Adding ability to define an initial state RDD for use with updateStateByKey PySpark. Added unit test and changed stateful_network_wordcount example to use initial RDD.
Author: Bryan Cutler <bjcutler@us.ibm.com>
Closes#10082 from BryanCutler/initial-rdd-updateStateByKey-SPARK-11713.
The original code does not properly handle the cases where the prefix is null, but suffix is not null - the suffix should be used but is not.
The fix is using StringBuilder to construct the proper file name.
Author: bomeng <bmeng@us.ibm.com>
Author: Bo Meng <mengbo@bos-macbook-pro.usca.ibm.com>
Closes#10185 from bomeng/SPARK-12136.
SPARK-12244:
Based on feedback from early users and personal experience attempting to explain it, the name trackStateByKey had two problem.
"trackState" is a completely new term which really does not give any intuition on what the operation is
the resultant data stream of objects returned by the function is called in docs as the "emitted" data for the lack of a better.
"mapWithState" makes sense because the API is like a mapping function like (Key, Value) => T with State as an additional parameter. The resultant data stream is "mapped data". So both problems are solved.
SPARK-12245:
From initial experiences, not having the key in the function makes it hard to return mapped stuff, as the whole information of the records is not there. Basically the user is restricted to doing something like mapValue() instead of map(). So adding the key as a parameter.
Author: Tathagata Das <tathagata.das1565@gmail.com>
Closes#10224 from tdas/rename.
The reason is that TrackStateRDDs generated by trackStateByKey expect the previous batch's TrackStateRDDs to have a partitioner. However, when recovery from DStream checkpoints, the RDDs recovered from RDD checkpoints do not have a partitioner attached to it. This is because RDD checkpoints do not preserve the partitioner (SPARK-12004).
While #9983 solves SPARK-12004 by preserving the partitioner through RDD checkpoints, there may be a non-zero chance that the saving and recovery fails. To be resilient, this PR repartitions the previous state RDD if the partitioner is not detected.
Author: Tathagata Das <tathagata.das1565@gmail.com>
Closes#9988 from tdas/SPARK-11932.
We need to make sure that the last entry is indeed the last entry in the queue.
Author: Burak Yavuz <brkyvz@gmail.com>
Closes#10110 from brkyvz/batch-wal-test-fix.
`ByteBuffer` doesn't guarantee all contents in `ByteBuffer.array` are valid. E.g, a ByteBuffer returned by `ByteBuffer.slice`. We should not use the whole content of `ByteBuffer` unless we know that's correct.
This patch fixed all places that use `ByteBuffer.array` incorrectly.
Author: Shixiong Zhu <shixiong@databricks.com>
Closes#10083 from zsxwing/bytebuffer-array.
This replaces https://github.com/apache/spark/pull/9696
Invoke Checkstyle and print any errors to the console, failing the step.
Use Google's style rules modified according to
https://cwiki.apache.org/confluence/display/SPARK/Spark+Code+Style+Guide
Some important checks are disabled (see TODOs in `checkstyle.xml`) due to
multiple violations being present in the codebase.
Suggest fixing those TODOs in a separate PR(s).
More on Checkstyle can be found on the [official website](http://checkstyle.sourceforge.net/).
Sample output (from [build 46345](https://amplab.cs.berkeley.edu/jenkins/job/SparkPullRequestBuilder/46345/consoleFull)) (duplicated because I run the build twice with different profiles):
> Checkstyle checks failed at following occurrences:
[ERROR] src/main/java/org/apache/spark/sql/execution/datasources/parquet/UnsafeRowParquetRecordReader.java:[217,7] (coding) MissingSwitchDefault: switch without "default" clause.
> [ERROR] src/main/java/org/apache/spark/sql/execution/datasources/parquet/SpecificParquetRecordReaderBase.java:[198,10] (modifier) ModifierOrder: 'protected' modifier out of order with the JLS suggestions.
> [ERROR] src/main/java/org/apache/spark/sql/execution/datasources/parquet/UnsafeRowParquetRecordReader.java:[217,7] (coding) MissingSwitchDefault: switch without "default" clause.
> [ERROR] src/main/java/org/apache/spark/sql/execution/datasources/parquet/SpecificParquetRecordReaderBase.java:[198,10] (modifier) ModifierOrder: 'protected' modifier out of order with the JLS suggestions.
> [error] running /home/jenkins/workspace/SparkPullRequestBuilder2/dev/lint-java ; received return code 1
Also fix some of the minor violations that didn't require sweeping changes.
Apologies for the previous botched PRs - I finally figured out the issue.
cr: JoshRosen, pwendell
> I state that the contribution is my original work, and I license the work to the project under the project's open source license.
Author: Dmitry Erastov <derastov@gmail.com>
Closes#9867 from dskrvk/master.
If `StreamingContext.stop()` is interrupted midway through the call, the context will be marked as stopped but certain state will have not been cleaned up. Because `state = STOPPED` will be set, subsequent `stop()` calls will be unable to finish stopping the context, preventing any new StreamingContexts from being created.
This patch addresses this issue by only marking the context as `STOPPED` once the `stop()` has successfully completed which allows `stop()` to be called a second time in order to finish stopping the context in case the original `stop()` call was interrupted.
I discovered this issue by examining logs from a failed Jenkins run in which this race condition occurred in `FailureSuite`, leaking an unstoppable context and causing all subsequent tests to fail.
Author: Josh Rosen <joshrosen@databricks.com>
Closes#9982 from JoshRosen/SPARK-12001.
The JobConf object created in `DStream.saveAsHadoopFiles` is used concurrently in multiple places:
* The JobConf is updated by `RDD.saveAsHadoopFile()` before the job is launched
* The JobConf is serialized as part of the DStream checkpoints.
These concurrent accesses (updating in one thread, while the another thread is serializing it) can lead to concurrentModidicationException in the underlying Java hashmap using in the internal Hadoop Configuration object.
The solution is to create a new JobConf in every batch, that is updated by `RDD.saveAsHadoopFile()`, while the checkpointing serializes the original JobConf.
Tests to be added in #9988 will fail reliably without this patch. Keeping this patch really small to make sure that it can be added to previous branches.
Author: Tathagata Das <tathagata.das1565@gmail.com>
Closes#10088 from tdas/SPARK-12087.
In StreamingListenerSuite."don't call ssc.stop in listener", after the main thread calls `ssc.stop()`, `StreamingContextStoppingCollector` may call `ssc.stop()` in the listener bus thread, which is a dead-lock. This PR updated `StreamingContextStoppingCollector` to only call `ssc.stop()` in the first batch to avoid the dead-lock.
Author: Shixiong Zhu <shixiong@databricks.com>
Closes#10011 from zsxwing/fix-test-deadlock.
The Python exception track in TransformFunction and TransformFunctionSerializer is not sent back to Java. Py4j just throws a very general exception, which is hard to debug.
This PRs adds `getFailure` method to get the failure message in Java side.
Author: Shixiong Zhu <shixiong@databricks.com>
Closes#9922 from zsxwing/SPARK-11935.
This solves the following exception caused when empty state RDD is checkpointed and recovered. The root cause is that an empty OpenHashMapBasedStateMap cannot be deserialized as the initialCapacity is set to zero.
