The main issue we are trying to solve is the memory bloat of the Driver when tasks request the map output statuses. This means with a large number of tasks you either need a huge amount of memory on Driver or you have to repartition to smaller number. This makes it really difficult to run over say 50000 tasks.
The main issues that cause the memory bloat are:
1) no flow control on sending the map output status responses. We serialize the map status output and then hand off to netty to send. netty is sending asynchronously and it can't send them fast enough to keep up with incoming requests so we end up with lots of copies of the serialized map output statuses sitting there and this causes huge bloat when you have 10's of thousands of tasks and map output status is in the 10's of MB.
2) When initial reduce tasks are started up, they all request the map output statuses from the Driver. These requests are handled by multiple threads in parallel so even though we check to see if we have a cached version, initially when we don't have a cached version yet, many of initial requests can all end up serializing the exact same map output statuses.
This patch does a couple of things:
- When the map output status size is over a threshold (default 512K) then it uses broadcast to send the map statuses. This means we no longer serialize a large map output status and thus we don't have issues with memory bloat. the messages sizes are now in the 300-400 byte range and the map status output are broadcast. If its under the threadshold it sends it as before, the message contains the DIRECT indicator now.
- synchronize the incoming requests to allow one thread to cache the serialized output and broadcast the map output status that can then be used by everyone else. This ensures we don't create multiple broadcast variables when we don't need to. To ensure this happens I added a second thread pool which the Dispatcher hands the requests to so that those threads can block without blocking the main dispatcher threads (which would cause things like heartbeats and such not to come through)
Note that some of design and code was contributed by mridulm
## How was this patch tested?
Unit tests and a lot of manually testing.
Ran with akka and netty rpc. Ran with both dynamic allocation on and off.
one of the large jobs I used to test this was a join of 15TB of data. it had 200,000 map tasks, and 20,000 reduce tasks. Executors ranged from 200 to 2000. This job ran successfully with 5GB of memory on the driver with these changes. Without these changes I was using 20GB and only had 500 reduce tasks. The job has 50mb of serialized map output statuses and took roughly the same amount of time for the executors to get the map output statuses as before.
Ran a variety of other jobs, from large wordcounts to small ones not using broadcasts.
Author: Thomas Graves <tgraves@staydecay.corp.gq1.yahoo.com>
Closes#12113 from tgravescs/SPARK-1239.
## What changes were proposed in this pull request?
Minor doc and code style fixes
## How was this patch tested?
local build
Author: Jacek Laskowski <jacek@japila.pl>
Closes#12928 from jaceklaskowski/SPARK-15152.
This patch has the new logic from #8512 that uses a parallel collection to compute partitions in UnionRDD. The rest of #8512 added an alternative code path for calculating splits in S3, but that isn't necessary to get the same speedup. The underlying problem wasn't that bulk listing wasn't used, it was that an extra FileStatus was retrieved for each file. The fix was just committed as [HADOOP-12810](https://issues.apache.org/jira/browse/HADOOP-12810). (I think the original commit also used a single prefix to enumerate all paths, but that isn't always helpful and it was removed in later versions so there is no need for SparkS3Utils.)
I tested this using the same table that piapiaozhexiu was using. Calculating splits for a 10-day period took 25 seconds with this change and HADOOP-12810, which is on par with the results from #8512.
Author: Ryan Blue <blue@apache.org>
Author: Cheolsoo Park <cheolsoop@netflix.com>
Closes#11242 from rdblue/SPARK-9926-parallelize-union-rdd.
## What changes were proposed in this pull request?
set log level to debug when check shouldRollover
## How was this patch tested?
It's tested manually.
Author: depend <depend@gmail.com>
Closes#12931 from depend/master.
## What changes were proposed in this pull request?
Don't re-queue a task if another attempt has already succeeded. This currently happens when a speculative task is denied from committing the result due to another copy of the task already having succeeded.
## How was this patch tested?
I'm running a job which has a fair bit of skew in the processing time across the tasks for speculation to trigger in the last quarter (default settings), causing many commit denied exceptions to be thrown. Previously, these tasks were then being retried over and over again until the stage possibly completes (despite using compute resources on these superfluous tasks). With this change (applied to the 1.6 branch), they no longer retry and the stage completes successfully without these extra task attempts.
Author: Jason Moore <jasonmoore2k@outlook.com>
Closes#12751 from jasonmoore2k/SPARK-14915.
## What changes were proposed in this pull request?
Replace com.sun.jersey with org.glassfish.jersey. Changes to the Spark Web UI code were required to compile. The changes were relatively standard Jersey migration things.
## How was this patch tested?
I did a manual test for the standalone web APIs. Although I didn't test the functionality of the security filter itself, the code that changed non-trivially is how we actually register the filter. I attached a debugger to the Spark master and verified that the SecurityFilter code is indeed invoked upon hitting /api/v1/applications.
Author: mcheah <mcheah@palantir.com>
Closes#12715 from mccheah/feature/upgrade-jersey.
## What changes were proposed in this pull request?
Removed the DeadCode as suggested.
Author: Abhinav Gupta <abhi.951990@gmail.com>
Closes#12829 from abhi951990/master.
Similar to https://github.com/apache/spark/pull/8639
This change rejects offers for 120s when reached `spark.cores.max` in coarse-grained mode to mitigate offer starvation. This prevents Mesos to send us offers again and again, starving other frameworks. This is especially problematic when running many small frameworks on the same Mesos cluster, e.g. many small Sparks streaming jobs, and cause the bigger spark jobs to stop receiving offers. By rejecting the offers for a long period of time, they become available to those other frameworks.
Author: Sebastien Rainville <sebastien@hopper.com>
Closes#10924 from sebastienrainville/master.
Remove history server functionality from standalone Master. Previously, the Master process rebuilt a SparkUI once the application was completed which sometimes caused problems, such as OOM, when the application event log is large (see SPARK-6270). Keeping this functionality out of the Master will help to simplify the process and increase stability.
Testing for this change included running core unit tests and manually running an application on a standalone cluster to verify that it completed successfully and that the Master UI functioned correctly. Also added 2 unit tests to verify killing an application and driver from MasterWebUI makes the correct request to the Master.
Author: Bryan Cutler <cutlerb@gmail.com>
Closes#10991 from BryanCutler/remove-history-master-SPARK-12299.
## What changes were proposed in this pull request?
We currently have a single suite that is very large, making it difficult to maintain and play with specific primitives. This patch reorganizes the file by creating multiple benchmark suites in a single package.
Most of the changes are straightforward move of code. On top of the code moving, I did:
1. Use SparkSession instead of SQLContext.
2. Turned most benchmark scenarios into a their own test cases, rather than having multiple scenarios in a single test case, which takes forever to run.
## How was this patch tested?
This is a test only change.
Author: Reynold Xin <rxin@databricks.com>
Closes#12891 from rxin/SPARK-15115.
## What changes were proposed in this pull request?
Currently only a list of users can be specified for view and modify acls. This change enables a group of admins/devs/users to be provisioned for viewing and modifying Spark jobs.
**Changes Proposed in the fix**
Three new corresponding config entries have been added where the user can specify the groups to be given access.
```
spark.admin.acls.groups
spark.modify.acls.groups
spark.ui.view.acls.groups
```
New config entries were added because specifying the users and groups explicitly is a better and cleaner way compared to specifying them in the existing config entry using a delimiter.
A generic trait has been introduced to provide the user to group mapping which makes it pluggable to support a variety of mapping protocols - similar to the one used in hadoop. A default unix shell based implementation has been provided.
Custom user to group mapping protocol can be specified and configured by the entry ```spark.user.groups.mapping```
**How the patch was Tested**
We ran different spark jobs setting the config entries in combinations of admin, modify and ui acls. For modify acls we tried killing the job stages from the ui and using yarn commands. For view acls we tried accessing the UI tabs and the logs. Headless accounts were used to launch these jobs and different users tried to modify and view the jobs to ensure that the groups mapping applied correctly.
Additional Unit tests have been added without modifying the existing ones. These test for different ways of setting the acls through configuration and/or API and validate the expected behavior.
Author: Dhruve Ashar <dhruveashar@gmail.com>
Closes#12760 from dhruve/impr/SPARK-4224.
