke Saruta] Add a test case to BlockFetcherIteratorSuite.scala for fetching block from remote from successfully
0654128 [Kousuke Saruta] Merge branch 'master' of git://git.apache.org/spark into SPARK-2583
ffaa83d [Kousuke Saruta] Merge branch 'master' of git://git.apache.org/spark into SPARK-2583
12d3de8 [Kousuke Saruta] Added BlockFetcherIteratorSuite.scala
4117b8f [Kousuke Saruta] Modified ConnectionManager to be alble to handle error during processing message
717c9c3 [Kousuke Saruta] Merge branch 'master' of git://git.apache.org/spark into SPARK-2583
6635467 [Kousuke Saruta] Merge branch 'master' of git://git.apache.org/spark into SPARK-2583
e2b8c4a [Kousuke Saruta] Modify to propagete error using ConnectionManager
I haven't tested this out on a cluster yet, but wanted to make sure the approach (passing ShuffleWriteMetrics down to DiskBlockObjectWriter) was ok
Author: Sandy Ryza <sandy@cloudera.com>
Closes#1481 from sryza/sandy-spark-2566 and squashes the following commits:
8090d88 [Sandy Ryza] Fix ExternalSorter
b2a62ed [Sandy Ryza] Fix more test failures
8be6218 [Sandy Ryza] Fix test failures and mark a couple variables private
c5e68e5 [Sandy Ryza] SPARK-2566. Update ShuffleWriteMetrics incrementally
JIRA issues:
- Main: [SPARK-2678](https://issues.apache.org/jira/browse/SPARK-2678)
- Related: [SPARK-2874](https://issues.apache.org/jira/browse/SPARK-2874)
Related PR:
- #1715
This PR is both a fix for SPARK-2874 and a workaround for SPARK-2678. Fixing SPARK-2678 completely requires some API level changes that need further discussion, and we decided not to include it in Spark 1.1 release. As currently SPARK-2678 only affects Spark SQL scripts, this workaround is enough for Spark 1.1. Command line option handling logic in bash scripts looks somewhat dirty and duplicated, but it helps to provide a cleaner user interface as well as retain full downward compatibility for now.
Author: Cheng Lian <lian.cs.zju@gmail.com>
Closes#1801 from liancheng/spark-2874 and squashes the following commits:
8045d7a [Cheng Lian] Make sure test suites pass
8493a9e [Cheng Lian] Using eval to retain quoted arguments
aed523f [Cheng Lian] Fixed typo in bin/spark-sql
f12a0b1 [Cheng Lian] Worked arount SPARK-2678
daee105 [Cheng Lian] Fixed usage messages of all Spark SQL related scripts
The goal of this PR is to allow users of Spark to write tight firewall rules for their clusters. This is currently not possible because Spark uses random ports in many places, notably the communication between executors and drivers. The changes in this PR are based on top of ash211's changes in #1107.
The list covered here may or may not be the complete set of port needed for Spark to operate perfectly. However, as of the latest commit there are no known sources of random ports (except in tests). I have not documented a few of the more obscure configs.
My spark-env.sh looks like this:
```
export SPARK_MASTER_PORT=6060
export SPARK_WORKER_PORT=7070
export SPARK_MASTER_WEBUI_PORT=9090
export SPARK_WORKER_WEBUI_PORT=9091
```
and my spark-defaults.conf looks like this:
```
spark.master spark://andrews-mbp:6060
spark.driver.port 5001
spark.fileserver.port 5011
spark.broadcast.port 5021
spark.replClassServer.port 5031
spark.blockManager.port 5041
spark.executor.port 5051
```
Author: Andrew Or <andrewor14@gmail.com>
Author: Andrew Ash <andrew@andrewash.com>
Closes#1777 from andrewor14/configure-ports and squashes the following commits:
621267b [Andrew Or] Merge branch 'master' of github.com:apache/spark into configure-ports
8a6b820 [Andrew Or] Use a random UI port during tests
7da0493 [Andrew Or] Fix tests
523c30e [Andrew Or] Add test for isBindCollision
b97b02a [Andrew Or] Minor fixes
c22ad00 [Andrew Or] Merge branch 'master' of github.com:apache/spark into configure-ports
93d359f [Andrew Or] Executors connect to wrong port when collision occurs
d502e5f [Andrew Or] Handle port collisions when creating Akka systems
a2dd05c [Andrew Or] Patrick's comment nit
86461e2 [Andrew Or] Remove spark.executor.env.port and spark.standalone.client.port
1d2d5c6 [Andrew Or] Fix ports for standalone cluster mode
cb3be88 [Andrew Or] Various doc fixes (broken link, format etc.)
e837cde [Andrew Or] Remove outdated TODOs
bfbab28 [Andrew Or] Merge branch 'master' of github.com:apache/spark into configure-ports
de1b207 [Andrew Or] Update docs to reflect new ports
b565079 [Andrew Or] Add spark.ports.maxRetries
2551eb2 [Andrew Or] Remove spark.worker.watcher.port
151327a [Andrew Or] Merge branch 'master' of github.com:apache/spark into configure-ports
9868358 [Andrew Or] Add a few miscellaneous ports
6016e77 [Andrew Or] Add spark.executor.port
8d836e6 [Andrew Or] Also document SPARK_{MASTER/WORKER}_WEBUI_PORT
4d9e6f3 [Andrew Or] Fix super subtle bug
3f8e51b [Andrew Or] Correct erroneous docs...
e111d08 [Andrew Or] Add names for UI services
470f38c [Andrew Or] Special case non-"Address already in use" exceptions
1d7e408 [Andrew Or] Treat 0 ports specially + return correct ConnectionManager port
ba32280 [Andrew Or] Minor fixes
6b550b0 [Andrew Or] Assorted fixes
73fbe89 [Andrew Or] Move start service logic to Utils
ec676f4 [Andrew Or] Merge branch 'SPARK-2157' of github.com:ash211/spark into configure-ports
038a579 [Andrew Ash] Trust the server start function to report the port the service started on
7c5bdc4 [Andrew Ash] Fix style issue
0347aef [Andrew Ash] Unify port fallback logic to a single place
24a4c32 [Andrew Ash] Remove type on val to match surrounding style
9e4ad96 [Andrew Ash] Reformat for style checker
5d84e0e [Andrew Ash] Document new port configuration options
066dc7a [Andrew Ash] Fix up HttpServer port increments
cad16da [Andrew Ash] Add fallover increment logic for HttpServer
c5a0568 [Andrew Ash] Fix ConnectionManager to retry with increment
b80d2fd [Andrew Ash] Make Spark's block manager port configurable
17c79bb [Andrew Ash] Add a configuration option for spark-shell's class server
f34115d [Andrew Ash] SPARK-1176 Add port configuration for HttpBroadcast
49ee29b [Andrew Ash] SPARK-1174 Add port configuration for HttpFileServer
1c0981a [Andrew Ash] Make port in HttpServer configurable
copied from original JIRA (https://issues.apache.org/jira/browse/SPARK-2294):
If an executor E is free, a task may be speculatively assigned to E when there are other tasks in the job that have not been launched (at all) yet. Similarly, a task without any locality preferences may be assigned to E when there was another NODE_LOCAL task that could have been scheduled.
This happens because TaskSchedulerImpl calls TaskSetManager.resourceOffer (which in turn calls TaskSetManager.findTask) with increasing locality levels, beginning with PROCESS_LOCAL, followed by NODE_LOCAL, and so on until the highest currently allowed level. Now, supposed NODE_LOCAL is the highest currently allowed locality level. The first time findTask is called, it will be called with max level PROCESS_LOCAL; if it cannot find any PROCESS_LOCAL tasks, it will try to schedule tasks with no locality preferences or speculative tasks. As a result, speculative tasks or tasks with no preferences may be scheduled instead of NODE_LOCAL tasks.
----
I added an additional parameter in resourceOffer and findTask, maxLocality, indicating when we should consider the tasks without locality preference
Author: CodingCat <zhunansjtu@gmail.com>
Closes#1313 from CodingCat/SPARK-2294 and squashes the following commits:
bf3f13b [CodingCat] rollback some forgotten changes
89f9bc0 [CodingCat] address matei's comments
18cae02 [CodingCat] add test case for node-local tasks
2ba6195 [CodingCat] fix failed test cases
87dd09e [CodingCat] fix style
9b9432f [CodingCat] remove hasNodeLocalOnlyTasks
fdd1573 [CodingCat] fix failed test cases
941a4fd [CodingCat] see my shocked face..........
f600085 [CodingCat] remove hasNodeLocalOnlyTasks checking
0b8a46b [CodingCat] test whether hasNodeLocalOnlyTasks affect the results
73ceda8 [CodingCat] style fix
b3a430b [CodingCat] remove fine granularity tracking for node-local only tasks
f9a2ad8 [CodingCat] simplify the logic in TaskSchedulerImpl
c8c1de4 [CodingCat] simplify the patch
be652ed [CodingCat] avoid unnecessary delay when we only have nopref tasks
dee9e22 [CodingCat] fix locality inversion bug in TaskManager by moving nopref branch
This patch adds support for giving accumulators user-visible names and displaying accumulator values in the web UI. This allows users to create custom counters that can display in the UI. The current approach displays both the accumulator deltas caused by each task and a "current" value of the accumulator totals for each stage, which gets update as tasks finish.
Currently in Spark developers have been extending the `TaskMetrics` functionality to provide custom instrumentation for RDD's. This provides a potentially nicer alternative of going through the existing accumulator framework (actually `TaskMetrics` and accumulators are on an awkward collision course as we add more features to the former). The current patch demo's how we can use the feature to provide instrumentation for RDD input sizes. The nice thing about going through accumulators is that users can read the current value of the data being tracked in their programs. This could be useful to e.g. decide to short-circuit a Spark stage depending on how things are going.