```
Job aborted due to stage failure: Task 0 in stage 6.0 failed 1 times, most recent failure: Lost task 0.0 in stage 6.0 (TID 20, localhost): java.lang.IllegalArgumentException: requirement failed: Invalid initial capacity
at scala.Predef$.require(Predef.scala:233)
at org.apache.spark.streaming.util.OpenHashMapBasedStateMap.<init>(StateMap.scala:96)
at org.apache.spark.streaming.util.OpenHashMapBasedStateMap.<init>(StateMap.scala:86)
at org.apache.spark.streaming.util.OpenHashMapBasedStateMap.readObject(StateMap.scala:291)
at sun.reflect.NativeMethodAccessorImpl.invoke0(Native Method)
at sun.reflect.NativeMethodAccessorImpl.invoke(NativeMethodAccessorImpl.java:57)
at sun.reflect.DelegatingMethodAccessorImpl.invoke(DelegatingMethodAccessorImpl.java:43)
at java.lang.reflect.Method.invoke(Method.java:606)
at java.io.ObjectStreamClass.invokeReadObject(ObjectStreamClass.java:1017)
at java.io.ObjectInputStream.readSerialData(ObjectInputStream.java:1893)
at java.io.ObjectInputStream.readOrdinaryObject(ObjectInputStream.java:1798)
at java.io.ObjectInputStream.readObject0(ObjectInputStream.java:1350)
at java.io.ObjectInputStream.defaultReadFields(ObjectInputStream.java:1990)
at java.io.ObjectInputStream.readSerialData(ObjectInputStream.java:1915)
at java.io.ObjectInputStream.readOrdinaryObject(ObjectInputStream.java:1798)
at java.io.ObjectInputStream.readObject0(ObjectInputStream.java:1350)
at java.io.ObjectInputStream.readObject(ObjectInputStream.java:370)
at org.apache.spark.serializer.JavaDeserializationStream.readObject(JavaSerializer.scala:76)
at org.apache.spark.serializer.DeserializationStream$$anon$1.getNext(Serializer.scala:181)
at org.apache.spark.util.NextIterator.hasNext(NextIterator.scala:73)
at scala.collection.Iterator$$anon$13.hasNext(Iterator.scala:371)
at scala.collection.Iterator$class.foreach(Iterator.scala:727)
at scala.collection.AbstractIterator.foreach(Iterator.scala:1157)
at scala.collection.generic.Growable$class.$plus$plus$eq(Growable.scala:48)
at scala.collection.mutable.ArrayBuffer.$plus$plus$eq(ArrayBuffer.scala:103)
at scala.collection.mutable.ArrayBuffer.$plus$plus$eq(ArrayBuffer.scala:47)
at scala.collection.TraversableOnce$class.to(TraversableOnce.scala:273)
at scala.collection.AbstractIterator.to(Iterator.scala:1157)
at scala.collection.TraversableOnce$class.toBuffer(TraversableOnce.scala:265)
at scala.collection.AbstractIterator.toBuffer(Iterator.scala:1157)
at scala.collection.TraversableOnce$class.toArray(TraversableOnce.scala:252)
at scala.collection.AbstractIterator.toArray(Iterator.scala:1157)
at org.apache.spark.rdd.RDD$$anonfun$collect$1$$anonfun$12.apply(RDD.scala:921)
at org.apache.spark.rdd.RDD$$anonfun$collect$1$$anonfun$12.apply(RDD.scala:921)
at org.apache.spark.SparkContext$$anonfun$runJob$5.apply(SparkContext.scala:1858)
at org.apache.spark.SparkContext$$anonfun$runJob$5.apply(SparkContext.scala:1858)
at org.apache.spark.scheduler.ResultTask.runTask(ResultTask.scala:66)
at org.apache.spark.scheduler.Task.run(Task.scala:88)
at org.apache.spark.executor.Executor$TaskRunner.run(Executor.scala:214)
at java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1145)
at java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:615)
at java.lang.Thread.run(Thread.java:744)
```
Author: Tathagata Das <tathagata.das1565@gmail.com>
Closes#9958 from tdas/SPARK-11979.
To make sure that all lineage is correctly truncated for TrackStateRDD when checkpointed.
Author: Tathagata Das <tathagata.das1565@gmail.com>
Closes#9831 from tdas/SPARK-11845.
stack trace of failure:
```
org.scalatest.exceptions.TestFailedDueToTimeoutException: The code passed to eventually never returned normally. Attempted 62 times over 1.006322071 seconds. Last failure message:
Argument(s) are different! Wanted:
writeAheadLog.write(
java.nio.HeapByteBuffer[pos=0 lim=124 cap=124],
10
);
-> at org.apache.spark.streaming.util.BatchedWriteAheadLogSuite$$anonfun$23$$anonfun$apply$mcV$sp$15.apply(WriteAheadLogSuite.scala:518)
Actual invocation has different arguments:
writeAheadLog.write(
java.nio.HeapByteBuffer[pos=0 lim=124 cap=124],
10
);
-> at org.apache.spark.streaming.util.WriteAheadLogSuite$BlockingWriteAheadLog.write(WriteAheadLogSuite.scala:756)
```
I believe the issue was that due to a race condition, the ordering of the events could be messed up in the final ByteBuffer, therefore the comparison fails.
By adding eventually between the requests, we make sure the ordering is preserved. Note that in real life situations, the ordering across threads will not matter.
Another solution would be to implement a custom mockito matcher that sorts and then compares the results, but that kind of sounds like overkill to me. Let me know what you think tdas zsxwing
Author: Burak Yavuz <brkyvz@gmail.com>
Closes#9790 from brkyvz/fix-flaky-2.
DStream checkpoint interval is by default set at max(10 second, batch interval). That's bad for large batch intervals where the checkpoint interval = batch interval, and RDDs get checkpointed every batch.
This PR is to set the checkpoint interval of trackStateByKey to 10 * batch duration.
Author: Tathagata Das <tathagata.das1565@gmail.com>
Closes#9805 from tdas/SPARK-11814.
The HP Fortify Opens Source Review team (https://www.hpfod.com/open-source-review-project) reported a handful of potential resource leaks that were discovered using their static analysis tool. We should fix the issues identified by their scan.
Author: Josh Rosen <joshrosen@databricks.com>
Closes#9455 from JoshRosen/fix-potential-resource-leaks.
Currently streaming foreachRDD Java API uses a function prototype requiring a return value of null. This PR deprecates the old method and uses VoidFunction to allow for more concise declaration. Also added VoidFunction2 to Java API in order to use in Streaming methods. Unit test is added for using foreachRDD with VoidFunction, and changes have been tested with Java 7 and Java 8 using lambdas.
Author: Bryan Cutler <bjcutler@us.ibm.com>
Closes#9488 from BryanCutler/foreachRDD-VoidFunction-SPARK-4557.
See discussion toward the tail of https://github.com/apache/spark/pull/9723
From zsxwing :
```
The user should not call stop or other long-time work in a listener since it will block the listener thread, and prevent from stopping SparkContext/StreamingContext.
I cannot see an approach since we need to stop the listener bus's thread before stopping SparkContext/StreamingContext totally.
```
Proposed solution is to prevent the call to StreamingContext#stop() in the listener bus's thread.
Author: tedyu <yuzhihong@gmail.com>
Closes#9741 from tedyu/master.
We will do checkpoint when generating a batch and completing a batch. When the processing time of a batch is greater than the batch interval, checkpointing for completing an old batch may run after checkpointing for generating a new batch. If this happens, checkpoint of an old batch actually has the latest information, so we want to recovery from it. This PR will use the latest checkpoint time as the file name, so that we can always recovery from the latest checkpoint file.
Author: Shixiong Zhu <shixiong@databricks.com>
Closes#9707 from zsxwing/fix-checkpoint.
Using batching on the driver for the WriteAheadLog should be an improvement for all environments and use cases. Users will be able to scale to much higher number of receivers with the BatchedWriteAheadLog. Therefore we should turn it on by default, and QA it in the QA period.
I've also added some tests to make sure the default configurations are correct regarding recent additions:
- batching on by default
- closeFileAfterWrite off by default
- parallelRecovery off by default
Author: Burak Yavuz <brkyvz@gmail.com>
Closes#9695 from brkyvz/enable-batch-wal.
Bug: Timestamp is not updated if there is data but the corresponding state is not updated. This is wrong, and timeout is defined as "no data for a while", not "not state update for a while".
Fix: Update timestamp when timestamp when timeout is specified, otherwise no need.
Also refactored the code for better testability and added unit tests.
