External Sorting for Aggregator and CoGroupedRDDs (Revisited)
(This pull request is re-opened from https://github.com/apache/incubator-spark/pull/303, which was closed because Jenkins / github was misbehaving)
The target issue for this patch is the out-of-memory exceptions triggered by aggregate operations such as reduce, groupBy, join, and cogroup. The existing AppendOnlyMap used by these operations resides purely in memory, and grows with the size of the input data until the amount of allocated memory is exceeded. Under large workloads, this problem is aggravated by the fact that OOM frequently occurs only after a very long (> 1 hour) map phase, in which case the entire job must be restarted.
The solution is to spill the contents of this map to disk once a certain memory threshold is exceeded. This functionality is provided by ExternalAppendOnlyMap, which additionally sorts this buffer before writing it out to disk, and later merges these buffers back in sorted order.
Under normal circumstances in which OOM is not triggered, ExternalAppendOnlyMap is simply a wrapper around AppendOnlyMap and incurs little overhead. Only when the memory usage is expected to exceed the given threshold does ExternalAppendOnlyMap spill to disk.
Aside from trivial formatting changes, use nulls instead of Options for
DiskMapIterator, and add documentation for spark.shuffle.externalSorting
and spark.shuffle.memoryFraction.
Also, set spark.shuffle.memoryFraction to 0.3, and spark.storage.memoryFraction = 0.6.
Yarn client addjar and misc fixes
Fix the addJar functionality in yarn-client mode, add support for the other options supported in yarn-standalone mode, set the application type on yarn in hadoop 2.X, add documentation, change heartbeat interval to be same code as the yarn-standalone so it doesn't take so long to get containers and exit.
Make DEBUG-level logs consummable.
Removes two things that caused issues with the debug logs:
(a) Internal polling in the DAGScheduler was polluting the logs.
(b) The Scala REPL logs were really noisy.
Removes two things that caused issues with the debug logs:
(a) Internal polling in the DAGScheduler was polluting the logs.
(b) The Scala REPL logs were really noisy.
Fix bug added when we changed AppDescription.maxCores to an Option
The Scala compiler warned about this -- we were comparing an Option against an integer now.
This is an alternative to the existing approach, which evenly distributes the
collective shuffle memory among all running tasks. In the new approach, each
thread requests a chunk of memory whenever its map is about to multiplicatively
grow. If there is sufficient memory in the global pool, the thread allocates it
and grows its map. Otherwise, it spills.
A danger with the previous approach is that a new task may quickly fill up its
map before old tasks finish spilling, potentially causing an OOM. This approach
prevents this scenario as it favors existing tasks over new tasks; any thread
that may step over the boundary of other threads defensively backs off and
starts spilling.
Testing through spark-perf reveals: (1) When no spills have occured, the
performance of external sorting using this memory management approach is
essentially the same as without external sorting. (2) When one or more spills
have occured, the performance of external sorting is a small multiple (3x) worse
Add some missing Java API methods
These are primarily for setting job groups, canceling jobs, and setting names on RDDs. Seemed like useful stuff to expose in Java.
Bug fixes for updating the RDD block's memory and disk usage information
Bug fixes for updating the RDD block's memory and disk usage information.
From the code context, we can find that the memSize and diskSize here are both always equal to the size of the block. Actually, they never be zero. Thus, the logic here is wrong for recording the block usage in BlockStatus, especially for the blocks which are dropped from memory to ensure space for the new input rdd blocks. I have tested it that this would cause the storage metrics shown in the Storage webpage wrong and misleading. With this patch, the metrics will be okay.
Finally, Merry Christmas, guys:)
SPARK-998: Support Launching Driver Inside of Standalone Mode
[NOTE: I need to bring the tests up to date with new changes, so for now they will fail]
This patch provides support for launching driver programs inside of a standalone cluster manager. It also supports monitoring and re-launching of driver programs which is useful for long running, recoverable applications such as Spark Streaming jobs. For those jobs, this patch allows a deployment mode which is resilient to the failure of any worker node, failure of a master node (provided a multi-master setup), and even failures of the applicaiton itself, provided they are recoverable on a restart. Driver information, such as the status and logs from a driver, is displayed in the UI
There are a few small TODO's here, but the code is generally feature-complete. They are:
- Bring tests up to date and add test coverage
- Restarting on failure should be optional and maybe off by default.