Author: Patrick Wendell <pwendell@gmail.com>
Closes#1309 from pwendell/metrics and squashes the following commits:
8815308 [Patrick Wendell] Merge remote-tracking branch 'apache/master' into HEAD
93fbe0f [Patrick Wendell] Other minor fixes
cc43f68 [Patrick Wendell] Updating unit tests
c991b1b [Patrick Wendell] Moving some code into the Accumulators class
9a9ba3c [Patrick Wendell] More merge fixes
c5ace9e [Patrick Wendell] More merge conflicts
1da15e3 [Patrick Wendell] Merge remote-tracking branch 'apache/master' into metrics
9860c55 [Patrick Wendell] Potential solution to posting listener events
0bb0e33 [Patrick Wendell] Remove "display" variable and assume display = name.isDefined
0ec4ac7 [Patrick Wendell] Java API's
e95bf69 [Patrick Wendell] Stash
be97261 [Patrick Wendell] Style fix
8407308 [Patrick Wendell] Removing examples in Hadoop and RDD class
64d405f [Patrick Wendell] Adding missing file
5d8b156 [Patrick Wendell] Changes based on Kay's review.
9f18bad [Patrick Wendell] Minor style changes and tests
7a63abc [Patrick Wendell] Adding Json serialization and responding to Reynold's feedback
ad85076 [Patrick Wendell] Example of using named accumulators for custom RDD metrics.
0b72660 [Patrick Wendell] Initial WIP example of supporing globally named accumulators.
It was easier to combine these 2 jira since they touch many of the same places. This pr adds the following:
- adds modify acls
- adds admin acls (list of admins/users that get added to both view and modify acls)
- modify Kill button on UI to take modify acls into account
- changes config name of spark.ui.acls.enable to spark.acls.enable since I choose poorly in original name. We keep backwards compatibility so people can still use spark.ui.acls.enable. The acls should apply to any web ui as well as any CLI interfaces.
- send view and modify acls information on to YARN so that YARN interfaces can use (yarn cli for killing applications for example).
Author: Thomas Graves <tgraves@apache.org>
Closes#1196 from tgravescs/SPARK-1890 and squashes the following commits:
8292eb1 [Thomas Graves] review comments
b92ec89 [Thomas Graves] remove unneeded variable from applistener
4c765f4 [Thomas Graves] Add in admin acls
72eb0ac [Thomas Graves] Add modify acls
This tracks memory properly if there are multiple spilling collections in the same task (which was a problem before), and also implements an algorithm that lets each thread grow up to 1 / 2N of the memory pool (where N is the number of threads) before spilling, which avoids an inefficiency with small spills we had before (some threads would spill many times at 0-1 MB because the pool was allocated elsewhere).
Author: Matei Zaharia <matei@databricks.com>
Closes#1707 from mateiz/spark-2711 and squashes the following commits:
debf75b [Matei Zaharia] Review comments
24f28f3 [Matei Zaharia] Small rename
c8f3a8b [Matei Zaharia] Update ShuffleMemoryManager to be able to partially grant requests
315e3a5 [Matei Zaharia] Some review comments
b810120 [Matei Zaharia] Create central manager to track memory for all spilling collections
Author: Reynold Xin <rxin@apache.org>
Closes#1772 from rxin/accumulator-dagscheduler and squashes the following commits:
6a58520 [Reynold Xin] [SPARK-2323] Exception in accumulator update should not crash DAGScheduler & SparkContext.
All these changes are from mridulm's work in #1609, but extracted here to fix this specific issue and make it easier to merge not 1.1. This particular set of changes is to make sure that we read exactly the right range of bytes from each spill file in EAOM: some serializers can write bytes after the last object (e.g. the TC_RESET flag in Java serialization) and that would confuse the previous code into reading it as part of the next batch. There are also improvements to cleanup to make sure files are closed.
In addition to bringing in the changes to ExternalAppendOnlyMap, I also copied them to the corresponding code in ExternalSorter and updated its test suite to test for the same issues.
Author: Matei Zaharia <matei@databricks.com>
Closes#1722 from mateiz/spark-2792 and squashes the following commits:
5d4bfb5 [Matei Zaharia] Make objectStreamReset counter count the last object written too
18fe865 [Matei Zaharia] Update docs on objectStreamReset
576ee83 [Matei Zaharia] Allow objectStreamReset to be 0
0374217 [Matei Zaharia] Remove super paranoid code to close file handles
bda37bb [Matei Zaharia] Implement Mridul's ExternalAppendOnlyMap fixes in ExternalSorter too
0d6dad7 [Matei Zaharia] Added Mridul's test changes for ExternalAppendOnlyMap
9a78e4b [Matei Zaharia] Add @mridulm's fixes to ExternalAppendOnlyMap for batch sizes
`spark.test.home` was introduced in #1734. This is fine for SBT but is failing maven tests. Either way it shouldn't throw an NPE.
Author: Andrew Or <andrewor14@gmail.com>
Closes#1739 from andrewor14/fix-spark-test-home and squashes the following commits:
ce2624c [Andrew Or] Do not throw NPE if spark.test.home is not set
When standalone Workers launch executors, they inherit the Spark home set by the driver. This means if the worker machines do not share the same directory structure as the driver node, the Workers will attempt to run scripts (e.g. bin/compute-classpath.sh) that do not exist locally and fail. This is a common scenario if the driver is launched from outside of the cluster.
The solution is to simply not pass the driver's Spark home to the Workers. This PR further makes an attempt to avoid overloading the usages of `spark.home`, which is now only used for setting executor Spark home on Mesos and in python.
This is based on top of #1392 and originally reported by YanTangZhai. Tested on standalone cluster.
Author: Andrew Or <andrewor14@gmail.com>
Closes#1734 from andrewor14/spark-home-reprise and squashes the following commits:
f71f391 [Andrew Or] Revert changes in python
1c2532c [Andrew Or] Merge branch 'master' of github.com:apache/spark into spark-home-reprise
188fc5d [Andrew Or] Avoid using spark.home where possible
09272b7 [Andrew Or] Always use Worker's working directory as spark home
The existing code in `StorageUtils` is not the most efficient. Every time we want to update an `RDDInfo` we end up iterating through all blocks on all block managers just to discard most of them. The symptoms manifest themselves in the bountiful UI bugs observed in the wild. Many of these bugs are caused by the slow consumption of events in `LiveListenerBus`, which frequently leads to the event queue overflowing and `SparkListenerEvent`s being dropped on the floor. The changes made in this PR avoid this by first filtering out only the blocks relevant to us before computing storage information from them.
It's worth a mention that this corner of the Spark code is also not very well-tested at all. The bulk of the changes in this PR (more than 60%) is actually test cases for the various logic in `StorageUtils.scala` as well as `StorageTab.scala`. These will eventually be extended to cover the various listeners that constitute the `SparkUI`.
Author: Andrew Or <andrewor14@gmail.com>
Closes#1679 from andrewor14/fix-drop-events and squashes the following commits:
f80c1fa [Andrew Or] Rewrite fold and reduceOption as sum
e132d69 [Andrew Or] Merge branch 'master' of github.com:apache/spark into fix-drop-events
14fa1c3 [Andrew Or] Simplify some code + update a few comments
a91be46 [Andrew Or] Make ExecutorsPage blazingly fast
bf6f09b [Andrew Or] Minor changes
8981de1 [Andrew Or] Merge branch 'master' of github.com:apache/spark into fix-drop-events
af19bc0 [Andrew Or] *UsedByRDD -> *UsedByRdd (minor)
6970bc8 [Andrew Or] Add extensive tests for StorageListener and the new code in StorageUtils
e080b9e [Andrew Or] Reduce run time of StorageUtils.updateRddInfo to near constant
2c3ef6a [Andrew Or] Actually filter out only the relevant RDDs
6fef86a [Andrew Or] Add extensive tests for new code in StorageStatus
b66b6b0 [Andrew Or] Use more efficient underlying data structures for blocks
6a7b7c0 [Andrew Or] Avoid chained operations on TraversableLike
a9ec384 [Andrew Or] Merge branch 'master' of github.com:apache/spark into fix-drop-events
b12fcd7 [Andrew Or] Fix tests + simplify sc.getRDDStorageInfo
da8e322 [Andrew Or] Merge branch 'master' of github.com:apache/spark into fix-drop-events
8e91921 [Andrew Or] Iterate through a filtered set of blocks when updating RDDInfo
7b2c4aa [Andrew Or] Rewrite blockLocationsFromStorageStatus + clean up method signatures
41fa50d [Andrew Or] Add a legacy constructor for StorageStatus
53af15d [Andrew Or] Refactor StorageStatus + add a bunch of tests
All changes from this PR are by mridulm and are drawn from his work in #1609. This patch is intended to fix all major issues related to shuffle file consolidation that mridulm found, while minimizing changes to the code, with the hope that it may be more easily merged into 1.1.
This patch is **not** intended as a replacement for #1609, which provides many additional benefits, including fixes to ExternalAppendOnlyMap, improvements to DiskBlockObjectWriter's API, and several new unit tests.
If it is feasible to merge #1609 for the 1.1 deadline, that is a preferable option.
Author: Aaron Davidson <aaron@databricks.com>
Closes#1678 from aarondav/consol and squashes the following commits:
53b3f6d [Aaron Davidson] Correct behavior when writing unopened file
701d045 [Aaron Davidson] Rebase with sort-based shuffle
9160149 [Aaron Davidson] SPARK-2532: Minimal shuffle consolidation fixes
getPreferredLocs traverses a dependency graph of partitions using depth first search. Given a complex dependency graph, the old implementation may explore a set of paths in the graph that is exponential in the number of nodes. By maintaining a set of visited nodes the new implementation avoids revisiting nodes, preventing exponential blowup.
Some comment and whitespace cleanups are also included.
Author: Aaron Staple <aaron.staple@gmail.com>
Closes#1362 from staple/SPARK-695 and squashes the following commits:
ecea0f3 [Aaron Staple] address review comments
751c661 [Aaron Staple] [SPARK-695] Add a unit test.
5adf326 [Aaron Staple] Replace getPreferredLocsInternal's HashMap argument with a simpler HashSet.
58e37d0 [Aaron Staple] Replace comment documenting NarrowDependency.