Author: Tathagata Das <tathagata.das1565@gmail.com>
Closes#9648 from tdas/SPARK-11681.
The support for closing WriteAheadLog files after writes was just merged in. Closing every file after a write is a very expensive operation as it creates many small files on S3. It's not necessary to enable it on HDFS anyway.
However, when you have many small files on S3, recovery takes very long. In addition, files start stacking up pretty quickly, and deletes may not be able to keep up, therefore deletes can also be parallelized.
This PR adds support for the two parallelization steps mentioned above, in addition to a couple more failures I encountered during recovery.
Author: Burak Yavuz <brkyvz@gmail.com>
Closes#9373 from brkyvz/par-recovery.
TODO
- [x] Add Java API
- [x] Add API tests
- [x] Add a function test
Author: Shixiong Zhu <shixiong@databricks.com>
Closes#9636 from zsxwing/java-track.
Should not create SparkContext in the constructor of `TrackStateRDDSuite`. This is a follow up PR for #9256 to fix the test for maven build.
Author: Shixiong Zhu <shixiong@databricks.com>
Closes#9668 from zsxwing/hotfix.
Several elements could be drained if the main thread is not fast enough. zsxwing warned me about a similar problem, but missed it here :( Submitting the fix using a waiter.
cc tdas
Author: Burak Yavuz <brkyvz@gmail.com>
Closes#9605 from brkyvz/fix-flaky-test.
Current updateStateByKey provides stateful processing in Spark Streaming. It allows the user to maintain per-key state and manage that state using an updateFunction. The updateFunction is called for each key, and it uses new data and existing state of the key, to generate an updated state. However, based on community feedback, we have learnt the following lessons.
* Need for more optimized state management that does not scan every key
* Need to make it easier to implement common use cases - (a) timeout of idle data, (b) returning items other than state
The high level idea that of this PR
* Introduce a new API trackStateByKey that, allows the user to update per-key state, and emit arbitrary records. The new API is necessary as this will have significantly different semantics than the existing updateStateByKey API. This API will have direct support for timeouts.
* Internally, the system will keep the state data as a map/list within the partitions of the state RDDs. The new data RDDs will be partitioned appropriately, and for all the key-value data, it will lookup the map/list in the state RDD partition and create a new list/map of updated state data. The new state RDD partition will be created based on the update data and if necessary, with old data.
Here is the detailed design doc. Please take a look and provide feedback as comments.
https://docs.google.com/document/d/1NoALLyd83zGs1hNGMm0Pc5YOVgiPpMHugGMk6COqxxE/edit#heading=h.ph3w0clkd4em
This is still WIP. Major things left to be done.
- [x] Implement basic functionality of state tracking, with initial RDD and timeouts
- [x] Unit tests for state tracking
- [x] Unit tests for initial RDD and timeout
- [ ] Unit tests for TrackStateRDD
- [x] state creating, updating, removing
- [ ] emitting
- [ ] checkpointing
- [x] Misc unit tests for State, TrackStateSpec, etc.
- [x] Update docs and experimental tags
Author: Tathagata Das <tathagata.das1565@gmail.com>
Closes#9256 from tdas/trackStateByKey.
Currently, when a DStream sets the scope for RDD generated by it, that scope is not allowed to be overridden by the RDD operations. So in case of `DStream.foreachRDD`, all the RDDs generated inside the foreachRDD get the same scope - `foreachRDD <time>`, as set by the `ForeachDStream`. So it is hard to debug generated RDDs in the RDD DAG viz in the Spark UI.
This patch allows the RDD operations inside `DStream.transform` and `DStream.foreachRDD` to append their own scopes to the earlier DStream scope.
I have also slightly tweaked how callsites are set such that the short callsite reflects the RDD operation name and line number. This tweak is necessary as callsites are not managed through scopes (which support nesting and overriding) and I didnt want to add another local property to control nesting and overriding of callsites.
## Before:
## After:
The code that was used to generate this is:
```
val lines = ssc.socketTextStream(args(0), args(1).toInt, StorageLevel.MEMORY_AND_DISK_SER)
val words = lines.flatMap(_.split(" "))
val wordCounts = words.map(x => (x, 1)).reduceByKey(_ + _)
wordCounts.foreachRDD { rdd =>
val temp = rdd.map { _ -> 1 }.reduceByKey( _ + _)
val temp2 = temp.map { _ -> 1}.reduceByKey(_ + _)
val count = temp2.count
println(count)
}
```
Note
- The inner scopes of the RDD operations map/reduceByKey inside foreachRDD is visible
- The short callsites of stages refers to the line number of the RDD ops rather than the same line number of foreachRDD in all three cases.
Author: Tathagata Das <tathagata.das1565@gmail.com>
Closes#9315 from tdas/SPARK-11361.
While sbt successfully compiles as it properly pulls the mockito dependency, maven builds have broken. We need this in ASAP.
tdas
Author: Burak Yavuz <brkyvz@gmail.com>
Closes#9584 from brkyvz/fix-master.
Expose executorId to `ReceiverInfo` and UI since it's helpful when there are multiple executors running in the same host. Screenshot:
<img width="1058" alt="screen shot 2015-11-02 at 10 52 19 am" src="https://cloud.githubusercontent.com/assets/1000778/10890968/2e2f5512-8150-11e5-8d9d-746e826b69e8.png">
Author: Shixiong Zhu <shixiong@databricks.com>
Author: zsxwing <zsxwing@gmail.com>
Closes#9418 from zsxwing/SPARK-11333.
Currently, StreamingListener is not Java friendly because it exposes some Scala collections to Java users directly, such as Option, Map.
This PR added a Java version of StreamingListener and a bunch of Java friendly classes for Java users.
Author: zsxwing <zsxwing@gmail.com>
Author: Shixiong Zhu <shixiong@databricks.com>
Closes#9420 from zsxwing/java-streaming-listener.
When using S3 as a directory for WALs, the writes take too long. The driver gets very easily bottlenecked when multiple receivers send AddBlock events to the ReceiverTracker. This PR adds batching of events in the ReceivedBlockTracker so that receivers don't get blocked by the driver for too long.
cc zsxwing tdas
Author: Burak Yavuz <brkyvz@gmail.com>
Closes#9143 from brkyvz/batch-wal-writes.
Just ignored `InputDStream`s that have null `rememberDuration` in `DStreamGraph.getMaxInputStreamRememberDuration`.
Author: Shixiong Zhu <shixiong@databricks.com>
Closes#9476 from zsxwing/SPARK-11511.
Currently Yarn AM proxy filter configuration is recovered from checkpoint file when Spark Streaming application is restarted, which will lead to some unwanted behaviors:
1. Wrong RM address if RM is redeployed from failure.
2. Wrong proxyBase, since app id is updated, old app id for proxyBase is wrong.
So instead of recovering from checkpoint file, these configurations should be reloaded each time when app started.
This problem only exists in Yarn cluster mode, for Yarn client mode, these configurations will be updated with RPC message `AddWebUIFilter`.
Please help to review tdas harishreedharan vanzin , thanks a lot.
Author: jerryshao <sshao@hortonworks.com>
Closes#9412 from jerryshao/SPARK-11457.
… ReceiverTracker and ReceiverSchedulingPolicy to use it
This PR includes the following changes:
1. Add a new preferred location format, `executor_<host>_<executorID>` (e.g., "executor_localhost_2"), to support specifying the executor locations for RDD.
2. Use the new preferred location format in `ReceiverTracker` to optimize the starting time of Receivers when there are multiple executors in a host.
The goal of this PR is to enable the streaming scheduler to place receivers (which run as tasks) in specific executors. Basically, I want to have more control on the placement of the receivers such that they are evenly distributed among the executors. We tried to do this without changing the core scheduling logic. But it does not allow specifying particular executor as preferred location, only at the host level. So if there are two executors in the same host, and I want two receivers to run on them (one on each executor), I cannot specify that. Current code only specifies the host as preference, which may end up launching both receivers on the same executor. We try to work around it but restarting a receiver when it does not launch in the desired executor and hope that next time it will be started in the right one. But that cause lots of restarts, and delays in correctly launching the receiver.