- See if we can re-use akka connections to facilitate clients behind a firewall
A sensible place to start for review would be to look at the `DriverClient` class which presents users the ability to launch their driver program. I've also added an example program (`DriverSubmissionTest`) that allows you to test this locally and play around with killing workers, etc. Most of the code is devoted to persisting driver state in the cluster manger, exposing it in the UI, and dealing correctly with various types of failures.
Instructions to test locally:
- `sbt/sbt assembly/assembly examples/assembly`
- start a local version of the standalone cluster manager
```
./spark-class org.apache.spark.deploy.client.DriverClient \
-j -Dspark.test.property=something \
-e SPARK_TEST_KEY=SOMEVALUE \
launch spark://10.99.1.14:7077 \
../path-to-examples-assembly-jar \
org.apache.spark.examples.DriverSubmissionTest 1000 some extra options --some-option-here -X 13
```
- Go in the UI and make sure it started correctly, look at the output etc
- Kill workers, the driver program, masters, etc.
Minor style cleanup. Mostly on indenting & line width changes.
Focused on the few important files since they are the files that new contributors usually read first.
Set boolean param name for call to SparkHadoopMapReduceUtil.newTaskAttemptID
Set boolean param name for call to SparkHadoopMapReduceUtil.newTaskAttemptID to make it clear which param being set.
Remove calls to deprecated mapred's OutputCommitter.cleanupJob
Since Hadoop 1.0.4 the mapred OutputCommitter.commitJob should do cleanup job via call to OutputCommitter.cleanupJob,
Remove SparkHadoopWriter.cleanup since it is used only by PairRDDFunctions.
In fact the implementation of mapred OutputCommitter.commitJob looks like this:
public void commitJob(JobContext jobContext) throws IOException {
cleanupJob(jobContext);
}
the mapred OutputCommitter.commitJob should do cleanup job.
In fact the implementation of mapred OutputCommitter.commitJob looks like this:
public void commitJob(JobContext jobContext) throws IOException {
cleanupJob(jobContext);
}
(The jobContext input argument is type of org.apache.hadoop.mapred.JobContext)
Get rid of `Either[ActorRef, ActorSelection]'
In this pull request, instead of returning an `Either[ActorRef, ActorSelection]`, `registerOrLookup` identifies the remote actor blockingly to obtain an `ActorRef`, or throws an exception if the remote actor doesn't exist or the lookup times out (configured by `spark.akka.lookupTimeout`). This function is only called when an `SparkEnv` is constructed (instantiating driver or executor), so the blocking call is considered acceptable. Executor side `ActorSelection`s/`ActorRef`s to driver side `MapOutputTrackerMasterActor` and `BlockManagerMasterActor` are affected by this pull request.
`ActorSelection` is dangerous and should be used with care. It's only absolutely safe to send messages via an `ActorSelection` when the remote actor is stateless, so that actor incarnation is irrelevant. But as pointed by @ScrapCodes in the comments below, executor exits immediately once the connection to the driver lost, `ActorSelection`s are not harmful in this scenario. So this pull request is mostly a code style patch.
Add way to limit default # of cores used by apps in standalone mode
Also documents the spark.deploy.spreadOut option, and fixes a config option that had a dash in its name.
To make this work I had to rename the defaults file. Otherwise
maven's pattern matching rules included it when trying to match
other log4j.properties files.
I also fixed a bug in the existing maven build where two
<transformers> tags were present in assembly/pom.xml
such that one overwrote the other.
Suggested small changes to Java code for slightly more standard style, encapsulation and in some cases performance
Sorry if this is too abrupt or not a welcome set of changes, but thought I'd see if I could contribute a little. I'm a Java developer and just getting seriously into Spark. So I thought I'd suggest a number of small changes to the couple Java parts of the code to make it a little tighter, more standard and even a bit faster.
Feel free to take all, some or none of this. Happy to explain any of it.