6751ced [Aaron Staple] Revert "Remove unused variable."
04c7097 [Aaron Staple] Fix indentation.
0030884 [Aaron Staple] Remove unused variable.
33f67c6 [Aaron Staple] Clarify comment.
4e42b46 [Aaron Staple] Remove apparently incorrect comment describing NarrowDependency.
65c2d3d [Aaron Staple] [SPARK-695] In DAGScheduler's getPreferredLocs, track set of visited partitions.
This is a sketch of a patch that allows the UI to show metrics for tasks that have not yet completed. It adds a heartbeat every 2 seconds from the executors to the driver, reporting metrics for all of the executor's tasks.
It still needs unit tests, polish, and cluster testing, but I wanted to put it up to get feedback on the approach.
Author: Sandy Ryza <sandy@cloudera.com>
Closes#1056 from sryza/sandy-spark-2099 and squashes the following commits:
93b9fdb [Sandy Ryza] Up heartbeat interval to 10 seconds and other tidying
132aec7 [Sandy Ryza] Heartbeat and HeartbeatResponse are already Serializable as case classes
38dffde [Sandy Ryza] Additional review feedback and restore test that was removed in BlockManagerSuite
51fa396 [Sandy Ryza] Remove hostname race, add better comments about threading, and some stylistic improvements
3084f10 [Sandy Ryza] Make TaskUIData a case class again
3bda974 [Sandy Ryza] Stylistic fixes
0dae734 [Sandy Ryza] SPARK-2099. Report progress while task is running.
[SPARK-2557](https://issues.apache.org/jira/browse/SPARK-2557)
Author: Ye Xianjin <advancedxy@gmail.com>
Closes#1464 from advancedxy/SPARK-2557 and squashes the following commits:
d844d67 [Ye Xianjin] add local-*-n-failures, bad-local-n, bad-local-n-failures test case
3bbc668 [Ye Xianjin] fix LOCAL_N_REGEX regular expression and make local_n_failures accept * as all cores on the computer
This patch simply uses the ExternalSorter class from sort-based shuffle.
Closes#931 and Closes#1090
Author: Matei Zaharia <matei@databricks.com>
Closes#1677 from mateiz/spark-983 and squashes the following commits:
96b3fda [Matei Zaharia] SPARK-983. Support external sorting in sortByKey()
Author: Kousuke Saruta <sarutak@oss.nttdata.co.jp>
Closes#1578 from sarutak/SPARK-2670 and squashes the following commits:
85c8938 [Kousuke Saruta] Removed useless results.put for fail fast
e8713cc [Kousuke Saruta] Merge branch 'master' of git://git.apache.org/spark into SPARK-2670
d353984 [Kousuke Saruta] Refined assertion messages in BlockFetcherIteratorSuite.scala
03bcb02 [Kousuke Saruta] Merge branch 'SPARK-2670' of github.com:sarutak/spark into SPARK-2670
5d05855 [Kousuke Saruta] Merge branch 'master' of git://git.apache.org/spark into SPARK-2670
4fca130 [Kousuke Saruta] Added test cases for BasicBlockFetcherIterator
b7b8250 [Kousuke Saruta] Modified BasicBlockFetchIterator to fail fast when local fetch error has been occurred
a3a9be1 [Kousuke Saruta] Modified BlockFetcherIterator for SPARK-2670
460dc01 [Kousuke Saruta] Merge branch 'master' of git://git.apache.org/spark into SPARK-2670
e310c0b [Kousuke Saruta] Modified BlockFetcherIterator to handle local fetch failure as fatch fail
Author: Sandy Ryza <sandy@cloudera.com>
Closes#1642 from sryza/sandy-spark-2738 and squashes the following commits:
a923e4e [Sandy Ryza] SPARK-2738. Remove redundant imports in BlockManagerSuite
...ags
Author: Sandy Ryza <sandy@cloudera.com>
Closes#1665 from sryza/sandy-spark-2664 and squashes the following commits:
0518c63 [Sandy Ryza] SPARK-2664. Deal with `--conf` options in spark-submit that relate to flags
This allows users to gain access to the InputSplit which backs each partition.
An alternative solution would have been to have a .withInputSplit() method which returns a new RDD[(InputSplit, (K, V))], but this is confusing because you could not cache this RDD or shuffle it, as InputSplit is not inherently serializable.
Author: Aaron Davidson <aaron@databricks.com>
Closes#973 from aarondav/hadoop and squashes the following commits:
9c9112b [Aaron Davidson] Add JavaAPISuite test
9942cd7 [Aaron Davidson] Add Java API
1284a3a [Aaron Davidson] SPARK-2028: Expose mapPartitionsWithInputSplit in HadoopRDD
The Java API's use of fake ClassTags doesn't seem to cause any problems for Java users, but it can lead to issues when passing JavaRDDs' underlying RDDs to Scala code (e.g. in the MLlib Java API wrapper code). If we call collect() on a Scala RDD with an incorrect ClassTag, this causes ClassCastExceptions when we try to allocate an array of the wrong type (for example, see SPARK-2197).
There are a few possible fixes here. An API-breaking fix would be to completely remove the fake ClassTags and require Java API users to pass java.lang.Class instances to all parallelize() calls and add returnClass fields to all Function implementations. This would be extremely verbose.
Instead, this patch adds internal APIs to "repair" a Scala RDD with an incorrect ClassTag by wrapping it and overriding its ClassTag. This should be okay for cases where the Scala code that calls collect() knows what type of array should be allocated, which is the case in the MLlib wrappers.
Author: Josh Rosen <joshrosen@apache.org>
Closes#1639 from JoshRosen/SPARK-2737 and squashes the following commits:
572b4c8 [Josh Rosen] Replace newRDD[T] with mapPartitions().
469d941 [Josh Rosen] Preserve partitioner in retag().
af78816 [Josh Rosen] Allow retag() to get classTag implicitly.
d1d54e6 [Josh Rosen] [SPARK-2737] Add retag() method for changing RDDs' ClassTags.
We resolve relative paths to the local `file:/` system for `--jars` and `--files` in spark submit (#853). We should do the same for the history server.
Author: Andrew Or <andrewor14@gmail.com>
Closes#1280 from andrewor14/hist-serv-fix and squashes the following commits:
13ff406 [Andrew Or] Merge branch 'master' of github.com:apache/spark into hist-serv-fix
b393e17 [Andrew Or] Strip trailing "/" from logging directory
622a471 [Andrew Or] Fix test in EventLoggingListenerSuite
0e20f71 [Andrew Or] Shift responsibility of resolving paths up one level
b037c0c [Andrew Or] Use resolved paths for everything in history server
c7e36ee [Andrew Or] Resolve paths for event logging too
40e3933 [Andrew Or] Resolve history server file paths
Author: Reynold Xin <rxin@apache.org>
Closes#1675 from rxin/unionrdd and squashes the following commits:
941d316 [Reynold Xin] Clear RDDs for checkpointing.
c9f05f2 [Reynold Xin] [SPARK-2758] UnionRDD's UnionPartition should not reference parent RDDs
This adds a new ShuffleManager based on sorting, as described in https://issues.apache.org/jira/browse/SPARK-2045. The bulk of the code is in an ExternalSorter class that is similar to ExternalAppendOnlyMap, but sorts key-value pairs by partition ID and can be used to create a single sorted file with a map task's output. (Longer-term I think this can take on the remaining functionality in ExternalAppendOnlyMap and replace it so we don't have code duplication.)
The main TODOs still left are:
- [x] enabling ExternalSorter to merge across spilled files
- [x] with an Ordering
- [x] without an Ordering, using the keys' hash codes
- [x] adding more tests (e.g. a version of our shuffle suite that runs on this)
- [x] rebasing on top of the size-tracking refactoring in #1165 when that is merged
- [x] disabling spilling if spark.shuffle.spill is set to false
Despite this though, this seems to work pretty well (running successfully in cases where the hash shuffle would OOM, such as 1000 reduce tasks on executors with only 1G memory), and it seems to be comparable in speed or faster than hash-based shuffle (it will create much fewer files for the OS to keep track of). So I'm posting it to get some early feedback.
After these TODOs are done, I'd also like to enable ExternalSorter to sort data within each partition by a key as well, which will allow us to use it to implement external spilling in reduce tasks in `sortByKey`.
Author: Matei Zaharia <matei@databricks.com>
Closes#1499 from mateiz/sort-based-shuffle and squashes the following commits:
bd841f9 [Matei Zaharia] Various review comments
d1c137fd [Matei Zaharia] Various review comments
a611159 [Matei Zaharia] Compile fixes due to rebase
62c56c8 [Matei Zaharia] Fix ShuffledRDD sometimes not returning Tuple2s.
f617432 [Matei Zaharia] Fix a failing test (seems to be due to change in SizeTracker logic)
9464d5f [Matei Zaharia] Simplify code and fix conflicts after latest rebase
0174149 [Matei Zaharia] Add cleanup behavior and cleanup tests for sort-based shuffle
eb4ee0d [Matei Zaharia] Remove customizable element type in ShuffledRDD
fa2e8db [Matei Zaharia] Allow nextBatchStream to be called after we're done looking at all streams
a34b352 [Matei Zaharia] Fix tracking of indices within a partition in SpillReader, and add test
03e1006 [Matei Zaharia] Add a SortShuffleSuite that runs ShuffleSuite with sort-based shuffle
3c7ff1f [Matei Zaharia] Obey the spark.shuffle.spill setting in ExternalSorter
ad65fbd [Matei Zaharia] Rebase on top of Aaron's Sorter change, and use Sorter in our buffer
44d2a93 [Matei Zaharia] Use estimateSize instead of atGrowThreshold to test collection sizes
5686f71 [Matei Zaharia] Optimize merging phase for in-memory only data:
5461cbb [Matei Zaharia] Review comments and more tests (e.g. tests with 1 element per partition)
e9ad356 [Matei Zaharia] Update ContextCleanerSuite to make sure shuffle cleanup tests use hash shuffle (since they were written for it)
c72362a [Matei Zaharia] Added bug fix and test for when iterators are empty
de1fb40 [Matei Zaharia] Make trait SizeTrackingCollection private[spark]
4988d16 [Matei Zaharia] tweak
c1b7572 [Matei Zaharia] Small optimization
ba7db7f [Matei Zaharia] Handle null keys in hash-based comparator, and add tests for collisions
ef4e397 [Matei Zaharia] Support for partial aggregation even without an Ordering
4b7a5ce [Matei Zaharia] More tests, and ability to sort data if a total ordering is given
e1f84be [Matei Zaharia] Fix disk block manager test
5a40a1c [Matei Zaharia] More tests
614f1b4 [Matei Zaharia] Add spill metrics to map tasks
cc52caf [Matei Zaharia] Add more error handling and tests for error cases
bbf359d [Matei Zaharia] More work
3a56341 [Matei Zaharia] More partial work towards sort-based shuffle
7a0895d [Matei Zaharia] Some more partial work towards sort-based shuffle
b615476 [Matei Zaharia] Scaffolding for sort-based shuffle
This is a resubmission of #1452. It was reverted because it broke the build.