So this change, would allow the streaming scheduler to specify the exact executor as the preferred location. Also this is not exposed to the user, only the streaming scheduler uses this.
Author: zsxwing <zsxwing@gmail.com>
Closes#9181 from zsxwing/executor-location.
Currently the Write Ahead Log in Spark Streaming flushes data as writes need to be made. S3 does not support flushing of data, data is written once the stream is actually closed.
In case of failure, the data for the last minute (default rolling interval) will not be properly written. Therefore we need a flag to close the stream after the write, so that we achieve read after write consistency.
cc tdas zsxwing
Author: Burak Yavuz <brkyvz@gmail.com>
Closes#9285 from brkyvz/caw-wal.
When use Kafka DirectStream API to create checkpoint and restore saved checkpoint when restart,
ClassNotFound exception would occur.
The reason for this error is that ObjectInputStreamWithLoader extends the ObjectInputStream class and override its resolveClass method. But Instead of Using Class.forName(desc,false,loader), Spark uses loader.loadClass(desc) to instance the class, which do not works with array class.
For example:
Class.forName("[Lorg.apache.spark.streaming.kafka.OffsetRange.",false,loader) works well while loader.loadClass("[Lorg.apache.spark.streaming.kafka.OffsetRange") would throw an class not found exception.
details of the difference between Class.forName and loader.loadClass can be found here.
http://bugs.java.com/view_bug.do?bug_id=6446627
Author: maxwell <maxwellzdm@gmail.com>
Author: DEMING ZHU <deming.zhu@linecorp.com>
Closes#8955 from maxwellzdm/master.
This patch refactors the MemoryManager class structure. After #9000, Spark had the following classes:
- MemoryManager
- StaticMemoryManager
- ExecutorMemoryManager
- TaskMemoryManager
- ShuffleMemoryManager
This is fairly confusing. To simplify things, this patch consolidates several of these classes:
- ShuffleMemoryManager and ExecutorMemoryManager were merged into MemoryManager.
- TaskMemoryManager is moved into Spark Core.
**Key changes and tasks**:
- [x] Merge ExecutorMemoryManager into MemoryManager.
- [x] Move pooling logic into Allocator.
- [x] Move TaskMemoryManager from `spark-unsafe` to `spark-core`.
- [x] Refactor the existing Tungsten TaskMemoryManager interactions so Tungsten code use only this and not both this and ShuffleMemoryManager.
- [x] Refactor non-Tungsten code to use the TaskMemoryManager instead of ShuffleMemoryManager.
- [x] Merge ShuffleMemoryManager into MemoryManager.
- [x] Move code
- [x] ~~Simplify 1/n calculation.~~ **Will defer to followup, since this needs more work.**
- [x] Port ShuffleMemoryManagerSuite tests.
- [x] Move classes from `unsafe` package to `memory` package.
- [ ] Figure out how to handle the hacky use of the memory managers in HashedRelation's broadcast variable construction.
- [x] Test porting and cleanup: several tests relied on mock functionality (such as `TestShuffleMemoryManager.markAsOutOfMemory`) which has been changed or broken during the memory manager consolidation
- [x] AbstractBytesToBytesMapSuite
- [x] UnsafeExternalSorterSuite
- [x] UnsafeFixedWidthAggregationMapSuite
- [x] UnsafeKVExternalSorterSuite
**Compatiblity notes**:
- This patch introduces breaking changes in `ExternalAppendOnlyMap`, which is marked as `DevloperAPI` (likely for legacy reasons): this class now cannot be used outside of a task.
Author: Josh Rosen <joshrosen@databricks.com>
Closes#9127 from JoshRosen/SPARK-10984.
The format of RDD's preferredLocations must be hostname but the format of Streaming Receiver's scheduling executors is hostport. So it doesn't work.
This PR converts `schedulerExecutors` to `hosts` before creating Receiver's RDD.
Author: zsxwing <zsxwing@gmail.com>
Closes#9075 from zsxwing/SPARK-11063.
The following deadlock may happen if shutdownHook and StreamingContext.stop are running at the same time.
```
Java stack information for the threads listed above:
===================================================
"Thread-2":
at org.apache.spark.streaming.StreamingContext.stop(StreamingContext.scala:699)
- waiting to lock <0x00000005405a1680> (a org.apache.spark.streaming.StreamingContext)
at org.apache.spark.streaming.StreamingContext.org$apache$spark$streaming$StreamingContext$$stopOnShutdown(StreamingContext.scala:729)
at org.apache.spark.streaming.StreamingContext$$anonfun$start$1.apply$mcV$sp(StreamingContext.scala:625)
at org.apache.spark.util.SparkShutdownHook.run(ShutdownHookManager.scala:266)
at org.apache.spark.util.SparkShutdownHookManager$$anonfun$runAll$1$$anonfun$apply$mcV$sp$1.apply$mcV$sp(ShutdownHookManager.scala:236)
at org.apache.spark.util.SparkShutdownHookManager$$anonfun$runAll$1$$anonfun$apply$mcV$sp$1.apply(ShutdownHookManager.scala:236)
at org.apache.spark.util.SparkShutdownHookManager$$anonfun$runAll$1$$anonfun$apply$mcV$sp$1.apply(ShutdownHookManager.scala:236)
at org.apache.spark.util.Utils$.logUncaughtExceptions(Utils.scala:1697)
at org.apache.spark.util.SparkShutdownHookManager$$anonfun$runAll$1.apply$mcV$sp(ShutdownHookManager.scala:236)
at org.apache.spark.util.SparkShutdownHookManager$$anonfun$runAll$1.apply(ShutdownHookManager.scala:236)
at org.apache.spark.util.SparkShutdownHookManager$$anonfun$runAll$1.apply(ShutdownHookManager.scala:236)
at scala.util.Try$.apply(Try.scala:161)
at org.apache.spark.util.SparkShutdownHookManager.runAll(ShutdownHookManager.scala:236)
- locked <0x00000005405b6a00> (a org.apache.spark.util.SparkShutdownHookManager)
at org.apache.spark.util.SparkShutdownHookManager$$anon$2.run(ShutdownHookManager.scala:216)
at org.apache.hadoop.util.ShutdownHookManager$1.run(ShutdownHookManager.java:54)
"main":
at org.apache.spark.util.SparkShutdownHookManager.remove(ShutdownHookManager.scala:248)
- waiting to lock <0x00000005405b6a00> (a org.apache.spark.util.SparkShutdownHookManager)
at org.apache.spark.util.ShutdownHookManager$.removeShutdownHook(ShutdownHookManager.scala:199)
at org.apache.spark.streaming.StreamingContext.stop(StreamingContext.scala:712)
- locked <0x00000005405a1680> (a org.apache.spark.streaming.StreamingContext)
at org.apache.spark.streaming.StreamingContext.stop(StreamingContext.scala:684)
- locked <0x00000005405a1680> (a org.apache.spark.streaming.StreamingContext)
at org.apache.spark.streaming.SessionByKeyBenchmark$.main(SessionByKeyBenchmark.scala:108)
at org.apache.spark.streaming.SessionByKeyBenchmark.main(SessionByKeyBenchmark.scala)
at sun.reflect.NativeMethodAccessorImpl.invoke0(Native Method)
at sun.reflect.NativeMethodAccessorImpl.invoke(NativeMethodAccessorImpl.java:62)
at sun.reflect.DelegatingMethodAccessorImpl.invoke(DelegatingMethodAccessorImpl.java:43)
at java.lang.reflect.Method.invoke(Method.java:497)
at org.apache.spark.deploy.SparkSubmit$.org$apache$spark$deploy$SparkSubmit$$runMain(SparkSubmit.scala:680)
at org.apache.spark.deploy.SparkSubmit$.doRunMain$1(SparkSubmit.scala:180)
at org.apache.spark.deploy.SparkSubmit$.submit(SparkSubmit.scala:205)
at org.apache.spark.deploy.SparkSubmit$.main(SparkSubmit.scala:120)
at org.apache.spark.deploy.SparkSubmit.main(SparkSubmit.scala)
```
This PR just moved `ShutdownHookManager.removeShutdownHook` out of `synchronized` to avoid deadlock.