```
[error] /pod/home/anovak/build/graphx/core/src/main/scala/org/apache/spark/storage/ShuffleBlockManager.scala:172: not enough arguments for constructor PrimitiveKeyOpenHashMap: (initialCapacity: Int)(implicit evidence$3: ClassManifest[Int], implicit evidence$4: ClassManifest[Int])org.apache.spark.util.collection.PrimitiveKeyOpenHashMap[Int,Int]
[error] private val mapIdToIndex = new PrimitiveKeyOpenHashMap[Int, Int]()
[error] ^
[info] No documentation generated with unsucessful compiler run
[error] one error found
[error] (core/compile:doc) Scaladoc generation failed
[error] Total time: 67 s, completed Jan 6, 2014 2:20:51 PM
```
In theory a no-argument constructor ought not to differ from one with a single argument that has a default value, but in practice there seems to be an issue.
Fix handling of empty SPARK_EXAMPLES_JAR
Currently if SPARK_EXAMPLES_JAR is left unset you get a null pointer exception when running the examples (atleast on spark on yarn). The null now gets turned into a string of "null" when its put into the SparkConf so addJar no longer properly ignores it. This fixes that so that it can be left unset.
Quiet ERROR-level Akka Logs
This fixes an issue I've seen where akka logs a bunch of things at ERROR level when connecting to a standalone cluster, even in the normal case. I noticed that even when lifecycle logging was disabled, the netty code inside of akka still logged away via akka's EndpointWriter class. There are also some other log streams that I think are new in akka 2.2.1 that I've disabled.
Finally, I added some better logging to the standalone client. This makes it more clear when a connection failure occurs what is going on. Previously it never explicitly said if a connection attempt had failed.
The commit messages here have some more detail.
Removing SPARK_EXAMPLES_JAR in the code
This re-writes all of the examples to use the `SparkContext.jarOfClass` mechanism for loading the examples jar. This necessary for environments like YARN and the Standalone mode where example programs will be submit from inside the cluster rather than at the client using `./spark-example`.
This still leaves SPARK_EXAMPLES_JAR in place in the shell scripts for setting up the classpath if `./spark-example` is run.
Although we can send messages via an ActorSelection, it would be better to identify the actor and obtain an ActorRef first, so that we can get informed earlier if the remote actor doesn't exist, and get rid of the annoying Either wrapper.
Without these it's a bit less clear what's going on for the user.
One thing I realize when doing this is that akka itself actually retries
the initial association. So the retry we currently have is redundant with
akka's.
I noticed when connecting to a standalone cluster Spark gives a bunch
of Akka ERROR logs that make it seem like something is failing.
This patch does two things:
1. Akka dead letter logging is turned on/off according to the existing
lifecycle spark property.
2. We explicitly silence akka's EndpointWriter log in log4j. This is necessary
because for some reason that log doesn't pick up on the lifecycle
logging settings. After a few hours of debugging this was the only solution
I found that worked.
Further, divide this threshold by the number of tasks running concurrently.
Note that this does not guard against the following scenario: a new task
quickly fills up its share of the memory before old tasks finish spilling
their contents, in which case the total memory used by such maps may exceed
what was specified. Currently, spark.shuffle.safetyFraction mitigates the
effect of this.
Remove erroneous FAILED state for killed tasks.
Currently, when tasks are killed, the Executor first sends a
status update for the task with a "KILLED" state, and then
sends a second status update with a "FAILED" state saying that
the task failed due to an exception. The second FAILED state is
misleading/unncessary, and occurs due to a NonLocalReturnControl
Exception that gets thrown due to the way we kill tasks. This
commit eliminates that problem.
I'm not at all sure that this is the best way to fix this problem,
so alternate suggestions welcome. @rxin guessing you're the right
person to look at this.
Improvements to DStream window ops and refactoring of Spark's CheckpointSuite
- Added a new RDD - PartitionerAwareUnionRDD. Using this RDD, one can take multiple RDDs partitioned by the same partitioner and unify them into a single RDD while preserving the partitioner. So m RDDs with p partitions each will be unified to a single RDD with p partitions and the same partitioner. The preferred location for each partition of the unified RDD will be the most common preferred location of the corresponding partitions of the parent RDDs. For example, location of partition 0 of the unified RDD will be where most of partition 0 of the parent RDDs are located.