Currently (as of Spark 1.0.1), Spark sends RDD object (which contains closures) using Akka along with the task itself to the executors. This is inefficient because all tasks in the same stage use the same RDD object, but we have to send RDD object multiple times to the executors. This is especially bad when a closure references some variable that is very large. The current design led to users having to explicitly broadcast large variables.
The patch uses broadcast to send RDD objects and the closures to executors, and use Akka to only send a reference to the broadcast RDD/closure along with the partition specific information for the task. For those of you who know more about the internals, Spark already relies on broadcast to send the Hadoop JobConf every time it uses the Hadoop input, because the JobConf is large.
The user-facing impact of the change include:
1. Users won't need to decide what to broadcast anymore, unless they would want to use a large object multiple times in different operations
2. Task size will get smaller, resulting in faster scheduling and higher task dispatch throughput.
In addition, the change will simplify some internals of Spark, eliminating the need to maintain task caches and the complex logic to broadcast JobConf (which also led to a deadlock recently).
A simple way to test this:
```scala
val a = new Array[Byte](1000*1000); scala.util.Random.nextBytes(a);
sc.parallelize(1 to 1000, 1000).map { x => a; x }.groupBy { x => a; x }.count
```
Numbers on 3 r3.8xlarge instances on EC2
```
master branch: 5.648436068 s, 4.715361895 s, 5.360161877 s
with this change: 3.416348793 s, 1.477846558 s, 1.553432156 s
```
Author: Reynold Xin <rxin@apache.org>
Closes#1498 from rxin/broadcast-task and squashes the following commits:
f7364db [Reynold Xin] Code review feedback.
f8535dc [Reynold Xin] Fixed the style violation.
252238d [Reynold Xin] Serialize the final task closure as well as ShuffleDependency in taskBinary.
111007d [Reynold Xin] Fix broadcast tests.
797c247 [Reynold Xin] Properly send SparkListenerStageSubmitted and SparkListenerStageCompleted.
bab1d8b [Reynold Xin] Check for NotSerializableException in submitMissingTasks.
cf38450 [Reynold Xin] Use TorrentBroadcastFactory.
991c002 [Reynold Xin] Use HttpBroadcast.
de779f8 [Reynold Xin] Fix TaskContextSuite.
cc152fc [Reynold Xin] Don't cache the RDD broadcast variable.
d256b45 [Reynold Xin] Fixed unit test failures. One more to go.
cae0af3 [Reynold Xin] [SPARK-2521] Broadcast RDD object (instead of sending it along with every task).
Author: Koert Kuipers <koert@tresata.com>
Closes#735 from koertkuipers/feat-kryo-max-buffersize and squashes the following commits:
15f6d81 [Koert Kuipers] change default for spark.kryoserializer.buffer.max.mb to 64mb and add some documentation
1bcc22c [Koert Kuipers] Merge branch 'master' into feat-kryo-max-buffersize
0c9f8eb [Koert Kuipers] make default for kryo max buffer size 16MB
143ec4d [Koert Kuipers] test resizable buffer in kryo Output
0732445 [Koert Kuipers] support setting maxCapacity to something different than capacity in kryo Output
The main thing was that spark configs were not propagated to the driver, and so applications that do not specify `master` or `appName` automatically failed. This PR fixes that and a couple of miscellaneous things that are related.
One thing that may or may not be an issue is that the jars must be available on the driver node. In `standalone-cluster` mode, this effectively means these jars must be available on all the worker machines, since the driver is launched on one of them. The semantics here are not the same as `yarn-cluster` mode, where all the relevant jars are uploaded to a distributed cache automatically and shipped to the containers. This is probably not a concern, but still worth a mention.
Author: Andrew Or <andrewor14@gmail.com>
Closes#1538 from andrewor14/standalone-cluster and squashes the following commits:
8c11a0d [Andrew Or] Clean up imports / comments (minor)
2678d13 [Andrew Or] Handle extraJavaOpts properly
7660547 [Andrew Or] Merge branch 'master' of github.com:apache/spark into standalone-cluster
6f64a9b [Andrew Or] Revert changes in YARN
2f2908b [Andrew Or] Fix tests
ed01491 [Andrew Or] Don't go overboard with escaping
8e105e1 [Andrew Or] Merge branch 'master' of github.com:apache/spark into standalone-cluster
b890949 [Andrew Or] Abstract usages of converting spark opts to java opts
79f63a3 [Andrew Or] Move sparkProps into javaOpts
78752f8 [Andrew Or] Fix tests
5a9c6c7 [Andrew Or] Fix line too long
c141a00 [Andrew Or] Don't display "unknown app" on driver log pages
d7e2728 [Andrew Or] Avoid deprecation warning in standalone Client
6ceb14f [Andrew Or] Allow relevant configs to propagate to standalone Driver
7f854bc [Andrew Or] Fix test
855256e [Andrew Or] Fix standalone-cluster mode
fd9da51 [Andrew Or] Formatting changes (minor)
As of Spark 1.0, RangePartitioner goes through data twice: once to compute the count and once to do sampling. As a result, to do sortByKey, Spark goes through data 3 times (once to count, once to sample, and once to sort).
`RangePartitioner` should go through data only once, collecting samples from input partitions as well as counting. If the data is balanced, this should give us a good sketch. If we see big partitions, we re-sample from them in order to collect enough items.
The downside is that we need to collect more from each partition in the first pass. An alternative solution is caching the intermediate result and decide whether to fetch the data after.
Author: Xiangrui Meng <meng@databricks.com>
Author: Reynold Xin <rxin@apache.org>
Closes#1562 from mengxr/range-partitioner and squashes the following commits:
6cc2551 [Xiangrui Meng] change foreach to for
eb39b08 [Xiangrui Meng] Merge branch 'master' into range-partitioner
eb95dd8 [Xiangrui Meng] separate sketching and determining bounds impl
c436d30 [Xiangrui Meng] fix binary metrics unit tests
db58a55 [Xiangrui Meng] add unit tests
a6e35d6 [Xiangrui Meng] minor update
60be09e [Xiangrui Meng] remove importance sampler
9ee9992 [Xiangrui Meng] update range partitioner to run only one job on roughly balanced data
cc12f47 [Xiangrui Meng] Merge remote-tracking branch 'apache/master' into range-part
06ac2ec [Xiangrui Meng] Merge remote-tracking branch 'apache/master' into range-part
17bcbf3 [Reynold Xin] Added seed.
badf20d [Reynold Xin] Renamed the method.
6940010 [Reynold Xin] Reservoir sampling implementation.
Implemented stratified sampling that guarantees exact sample size using ScaRSR with two passes over the RDD for sampling without replacement and three passes for sampling with replacement.
Author: Doris Xin <doris.s.xin@gmail.com>
Author: Xiangrui Meng <meng@databricks.com>
Closes#1025 from dorx/stratified and squashes the following commits:
245439e [Doris Xin] moved minSamplingRate to getUpperBound
eaf5771 [Doris Xin] bug fixes.
17a381b [Doris Xin] fixed a merge issue and a failed unit
ea7d27f [Doris Xin] merge master
b223529 [Xiangrui Meng] use approx bounds for poisson fix poisson mean for waitlisting add unit tests for Java
b3013a4 [Xiangrui Meng] move math3 back to test scope
eecee5f [Doris Xin] Merge branch 'master' into stratified
f4c21f3 [Doris Xin] Reviewer comments
a10e68d [Doris Xin] style fix
a2bf756 [Doris Xin] Merge branch 'master' into stratified
680b677 [Doris Xin] use mapPartitionWithIndex instead
9884a9f [Doris Xin] style fix
bbfb8c9 [Doris Xin] Merge branch 'master' into stratified
ee9d260 [Doris Xin] addressed reviewer comments
6b5b10b [Doris Xin] Merge branch 'master' into stratified
254e03c [Doris Xin] minor fixes and Java API.
4ad516b [Doris Xin] remove unused imports from PairRDDFunctions
bd9dc6e [Doris Xin] unit bug and style violation fixed
1fe1cff [Doris Xin] Changed fractionByKey to a map to enable arg check
944a10c [Doris Xin] [SPARK-2145] Add lower bound on sampling rate
0214a76 [Doris Xin] cleanUp
90d94c0 [Doris Xin] merge master
9e74ab5 [Doris Xin] Separated out most of the logic in sampleByKey
7327611 [Doris Xin] merge master
50581fc [Doris Xin] added a TODO for logging in python
46f6c8c [Doris Xin] fixed the NPE caused by closures being cleaned before being passed into the aggregate function
7e1a481 [Doris Xin] changed the permission on SamplingUtil
1d413ce [Doris Xin] fixed checkstyle issues
9ee94ee [Doris Xin] [SPARK-2082] stratified sampling in PairRDDFunctions that guarantees exact sample size
e3fd6a6 [Doris Xin] Merge branch 'master' into takeSample
7cab53a [Doris Xin] fixed import bug in rdd.py
ffea61a [Doris Xin] SPARK-1939: Refactor takeSample method in RDD
1441977 [Doris Xin] SPARK-1939 Refactor takeSample method in RDD to use ScaSRS
**Problem.** When caching, we currently unroll the entire RDD partition before making sure we have enough free memory. This is a common cause for OOMs especially when (1) the BlockManager has little free space left in memory, and (2) the partition is large.