Author: zsxwing <zsxwing@gmail.com>
Closes#9116 from zsxwing/stop-deadlock.
This patch fixes:
1. Guard out against NPEs in `TransformedDStream` when parent DStream returns None instead of empty RDD.
2. Verify some input streams which will potentially return None.
3. Add unit test to verify the behavior when input stream returns None.
cc tdas , please help to review, thanks a lot :).
Author: jerryshao <sshao@hortonworks.com>
Closes#9070 from jerryshao/SPARK-11060.
should pick into spark 1.5.2 also.
https://issues.apache.org/jira/browse/SPARK-10619
looks like this was broken by commit: fb1d06fc24 (diff-b8adb646ef90f616c34eb5c98d1ebd16)
It looks like somethings were change to use the UIUtils.listingTable but executor page wasn't converted so when it removed sortable from the UIUtils. TABLE_CLASS_NOT_STRIPED it broke this page.
Simply add the sortable tag back in and it fixes both active UI and the history server UI.
Author: Tom Graves <tgraves@yahoo-inc.com>
Closes#9101 from tgravescs/SPARK-10619.
Currently, the ```TransformedDStream``` will using ```Some(transformFunc(parentRDDs, validTime))``` as compute return value, when the ```transformFunc``` somehow returns null as return value, the followed operator will have NullPointerExeception.
This fix uses the ```Option()``` instead of ```Some()``` to deal with the possible null value. When ```transformFunc``` returns ```null```, the option will transform null to ```None```, the downstream can handle ```None``` correctly.
NOTE (2015-09-25): The latest fix will check the return value of transform function, if it is ```NULL```, a spark exception will be thrown out
Author: Jacker Hu <gt.hu.chang@gmail.com>
Author: jhu-chang <gt.hu.chang@gmail.com>
Closes#8881 from jhu-chang/Fix_Transform.
This patch introduces a `MemoryManager` that is the central arbiter of how much memory to grant to storage and execution. This patch is primarily concerned only with refactoring while preserving the existing behavior as much as possible.
This is the first step away from the existing rigid separation of storage and execution memory, which has several major drawbacks discussed on the [issue](https://issues.apache.org/jira/browse/SPARK-10956). It is the precursor of a series of patches that will attempt to address those drawbacks.
Author: Andrew Or <andrew@databricks.com>
Author: Josh Rosen <joshrosen@databricks.com>
Author: andrewor14 <andrew@databricks.com>
Closes#9000 from andrewor14/memory-manager.
Dynamic allocation can be painful for streaming apps and can lose data. Log a warning for streaming applications if dynamic allocation is enabled.
Author: Hari Shreedharan <hshreedharan@apache.org>
Closes#8998 from harishreedharan/ss-log-error and squashes the following commits:
462b264 [Hari Shreedharan] Improve log message.
2733d94 [Hari Shreedharan] Minor change to warning message.
eaa48cc [Hari Shreedharan] Log a warning instead of failing the application if dynamic allocation is enabled.
725f090 [Hari Shreedharan] Add config parameter to allow dynamic allocation if the user explicitly sets it.
b3f9a95 [Hari Shreedharan] Disable dynamic allocation and kill app if it is enabled.
a4a5212 [Hari Shreedharan] [streaming] SPARK-10955. Disable dynamic allocation for Streaming applications.
This PR implements the following features for both `master` and `branch-1.5`.
1. Display the failed output op count in the batch list
2. Display the failure reason of output op in the batch detail page
Screenshots:
<img width="1356" alt="1" src="https://cloud.githubusercontent.com/assets/1000778/10198387/5b2b97ec-67ce-11e5-81c2-f818b9d2f3ad.png">
<img width="1356" alt="2" src="https://cloud.githubusercontent.com/assets/1000778/10198388/5b76ac14-67ce-11e5-8c8b-de2683c5b485.png">
There are still two remaining problems in the UI.
1. If an output operation doesn't run any spark job, we cannot get the its duration since now it's the sum of all jobs' durations.
2. If an output operation doesn't run any spark job, we cannot get the description since it's the latest job's call site.
We need to add new `StreamingListenerEvent` about output operations to fix them. So I'd like to fix them only for `master` in another PR.
Author: zsxwing <zsxwing@gmail.com>
Closes#8950 from zsxwing/batch-failure.
Add output operation events to StreamingListener so as to implement the following UI features:
1. Progress bar of a batch in the batch list.
2. Be able to display output operation `description` and `duration` when there is no spark job in a Streaming job.
Author: zsxwing <zsxwing@gmail.com>
Closes#8958 from zsxwing/output-operation-events.
I was reading throught the scheduler and found this small mistake.
Author: Guillaume Poulin <guillaume@hopper.com>
Closes#8966 from gpoulin/remember_duration_typo.
This PR just reverted 02144d6745 to remerge #6457 and also included the commits in #8905.
Author: zsxwing <zsxwing@gmail.com>
Closes#8944 from zsxwing/SPARK-6028.
Slightly modified version of #8818, all credit goes to zsxwing
Author: zsxwing <zsxwing@gmail.com>
Author: Tathagata Das <tathagata.das1565@gmail.com>
Closes#8892 from tdas/SPARK-10692.
Fixed the following failure in https://amplab.cs.berkeley.edu/jenkins/job/NewSparkPullRequestBuilder/1787/testReport/junit/org.apache.spark.streaming/CheckpointSuite/recovery_maintains_rate_controller/
```
sbt.ForkMain$ForkError: The code passed to eventually never returned normally. Attempted 660 times over 10.000044392000001 seconds. Last failure message: 9223372036854775807 did not equal 200.
at org.scalatest.concurrent.Eventually$class.tryTryAgain$1(Eventually.scala:420)
at org.scalatest.concurrent.Eventually$class.eventually(Eventually.scala:438)
at org.scalatest.concurrent.Eventually$.eventually(Eventually.scala:478)
at org.scalatest.concurrent.Eventually$class.eventually(Eventually.scala:336)
at org.scalatest.concurrent.Eventually$.eventually(Eventually.scala:478)
at org.apache.spark.streaming.CheckpointSuite$$anonfun$15.apply$mcV$sp(CheckpointSuite.scala:413)
at org.apache.spark.streaming.CheckpointSuite$$anonfun$15.apply(CheckpointSuite.scala:396)
at org.apache.spark.streaming.CheckpointSuite$$anonfun$15.apply(CheckpointSuite.scala:396)
at org.scalatest.Transformer$$anonfun$apply$1.apply$mcV$sp(Transformer.scala:22)
at org.scalatest.OutcomeOf$class.outcomeOf(OutcomeOf.scala:85)
at org.scalatest.OutcomeOf$.outcomeOf(OutcomeOf.scala:104)
at org.scalatest.Transformer.apply(Transformer.scala:22)
```
In this test, it calls `advanceTimeWithRealDelay(ssc, 2)` to run two batch jobs. However, one race condition is these two jobs can finish before the receiver is registered. Then `UpdateRateLimit` won't be sent to the receiver and `getDefaultBlockGeneratorRateLimit` cannot be updated.