## What changes were proposed in this pull request?
This patch changes our micro-benchmark util to allow setting different iteration numbers for different test cases. For some of our benchmarks, turning off whole-stage codegen can make the runtime 20X slower, making it very difficult to run a large number of times without substantially shortening the input cardinality.
With this change, I set the default num iterations to 2 for whole stage codegen off, and 5 for whole stage codegen on. I also updated some results.
## How was this patch tested?
N/A - this is a test util.
Author: Reynold Xin <rxin@databricks.com>
Closes#12884 from rxin/SPARK-15107.
## What changes were proposed in this pull request?
Fix SparkSubmit to allow non-local python uris
## How was this patch tested?
Manually tested with mesos-spark-dispatcher
Author: Timothy Chen <tnachen@gmail.com>
Closes#12403 from tnachen/enable_remote_python.
Otherwise get logs that look like this (note no space before NODE_LOCAL)
```
INFO [2016-05-03 21:18:51,477] org.apache.spark.scheduler.TaskSetManager: Starting task 0.0 in stage 101.0 (TID 7029, localhost, partition 0,NODE_LOCAL, 1894 bytes)
```
Author: Andrew Ash <andrew@andrewash.com>
Closes#12880 from ash211/patch-7.
## What changes were proposed in this pull request?
coalesce doesn't handle UnionRDD with partial locality properly. I had a user who had a UnionRDD that was made up of mapPartitionRDD without preferred locations and a checkpointedRDD with preferred locations (getting from hdfs). It took the driver over 20 minutes to setup the groups and put the partitions into those groups before it even started any tasks. Even perhaps worse is it didn't end up with the number of partitions he was asking for because it didn't put a partition in each of the groups properly.
The changes in this patch get rid of a n^2 while loop that was causing the 20 minutes, it properly distributes the partitions to have at least one per group, and it changes from using the rotation iterator which got the preferred locations many times to get all the preferred locations once up front.
Note that the n^2 while loop that I removed in setupGroups took so long because all of the partitions with preferred locations were already assigned to group, so it basically looped through every single one and wasn't ever able to assign it. At the time I had 960 partitions with preferred locations and 1020 without and did the outer while loop 319 times because that is the # of groups left to create. Note that each of those times through the inner while loop is going off to hdfs to get the block locations, so this is extremely inefficient.
## How was the this patch tested?
Added unit tests for this case and ran existing ones that applied to make sure no regressions.
Also manually tested on the users production job to make sure it fixed their issue. It created the proper number of partitions and now it takes about 6 seconds rather then 20 minutes.
I did also run some basic manual tests with spark-shell doing coalesced to smaller number, same number, and then greater with shuffle.
Author: Thomas Graves <tgraves@prevailsail.corp.gq1.yahoo.com>
Closes#11327 from tgravescs/SPARK-11316.
## What changes were proposed in this pull request?
Resetting the task interruption status before updating the task status.
## How was this patch tested?
I have verified it manually by running multiple applications, Executor doesn't crash and updates the status to the driver without any exceptions with the patch changes.
Author: Devaraj K <devaraj@apache.org>
Closes#12031 from devaraj-kavali/SPARK-14234.
## What changes were proposed in this pull request?
In some cases, fine-grained lock have race condition with class-loader lock and have caused dead lock issue. It is safe to drop this fine grained lock and load all classes by single class-loader lock.
Author: Zheng Tan <zheng.tan@hulu.com>
Closes#12857 from tankkyo/master.
## What changes were proposed in this pull request?
Added tests for ListAccumulator and LegacyAccumulatorWrapper, test for ListAccumulator is one similar to old Collection Accumulators
## How was this patch tested?
Ran tests locally.
cc rxin
Author: Sandeep Singh <sandeep@techaddict.me>
Closes#12862 from techaddict/SPARK-15082.
## What changes were proposed in this pull request?
Remove AccumulatorV2.localValue and keep only value
## How was this patch tested?
existing tests
Author: Sandeep Singh <sandeep@techaddict.me>
Closes#12865 from techaddict/SPARK-15087.
## What changes were proposed in this pull request?
This patch moves AccumulatorV2 and subclasses into util package.
## How was this patch tested?
Updated relevant tests.
Author: Reynold Xin <rxin@databricks.com>
Closes#12863 from rxin/SPARK-15081.
## What changes were proposed in this pull request?
When ALS is run with a checkpoint interval, during the checkpoint materialize the current state and cleanup the previous shuffles (non-blocking).
## How was this patch tested?
Existing ALS unit tests, new ALS checkpoint cleanup unit tests added & shuffle files checked after ALS w/checkpointing run.
Author: Holden Karau <holden@us.ibm.com>
Author: Holden Karau <holden@pigscanfly.ca>
Closes#11919 from holdenk/SPARK-6717-clear-shuffle-files-after-checkpointing-in-ALS.
## What changes were proposed in this pull request?
This patch removes AverageAccumulator and adds the ability to compute average to LongAccumulator and DoubleAccumulator. The patch also improves documentation for the two accumulators.
## How was this patch tested?
Added unit tests for this.
Author: Reynold Xin <rxin@databricks.com>
Closes#12858 from rxin/SPARK-15079.
## What changes were proposed in this pull request?
This updates the java-/scala- doc for setLocalProperty to document heritability of localProperties. This also adds tests for that behaviour.
## How was this patch tested?
Tests pass. New tests were added.
Author: Marcin Tustin <marcin.tustin@gmail.com>
Closes#12455 from marcintustin/SPARK-14685.
## What changes were proposed in this pull request?
This patch deprecates the old accumulator API.
## How was this patch tested?
N/A
Author: Reynold Xin <rxin@databricks.com>
Closes#12832 from rxin/SPARK-15054.
## What changes were proposed in this pull request?
initialize SparkSubmitArgument#files first from spark-submit arguments then from properties file, so that sys property spark.yarn.dist.files will be set correctly.
```
OptionAssigner(args.files, YARN, ALL_DEPLOY_MODES, sysProp = "spark.yarn.dist.files"),
```
## How was this patch tested?
manul test. file defined in properties file is also distributed to driver in yarn-cluster mode.
Author: Jeff Zhang <zjffdu@apache.org>
Closes#12656 from zjffdu/SPARK-14845.
## What changes were proposed in this pull request?
NewAccumulator isn't the best name if we ever come up with v3 of the API.
## How was this patch tested?
Updated tests to reflect the change.
Author: Reynold Xin <rxin@databricks.com>
Closes#12827 from rxin/SPARK-15049.
#### What changes were proposed in this pull request?
This PR removes three methods the were deprecated in 1.6.0:
- `PortableDataStream.close()`
- `LinearRegression.weights`
- `LogisticRegression.weights`
The rationale for doing this is that the impact is small and that Spark 2.0 is a major release.
#### How was this patch tested?
Compilation succeded.
Author: Herman van Hovell <hvanhovell@questtec.nl>
Closes#12732 from hvanhovell/SPARK-14952.
## What changes were proposed in this pull request?
This patch removes some code that are no longer relevant -- mainly HiveSessionState.setDefaultOverrideConfs.
## How was this patch tested?
N/A
Author: Reynold Xin <rxin@databricks.com>
Closes#12806 from rxin/SPARK-15028.
## What changes were proposed in this pull request?
The RPC framework will not serialize and deserialize messages in local mode, we should not call `acc.value` when receive heartbeat message, because the serialization hook of new accumulator may not be triggered and the `atDriverSide` flag may not be set.
## How was this patch tested?
tested it locally via spark shell
Author: Wenchen Fan <wenchen@databricks.com>
Closes#12795 from cloud-fan/bug.
## What changes were proposed in this pull request?
This PR proposes to use ConcurrentHashMap in place of HashMap for NewAccumulator.originals
This should result in better performance.
## How was this patch tested?
Existing unit test suite
cloud-fan
Author: tedyu <yuzhihong@gmail.com>
Closes#12776 from tedyu/master.
## What changes were proposed in this pull request?
dapply() applies an R function on each partition of a DataFrame and returns a new DataFrame.
The function signature is:
dapply(df, function(localDF) {}, schema = NULL)
R function input: local data.frame from the partition on local node
R function output: local data.frame
Schema specifies the Row format of the resulting DataFrame. It must match the R function's output.
If schema is not specified, each partition of the result DataFrame will be serialized in R into a single byte array. Such resulting DataFrame can be processed by successive calls to dapply().