**Solution.** We maintain a global memory pool of `M` bytes shared across all threads, similar to the way we currently manage memory for shuffle aggregation. Then, while we unroll each partition, periodically check if there is enough space to continue. If not, drop enough RDD blocks to ensure we have at least `M` bytes to work with, then try again. If we still don't have enough space to unroll the partition, give up and drop the block to disk directly if applicable.
**New configurations.**
- `spark.storage.bufferFraction` - the value of `M` as a fraction of the storage memory. (default: 0.2)
- `spark.storage.safetyFraction` - a margin of safety in case size estimation is slightly off. This is the equivalent of the existing `spark.shuffle.safetyFraction`. (default 0.9)
For more detail, see the [design document](https://issues.apache.org/jira/secure/attachment/12651793/spark-1777-design-doc.pdf). Tests pending for performance and memory usage patterns.
Author: Andrew Or <andrewor14@gmail.com>
Closes#1165 from andrewor14/them-rdd-memories and squashes the following commits:
e77f451 [Andrew Or] Merge branch 'master' of github.com:apache/spark into them-rdd-memories
c7c8832 [Andrew Or] Simplify logic + update a few comments
269d07b [Andrew Or] Very minor changes to tests
6645a8a [Andrew Or] Merge branch 'master' of github.com:apache/spark into them-rdd-memories
b7e165c [Andrew Or] Add new tests for unrolling blocks
f12916d [Andrew Or] Slightly clean up tests
71672a7 [Andrew Or] Update unrollSafely tests
369ad07 [Andrew Or] Correct ensureFreeSpace and requestMemory behavior
f4d035c [Andrew Or] Allow one thread to unroll multiple blocks
a66fbd2 [Andrew Or] Rename a few things + update comments
68730b3 [Andrew Or] Fix weird scalatest behavior
e40c60d [Andrew Or] Fix MIMA excludes
ff77aa1 [Andrew Or] Fix tests
1a43c06 [Andrew Or] Merge branch 'master' of github.com:apache/spark into them-rdd-memories
b9a6eee [Andrew Or] Simplify locking behavior on unrollMemoryMap
ed6cda4 [Andrew Or] Formatting fix (super minor)
f9ff82e [Andrew Or] putValues -> putIterator + putArray
beb368f [Andrew Or] Merge branch 'master' of github.com:apache/spark into them-rdd-memories
8448c9b [Andrew Or] Fix tests
a49ba4d [Andrew Or] Do not expose unroll memory check period
69bc0a5 [Andrew Or] Always synchronize on putLock before unrollMemoryMap
3f5a083 [Andrew Or] Simplify signature of ensureFreeSpace
dce55c8 [Andrew Or] Merge branch 'master' of github.com:apache/spark into them-rdd-memories
8288228 [Andrew Or] Synchronize put and unroll properly
4f18a3d [Andrew Or] bufferFraction -> unrollFraction
28edfa3 [Andrew Or] Update a few comments / log messages
728323b [Andrew Or] Do not synchronize every 1000 elements
5ab2329 [Andrew Or] Merge branch 'master' of github.com:apache/spark into them-rdd-memories
129c441 [Andrew Or] Fix bug: Use toArray rather than array
9a65245 [Andrew Or] Update a few comments + minor control flow changes
57f8d85 [Andrew Or] Merge branch 'master' of github.com:apache/spark into them-rdd-memories
abeae4f [Andrew Or] Add comment clarifying the MEMORY_AND_DISK case
3dd96aa [Andrew Or] AppendOnlyBuffer -> Vector (+ a few small changes)
f920531 [Andrew Or] Merge branch 'master' of github.com:apache/spark into them-rdd-memories
0871835 [Andrew Or] Add an effective storage level interface to BlockManager
64e7d4c [Andrew Or] Add/modify a few comments (minor)
8af2f35 [Andrew Or] Merge branch 'master' of github.com:apache/spark into them-rdd-memories
4f4834e [Andrew Or] Use original storage level for blocks dropped to disk
ecc8c2d [Andrew Or] Fix binary incompatibility
24185ea [Andrew Or] Avoid dropping a block back to disk if reading from disk
2b7ee66 [Andrew Or] Fix bug in SizeTracking*
9b9a273 [Andrew Or] Fix tests
20eb3e5 [Andrew Or] Merge branch 'master' of github.com:apache/spark into them-rdd-memories
649bdb3 [Andrew Or] Document spark.storage.bufferFraction
a10b0e7 [Andrew Or] Add initial memory request threshold + rename a few things
e9c3cb0 [Andrew Or] cacheMemoryMap -> unrollMemoryMap
198e374 [Andrew Or] Unfold -> unroll
0d50155 [Andrew Or] Merge branch 'master' of github.com:apache/spark into them-rdd-memories
d9d02a8 [Andrew Or] Remove unused param in unfoldSafely
ec728d8 [Andrew Or] Add tests for safe unfolding of blocks
22b2209 [Andrew Or] Merge branch 'master' of github.com:apache/spark into them-rdd-memories
078eb83 [Andrew Or] Add check for hasNext in PrimitiveVector.iterator
0871535 [Andrew Or] Fix tests in BlockManagerSuite
d68f31e [Andrew Or] Safely unfold blocks for all memory puts
5961f50 [Andrew Or] Fix tests
195abd7 [Andrew Or] Refactor: move unfold logic to MemoryStore
1e82d00 [Andrew Or] Merge branch 'master' of github.com:apache/spark into them-rdd-memories
3ce413e [Andrew Or] Merge branch 'master' of github.com:apache/spark into them-rdd-memories
d5dd3b4 [Andrew Or] Free buffer memory in finally
ea02eec [Andrew Or] Fix tests
b8e1d9c [Andrew Or] Merge branch 'master' of github.com:apache/spark into them-rdd-memories
a8704c1 [Andrew Or] Merge branch 'master' of github.com:apache/spark into them-rdd-memories
e1b8b25 [Andrew Or] Merge branch 'master' of github.com:apache/spark into them-rdd-memories
87aa75c [Andrew Or] Fix mima excludes again (typo)
11eb921 [Andrew Or] Clarify comment (minor)
50cae44 [Andrew Or] Remove now duplicate mima exclude
7de5ef9 [Andrew Or] Merge branch 'master' of github.com:apache/spark into them-rdd-memories
df47265 [Andrew Or] Fix binary incompatibility
6d05a81 [Andrew Or] Merge branch 'master' of github.com:apache/spark into them-rdd-memories
f94f5af [Andrew Or] Update a few comments (minor)
776aec9 [Andrew Or] Prevent OOM if a single RDD partition is too large
bbd3eea [Andrew Or] Fix CacheManagerSuite to use Array
97ea499 [Andrew Or] Change BlockManager interface to use Arrays
c12f093 [Andrew Or] Add SizeTrackingAppendOnlyBuffer and tests
This will decrease object allocation from the "update" closure used in map.changeValue.
Author: Matei Zaharia <matei@databricks.com>
Closes#1607 from mateiz/spark-2684 and squashes the following commits:
b7d89e6 [Matei Zaharia] Add insertAll for Iterables too, and fix some code style
561fc97 [Matei Zaharia] Update ExternalAppendOnlyMap to take an iterator as input
Added emptyRDD method to Java API with tests.
Author: bpaulin <bob@bobpaulin.com>
Closes#1597 from bobpaulin/SPARK-2279 and squashes the following commits:
5ad57c2 [bpaulin] [SPARK-2279] Added emptyRDD method to Java API
This is part of the scheduler cleanup/refactoring effort to make the scheduler code easier to maintain.
@kayousterhout @markhamstra please take a look ...
Author: Reynold Xin <rxin@apache.org>
Closes#1561 from rxin/dagSchedulerHashMaps and squashes the following commits:
1c44e15 [Reynold Xin] Clear pending tasks in submitMissingTasks.
620a0d1 [Reynold Xin] Use filterKeys.
5b54404 [Reynold Xin] Code review feedback.
c1e9a1c [Reynold Xin] Removed some HashMaps from DAGScheduler by storing information in Stage.
Due to problems with when we update runningStages (in DAGScheduler.scala)
and how we decide to send a SparkListenerStageCompleted message to
SparkListeners, sometimes stages can be shown as "running" in the UI forever
(even after they have failed). This issue can manifest when stages are
resubmitted with 0 tasks, or when the DAGScheduler catches non-serializable
tasks. The problem also resulted in a (small) memory leak in the DAGScheduler,
where stages can stay in runningStages forever. This commit fixes
that problem and adds a unit test.
Thanks tsudukim for helping to look into this issue!
cc markhamstra rxin
Author: Kay Ousterhout <kayousterhout@gmail.com>
Closes#1566 from kayousterhout/dag_fix and squashes the following commits:
217d74b [Kay Ousterhout] [SPARK-1726] [SPARK-2567] Eliminate zombie stages in UI.
Renaming `short` to `shortForm` and `long` to `longForm`.
JIRA: https://issues.apache.org/jira/browse/SPARK-2683
Author: Yin Huai <huai@cse.ohio-state.edu>
Closes#1585 from yhuai/SPARK-2683 and squashes the following commits:
5ddb843 [Yin Huai] "short" and "long" are Java keyworks. In order to generate javadoc, renaming "short" to "shortForm" and "long" to "longForm".
JIRA: https://issues.apache.org/jira/browse/SPARK-2657
Our current code uses ArrayBuffers for each group of values in groupBy, as well as for the key's elements in CoGroupedRDD. ArrayBuffers have a lot of overhead if there are few values in them, which is likely to happen in cases such as join. In particular, they have a pointer to an Object[] of size 16 by default, which is 24 bytes for the array header + 128 for the pointers in there, plus at least 32 for the ArrayBuffer data structure. This patch replaces the per-group buffers with a CompactBuffer class that can store up to 2 elements more efficiently (in fields of itself) and acts like an ArrayBuffer beyond that. For a key's elements in CoGroupedRDD, we use an Array of CompactBuffers instead of an ArrayBuffer of ArrayBuffers.