- Improved the performance of DStream's reduceByKeyAndWindow and groupByKeyAndWindow. Both these operations work by doing per-batch reduceByKey/groupByKey and then using PartitionerAwareUnionRDD to union the RDDs across the window. This eliminates a shuffle related to the window operation, which can reduce batch processing time by 30-40% for simple workloads.
- Fixed bugs and simplified Spark's CheckpointSuite. Some of the tests were incorrect and unreliable. Added missing tests for ZippedRDD. I can go into greater detail if necessary.
- Added mapSideCombine option to combineByKeyAndWindow.
SPARK-991: Report information gleaned from a Python stacktrace in the UI
Scala:
- Added setCallSite/clearCallSite to SparkContext and JavaSparkContext.
These functions mutate a LocalProperty called "externalCallSite."
- Add a wrapper, getCallSite, that checks for an externalCallSite and, if
none is found, calls the usual Utils.formatSparkCallSite.
- Change everything that calls Utils.formatSparkCallSite to call
getCallSite instead. Except getCallSite.
- Add wrappers to setCallSite/clearCallSite wrappers to JavaSparkContext.
Python:
- Add a gruesome hack to rdd.py that inspects the traceback and guesses
what you want to see in the UI.
- Add a RAII wrapper around said gruesome hack that calls
setCallSite/clearCallSite as appropriate.
- Wire said RAII wrapper up around three calls into the Scala code.
I'm not sure that I hit all the spots with the RAII wrapper. I'm also
not sure that my gruesome hack does exactly what we want.
One could also approach this change by refactoring
runJob/submitJob/runApproximateJob to take a call site, then threading
that parameter through everything that needs to know it.
One might object to the pointless-looking wrappers in JavaSparkContext.
Unfortunately, I can't directly access the SparkContext from
Python---or, if I can, I don't know how---so I need to wrap everything
that matters in JavaSparkContext.
Conflicts:
core/src/main/scala/org/apache/spark/api/java/JavaSparkContext.scala
Currently, when tasks are killed, the Executor first sends a
status update for the task with a "KILLED" state, and then
sends a second status update with a "FAILED" state saying that
the task failed due to an exception. The second FAILED state is
misleading/unncessary, and occurs due to a NonLocalReturnControl
Exception that gets thrown due to the way we kill tasks. This
commit eliminates that problem.
Also replaced SparkConf.getOrElse with just a "get" that takes a default
value, and added getInt, getLong, etc to make code that uses this
simpler later on.
Approximate distinct count
Added countApproxDistinct() to RDD and countApproxDistinctByKey() to PairRDDFunctions to approximately count distinct number of elements and distinct number of values per key, respectively. Both functions use HyperLogLog from stream-lib for counting. Both functions take a parameter that controls the trade-off between accuracy and memory consumption. Also added Scala docs and test suites for both methods.
Bug fixes for file input stream and checkpointing
- Fixed bugs in the file input stream that led the stream to fail due to transient HDFS errors (listing files when a background thread it deleting fails caused errors, etc.)
- Updated Spark's CheckpointRDD and Streaming's CheckpointWriter to use SparkContext.hadoopConfiguration, to allow checkpoints to be written to any HDFS compatible store requiring special configuration.
- Changed the API of SparkContext.setCheckpointDir() - eliminated the unnecessary 'useExisting' parameter. Now SparkContext will always create a unique subdirectory within the user specified checkpoint directory. This is to ensure that previous checkpoint files are not accidentally overwritten.
- Fixed bug where setting checkpoint directory as a relative local path caused the checkpointing to fail.
This gives us a couple advantages:
- Uses spark.local.dir and randomly selects a directory/disk.
- Ensure files are deleted on normal DiskBlockManager cleanup.
- Availability of same stats as usual DiskBlockObjectWriter (currenty unused).
Also enable basic cleanup when iterator is fully drained.
Still requires cleanup for operations that fail or don't go through all elements.