## How was this patch tested?
SparkR unit tests.
Author: Sun Rui <rui.sun@intel.com>
Author: Sun Rui <sunrui2016@gmail.com>
Closes#12493 from sun-rui/SPARK-12919.
## What changes were proposed in this pull request?
`AccumulatorContext` is not thread-safe, that's why all of its methods are synchronized. However, there is one exception: the `AccumulatorContext.originals`. `NewAccumulator` use it to check if it's registered, which is wrong as it's not synchronized.
This PR mark `AccumulatorContext.originals` as `private` and now all access to `AccumulatorContext` is synchronized.
## How was this patch tested?
I verified it locally. To be safe, we can let jenkins test it many times to make sure this problem is gone.
Author: Wenchen Fan <wenchen@databricks.com>
Closes#12773 from cloud-fan/debug.
## What changes were proposed in this pull request?
This PR adds `since` tag into the matrix and vector classes in spark-mllib-local.
## How was this patch tested?
Scala-style checks passed.
Author: Pravin Gadakh <prgadakh@in.ibm.com>
Closes#12416 from pravingadakh/SPARK-14613.
https://issues.apache.org/jira/browse/SPARK-14935
In DistributedSuite, the "local-cluster format" test actually launches a bunch of clusters, but this doesn't seem necessary for what should just be a unit test of a regex. We should clean up the code so that this is testable without actually launching a cluster, which should buy us about 20 seconds per build.
Passed unit test on my local machine
Author: Xin Ren <iamshrek@126.com>
Closes#12744 from keypointt/SPARK-14935.
## What changes were proposed in this pull request?
This is a proposal to print the Spark Driver UI link when spark-shell is launched.
## How was this patch tested?
Launched spark-shell in local mode and cluster mode. Spark-shell console output included following line:
"Spark context Web UI available at <Spark web url>"
Author: Ergin Seyfe <eseyfe@fb.com>
Closes#12341 from seyfe/spark_console_display_webui_link.
## What changes were proposed in this pull request?
This PR introduces a new accumulator API which is much simpler than before:
1. the type hierarchy is simplified, now we only have an `Accumulator` class
2. Combine `initialValue` and `zeroValue` concepts into just one concept: `zeroValue`
3. there in only one `register` method, the accumulator registration and cleanup registration are combined.
4. the `id`,`name` and `countFailedValues` are combined into an `AccumulatorMetadata`, and is provided during registration.
`SQLMetric` is a good example to show the simplicity of this new API.
What we break:
1. no `setValue` anymore. In the new API, the intermedia type can be different from the result type, it's very hard to implement a general `setValue`
2. accumulator can't be serialized before registered.
Problems need to be addressed in follow-ups:
1. with this new API, `AccumulatorInfo` doesn't make a lot of sense, the partial output is not partial updates, we need to expose the intermediate value.
2. `ExceptionFailure` should not carry the accumulator updates. Why do users care about accumulator updates for failed cases? It looks like we only use this feature to update the internal metrics, how about we sending a heartbeat to update internal metrics after the failure event?
3. the public event `SparkListenerTaskEnd` carries a `TaskMetrics`. Ideally this `TaskMetrics` don't need to carry external accumulators, as the only method of `TaskMetrics` that can access external accumulators is `private[spark]`. However, `SQLListener` use it to retrieve sql metrics.
## How was this patch tested?
existing tests
Author: Wenchen Fan <wenchen@databricks.com>
Closes#12612 from cloud-fan/acc.
## What changes were proposed in this pull request?
The current signal handlers have a subtle bug that stops evaluating registered actions as soon as one of them returns true, this is because `forall` is short-circuited.
This PR adds a strict mapping stage before evaluating returned result.
There are no known occurrences of the bug and this is a preemptive fix.
## How was this patch tested?
As with the original introduction of signal handlers, this was tested manually (unit testing with signals is not straightforward).
Author: Jakob Odersky <jakob@odersky.com>
Closes#12745 from jodersky/SPARK-10001-hotfix.
In local profiling, I noticed SizeEstimator spending tons of time estimating the size of objects which contain TypeTag or ClassTag fields. The problem with these tags is that they reference global Scala reflection objects, which, in turn, reference many singletons, such as TestHive. This throws off the accuracy of the size estimation and wastes tons of time traversing a huge object graph.
As a result, I think that SizeEstimator should ignore any classes in the `scala.reflect` package.
Author: Josh Rosen <joshrosen@databricks.com>
Closes#12741 from JoshRosen/ignore-scala-reflect-in-size-estimator.
## What changes were proposed in this pull request?
With the addition of ExternalClusterManager(ECM) interface in PR #11723, any cluster manager can now be integrated with Spark. It was suggested in ExternalClusterManager PR that one of the existing cluster managers should start using the new interface to ensure that the API is correct. Ideally, all the existing cluster managers should eventually use the ECM interface but as a first step yarn will now use the ECM interface. This PR refactors YARN code from SparkContext.createTaskScheduler function into YarnClusterManager that implements ECM interface.
## How was this patch tested?
Since this is refactoring, no new tests has been added. Existing tests have been run. Basic manual testing with YARN was done too.
Author: Hemant Bhanawat <hemant@snappydata.io>
Closes#12641 from hbhanawat/yarnClusterMgr.
## What changes were proposed in this pull request?
Pruned off white spaces present in the user provided comma separated list of classes for **spark.kryo.classesToRegister** and **spark.kryo.registrator**.
## How was this patch tested?
Manual tests
Author: Victor Chima <blazy2k9@gmail.com>
Closes#12701 from blazy2k9/master.
## What changes were proposed in this pull request?
Since `LZ4BlockInputStream.java` is not licensed to Apache Software Foundation (ASF), the Apache License header of that file is not monitored until now.
This PR aims to enable RAT checking on `LZ4BlockInputStream.java` by excluding from `dev/.rat-excludes`.
This will prevent accidental removal of Apache License header from that file.
## How was this patch tested?
Pass the Jenkins tests (Specifically, RAT check stage).
Author: Dongjoon Hyun <dongjoon@apache.org>
Closes#12677 from dongjoon-hyun/minor_rat_exclusion_file.
## What changes were proposed in this pull request?
Use Long.parseLong which returns a primative.
Use a series of appends() reduces the creation of an extra StringBuilder type
## How was this patch tested?
Unit tests
Author: Azeem Jiva <azeemj@gmail.com>
Closes#12520 from javawithjiva/minor.
## What changes were proposed in this pull request?
Handling exception for the below mentioned issue
```
➜ spark git:(master) ✗ ./bin/spark-shell -c spark.scheduler.mode=NONE
16/04/25 09:15:00 ERROR SparkContext: Error initializing SparkContext.
scala.MatchError: NONE (of class scala.Enumeration$Val)
at org.apache.spark.scheduler.Pool.<init>(Pool.scala:53)
at org.apache.spark.scheduler.TaskSchedulerImpl.initialize(TaskSchedulerImpl.scala:131)
at org.apache.spark.SparkContext$.org$apache$spark$SparkContext$$createTaskScheduler(SparkContext.scala:2352)
at org.apache.spark.SparkContext.<init>(SparkContext.scala:492)
```
The exception now looks like
```
java.lang.RuntimeException: The scheduler mode NONE is not supported by Spark.
```
## How was this patch tested?
manual tests
Author: Subhobrata Dey <sbcd90@gmail.com>
Closes#12666 from sbcd90/schedulerModeIssue.
## What changes were proposed in this pull request?
SPARK-12130 make 2.0 shuffle service incompatible with 1.x. So from discussion: [http://apache-spark-developers-list.1001551.n3.nabble.com/YARN-Shuffle-service-and-its-compatibility-td17222.html](url) we should maintain compatibility between Spark 1.x and Spark 2.x's shuffle service.
I put string comparison into executor's register at first avoid string comparison in getBlockData every time.
## How was this patch tested?
N/A
Author: Lianhui Wang <lianhuiwang09@gmail.com>
Closes#12568 from lianhuiwang/SPARK-14731.
## What changes were proposed in this pull request?
Implement the same memory size validations for the StaticMemoryManager (Legacy) as the UnifiedMemoryManager has.
## How was this patch tested?
Manual tests were done in CDH cluster.