There are some changes throughout the code to deal with CoGroupedRDD returning Array instead. We can also decide not to do that but CoGroupedRDD is a `DeveloperAPI` so I think it's okay to change it here.
Author: Matei Zaharia <matei@databricks.com>
Closes#1555 from mateiz/compact-groupby and squashes the following commits:
845a356 [Matei Zaharia] Lower initial size of CompactBuffer's vector to 8
07621a7 [Matei Zaharia] Review comments
0c1cd12 [Matei Zaharia] Don't use varargs in CompactBuffer.apply
bdc8a39 [Matei Zaharia] Small tweak to +=, and typos
f61f040 [Matei Zaharia] Fix line lengths
59da88b0 [Matei Zaharia] Fix line lengths
197cde8 [Matei Zaharia] Make CompactBuffer extend Seq to make its toSeq more efficient
775110f [Matei Zaharia] Change CoGroupedRDD to give (K, Array[Iterable[_]]) to avoid wrappers
9b4c6e8 [Matei Zaharia] Use CompactBuffer in CoGroupedRDD
ed577ab [Matei Zaharia] Use CompactBuffer in groupByKey
10f0de1 [Matei Zaharia] A CompactBuffer that's more memory-efficient than ArrayBuffer for small buffers
...spark-submit
The PR allows invocations like
spark-submit --class org.MyClass --spark.shuffle.spill false myjar.jar
Author: Sandy Ryza <sandy@cloudera.com>
Closes#1253 from sryza/sandy-spark-2310 and squashes the following commits:
1dc9855 [Sandy Ryza] More doc and cleanup
00edfb9 [Sandy Ryza] Review comments
91b244a [Sandy Ryza] Change format to --conf PROP=VALUE
8fabe77 [Sandy Ryza] SPARK-2310. Support arbitrary Spark properties on the command line with spark-submit
Hi mridulm, I just think of this issue of [#1212](https://github.com/apache/spark/pull/1212): I added FakeRackUtil to hold the host -> rack mapping. It should be cleaned up after use so that it won't mess up with test cases others may add later.
Really sorry about this.
Author: Rui Li <rui.li@intel.com>
Closes#1454 from lirui-intel/SPARK-2277-fix-UT and squashes the following commits:
f8ea25c [Rui Li] SPARK-2277: clear host->rack info properly
The name `preservesPartitioning` is ambiguous: 1) preserves the indices of partitions, 2) preserves the partitioner. The latter is correct and `preservesPartitioning` should really be called `preservesPartitioner` to avoid confusion. Unfortunately, this is already part of the API and we cannot change. We should be clear in the doc and fix wrong usages.
This PR
1. adds notes in `maPartitions*`,
2. makes `RDD.sample` preserve partitioner,
3. changes `preservesPartitioning` to false in `RDD.zip` because the keys of the first RDD are no longer the keys of the zipped RDD,
4. fixes some wrong usages in MLlib.
Author: Xiangrui Meng <meng@databricks.com>
Closes#1526 from mengxr/preserve-partitioner and squashes the following commits:
b361e65 [Xiangrui Meng] update doc based on pwendell's comments
3b1ba19 [Xiangrui Meng] update doc
357575c [Xiangrui Meng] fix unit test
20b4816 [Xiangrui Meng] Merge branch 'master' into preserve-partitioner
d1caa65 [Xiangrui Meng] add doc to explain preservesPartitioning fix wrong usage of preservesPartitioning make sample preserse partitioning
### Why and what?
Currently, the AppendOnlyMap performs an "in-place" sort by converting its array of [key, value, key, value] pairs into a an array of [(key, value), (key, value)] pairs. However, this causes us to allocate many Tuple2 objects, which come at a nontrivial overhead.
This patch adds a Sorter API, intended for in memory sorts, which simply ports the Android Timsort implementation (available under Apache v2) and abstracts the interface in a way which introduces no more than 1 virtual function invocation of overhead at each abstraction point.
Please compare our port of the Android Timsort sort with the original implementation: http://www.diffchecker.com/wiwrykcl
### Memory implications
An AppendOnlyMap contains N kv pairs, which results in roughly 2N elements within its underlying array. Each of these elements is 4 bytes wide in a [compressed OOPS](https://wikis.oracle.com/display/HotSpotInternals/CompressedOops) system, which is the default.
Today's approach immediately allocates N Tuple2 objects, which take up 24N bytes in total (exposed via YourKit), and undergoes a Java sort. The Java 6 version immediately copies the entire array (4N bytes here), while the Java 7 version has a worst-case allocation of half the array (2N bytes).
This results in a worst-case sorting overhead of 24N + 2N = 26N bytes (for Java 7).
The Sorter does not require allocating any tuples, but since it uses Timsort, it may copy up to half the entire array in the worst case.
This results in a worst-case sorting overhead of 4N bytes.
Thus, we have reduced the worst-case overhead of the sort by roughly 22 bytes times the number of elements.
### Performance implications
As the destructiveSortedIterator is used for spilling in an ExternalAppendOnlyMap, the purpose of this patch is to provide stability by reducing memory usage rather than improve performance. However, because it implements Timsort, it also brings a substantial performance boost over our prior implementation.
Here are the results of a microbenchmark that sorted 25 million, randomly distributed (Float, Int) pairs. The Java Arrays.sort() tests were run **only on the keys**, and thus moved less data. Our current implementation is called "Tuple-sort using Arrays.sort()" while the new implementation is "KV-array using Sorter".
<table>
<tr><th>Test</th><th>First run (JDK6)</th><th>Average of 10 (JDK6)</th><th>First run (JDK7)</th><th>Average of 10 (JDK7)</th></tr>
<tr><td>primitive Arrays.sort()</td><td>3216 ms</td><td>1190 ms</td><td>2724 ms</td><td>131 ms (!!)</td></tr>
<tr><td>Arrays.sort()</td><td>18564 ms</td><td>2006 ms</td><td>13201 ms</td><td>878 ms</td></tr>
<tr><td>Tuple-sort using Arrays.sort()</td><td>31813 ms</td><td>3550 ms</td><td>20990 ms</td><td>1919 ms</td></tr>
<tr><td><b>KV-array using Sorter</b></td><td></td><td></td><td><b>15020 ms</b></td><td><b>834 ms</b></td></tr>
</table>
The results show that this Sorter performs exactly as expected (after the first run) -- it is as fast as the Java 7 Arrays.sort() (which shares the same algorithm), but is significantly faster than the Tuple-sort on Java 6 or 7.
In short, this patch should significantly improve performance for users running either Java 6 or 7.
Author: Aaron Davidson <aaron@databricks.com>
Closes#1502 from aarondav/sort and squashes the following commits:
652d936 [Aaron Davidson] Update license, move Sorter to java src
a7b5b1c [Aaron Davidson] fix licenses
5c0efaf [Aaron Davidson] Update tmpLength
ec395c8 [Aaron Davidson] Ignore benchmark (again) and fix docs
034bf10 [Aaron Davidson] Change to Apache v2 Timsort
b97296c [Aaron Davidson] Don't try to run benchmark on Jenkins + private[spark]
6307338 [Aaron Davidson] SPARK-2047: Introduce an in-mem Sorter, and use it to reduce mem usage
Author: Sandy Ryza <sandy@cloudera.com>
Closes#1474 from sryza/sandy-spark-2564 and squashes the following commits:
35b8388 [Sandy Ryza] Fix compile error on upmerge
7b985fb [Sandy Ryza] Fix test compile error
43f79e6 [Sandy Ryza] SPARK-2564. ShuffleReadMetrics.totalBlocksRead is redundant
We should fix this in branch-1.0 as well.
Author: Reynold Xin <rxin@apache.org>
Closes#1500 from rxin/rangePartitioner and squashes the following commits:
c0a94f5 [Reynold Xin] [SPARK-2598] RangePartitioner's binary search does not use the given Ordering.
Currently (as of Spark 1.0.1), Spark sends RDD object (which contains closures) using Akka along with the task itself to the executors. This is inefficient because all tasks in the same stage use the same RDD object, but we have to send RDD object multiple times to the executors. This is especially bad when a closure references some variable that is very large. The current design led to users having to explicitly broadcast large variables.
The patch uses broadcast to send RDD objects and the closures to executors, and use Akka to only send a reference to the broadcast RDD/closure along with the partition specific information for the task. For those of you who know more about the internals, Spark already relies on broadcast to send the Hadoop JobConf every time it uses the Hadoop input, because the JobConf is large.
The user-facing impact of the change include:
1. Users won't need to decide what to broadcast anymore, unless they would want to use a large object multiple times in different operations
2. Task size will get smaller, resulting in faster scheduling and higher task dispatch throughput.
In addition, the change will simplify some internals of Spark, eliminating the need to maintain task caches and the complex logic to broadcast JobConf (which also led to a deadlock recently).
A simple way to test this:
```scala
val a = new Array[Byte](1000*1000); scala.util.Random.nextBytes(a);
sc.parallelize(1 to 1000, 1000).map { x => a; x }.groupBy { x => a; x }.count
```
Numbers on 3 r3.8xlarge instances on EC2
```
master branch: 5.648436068 s, 4.715361895 s, 5.360161877 s
with this change: 3.416348793 s, 1.477846558 s, 1.553432156 s
```
Author: Reynold Xin <rxin@apache.org>
Closes#1452 from rxin/broadcast-task and squashes the following commits:
762e0be [Reynold Xin] Warn large broadcasts.
ade6eac [Reynold Xin] Log broadcast size.
c3b6f11 [Reynold Xin] Added a unit test for clean up.
754085f [Reynold Xin] Explain why broadcasting serialized copy of the task.
04b17f0 [Reynold Xin] [SPARK-2521] Broadcast RDD object once per TaskSet (instead of sending it for every task).