Here are the logs related to this issue:
```
15/09/22 19:28:26.154 pool-1-thread-1-ScalaTest-running-CheckpointSuite INFO CheckpointSuite: Manual clock before advancing = 2500
15/09/22 19:28:26.869 JobScheduler INFO JobScheduler: Finished job streaming job 3000 ms.0 from job set of time 3000 ms
15/09/22 19:28:26.869 JobScheduler INFO JobScheduler: Total delay: 1442975303.869 s for time 3000 ms (execution: 0.711 s)
15/09/22 19:28:26.873 JobScheduler INFO JobScheduler: Finished job streaming job 3500 ms.0 from job set of time 3500 ms
15/09/22 19:28:26.873 JobScheduler INFO JobScheduler: Total delay: 1442975303.373 s for time 3500 ms (execution: 0.004 s)
15/09/22 19:28:26.879 sparkDriver-akka.actor.default-dispatcher-3 INFO ReceiverTracker: Registered receiver for stream 0 from localhost:57749
15/09/22 19:28:27.154 pool-1-thread-1-ScalaTest-running-CheckpointSuite INFO CheckpointSuite: Manual clock after advancing = 3500
```
`advanceTimeWithRealDelay(ssc, 2)` triggered job 3000ms and 3500ms but the receiver was registered after job 3000ms and 3500ms finished.
So we should make sure the receiver online before running `advanceTimeWithRealDelay(ssc, 2)`.
Author: zsxwing <zsxwing@gmail.com>
Closes#8877 from zsxwing/SPARK-10769.
`blockIntervalTimer.stop(interruptTimer = false)` doesn't guarantee calling `updateCurrentBuffer`. So it's possible that `blockIntervalTimer` will exit when `updateCurrentBuffer` is not empty. Then the data in `currentBuffer` will be lost.
To reproduce it, you can add `Thread.sleep(200)` in this line (69c9c17716/streaming/src/main/scala/org/apache/spark/streaming/util/RecurringTimer.scala (L100)) and run `StreamingContexSuite`.
I cannot write a unit test to reproduce it because I cannot find an approach to force `RecurringTimer` suspend at this line for a few milliseconds.
There was a failure in Jenkins here: https://amplab.cs.berkeley.edu/jenkins/job/SparkPullRequestBuilder/41455/console
This PR updates RecurringTimer to make sure `stop(interruptTimer = false)` will call `callback` at least once after the `stop` method is called.
Author: zsxwing <zsxwing@gmail.com>
Closes#8417 from zsxwing/SPARK-10224.
The job group, and job descriptions information is passed through thread local properties, and get inherited by child threads. In case of spark streaming, the streaming jobs inherit these properties from the thread that called streamingContext.start(). This may not make sense.
1. Job group: This is mainly used for cancelling a group of jobs together. It does not make sense to cancel streaming jobs like this, as the effect will be unpredictable. And its not a valid usecase any way, to cancel a streaming context, call streamingContext.stop()
2. Job description: This is used to pass on nice text descriptions for jobs to show up in the UI. The job description of the thread that calls streamingContext.start() is not useful for all the streaming jobs, as it does not make sense for all of the streaming jobs to have the same description, and the description may or may not be related to streaming.
The solution in this PR is meant for the Spark master branch, where local properties are inherited by cloning the properties. The job group and job description in the thread that starts the streaming scheduler are explicitly removed, so that all the subsequent child threads does not inherit them. Also, the starting is done in a new child thread, so that setting the job group and description for streaming, does not change those properties in the thread that called streamingContext.start().
Author: Tathagata Das <tathagata.das1565@gmail.com>
Closes#8781 from tdas/SPARK-10649.
This change does two things:
- tag a few tests and adds the mechanism in the build to be able to disable those tags,
both in maven and sbt, for both junit and scalatest suites.
- add some logic to run-tests.py to disable some tags depending on what files have
changed; that's used to disable expensive tests when a module hasn't explicitly
been changed, to speed up testing for changes that don't directly affect those
modules.
Author: Marcelo Vanzin <vanzin@cloudera.com>
Closes#8437 from vanzin/test-tags.
Move .java files in `src/main/scala` to `src/main/java` root, except for `package-info.java` (to stay next to package.scala)
Author: Sean Owen <sowen@cloudera.com>
Closes#8736 from srowen/SPARK-10576.
Fix a few Java API test style issues: unused generic types, exceptions, wrong assert argument order
Author: Sean Owen <sowen@cloudera.com>
Closes#8706 from srowen/SPARK-10547.
The bulk of the changes are on `transient` annotation on class parameter. Often the compiler doesn't generate a field for this parameters, so the the transient annotation would be unnecessary.
But if the class parameter are used in methods, then fields are created. So it is safer to keep the annotations.
The remainder are some potential bugs, and deprecated syntax.
Author: Luc Bourlier <luc.bourlier@typesafe.com>
Closes#8433 from skyluc/issue/sbt-2.11.
Output a warning when serializing QueueInputDStream rather than throwing an exception to allow unit tests use it. Moreover, this PR also throws an better exception when deserializing QueueInputDStream to make the user find out the problem easily. The previous exception is hard to understand: https://issues.apache.org/jira/browse/SPARK-8553
Author: zsxwing <zsxwing@gmail.com>
Closes#8624 from zsxwing/SPARK-10071 and squashes the following commits:
847cfa8 [zsxwing] Output a warning when writing QueueInputDStream and throw a better exception when reading QueueInputDStream
We introduced the Netty network module for shuffle in Spark 1.2, and has turned it on by default for 3 releases. The old ConnectionManager is difficult to maintain. If we merge the patch now, by the time it is released, it would be 1 yr for which ConnectionManager is off by default. It's time to remove it.
Author: Reynold Xin <rxin@databricks.com>
Closes#8161 from rxin/SPARK-9767.
`deregisterReceiver` should not remove `ReceiverTrackingInfo`. Otherwise, it will throw `java.util.NoSuchElementException: key not found` when restarting it.
Author: zsxwing <zsxwing@gmail.com>
Closes#8538 from zsxwing/SPARK-10369.
Replace `JavaConversions` implicits with `JavaConverters`
Most occurrences I've seen so far are necessary conversions; a few have been avoidable. None are in critical code as far as I see, yet.
Author: Sean Owen <sowen@cloudera.com>
Closes#8033 from srowen/SPARK-9613.
When write ahead log is not enabled, a recovered streaming driver still tries to run jobs using pre-failure block ids, and fails as the block do not exists in-memory any more (and cannot be recovered as receiver WAL is not enabled).
This occurs because the driver-side WAL of ReceivedBlockTracker is recovers that past block information, and ReceiveInputDStream creates BlockRDDs even if those blocks do not exist.
The solution in this PR is to filter out block ids that do not exist before creating the BlockRDD. In addition, it adds unit tests to verify other logic in ReceiverInputDStream.
Author: Tathagata Das <tathagata.das1565@gmail.com>
Closes#8405 from tdas/SPARK-10210.
This PR fixes the following cases for `ReceiverSchedulingPolicy`.
1) Assume there are 4 executors: host1, host2, host3, host4, and 5 receivers: r1, r2, r3, r4, r5. Then `ReceiverSchedulingPolicy.scheduleReceivers` will return (r1 -> host1, r2 -> host2, r3 -> host3, r4 -> host4, r5 -> host1).
Let's assume r1 starts at first on `host1` as `scheduleReceivers` suggested, and try to register with ReceiverTracker. But the previous `ReceiverSchedulingPolicy.rescheduleReceiver` will return (host2, host3, host4) according to the current executor weights (host1 -> 1.0, host2 -> 0.5, host3 -> 0.5, host4 -> 0.5), so ReceiverTracker will reject `r1`. This is unexpected since r1 is starting exactly where `scheduleReceivers` suggested.
This case can be fixed by ignoring the information of the receiver that is rescheduling in `receiverTrackingInfoMap`.
2) Assume there are 3 executors (host1, host2, host3) and each executors has 3 cores, and 3 receivers: r1, r2, r3. Assume r1 is running on host1. Now r2 is restarting, the previous `ReceiverSchedulingPolicy.rescheduleReceiver` will always return (host1, host2, host3). So it's possible that r2 will be scheduled to host1 by TaskScheduler. r3 is similar. Then at last, it's possible that there are 3 receivers running on host1, while host2 and host3 are idle.