Test with small executor memory:
`
spark-submit --class org.apache.spark.examples.SparkPi --deploy-mode client --master yarn --executor-memory 15m --conf spark.memory.useLegacyMode=true /opt/cloudera/parcels/CDH/lib/spark/examples/lib/spark-examples*.jar 10
`
Exception thrown:
```
ERROR spark.SparkContext: Error initializing SparkContext.
java.lang.IllegalArgumentException: Executor memory 15728640 must be at least 471859200. Please increase executor memory using the --executor-memory option or spark.executor.memory in Spark configuration.
at org.apache.spark.memory.StaticMemoryManager$.org$apache$spark$memory$StaticMemoryManager$$getMaxExecutionMemory(StaticMemoryManager.scala:127)
at org.apache.spark.memory.StaticMemoryManager.<init>(StaticMemoryManager.scala:46)
at org.apache.spark.SparkEnv$.create(SparkEnv.scala:352)
at org.apache.spark.SparkEnv$.createDriverEnv(SparkEnv.scala:193)
at org.apache.spark.SparkContext.createSparkEnv(SparkContext.scala:289)
at org.apache.spark.SparkContext.<init>(SparkContext.scala:462)
at org.apache.spark.examples.SparkPi$.main(SparkPi.scala:29)
at org.apache.spark.examples.SparkPi.main(SparkPi.scala)
at sun.reflect.NativeMethodAccessorImpl.invoke0(Native Method)
at sun.reflect.NativeMethodAccessorImpl.invoke(NativeMethodAccessorImpl.java:57)
at sun.reflect.DelegatingMethodAccessorImpl.invoke(DelegatingMethodAccessorImpl.java:43)
at java.lang.reflect.Method.invoke(Method.java:606)
at org.apache.spark.deploy.SparkSubmit$.org$apache$spark$deploy$SparkSubmit$$runMain(SparkSubmit.scala:731)
at org.apache.spark.deploy.SparkSubmit$.doRunMain$1(SparkSubmit.scala:181)
at org.apache.spark.deploy.SparkSubmit$.submit(SparkSubmit.scala:206)
at org.apache.spark.deploy.SparkSubmit$.main(SparkSubmit.scala:121)
at org.apache.spark.deploy.SparkSubmit.main(SparkSubmit.scala)
```
Author: Peter Ableda <peter.ableda@cloudera.com>
Closes#12395 from peterableda/SPARK-14636.
## What changes were proposed in this pull request?
Change default logging to WARN for pyspark shell and sparkR shell for a much cleaner environment.
## How was this patch tested?
Manually running pyspark and sparkR shell
Author: felixcheung <felixcheung_m@hotmail.com>
Closes#12648 from felixcheung/pylogging.
## What changes were proposed in this pull request?
Spark uses `NewLineAtEofChecker` rule in Scala by ScalaStyle. And, most Java code also comply with the rule. This PR aims to enforce the same rule `NewlineAtEndOfFile` by CheckStyle explicitly. Also, this fixes lint-java errors since SPARK-14465. The followings are the items.
- Adds a new line at the end of the files (19 files)
- Fixes 25 lint-java errors (12 RedundantModifier, 6 **ArrayTypeStyle**, 2 LineLength, 2 UnusedImports, 2 RegexpSingleline, 1 ModifierOrder)
## How was this patch tested?
After the Jenkins test succeeds, `dev/lint-java` should pass. (Currently, Jenkins dose not run lint-java.)
```bash
$ dev/lint-java
Using `mvn` from path: /usr/local/bin/mvn
Checkstyle checks passed.
```
Author: Dongjoon Hyun <dongjoon@apache.org>
Closes#12632 from dongjoon-hyun/SPARK-14868.
## What changes were proposed in this pull request?
Java `sampleByKey` methods should accept `Map` with `java.lang.Double` values
## How was this patch tested?
Existing (updated) Jenkins tests
Author: Sean Owen <sowen@cloudera.com>
Closes#12637 from srowen/SPARK-14873.
## What changes were proposed in this pull request?
1. Fix the "spill size" of TungstenAggregate and Sort
2. Rename "data size" to "peak memory" to match the actual meaning (also consistent with task metrics)
3. Added "data size" for ShuffleExchange and BroadcastExchange
4. Added some timing for Sort, Aggregate and BroadcastExchange (this requires another patch to work)
## How was this patch tested?
Existing tests.
Author: Davies Liu <davies@databricks.com>
Closes#12425 from davies/fix_metrics.
## What changes were proposed in this pull request?
This is a follow-up to #12557, with the following changes:
1. Fixes some of the style issues.
2. Merges Signaling and SignalLogger into a new class called SignalUtils. It was pretty confusing to have Signaling and Signal in one file, and it was also confusing to have two classes named Signaling and one called the other.
3. Made logging registration idempotent.
## How was this patch tested?
N/A.
Author: Reynold Xin <rxin@databricks.com>
Closes#12605 from rxin/SPARK-10001.
## What changes were proposed in this pull request?
Implement some `hashCode` and `equals` together in order to enable the scalastyle.
This is a first batch, I will continue to implement them but I wanted to know your thoughts.
Author: Joan <joan@goyeau.com>
Closes#12157 from joan38/SPARK-6429-HashCode-Equals.
## What changes were proposed in this pull request?
Improve signal handling to allow interrupting running tasks from the REPL (with Ctrl+C).
If no tasks are running or Ctrl+C is pressed twice, the signal is forwarded to the default handler resulting in the usual termination of the application.
This PR is a rewrite of -- and therefore closes#8216 -- as per piaozhexiu's request
## How was this patch tested?
Signal handling is not easily testable therefore no unit tests were added. Nevertheless, the new functionality is implemented in a best-effort approach, soft-failing in case signals aren't available on a specific OS.
Author: Jakob Odersky <jakob@odersky.com>
Closes#12557 from jodersky/SPARK-10001-sigint.
## What changes were proposed in this pull request?
Spark currently uses TimSort for all in-memory sorts, including sorts done for shuffle. One low-hanging fruit is to use radix sort when possible (e.g. sorting by integer keys). This PR adds a radix sort implementation to the unsafe sort package and switches shuffles and sorts to use it when possible.
The current implementation does not have special support for null values, so we cannot radix-sort `LongType`. I will address this in a follow-up PR.
## How was this patch tested?
Unit tests, enabling radix sort on existing tests. Microbenchmark results:
```
Running benchmark: radix sort 25000000
Java HotSpot(TM) 64-Bit Server VM 1.8.0_66-b17 on Linux 3.13.0-44-generic
Intel(R) Core(TM) i7-4600U CPU 2.10GHz
radix sort 25000000: Best/Avg Time(ms) Rate(M/s) Per Row(ns) Relative
-------------------------------------------------------------------------------------------
reference TimSort key prefix array 15546 / 15859 1.6 621.9 1.0X
reference Arrays.sort 2416 / 2446 10.3 96.6 6.4X
radix sort one byte 133 / 137 188.4 5.3 117.2X
radix sort two bytes 255 / 258 98.2 10.2 61.1X
radix sort eight bytes 991 / 997 25.2 39.6 15.7X
radix sort key prefix array 1540 / 1563 16.2 61.6 10.1X
```
I also ran a mix of the supported TPCDS queries and compared TimSort vs RadixSort metrics. The overall benchmark ran ~10% faster with radix sort on. In the breakdown below, the radix-enabled sort phases averaged about 20x faster than TimSort, however sorting is only a small fraction of the overall runtime. About half of the TPCDS queries were able to take advantage of radix sort.
```
TPCDS on master: 2499s real time, 8185s executor
- 1171s in TimSort, avg 267 MB/s
(note the /s accounting is weird here since dataSize counts the record sizes too)
TPCDS with radix enabled: 2294s real time, 7391s executor
- 596s in TimSort, avg 254 MB/s
- 26s in radix sort, avg 4.2 GB/s
```
cc davies rxin
Author: Eric Liang <ekl@databricks.com>
Closes#12490 from ericl/sort-benchmark.
## What changes were proposed in this pull request?
In general, `onDisconnected` is for dealing with unexpected network disconnections. When RpcEnv.shutdown is called, the disconnections are expected so RpcEnv should not fire these events.
This PR moves `dispatcher.stop()` above closing the connections so that when stopping RpcEnv, the endpoints won't receive `onDisconnected` events.
In addition, Outbox should not close the client since it will be reused by others. This PR fixes it as well.