Changed naming of StageCompleted event to be consistent
The rest of the SparkListener events are named with "SparkListener"
as the prefix of the name; this commit renames the StageCompleted
event to SparkListenerStageCompleted for consistency.
1. Adds a default log4j file that gets loaded if users haven't specified a log4j file.
2. Isolates use of the tools assembly jar. I found this produced SLF4J warnings
after building with SBT (and I've seen similar warnings on the mailing list).
- Got rid of global SparkContext.globalConf
- Pass SparkConf to serializers and compression codecs
- Made SparkConf public instead of private[spark]
- Improved API of SparkContext and SparkConf
- Switched executor environment vars to be passed through SparkConf
- Fixed some places that were still using system properties
- Fixed some tests, though others are still failing
This still fails several tests in core, repl and streaming, likely due
to properties not being set or cleared correctly (some of the tests run
fine in isolation).
Removed unused OtherFailure TaskEndReason.
The OtherFailure TaskEndReason was added by @mateiz 3 years ago in this commit: 24a1e7f838
Unless I am missing something, it doesn't seem to have been used then, and is not used now, so seems safe for deletion.
The rest of the SparkListener events are named with "SparkListener"
as the prefix of the name; this commit renames the StageCompleted
event to SparkListenerStageCompleted for consistency.
Deduplicate Local and Cluster schedulers.
The code in LocalScheduler/LocalTaskSetManager was nearly identical
to the code in ClusterScheduler/ClusterTaskSetManager. The redundancy
made making updating the schedulers unnecessarily painful and error-
prone. This commit combines the two into a single TaskScheduler/
TaskSetManager.
Unfortunately the diff makes this change look much more invasive than it is -- TaskScheduler.scala is only superficially changed (names updated, overrides removed) from the old ClusterScheduler.scala, and the same with
TaskSetManager.scala.
Thanks @rxin for suggesting this change!
Clean up shuffle files once their metadata is gone
Previously, we would only clean the in-memory metadata for consolidated shuffle files.
Additionally, fixes a bug where the Metadata Cleaner was ignoring type-specific TTLs.
Refactored the streaming scheduler and added StreamingListener interface
- Refactored the streaming scheduler for cleaner code. Specifically, the JobManager was renamed to JobScheduler, as it does the actual scheduling of Spark jobs to the SparkContext. The earlier Scheduler was renamed to JobGenerator, as it actually generates the jobs from the DStreams. The JobScheduler starts the JobGenerator. Also, moved all the scheduler related code from spark.streaming to spark.streaming.scheduler package.
- Implemented the StreamingListener interface, similar to SparkListener. The streaming version of StatusReportListener prints the batch processing time statistics (for now). Added StreamingListernerSuite to test it.
- Refactored streaming TestSuiteBase for deduping code in the other streaming testsuites.
Track and report task result serialisation time.
- DirectTaskResult now has a ByteBuffer valueBytes instead of a T value.
- DirectTaskResult now has a member function T value() that deserialises valueBytes.
- Executor serialises value into a ByteBuffer and passes it to DTR's ctor.
- Executor tracks the time taken to do so and puts it in a new field in TaskMetrics.
- StagePage now reports serialisation time from TaskMetrics along with the other things it reported.
Previously, we would only clean the in-memory metadata for consolidated
shuffle files.
Additionally, fixes a bug where the Metadata Cleaner was ignoring type-
specific TTLs.
Add collectPartition to JavaRDD interface.
This interface is useful for implementing `take` from other language frontends where the data is serialized. Also remove `takePartition` from PythonRDD and use `collectPartition` in rdd.py.
Thanks @concretevitamin for the original change and tests.
Change the implementation to use runJob instead of PartitionPruningRDD.
Also update the unit tests and the python take implementation
to use the new interface.
despite having a low number of nodes and relatively small workload (16 nodes, <1.5 TB data).
This would cause an entire job to fail at the beginning of the reduce phase.
There is no particular reason for this value to be small as a timeout should only occur
in an exceptional situation.
Also centralized the reading of spark.akka.askTimeout to AkkaUtils (surely this can later
be cleaned up to use Typesafe).
Finally, deleted some lurking implicits. If anyone can think of a reason they should still
be there, please let me know.