Currently, shuffle read metrics are incorrectly reported when stages have multiple shuffle dependencies (they are set to be the metrics from just one of the shuffle dependencies, rather than the accumulated metrics from all of the shuffle dependencies). This fixes that problem, and should probably be back-ported to the 0.9 branch.
Thanks ryanra for discovering this problem!
cc rxin andrewor14
Author: Kay Ousterhout <kayousterhout@gmail.com>
Closes#1476 from kayousterhout/join_bug and squashes the following commits:
0203a16 [Kay Ousterhout] Fix broken unit tests.
f463c2e [Kay Ousterhout] [SPARK-2571] Correctly report shuffle read metrics.
This is going to be used in https://issues.apache.org/jira/browse/SPARK-2568
Author: Reynold Xin <rxin@apache.org>
Closes#1478 from rxin/reservoirSample and squashes the following commits:
17bcbf3 [Reynold Xin] Added seed.
badf20d [Reynold Xin] Renamed the method.
6940010 [Reynold Xin] Reservoir sampling implementation.
This should reduce memory usage for the web ui as well as slightly increase its speed in draining the UI event queue.
@andrewor14
Author: Reynold Xin <rxin@apache.org>
Closes#1262 from rxin/ui-consolidate-hashtables and squashes the following commits:
1ac3f97 [Reynold Xin] Oops. Properly handle description.
f5736ad [Reynold Xin] Code review comments.
b8828dc [Reynold Xin] Merge branch 'master' into ui-consolidate-hashtables
7a7b6c4 [Reynold Xin] Revert css change.
f959bb8 [Reynold Xin] [SPARK-2299] Consolidate various stageIdTo* hash maps in JobProgressListener to speed it up.
63256f5 [Reynold Xin] [SPARK-2320] Reduce <pre> block font size.
If the first pass of CoalescedRDD does not find the target number of locations AND the second pass finds new locations, an exception is thrown, as "groupHash.get(nxt_replica).get" is not valid.
The fix is just to add an ArrayBuffer to groupHash for that replica if it didn't already exist.
Author: Aaron Davidson <aaron@databricks.com>
Closes#1337 from aarondav/2412 and squashes the following commits:
f587b5d [Aaron Davidson] getOrElseUpdate
3ad8a3c [Aaron Davidson] [SPARK-2412] CoalescedRDD throws exception with certain pref locs
We use TID to indicate task logging. However, TID itself does not capture stage or retries, making it harder to correlate with the application itself. This pull request changes all logging messages for tasks to include both the TID and the stage id, stage attempt, task id, and task attempt. I've consulted various people but unfortunately this is a really hard task.
Driver log looks like:
```
14/06/28 18:53:29 INFO DAGScheduler: Submitting 10 missing tasks from Stage 0 (MappedRDD[1] at map at <console>:13)
14/06/28 18:53:29 INFO TaskSchedulerImpl: Adding task set 0.0 with 10 tasks
14/06/28 18:53:29 INFO TaskSetManager: Re-computing pending task lists.
14/07/15 19:44:40 INFO TaskSetManager: Starting task 0.0 in stage 1.0 (TID 0, localhost, PROCESS_LOCAL, 1855 bytes)
14/07/15 19:44:40 INFO TaskSetManager: Starting task 1.0 in stage 1.0 (TID 1, localhost, PROCESS_LOCAL, 1855 bytes)
14/07/15 19:44:40 INFO TaskSetManager: Starting task 2.0 in stage 1.0 (TID 2, localhost, PROCESS_LOCAL, 1855 bytes)
14/07/15 19:44:40 INFO TaskSetManager: Starting task 3.0 in stage 1.0 (TID 3, localhost, PROCESS_LOCAL, 1855 bytes)
14/07/15 19:44:40 INFO TaskSetManager: Starting task 4.0 in stage 1.0 (TID 4, localhost, PROCESS_LOCAL, 1855 bytes)
14/07/15 19:44:40 INFO TaskSetManager: Starting task 5.0 in stage 1.0 (TID 5, localhost, PROCESS_LOCAL, 1855 bytes)
14/07/15 19:44:40 INFO TaskSetManager: Starting task 6.0 in stage 1.0 (TID 6, localhost, PROCESS_LOCAL, 1855 bytes)
...
14/07/15 19:44:40 INFO TaskSetManager: Finished task 1.0 in stage 1.0 (TID 1) in 64 ms on localhost (4/10)
14/07/15 19:44:40 INFO TaskSetManager: Finished task 4.0 in stage 1.0 (TID 4) in 63 ms on localhost (5/10)
14/07/15 19:44:40 INFO TaskSetManager: Finished task 2.0 in stage 1.0 (TID 2) in 63 ms on localhost (6/10)
14/07/15 19:44:40 INFO TaskSetManager: Finished task 7.0 in stage 1.0 (TID 7) in 62 ms on localhost (7/10)
14/07/15 19:44:40 INFO TaskSetManager: Finished task 6.0 in stage 1.0 (TID 6) in 63 ms on localhost (8/10)
14/07/15 19:44:40 INFO TaskSetManager: Finished task 9.0 in stage 1.0 (TID 9) in 8 ms on localhost (9/10)
14/07/15 19:44:40 INFO TaskSetManager: Finished task 8.0 in stage 1.0 (TID 8) in 9 ms on localhost (10/10)
```
Executor log looks like
```
14/07/15 19:44:40 INFO Executor: Running task 0.0 in stage 1.0 (TID 0)
14/07/15 19:44:40 INFO Executor: Running task 3.0 in stage 1.0 (TID 3)
14/07/15 19:44:40 INFO Executor: Running task 1.0 in stage 1.0 (TID 1)
14/07/15 19:44:40 INFO Executor: Running task 4.0 in stage 1.0 (TID 4)
14/07/15 19:44:40 INFO Executor: Running task 2.0 in stage 1.0 (TID 2)
14/07/15 19:44:40 INFO Executor: Running task 5.0 in stage 1.0 (TID 5)
14/07/15 19:44:40 INFO Executor: Running task 6.0 in stage 1.0 (TID 6)
14/07/15 19:44:40 INFO Executor: Running task 7.0 in stage 1.0 (TID 7)
14/07/15 19:44:40 INFO Executor: Finished task 3.0 in stage 1.0 (TID 3). 847 bytes result sent to driver
14/07/15 19:44:40 INFO Executor: Finished task 2.0 in stage 1.0 (TID 2). 847 bytes result sent to driver
14/07/15 19:44:40 INFO Executor: Finished task 0.0 in stage 1.0 (TID 0). 847 bytes result sent to driver
14/07/15 19:44:40 INFO Executor: Finished task 1.0 in stage 1.0 (TID 1). 847 bytes result sent to driver
14/07/15 19:44:40 INFO Executor: Finished task 5.0 in stage 1.0 (TID 5). 847 bytes result sent to driver
14/07/15 19:44:40 INFO Executor: Finished task 4.0 in stage 1.0 (TID 4). 847 bytes result sent to driver
14/07/15 19:44:40 INFO Executor: Finished task 6.0 in stage 1.0 (TID 6). 847 bytes result sent to driver
14/07/15 19:44:40 INFO Executor: Finished task 7.0 in stage 1.0 (TID 7). 847 bytes result sent to driver
```
Author: Reynold Xin <rxin@apache.org>
Closes#1259 from rxin/betterTaskLogging and squashes the following commits:
c28ada1 [Reynold Xin] Fix unit test failure.
987d043 [Reynold Xin] Updated log messages.
c6cfd46 [Reynold Xin] Merge branch 'master' into betterTaskLogging
b7b1bcc [Reynold Xin] Fixed a typo.
f9aba3c [Reynold Xin] Made it compile.
f8a5c06 [Reynold Xin] Merge branch 'master' into betterTaskLogging
07264e6 [Reynold Xin] Defensive check against unknown TaskEndReason.
76bbd18 [Reynold Xin] FailureSuite not serializable reporting.
4659b20 [Reynold Xin] Remove unused variable.
53888e3 [Reynold Xin] [SPARK-2317] Improve task logging.
Author: Reynold Xin <rxin@apache.org>
Closes#1433 from rxin/compile-warning and squashes the following commits:
8d0b890 [Reynold Xin] Remove some compiler warnings.
Hi mateiz, I've created [SPARK-2277](https://issues.apache.org/jira/browse/SPARK-2277) to make TaskScheduler track hosts on each rack. Please help to review, thanks.
Author: Rui Li <rui.li@intel.com>
Closes#1212 from lirui-intel/trackHostOnRack and squashes the following commits:
2b4bd0f [Rui Li] SPARK-2277: refine UT
fbde838 [Rui Li] SPARK-2277: add UT
7bbe658 [Rui Li] SPARK-2277: rename the method
5e4ef62 [Rui Li] SPARK-2277: remove unnecessary import
79ac750 [Rui Li] SPARK-2277: make TaskScheduler track hosts on rack
This reduces shuffle compression memory usage by 3x.
Author: Reynold Xin <rxin@apache.org>
Closes#1415 from rxin/snappy and squashes the following commits:
06c1a01 [Reynold Xin] SPARK-2469: Use Snappy (instead of LZF) for default shuffle compression codec.
Based on Greg Bowyer's patch from JIRA https://issues.apache.org/jira/browse/SPARK-2399
Author: Reynold Xin <rxin@apache.org>
Closes#1416 from rxin/lz4 and squashes the following commits:
6c8fefe [Reynold Xin] Fixed typo.
8a14d38 [Reynold Xin] [SPARK-2399] Add support for LZ4 compression.
Just move some test suite to corresponding package
Author: Daoyuan <daoyuan.wang@intel.com>
Closes#1401 from adrian-wang/movetestfiles and squashes the following commits:
d1a6803 [Daoyuan] move some test file to match src code
This makes it possible to read classes from the object file which were specified in the user-provided jars. (By default ObjectInputStream uses latestUserDefinedLoader, which may or may not be the right one.)