This issue can be fixed by returning only executors that have the minimum wight rather than returning at least 3 executors.
Author: zsxwing <zsxwing@gmail.com>
Closes#8340 from zsxwing/fix-receiver-scheduling.
In addition, some random cleanup of import ordering
Author: Tathagata Das <tathagata.das1565@gmail.com>
Closes#8387 from tdas/SPARK-9791 and squashes the following commits:
67f3ee9 [Tathagata Das] Change private class to private[package] class to prevent them from showing up in the docs
The current code only checks checkpoint files in local filesystem, and always tries to create a new Python SparkContext (even if one already exists). The solution is to do the following:
1. Use the same code path as Java to check whether a valid checkpoint exists
2. Create a new Python SparkContext only if there no active one.
There is not test for the path as its hard to test with distributed filesystem paths in a local unit test. I am going to test it with a distributed file system manually to verify that this patch works.
Author: Tathagata Das <tathagata.das1565@gmail.com>
Closes#8366 from tdas/SPARK-10142 and squashes the following commits:
3afa666 [Tathagata Das] Added tests
2dd4ae5 [Tathagata Das] Added the check to not create a context if one already exists
9bf151b [Tathagata Das] Made python checkpoint recovery use java to find the checkpoint files
Recovering Kinesis sequence numbers from WAL leads to classnotfoundexception because the ObjectInputStream does not use the correct classloader and the SequenceNumberRanges class (in streaming-kinesis-asl package) cannot be found (added through spark-submit) while deserializing. The solution is to use `Thread.currentThread().getContextClassLoader` while deserializing.
Author: Tathagata Das <tathagata.das1565@gmail.com>
Closes#8328 from tdas/SPARK-10128 and squashes the following commits:
f19b1c2 [Tathagata Das] Used correct classloader to deserialize WAL data
Because `lazy val` uses `this` lock, if JobGenerator.stop and JobGenerator.doCheckpoint (JobGenerator.shouldCheckpoint has not yet been initialized) run at the same time, it may hang.
Here are the stack traces for the deadlock:
```Java
"pool-1-thread-1-ScalaTest-running-StreamingListenerSuite" #11 prio=5 os_prio=31 tid=0x00007fd35d094800 nid=0x5703 in Object.wait() [0x000000012ecaf000]
java.lang.Thread.State: WAITING (on object monitor)
at java.lang.Object.wait(Native Method)
at java.lang.Thread.join(Thread.java:1245)
- locked <0x00000007b5d8d7f8> (a org.apache.spark.util.EventLoop$$anon$1)
at java.lang.Thread.join(Thread.java:1319)
at org.apache.spark.util.EventLoop.stop(EventLoop.scala:81)
at org.apache.spark.streaming.scheduler.JobGenerator.stop(JobGenerator.scala:155)
- locked <0x00000007b5d8cea0> (a org.apache.spark.streaming.scheduler.JobGenerator)
at org.apache.spark.streaming.scheduler.JobScheduler.stop(JobScheduler.scala:95)
- locked <0x00000007b5d8ced8> (a org.apache.spark.streaming.scheduler.JobScheduler)
at org.apache.spark.streaming.StreamingContext.stop(StreamingContext.scala:687)
"JobGenerator" #67 daemon prio=5 os_prio=31 tid=0x00007fd35c3b9800 nid=0x9f03 waiting for monitor entry [0x0000000139e4a000]
java.lang.Thread.State: BLOCKED (on object monitor)
at org.apache.spark.streaming.scheduler.JobGenerator.shouldCheckpoint$lzycompute(JobGenerator.scala:63)
- waiting to lock <0x00000007b5d8cea0> (a org.apache.spark.streaming.scheduler.JobGenerator)
at org.apache.spark.streaming.scheduler.JobGenerator.shouldCheckpoint(JobGenerator.scala:63)
at org.apache.spark.streaming.scheduler.JobGenerator.doCheckpoint(JobGenerator.scala:290)
at org.apache.spark.streaming.scheduler.JobGenerator.org$apache$spark$streaming$scheduler$JobGenerator$$processEvent(JobGenerator.scala:182)
at org.apache.spark.streaming.scheduler.JobGenerator$$anon$1.onReceive(JobGenerator.scala:83)
at org.apache.spark.streaming.scheduler.JobGenerator$$anon$1.onReceive(JobGenerator.scala:82)
at org.apache.spark.util.EventLoop$$anon$1.run(EventLoop.scala:48)
```
I can use this patch to produce this deadlock: 8a88f28d13
And a timeout build in Jenkins due to this deadlock: https://amplab.cs.berkeley.edu/jenkins/job/NewSparkPullRequestBuilder/1654/
This PR initializes `checkpointWriter` before `eventLoop` uses it to avoid this deadlock.
Author: zsxwing <zsxwing@gmail.com>
Closes#8326 from zsxwing/SPARK-10125.
Small changes
- Renamed conf spark.streaming.backpressure.{enable --> enabled}
- Change Java Deprecated annotations to Scala deprecated annotation with more information.
Author: Tathagata Das <tathagata.das1565@gmail.com>
Closes#8299 from tdas/SPARK-9967.
Generated blocks are inserted into an ArrayBlockingQueue, and another thread pulls stuff from the ArrayBlockingQueue and pushes it into BlockManager. Now if that queue fills up to capacity (default is 10 blocks), then the inserting into queue (done in the function updateCurrentBuffer) get blocked inside a synchronized block. However, the thread that is pulling blocks from the queue uses the same lock to check the current (active or stopped) while pulling from the queue. Since the block generating threads is blocked (as the queue is full) on the lock, this thread that is supposed to drain the queue gets blocked. Ergo, deadlock.
Solution: Moved blocking call to ArrayBlockingQueue outside the synchronized to prevent deadlock.
Author: Tathagata Das <tathagata.das1565@gmail.com>
Closes#8257 from tdas/SPARK-10072.
Failures in streaming.FailureSuite can leak StreamingContext and SparkContext which fails all subsequent tests
Author: Tathagata Das <tathagata.das1565@gmail.com>
Closes#8289 from tdas/SPARK-10098.
When the rate limiter is actually limiting the rate at which data is inserted into the buffer, the synchronized block of BlockGenerator.addData stays blocked for long time. This causes the thread switching the buffer and generating blocks (synchronized with addData) to starve and not generate blocks for seconds. The correct solution is to not block on the rate limiter within the synchronized block for adding data to the buffer.
Author: Tathagata Das <tathagata.das1565@gmail.com>
Closes#8204 from tdas/SPARK-9968 and squashes the following commits:
8cbcc1b [Tathagata Das] Removed unused val
a73b645 [Tathagata Das] Reduced time spent within synchronized block
1. The rate estimator should not estimate any rate when there are no records in the batch, as there is no data to estimate the rate. In the current state, it estimates and set the rate to zero. That is incorrect.
2. The rate estimator should not never set the rate to zero under any circumstances. Otherwise the system will stop receiving data, and stop generating useful estimates (see reason 1). So the fix is to define a parameters that sets a lower bound on the estimated rate, so that the system always receives some data.
Author: Tathagata Das <tathagata.das1565@gmail.com>
Closes#8199 from tdas/SPARK-9966 and squashes the following commits:
829f793 [Tathagata Das] Fixed unit test and added comments
3a994db [Tathagata Das] Added min rate and updated tests in PIDRateEstimator
Refactor Utils class and create ShutdownHookManager.
NOTE: Wasn't able to run /dev/run-tests on windows machine.