## How was this patch tested?
test("SPARK-14699: RpcEnv.shutdown should not fire onDisconnected events")
Author: Shixiong Zhu <shixiong@databricks.com>
Closes#12481 from zsxwing/SPARK-14699.
## What changes were proposed in this pull request?
In #9241 It implemented a mechanism to call spill() on those SQL operators that support spilling if there is not enough memory for execution.
But ExternalSorter and AppendOnlyMap in Spark core are not worked. So this PR make them benefit from #9241. Now when there is not enough memory for execution, it can get memory by spilling ExternalSorter and AppendOnlyMap in Spark core.
## How was this patch tested?
add two unit tests for it.
Author: Lianhui Wang <lianhuiwang09@gmail.com>
Closes#10024 from lianhuiwang/SPARK-4452-2.
## What changes were proposed in this pull request?
The patch makes event log processing multi threaded.
## How was this patch tested?
Existing tests pass, there is no new tests needed to test the functionality as this is a perf improvement. I tested the patch locally by generating one big event log (big1), one small event log(small1) and again a big event log(big2). Without this patch UI does not render any app for almost 30 seconds and then big2 and small1 appears. another 30 second delay and finally big1 also shows up in UI. With this change small1 shows up immediately and big1 and big2 comes up in 30 seconds. Locally it also displays them in the correct order in the UI.
Author: Parth Brahmbhatt <pbrahmbhatt@netflix.com>
Closes#11800 from Parth-Brahmbhatt/SPARK-13988.
In o.a.s.deploy.worker.Worker.scala, when receiving a KillExecutor message from an invalid Master, fixed typo by changing the log message to read "..attemped to kill executor.."
Author: Bryan Cutler <cutlerb@gmail.com>
Closes#12546 from BryanCutler/worker-killexecutor-log-message.
## What changes were proposed in this pull request?
the `Accumulable.internal` flag is only used to avoid registering internal accumulators for 2 certain cases:
1. `TaskMetrics.createTempShuffleReadMetrics`: the accumulators in the temp shuffle read metrics should not be registered.
2. `TaskMetrics.fromAccumulatorUpdates`: the created task metrics is only used to post event, accumulators inside it should not be registered.
For 1, we can create a `TempShuffleReadMetrics` that don't create accumulators, just keep the data and merge it at last.
For 2, we can un-register these accumulators immediately.
TODO: remove `internal` flag in `AccumulableInfo` with followup PR
## How was this patch tested?
existing tests.
Author: Wenchen Fan <wenchen@databricks.com>
Closes#12525 from cloud-fan/acc.
This change avoids using the environment to pass this information, since
with many jars it's easy to hit limits on certain OSes. Instead, it encodes
the information into the Spark configuration propagated to the AM.
The first problem that needed to be solved is a chicken & egg issue: the
config file is distributed using the cache, and it needs to contain information
about the files that are being distributed. To solve that, the code now treats
the config archive especially, and uses slightly different code to distribute
it, so that only its cache path needs to be saved to the config file.
The second problem is that the extra information would show up in the Web UI,
which made the environment tab even more noisy than it already is when lots
of jars are listed. This is solved by two changes: the list of cached files
is now read only once in the AM, and propagated down to the ExecutorRunnable
code (which actually sends the list to the NMs when starting containers). The
second change is to unset those config entries after the list is read, so that
the SparkContext never sees them.
Tested with both client and cluster mode by running "run-example SparkPi". This
uploads a whole lot of files when run from a build dir (instead of a distribution,
where the list is cleaned up), and I verified that the configs do not show
up in the UI.
Author: Marcelo Vanzin <vanzin@cloudera.com>
Closes#12487 from vanzin/SPARK-14602.
## What changes were proposed in this pull request?
This PR has two main changes.
1. Move Hive-specific methods from HiveContext to HiveSessionState, which help the work of removing HiveContext.
2. Create a SparkSession Class, which will later be the entry point of Spark SQL users.
## How was this patch tested?
Existing tests
This PR is trying to fix test failures of https://github.com/apache/spark/pull/12485.
Author: Andrew Or <andrew@databricks.com>
Author: Yin Huai <yhuai@databricks.com>
Closes#12522 from yhuai/spark-session.
## What changes were proposed in this pull request?
This proposal removes the class `HttpServer`, with the changing of internal file/jar/class transmission to RPC layer, currently there's no code using this `HttpServer`, so here propose to remove it.
## How was this patch tested?
Unit test is verified locally.
Author: jerryshao <sshao@hortonworks.com>
Closes#12526 from jerryshao/SPARK-14725.
Updated the log page by replacing the current pagination with a javascript-based infinite scroll solution
Author: Alex Bozarth <ajbozart@us.ibm.com>
Closes#10910 from ajbozarth/spark8171.
## What changes were proposed in this pull request?
- replaced `FileSystem.get(conf)` calls with `path.getFileSystem(conf)`
## How was this patch tested?
N/A
Author: Liwei Lin <lwlin7@gmail.com>
Closes#12450 from lw-lin/fix-fs-get.
## What changes were proposed in this pull request?
The DAG visualization can cause an OOM when generating the DOT file.
This happens because clusters are not correctly deduped by a contains
check because they use the default equals implementation. This adds a
working equals implementation.
## How was this patch tested?
This adds a test suite that checks the new equals implementation.
Author: Ryan Blue <blue@apache.org>
Closes#12437 from rdblue/SPARK-14679-fix-ui-oom.
## What changes were proposed in this pull request?
Before this PR, we create accumulators at driver side(and register them) and send them to executor side, then we create `TaskMetrics` with these accumulators at executor side.
After this PR, we will create `TaskMetrics` at driver side and send it to executor side, so that we can create accumulators inside `TaskMetrics` directly, which is cleaner.
## How was this patch tested?
existing tests.
Author: Wenchen Fan <wenchen@databricks.com>
Closes#12472 from cloud-fan/acc.
Add R API for `read.jdbc`, `write.jdbc`.
Tested this quite a bit manually with different combinations of parameters. It's not clear if we could have automated tests in R for this - Scala `JDBCSuite` depends on Java H2 in-memory database.
Refactored some code into util so they could be tested.
Core's R SerDe code needs to be updated to allow access to java.util.Properties as `jobj` handle which is required by DataFrameReader/Writer's `jdbc` method. It would be possible, though more code to add a `sql/r/SQLUtils` helper function.
Tested:
```
# with postgresql
../bin/sparkR --driver-class-path /usr/share/java/postgresql-9.4.1207.jre7.jar
# read.jdbc
df <- read.jdbc(sqlContext, "jdbc:postgresql://localhost/db", "films2", user = "user", password = "12345")
df <- read.jdbc(sqlContext, "jdbc:postgresql://localhost/db", "films2", user = "user", password = 12345)
# partitionColumn and numPartitions test
df <- read.jdbc(sqlContext, "jdbc:postgresql://localhost/db", "films2", partitionColumn = "did", lowerBound = 0, upperBound = 200, numPartitions = 4, user = "user", password = 12345)
a <- SparkR:::toRDD(df)
SparkR:::getNumPartitions(a)
[1] 4
SparkR:::collectPartition(a, 2L)
# defaultParallelism test
df <- read.jdbc(sqlContext, "jdbc:postgresql://localhost/db", "films2", partitionColumn = "did", lowerBound = 0, upperBound = 200, user = "user", password = 12345)
SparkR:::getNumPartitions(a)
[1] 2
# predicates test
df <- read.jdbc(sqlContext, "jdbc:postgresql://localhost/db", "films2", predicates = list("did<=105"), user = "user", password = 12345)
count(df) == 1
# write.jdbc, default save mode "error"
irisDf <- as.DataFrame(sqlContext, iris)
write.jdbc(irisDf, "jdbc:postgresql://localhost/db", "films2", user = "user", password = "12345")
"error, already exists"
write.jdbc(irisDf, "jdbc:postgresql://localhost/db", "iris", user = "user", password = "12345")
```
Author: felixcheung <felixcheung_m@hotmail.com>
Closes#10480 from felixcheung/rreadjdbc.
## What changes were proposed in this pull request?
The current benchmark framework runs a code block for several iterations and reports statistics. However there is no way to exclude per-iteration setup time from the overall results. This PR adds a timer control object passed into the closure that can be used for this purpose.