I created this because I ran into the following problem. I have x:RDD[X] with X being defined in the jar that I provide to SparkContext. I save it with x.saveAsObjectFile("x"). I try to load it with sc.objectFile\[X\]("x"). It fails with ClassNotFoundException.
After a good while of debugging I figured out that Utils.deserialize() most likely uses the ClassLoader of Utils. This is the bootstrap ClassLoader, so it is not aware of the dynamically added jars. This patch fixes the issue.
A more robust fix would be to always default to Thread.currentThread.getContextClassLoader. This would prevent this problem from biting anyone in the future. It would be a bit harder to test though. On the topic of testing, if you'd like to see tests for this, I will need some hand-holding. Thanks!
Author: Daniel Darabos <darabos.daniel@gmail.com>
Closes#181 from darabos/master and squashes the following commits:
45a011a [Daniel Darabos] Add test for SPARK-1877. (Fixed in 52eb54d.)
e13e090 [Daniel Darabos] Merge branch 'master' of https://github.com/apache/spark
61fe0d0 [Daniel Darabos] Fix style (line too long).
1b5df2c [Daniel Darabos] Use the Executor's ClassLoader in sc.objectFile(). This makes it possible to read classes from the object file which were specified in the user-provided jars. (By default ObjectInputStream uses latestUserDefinedLoader, which may or may not be the right one.)
Author: witgo <witgo@qq.com>
Closes#1153 from witgo/expectResult and squashes the following commits:
97541d8 [witgo] merge master
ead26e7 [witgo] Resolve sbt warnings during build
**Problem.** The existing code in `ExecutorPage.scala` requires a linear scan through all the blocks to filter out the uncached ones. Every refresh could be expensive if there are many blocks and many executors.
**Solution.** The proper semantics should be the following: `StorageStatusListener` should contain only block statuses that are cached. This means as soon as a block is unpersisted by any mean, its status should be removed. This is reflected in the changes made in `StorageStatusListener.scala`.
Further, the `StorageTab` must stop relying on the `StorageStatusListener` changing a dropped block's status to `StorageLevel.NONE` (which no longer happens). This is reflected in the changes made in `StorageTab.scala` and `StorageUtils.scala`.
----------
If you have been following this chain of PRs like pwendell, you will quickly notice that this reverts the changes in #1249, which reverts the changes in #1080. In other words, we are adding back the changes from #1080, and fixing SPARK-2307 on top of those changes. Please ask questions if you are confused.
Author: Andrew Or <andrewor14@gmail.com>
Closes#1255 from andrewor14/storage-ui-fix-reprise and squashes the following commits:
45416fa [Andrew Or] Merge branch 'master' of github.com:apache/spark into storage-ui-fix-reprise
a82ea25 [Andrew Or] Add tests for StorageStatusListener
8773b01 [Andrew Or] Update comment / minor changes
3afde3f [Andrew Or] Correctly report the number of blocks on SparkUI
It did not handle null keys very gracefully before.
Author: Andrew Or <andrewor14@gmail.com>
Closes#1288 from andrewor14/fix-external and squashes the following commits:
312b8d8 [Andrew Or] Abstract key hash code
ed5adf9 [Andrew Or] Fix NPE for ExternalAppendOnlyMap
This functionality was added in an earlier commit but shortly
after was removed due to a bad git merge (totally my fault).
Author: Kay Ousterhout <kayousterhout@gmail.com>
Closes#1149 from kayousterhout/warning_bug and squashes the following commits:
3f1bb00 [Kay Ousterhout] Fixed Json tests
462a664 [Kay Ousterhout] Removed task set name from warning message
e89b2f6 [Kay Ousterhout] Fixed Json tests.
7af424c [Kay Ousterhout] Emit warning when task size exceeds a threshold.
This should go into 1.0.1.
Author: Reynold Xin <rxin@apache.org>
Closes#1264 from rxin/SPARK-2322 and squashes the following commits:
c77c07f [Reynold Xin] Added comment to SparkDriverExecutionException and a test case for accumulator.
5d8d920 [Reynold Xin] [SPARK-2322] Exception in resultHandler could crash DAGScheduler and shutdown SparkContext.
These commits cause `ClosureCleaner.clean` to attempt to serialize the cleaned closure with the default closure serializer and throw a `SparkException` if doing so fails. This behavior is enabled by default but can be disabled at individual callsites of `SparkContext.clean`.
Commit 98e01ae8 fixes some no-op assertions in `GraphSuite` that this work exposed; I'm happy to put that in a separate PR if that would be more appropriate.
Author: William Benton <willb@redhat.com>
Closes#143 from willb/spark-897 and squashes the following commits:
bceab8a [William Benton] Commented DStream corner cases for serializability checking.
64d04d2 [William Benton] FailureSuite now checks both messages and causes.
3b3f74a [William Benton] Stylistic and doc cleanups.
b215dea [William Benton] Fixed spurious failures in ImplicitOrderingSuite
be1ecd6 [William Benton] Don't check serializability of DStream transforms.
abe816b [William Benton] Make proactive serializability checking optional.
5bfff24 [William Benton] Adds proactive closure-serializablilty checking
ed2ccf0 [William Benton] Test cases for SPARK-897.
Details can be see in [SPARK-2104](https://issues.apache.org/jira/browse/SPARK-2104). This work is based on Reynold's work, add some unit tests to validate the issue.
@rxin , would you please take a look at this PR, thanks a lot.
Author: jerryshao <saisai.shao@intel.com>
Closes#1245 from jerryshao/SPARK-2104 and squashes the following commits:
c8ee362 [jerryshao] Make field partitions transient
2b41917 [jerryshao] Minor changes
47d763c [jerryshao] Fix task serializing issue when sort with Java non serializable class
This commit adds a new metric in TaskMetrics to record
the input data size and displays this information in the UI.
An earlier version of this commit also added the read time,
which can be useful for diagnosing straggler problems,
but unfortunately that change introduced a significant performance
regression for jobs that don't do much computation. In order to
track read time, we'll need to do sampling.
The screenshots below show the UI with the new "Input" field,
which I added to the stage summary page, the executor summary page,
and the per-stage page.
Author: Kay Ousterhout <kayousterhout@gmail.com>
Closes#962 from kayousterhout/read_metrics and squashes the following commits:
f13b67d [Kay Ousterhout] Correctly format input bytes on executor page
8b70cde [Kay Ousterhout] Added comment about potential inaccuracy of bytesRead
d1016e8 [Kay Ousterhout] Udated SparkListenerSuite test
8461492 [Kay Ousterhout] Miniscule style fix
ae04d99 [Kay Ousterhout] Remove input metrics for parallel collections
719f19d [Kay Ousterhout] Style fixes
bb6ec62 [Kay Ousterhout] Small fixes
869ac7b [Kay Ousterhout] Updated Json tests
44a0301 [Kay Ousterhout] Fixed accidentally added line
4bd0568 [Kay Ousterhout] Added input source, renamed Hdfs to Hadoop.
f27e535 [Kay Ousterhout] Updates based on review comments and to fix rebase
bf41029 [Kay Ousterhout] Updated Json tests to pass
0fc33e0 [Kay Ousterhout] Added explicit backward compatibility test
4e52925 [Kay Ousterhout] Added Json output and associated tests.
365400b [Kay Ousterhout] [SPARK-1683] Track task read metrics.
The following code is very likely to throw an exception:
~~~
val rdd = sc.parallelize(0 until 111, 10).sample(false, 0.1)
rdd.zip(rdd).count()
~~~
because the same random number generator is used in compute partitions.
Author: Xiangrui Meng <meng@databricks.com>
Closes#1229 from mengxr/fix-sample and squashes the following commits:
f1ee3d7 [Xiangrui Meng] fix concurrency issues in random sampler
New UI:
Author: Reynold Xin <rxin@apache.org>
Closes#1236 from rxin/ui-task-attempt and squashes the following commits:
3b645dd [Reynold Xin] Expose attemptId in Stage.
c0474b1 [Reynold Xin] Beefed up unit test.
c404bdd [Reynold Xin] Fix ReplayListenerSuite.
f56be4b [Reynold Xin] Fixed JsonProtocolSuite.
e29e0f7 [Reynold Xin] Minor update.
5e4354a [Reynold Xin] [SPARK-2297][UI] Make task attempt and speculation more explicit in UI.
FetchFailedException used to have a Throwable field, but in reality we never propagate any of the throwable/exceptions back to the driver because Executor explicitly looks for FetchFailedException and then sends FetchFailed as the TaskEndReason.
This pull request removes the throwable and adds a MetadataFetchFailedException that extends FetchFailedException (so now MapOutputTracker throws MetadataFetchFailedException instead).
Author: Reynold Xin <rxin@apache.org>
Closes#1227 from rxin/metadataFetchException and squashes the following commits:
5cb1e0a [Reynold Xin] MetadataFetchFailedException extends FetchFailedException.
8861ee2 [Reynold Xin] Throw MetadataFetchFailedException in MapOutputTracker.
Previously only tasks failed with ExceptionFailure reason was marked as failure.
Author: Reynold Xin <rxin@apache.org>
Closes#1224 from rxin/SPARK-2284 and squashes the following commits:
be79dbd [Reynold Xin] [SPARK-2284][UI] Mark all failed tasks as failures.
This is a fixed up version of #686 (cc @markhamstra @pwendell). The last commit (the only one I authored) reflects the changes I made from Mark's original patch.
Author: Mark Hamstra <markhamstra@gmail.com>
Author: Kay Ousterhout <kayousterhout@gmail.com>
Closes#1219 from kayousterhout/mark-SPARK-1749 and squashes the following commits:
42dfa7e [Kay Ousterhout] Got rid of terrible double-negative name
80b3205 [Kay Ousterhout] Don't notify listeners of job failure if it wasn't successfully cancelled.
d156d33 [Mark Hamstra] Do nothing in no-kill submitTasks
9312baa [Mark Hamstra] code review update
cc353c8 [Mark Hamstra] scalastyle
e61f7f8 [Mark Hamstra] Catch UnsupportedOperationException when DAGScheduler tries to cancel a job on a SchedulerBackend that does not implement killTask