Manual tests were conducted locally using custom log4j.properties file with Redis appender and logstash formatter (bundled in the fat-jar submitted to spark)
ex:
log4j.rootCategory=WARN,console,redis
log4j.appender.console=org.apache.log4j.ConsoleAppender
log4j.appender.console.target=System.err
log4j.appender.console.layout=org.apache.log4j.PatternLayout
log4j.appender.console.layout.ConversionPattern=%d{yy/MM/dd HH:mm:ss} %p %c{1}: %m%n
log4j.logger.org.eclipse.jetty=WARN
log4j.logger.org.eclipse.jetty.util.component.AbstractLifeCycle=ERROR
log4j.logger.org.apache.spark.repl.SparkIMain$exprTyper=INFO
log4j.logger.org.apache.spark.repl.SparkILoop$SparkILoopInterpreter=INFO
log4j.logger.org.apache.spark.graphx.Pregel=INFO
log4j.appender.redis=com.ryantenney.log4j.FailoverRedisAppender
log4j.appender.redis.endpoints=hostname:port
log4j.appender.redis.key=mykey
log4j.appender.redis.alwaysBatch=false
log4j.appender.redis.layout=net.logstash.log4j.JSONEventLayoutV1
Author: michellemay <mlemay@gmail.com>
Closes#8109 from michellemay/SPARK-9826.
Spark streaming deletes the temp file and backup files without checking if they exist or not
Author: Hao Zhu <viadeazhu@gmail.com>
Closes#8082 from viadea/master and squashes the following commits:
242d05f [Hao Zhu] [SPARK-9801][Streaming]No need to check the existence of those files
fd143f2 [Hao Zhu] [SPARK-9801][Streaming]Check if backupFile exists before deleting backupFile files.
087daf0 [Hao Zhu] SPARK-9801
Because `JobScheduler.stop(false)` may set `eventLoop` to null when `JobHandler` is running, then it's possible that when `post` is called, `eventLoop` happens to null.
This PR fixed this bug and also set threads in `jobExecutor` to `daemon`.
Author: zsxwing <zsxwing@gmail.com>
Closes#7960 from zsxwing/fix-npe and squashes the following commits:
b0864c4 [zsxwing] Fix a potential NPE in Streaming JobScheduler
In some receivers, instead of using the default `BlockGenerator` in `ReceiverSupervisorImpl`, custom generator with their custom listeners are used for reliability (see [`ReliableKafkaReceiver`](https://github.com/apache/spark/blob/master/external/kafka/src/main/scala/org/apache/spark/streaming/kafka/ReliableKafkaReceiver.scala#L99) and [updated `KinesisReceiver`](https://github.com/apache/spark/pull/7825/files)). These custom generators do not receive rate updates. This PR modifies the code to allow custom `BlockGenerator`s to be created through the `ReceiverSupervisorImpl` so that they can be kept track and rate updates can be applied.
In the process, I did some simplification, and de-flaki-fication of some rate controller related tests. In particular.
- Renamed `Receiver.executor` to `Receiver.supervisor` (to match `ReceiverSupervisor`)
- Made `RateControllerSuite` faster (by increasing batch interval) and less flaky
- Changed a few internal API to return the current rate of block generators as Long instead of Option\[Long\] (was inconsistent at places).
- Updated existing `ReceiverTrackerSuite` to test that custom block generators get rate updates as well.
Author: Tathagata Das <tathagata.das1565@gmail.com>
Closes#7913 from tdas/SPARK-9556 and squashes the following commits:
41d4461 [Tathagata Das] fix scala style
eb9fd59 [Tathagata Das] Updated kinesis receiver
d24994d [Tathagata Das] Updated BlockGeneratorSuite to use manual clock in BlockGenerator
d70608b [Tathagata Das] Updated BlockGenerator with states and proper synchronization
f6bd47e [Tathagata Das] Merge remote-tracking branch 'apache-github/master' into SPARK-9556
31da173 [Tathagata Das] Fix bug
12116df [Tathagata Das] Add BlockGeneratorSuite
74bd069 [Tathagata Das] Fix style
989bb5c [Tathagata Das] Made BlockGenerator fail is used after stop, and added better unit tests for it
3ff618c [Tathagata Das] Fix test
b40eff8 [Tathagata Das] slight refactoring
f0df0f1 [Tathagata Das] Scala style fixes
51759cb [Tathagata Das] Refactored rate controller tests and added the ability to update rate of any custom block generator
This PR is the second one in the larger issue of making the Kinesis integration reliable and provide WAL-free at-least once guarantee. It is based on the design doc - https://docs.google.com/document/d/1k0dl270EnK7uExrsCE7jYw7PYx0YC935uBcxn3p0f58/edit
In this PR, I have updated the Kinesis Receiver to do the following.
- Control the block generation, by creating its own BlockGenerator with own callback methods and using it to keep track of the ranges of sequence numbers that go into each block.
- More specifically, as the KinesisRecordProcessor provides small batches of records, the records are atomically inserted into the block (that is, either the whole batch is in the block, or not). Accordingly the sequence number range of the batch is recorded. Since there may be many batches added to a block, the receiver tracks all the range of sequence numbers that is added to a block.
- When the block is ready to be pushed, the block is pushed and the ranges are reported as metadata of the block. In addition, the ranges are used to find out the latest sequence number for each shard that can be checkpointed through the DynamoDB.
- Periodically, each KinesisRecordProcessor checkpoints the latest successfully stored sequence number for it own shard.
- The array of ranges in the block metadata is used to create KinesisBackedBlockRDDs. The ReceiverInputDStream has been slightly refactored to allow the creation of KinesisBackedBlockRDDs instead of the WALBackedBlockRDDs.
Things to be done
- [x] Add new test to verify that the sequence numbers are recovered.
Author: Tathagata Das <tathagata.das1565@gmail.com>
Closes#7825 from tdas/kinesis-receiver and squashes the following commits:
2159be9 [Tathagata Das] Fixed bug
569be83 [Tathagata Das] Fix scala style issue
bf31e22 [Tathagata Das] Added more documentation to make the kinesis test endpoint more configurable
3ad8361 [Tathagata Das] Merge remote-tracking branch 'apache-github/master' into kinesis-receiver
c693a63 [Tathagata Das] Removed unnecessary constructor params from KinesisTestUtils
e1f1d0a [Tathagata Das] Addressed PR comments
b9fa6bf [Tathagata Das] Fix serialization issues
f8b7680 [Tathagata Das] Updated doc
33fe43a [Tathagata Das] Added more tests
7997138 [Tathagata Das] Fix style errors
a806710 [Tathagata Das] Fixed unit test and use KinesisInputDStream
40a1709 [Tathagata Das] Fixed KinesisReceiverSuite tests
7e44df6 [Tathagata Das] Added documentation and fixed checkpointing
096383f [Tathagata Das] Added test, and addressed some of the comments.
84a7892 [Tathagata Das] fixed scala style issue
e19e37d [Tathagata Das] Added license
1cd7b66 [Tathagata Das] Updated kinesis receiver
Enable most javac lint warnings; fix a lot of build warnings. In a few cases, touch up surrounding code in the process.
I'll explain several of the changes inline in comments.
Author: Sean Owen <sowen@cloudera.com>
Closes#7862 from srowen/SPARK-9534 and squashes the following commits:
ea51618 [Sean Owen] Enable most javac lint warnings; fix a lot of build warnings. In a few cases, touch up surrounding code in the process.
Rename configuration `spark.streaming.minRememberDuration` to `spark.streaming.fileStream.minRememberDuration`
Author: Sameer Abhyankar <sabhyankar@sabhyankar-MBP.local>
Author: Sameer Abhyankar <sabhyankar@sabhyankar-MBP.Samavihome>
Closes#7740 from sabhyankar/spark_branch_9056 and squashes the following commits:
d5b2f1f [Sameer Abhyankar] Correct deprecated version to 1.5
1268133 [Sameer Abhyankar] Add {} and indentation
ddf9844 [Sameer Abhyankar] Change 4 space indentation to 2 space indentation
1819b5f [Sameer Abhyankar] Use spark.streaming.fileStream.minRememberDuration property in lieu of spark.streaming.minRememberDuration