## How was this patch tested?
Existing benchmark code. Also see https://github.com/apache/spark/pull/12490
Author: Eric Liang <ekl@databricks.com>
Closes#12502 from ericl/spark-14733.
## What changes were proposed in this pull request?
This PR adds support for specifying an optional custom coalescer to the `coalesce()` method. Currently I have only added this feature to the `RDD` interface, and once we sort out the details we can proceed with adding this feature to the other APIs (`Dataset` etc.)
## How was this patch tested?
Added a unit test for this functionality.
/cc rxin (per our discussion on the mailing list)
Author: Nezih Yigitbasi <nyigitbasi@netflix.com>
Closes#11865 from nezihyigitbasi/custom_coalesce_policy.
## What changes were proposed in this pull request?
This PR returns correct processor name in ```/proc/cpuinfo``` on Linux from ```Benchmark.getPorcessorName()```. Now, this return ```Unknown processor```.
Since ```Utils.executeAndGetOutput(Seq("which", "grep"))``` return ```/bin/grep\n```, it is failed to execute ```/bin/grep\n```. This PR strips ```\n``` at the end of the line of a result of ```Utils.executeAndGetOutput()```
Before applying this PR
````
Java HotSpot(TM) 64-Bit Server VM 1.8.0_66-b17 on Linux 2.6.32-504.el6.x86_64
Unknown processor
back-to-back filter: Best/Avg Time(ms) Rate(M/s) Per Row(ns) Relative
-------------------------------------------------------------------------------------------
Dataset 472 / 503 21.2 47.2 1.0X
DataFrame 51 / 58 198.0 5.1 9.3X
RDD 189 / 211 52.8 18.9 2.5X
````
After applying this PR
```
Java HotSpot(TM) 64-Bit Server VM 1.8.0_66-b17 on Linux 2.6.32-504.el6.x86_64
Intel(R) Xeon(R) CPU E5-2667 v2 3.30GHz
back-to-back filter: Best/Avg Time(ms) Rate(M/s) Per Row(ns) Relative
-------------------------------------------------------------------------------------------
Dataset 490 / 502 20.4 49.0 1.0X
DataFrame 55 / 61 183.4 5.5 9.0X
RDD 210 / 237 47.7 21.0 2.3X
```
## How was this patch tested?
Run Benchmark programs on Linux by hand
Author: Kazuaki Ishizaki <ishizaki@jp.ibm.com>
Closes#12411 from kiszk/SPARK-14656.
When `Await.result` throws an exception which originated from a different thread, the resulting stacktrace doesn't include the path leading to the `Await.result` call itself, making it difficult to identify the impact of these exceptions. For example, I've seen cases where broadcast cleaning errors propagate to the main thread and crash it but the resulting stacktrace doesn't include any of the main thread's code, making it difficult to pinpoint which exception crashed that thread.
This patch addresses this issue by explicitly catching, wrapping, and re-throwing exceptions that are thrown by `Await.result`.
I tested this manually using 16b31c8251, a patch which reproduces an issue where an RPC exception which occurs while unpersisting RDDs manages to crash the main thread without any useful stacktrace, and verified that informative, full stacktraces were generated after applying the fix in this PR.
/cc rxin nongli yhuai anabranch
Author: Josh Rosen <joshrosen@databricks.com>
Closes#12433 from JoshRosen/wrap-and-rethrow-await-exceptions.
## What changes were proposed in this pull request?
This PR adds exit code parameter to exitExecutor() so that caller can specify different exit code.
## How was this patch tested?
Existing test
rxin hbhanawat
Author: tedyu <yuzhihong@gmail.com>
Closes#12457 from tedyu/master.
## What changes were proposed in this pull request?
The sort shuffle manager has been the default since Spark 1.2. It is time to remove the old hash shuffle manager.
## How was this patch tested?
Removed some tests related to the old manager.
Author: Reynold Xin <rxin@databricks.com>
Closes#12423 from rxin/SPARK-14667.
https://issues.apache.org/jira/browse/SPARK-13227
It might confuse the future developers when they use OpenHashMap.apply() with a numeric value type.
null.asInstance[Int], null.asInstance[Long], null.asInstace[Float] and null.asInstance[Double] will return 0/0.0/0L, which might confuse the developer if the value set contains 0/0.0/0L with an existing key
The current patch only adds the comments describing the issue, with the respect to apply the minimum changes to the code base
The more direct, yet more aggressive, approach is use Option as the return type
andrewor14 JoshRosen any thoughts about how to avoid the potential issue?
Author: CodingCat <zhunansjtu@gmail.com>
Closes#11107 from CodingCat/SPARK-13227.
## What changes were proposed in this pull request?
This PR is a follow up for https://github.com/apache/spark/pull/12417, now we always track input/output/shuffle metrics in spark JSON protocol and status API.
Most of the line changes are because of re-generating the gold answer for `HistoryServerSuite`, and we add a lot of 0 values for read/write metrics.
## How was this patch tested?
existing tests.
Author: Wenchen Fan <wenchen@databricks.com>
Closes#12462 from cloud-fan/follow.
## What changes were proposed in this pull request?
When there are multiple tests running, "NettyBlockTransferServiceSuite.can bind to a specific port twice and the second increments" may fail.
E.g., assume there are 2 tests running. Here are the execution order to reproduce the test failure.
| Execution Order | Test 1 | Test 2 |
| ------------- | ------------- | ------------- |
| 1 | service0 binds to 17634 | |
| 2 | | service0 binds to 17635 (17634 is occupied) |
| 3 | service1 binds to 17636 | |
| 4 | pass test | |
| 5 | service0.close (release 17634) | |
| 6 | | service1 binds to 17634 |
| 7 | | `service1.port should be (service0.port + 1)` fails (17634 != 17635 + 1) |
Here is an example in Jenkins: https://amplab.cs.berkeley.edu/jenkins/view/Spark%20QA%20Test/job/spark-master-test-maven-hadoop-2.2/786/testReport/junit/org.apache.spark.network.netty/NettyBlockTransferServiceSuite/can_bind_to_a_specific_port_twice_and_the_second_increments/
This PR makes two changes:
- Use a random port between 17634 and 27634 to reduce the possibility of port conflicts.
- Make `service1` use `service0.port` to bind to avoid the above race condition.
## How was this patch tested?
Jenkins unit tests.
Author: Shixiong Zhu <shixiong@databricks.com>
Closes#12477 from zsxwing/SPARK-14713.
## What changes were proposed in this pull request?
This commit adds support for pluggable cluster manager. And also allows a cluster manager to clean up tasks without taking the parent process down.
To plug a new external cluster manager, ExternalClusterManager trait should be implemented. It returns task scheduler and backend scheduler that will be used by SparkContext to schedule tasks. An external cluster manager is registered using the java.util.ServiceLoader mechanism (This mechanism is also being used to register data sources like parquet, json, jdbc etc.). This allows auto-loading implementations of ExternalClusterManager interface.
Currently, when a driver fails, executors exit using system.exit. This does not bode well for cluster managers that would like to reuse the parent process of an executor. Hence,
1. Moving system.exit to a function that can be overriden in subclasses of CoarseGrainedExecutorBackend.
2. Added functionality of killing all the running tasks in an executor.
## How was this patch tested?
ExternalClusterManagerSuite.scala was added to test this patch.
Author: Hemant Bhanawat <hemant@snappydata.io>
Closes#11723 from hbhanawat/pluggableScheduler.
## What changes were proposed in this pull request?
This PR removes
- Inappropriate type notations
For example, from
```scala
words.foreachRDD { (rdd: RDD[String], time: Time) =>
...
```
to
```scala
words.foreachRDD { (rdd, time) =>
...
```
- Extra anonymous closure within functional transformations.
For example,
```scala
.map(item => {
...
})
```
which can be just simply as below:
```scala
.map { item =>
...
}
```
and corrects some obvious style nits.
## How was this patch tested?
This was tested after adding rules in `scalastyle-config.xml`, which ended up with not finding all perfectly.
The rules applied were below:
- For the first correction,
```xml
<check customId="NoExtraClosure" level="error" class="org.scalastyle.file.RegexChecker" enabled="true">
<parameters><parameter name="regex">(?m)\.[a-zA-Z_][a-zA-Z0-9]*\(\s*[^,]+s*=>\s*\{[^\}]+\}\s*\)</parameter></parameters>
</check>
```
```xml
<check customId="NoExtraClosure" level="error" class="org.scalastyle.file.RegexChecker" enabled="true">
<parameters><parameter name="regex">\.[a-zA-Z_][a-zA-Z0-9]*\s*[\{|\(]([^\n>,]+=>)?\s*\{([^()]|(?R))*\}^[,]</parameter></parameters>
</check>
```
- For the second correction
```xml
<check customId="TypeNotation" level="error" class="org.scalastyle.file.RegexChecker" enabled="true">
<parameters><parameter name="regex">\.[a-zA-Z_][a-zA-Z0-9]*\s*[\{|\(]\s*\([^):]*:R))*\}^[,]</parameter></parameters>
</check>
```
**Those rules were not added**
Author: hyukjinkwon <gurwls223@gmail.com>
Closes#12413 from HyukjinKwon/SPARK-style.
## What changes were proposed in this pull request?
Part of the reason why TaskMetrics and its callers are complicated are due to the optional metrics we collect, including input, output, shuffle read, and shuffle write. I think we can always track them and just assign 0 as the initial values. It is usually very obvious whether a task is supposed to read any data or not. By always tracking them, we can remove a lot of map, foreach, flatMap, getOrElse(0L) calls throughout Spark.
This patch also changes a few behaviors.
1. Removed the distinction of data read/write methods (e.g. Hadoop, Memory, Network, etc).
2. Accumulate all data reads and writes, rather than only the first method. (Fixes SPARK-5225)
## How was this patch tested?
existing tests.
This is bases on https://github.com/apache/spark/pull/12388, with more test fixes.
Author: Reynold Xin <rxin@databricks.com>
Author: Wenchen Fan <wenchen@databricks.com>
Closes#12417 from cloud-fan/metrics-refactor.
## What changes were proposed in this pull request?
Round memory bytes and convert it to Long to it’s original type. This change fixes the formatting issue in the Exception message.
## How was this patch tested?
Manual tests were done in CDH cluster.
Author: Peter Ableda <peter.ableda@cloudera.com>
Closes#12392 from peterableda/SPARK-14633.
## What changes were proposed in this pull request?
Removing references to assembly jar in documentation.
Adding an additional (previously undocumented) usage of spark-submit to run examples.
## How was this patch tested?
Ran spark-submit usage to ensure formatting was fine. Ran examples using SparkSubmit.
Author: Mark Grover <mark@apache.org>
Closes#12365 from markgrover/spark-14601.
## What changes were proposed in this pull request?
When we clean a closure, if its outermost parent is not a closure, we won't clone and clean it as cloning user's objects is dangerous. However, if it's a REPL line object, which may carry a lot of unnecessary references(like hadoop conf, spark conf, etc.), we should clean it as it's not a user object.
This PR improves the check for user's objects to exclude REPL line object.
## How was this patch tested?
existing tests.
Author: Wenchen Fan <wenchen@databricks.com>
Closes#12327 from cloud-fan/closure.
## What changes were proposed in this pull request?
This patch removes some of the deprecated APIs in TaskMetrics. This is part of my bigger effort to simplify accumulators and task metrics.
## How was this patch tested?
N/A - only removals
Author: Reynold Xin <rxin@databricks.com>
Closes#12375 from rxin/SPARK-14617.
## What changes were proposed in this pull request?
When there are multiple attempts for a stage, we currently only reset internal accumulator values if all the tasks are resubmitted. It would make more sense to reset the accumulator values for each stage attempt. This will allow us to eventually get rid of the internal flag in the Accumulator class. This is part of my bigger effort to simplify accumulators and task metrics.
## How was this patch tested?
Covered by existing tests.
Author: Reynold Xin <rxin@databricks.com>
Closes#12378 from rxin/SPARK-14619.
## What changes were proposed in this pull request?
Move json4s, breeze dependency declaration into parent
## How was this patch tested?
Should be no functional change, but Jenkins tests will test that.
Author: Sean Owen <sowen@cloudera.com>
Closes#12390 from srowen/SPARK-14612.
## What changes were proposed in this pull request?
Currently many public abstract methods (in abstract classes as well as traits) don't declare return types explicitly, such as in [o.a.s.streaming.dstream.InputDStream](https://github.com/apache/spark/blob/master/streaming/src/main/scala/org/apache/spark/streaming/dstream/InputDStream.scala#L110):
```scala
def start() // should be: def start(): Unit
def stop() // should be: def stop(): Unit
```
These methods exist in core, sql, streaming; this PR fixes them.
## How was this patch tested?
N/A
## Which piece of scala style rule led to the changes?
the rule was added separately in https://github.com/apache/spark/pull/12396
Author: Liwei Lin <lwlin7@gmail.com>
Closes#12389 from lw-lin/public-abstract-methods.
## What changes were proposed in this pull request?
I was trying to understand the accumulator and metrics update source code and these two classes don't really need to be case classes. It would also be more consistent with other UI classes if they are not case classes. This is part of my bigger effort to simplify accumulators and task metrics.
## How was this patch tested?
This is a straightforward refactoring without behavior change.
Author: Reynold Xin <rxin@databricks.com>
Closes#12386 from rxin/SPARK-14625.
## What changes were proposed in this pull request?
This PR removes extra anonymous closure within functional transformations.
For example,
```scala
.map(item => {
...
})
```
which can be just simply as below:
```scala
.map { item =>
...
}
```
## How was this patch tested?
Related unit tests and `sbt scalastyle`.
Author: hyukjinkwon <gurwls223@gmail.com>
Closes#12382 from HyukjinKwon/minor-extra-closers.
## What changes were proposed in this pull request?
Old `HadoopFsRelation` API includes `buildInternalScan()` which uses `SqlNewHadoopRDD` in `ParquetRelation`.
Because now the old API is removed, `SqlNewHadoopRDD` is not used anymore.
So, this PR removes `SqlNewHadoopRDD` and several unused imports.
This was discussed in https://github.com/apache/spark/pull/12326.
## How was this patch tested?
Several related existing unit tests and `sbt scalastyle`.
Author: hyukjinkwon <gurwls223@gmail.com>
Closes#12354 from HyukjinKwon/SPARK-14596.
This patch makes the postStartHook throw an IllegalStateException if the SparkContext is shutdown while it is waiting for the backend to be ready
Author: Charles Allen <charles@allen-net.com>
Closes#12301 from drcrallen/SPARK-14537.
## What changes were proposed in this pull request?
- updated `OFF_HEAP` semantics for `StorageLevels.java`
- updated `OFF_HEAP` semantics for `storagelevel.py`
## How was this patch tested?
no need to test
Author: Liwei Lin <lwlin7@gmail.com>
Closes#12126 from lw-lin/storagelevel.py.
## What changes were proposed in this pull request?
Fix memory leak in the Sorter. When the UnsafeExternalSorter spills the data to disk, it does not free up the underlying pointer array. As a result, we see a lot of executor OOM and also memory under utilization.
This is a regression partially introduced in PR https://github.com/apache/spark/pull/9241
## How was this patch tested?
Tested by running a job and observed around 30% speedup after this change.
Author: Sital Kedia <skedia@fb.com>
Closes#12285 from sitalkedia/executor_oom.
## What changes were proposed in this pull request?
This PR improve the performance of SQL UI by:
1) remove the details column in all executions page (the first page in SQL tab). We can check the details by enter the execution page.
2) break-all is super slow in Chrome recently, so switch to break-word.
3) Using "display: none" to hide a block.
4) using one js closure for for all the executions, not one for each.
5) remove the height limitation of details, don't need to scroll it in the tiny window.
## How was this patch tested?
Exists tests.
Author: Davies Liu <davies@databricks.com>
Closes#12311 from davies/ui_perf.
## What changes were proposed in this pull request?
Shutting down `QueuedThreadPool` used by Jetty `Server` to avoid threads leakage after SparkContext is stopped.
Note: If this fix is going to apply to the `branch-1.6`, one more patch on the `NettyRpcEnv` class is needed so that the `NettyRpcEnv._fileServer.shutdown` is called in the `NettyRpcEnv.cleanup` method. This is due to the removal of `_fileServer` field in the `NettyRpcEnv` class in the master branch. Please advice if a second PR is necessary for bring this fix back to `branch-1.6`
## How was this patch tested?
Ran the ./dev/run-tests locally
Author: Terence Yim <terence@cask.co>
Closes#12318 from chtyim/fixes/SPARK-14513-thread-leak.
## What changes were proposed in this pull request?
According to the [Spark Code Style Guide](https://cwiki.apache.org/confluence/display/SPARK/Spark+Code+Style+Guide) and [Scala Style Guide](http://docs.scala-lang.org/style/control-structures.html#curlybraces), we had better enforce the following rule.
```
case: Always omit braces in case clauses.
```
This PR makes a new ScalaStyle rule, 'OmitBracesInCase', and enforces it to the code.
## How was this patch tested?
Pass the Jenkins tests (including Scala style checking)
Author: Dongjoon Hyun <dongjoon@apache.org>
Closes#12280 from dongjoon-hyun/SPARK-14508.
## What changes were proposed in this pull request?
This adds a new API call `TaskContext.getLocalProperty` for getting properties set in the driver from executors. These local properties are automatically propagated from the driver to executors. For streaming, the context for streaming tasks will be the initial driver context when ssc.start() is called.
## How was this patch tested?
Unit tests.
cc JoshRosen
Author: Eric Liang <ekl@databricks.com>
Closes#12248 from ericl/sc-2813.
## What changes were proposed in this pull request?
When deciding whether a CommitDeniedException caused a task to fail, consider the root cause of the Exception.
## How was this patch tested?
Added a test suite for the component that extracts the root cause of the error.
Made a distribution after cherry-picking this commit to branch-1.6 and used to run our Spark application that would quite often fail due to the CommitDeniedException.
Author: Jason Moore <jasonmoore2k@outlook.com>
Closes#12228 from jasonmoore2k/SPARK-14357.
## What changes were proposed in this pull request?
Currently, `checkstyle` is configured to check the files under `src/main/java`. However, Spark has Java files in `src/main/scala`, too. This PR fixes the following configuration in `pom.xml` and the unchecked-so-far violations on those files.
```xml
-<sourceDirectory>${basedir}/src/main/java</sourceDirectory>
+<sourceDirectories>${basedir}/src/main/java,${basedir}/src/main/scala</sourceDirectories>
```
## How was this patch tested?
After passing the Jenkins build and manually `dev/lint-java`. (Note that Jenkins does not run `lint-java`)
Author: Dongjoon Hyun <dongjoon@apache.org>
Closes#12242 from dongjoon-hyun/SPARK-14465.
## What changes were proposed in this pull request?
Currently, we use java HashMap for HashedRelation if the key could fit within a Long. The java HashMap and CompactBuffer are not memory efficient, the memory used by them is also accounted accurately.
This PR introduce a LongToUnsafeRowMap (similar to BytesToBytesMap) for better memory efficiency and performance.
This PR reopen#12190 to fix bugs.
## How was this patch tested?
Existing tests.
Author: Davies Liu <davies@databricks.com>
Closes#12278 from davies/long_map3.
## What changes were proposed in this pull request?
This patch adds support for better handling of exceptions inside catch blocks if the code within the block throws an exception. For instance here is the code in a catch block before this change in `WriterContainer.scala`:
```scala
logError("Aborting task.", cause)
// call failure callbacks first, so we could have a chance to cleanup the writer.
TaskContext.get().asInstanceOf[TaskContextImpl].markTaskFailed(cause)
if (currentWriter != null) {
currentWriter.close()
}
abortTask()
throw new SparkException("Task failed while writing rows.", cause)
```
If `markTaskFailed` or `currentWriter.close` throws an exception, we currently lose the original cause. This PR fixes this problem by implementing a utility function `Utils.tryWithSafeCatch` that suppresses (`Throwable.addSuppressed`) the exception that are thrown within the catch block and rethrowing the original exception.
## How was this patch tested?
No new functionality added
Author: Sameer Agarwal <sameer@databricks.com>
Closes#12234 from sameeragarwal/fix-exception.
## What changes were proposed in this pull request?
Here is why SPARK-14437 happens:
BlockManagerId is created using NettyBlockTransferService.hostName which comes from `customHostname`. And `Executor` will set `customHostname` to the hostname which is detected by the driver. However, the driver may not be able to detect the correct address in some complicated network (Netty's Channel.remoteAddress doesn't always return a connectable address). In such case, `BlockManagerId` will be created using a wrong hostname.
To fix this issue, this PR uses `hostname` provided by `SparkEnv.create` to create `NettyBlockTransferService` and set `NettyBlockTransferService.hostname` to this one directly. A bonus of this approach is NettyBlockTransferService won't bound to `0.0.0.0` which is much safer.
## How was this patch tested?
Manually checked the bound address using local-cluster.
Author: Shixiong Zhu <shixiong@databricks.com>
Closes#12240 from zsxwing/SPARK-14437.
Currently all `SparkFirehoseListener` implementations are broken since we expect listeners to extend `SparkListener`, while the fire hose only extends `SparkListenerInterface`. This changes the addListener function and the config based injection to use the interface instead.
The existing tests in SparkListenerSuite are improved such that they would have caught this.
Follow-up to #12142
Author: Michael Armbrust <michael@databricks.com>
Closes#12227 from marmbrus/fixListener.
## What changes were proposed in this pull request?
`OutputCommitCoordinator` was introduced to deal with concurrent task attempts racing to write output, leading to data loss or corruption. For more detail, read the [JIRA description](https://issues.apache.org/jira/browse/SPARK-14468).
Before: `OutputCommitCoordinator` is enabled only if speculation is enabled.
After: `OutputCommitCoordinator` is always enabled.
Users may still disable this through `spark.hadoop.outputCommitCoordination.enabled`, but they really shouldn't...
## How was this patch tested?
`OutputCommitCoordinator*Suite`
Author: Andrew Or <andrew@databricks.com>
Closes#12244 from andrewor14/always-occ.
## What changes were proposed in this pull request?
Currently Spark clients are started with the same memory setting for Xms and Xms leading to reserving unnecessary higher amounts of memory.
This behavior is changed and the clients can now specify an initial heap size using the extraJavaOptions in the config for driver,executor and am individually.
Note, that only -Xms can be provided through this config option, if the client wants to set the max size(-Xmx), this has to be done via the *.memory configuration knobs which are currently supported.
## How was this patch tested?
Monitored executor and yarn logs in debug mode to verify the commands through which they are being launched in client and cluster mode. The driver memory was verified locally using jps -v. Setting up -Xmx parameter in the javaExtraOptions raises exception with the info provided.
Author: Dhruve Ashar <dhruveashar@gmail.com>
Closes#12115 from dhruve/impr/SPARK-12384.
## What changes were proposed in this pull request?
The Spark UI (both active and history) should show the user who ran the application somewhere when you are in the application view. This was added under the Jobs view by total uptime and scheduler mode.
## How was this patch tested?
Manual testing
<img width="191" alt="username" src="https://cloud.githubusercontent.com/assets/13952758/14222830/6d1fe542-f82a-11e5-885f-c05ee2cdf857.png">
Author: Alex Bozarth <ajbozart@us.ibm.com>
Closes#12123 from ajbozarth/spark14245.
The current package name uses a dash, which is a little weird but seemed
to work. That is, until a new test tried to mock a class that references
one of those shaded types, and then things started failing.
Most changes are just noise to fix the logging configs.
For reference, SPARK-8815 also raised this issue, although at the time it
did not cause any issues in Spark, so it was not addressed.
Author: Marcelo Vanzin <vanzin@cloudera.com>
Closes#11941 from vanzin/SPARK-14134.
## What changes were proposed in this pull request?
Send `RegisterExecutorResponse` using `executorRef` in order to make sure RegisterExecutorResponse and LaunchTask are both sent using the same channel. Then RegisterExecutorResponse will always arrive before LaunchTask
## How was this patch tested?
Existing unit tests
Closes#12078
Author: Shixiong Zhu <shixiong@databricks.com>
Closes#12211 from zsxwing/SPARK-13112.
## What changes were proposed in this pull request?
According to the [Spark Code Style Guide](https://cwiki.apache.org/confluence/display/SPARK/Spark+Code+Style+Guide#SparkCodeStyleGuide-Indentation), this PR adds a new scalastyle rule to prevent the followings.
```
/** In Spark, we don't use the ScalaDoc style so this
* is not correct.
*/
```
## How was this patch tested?
Pass the Jenkins tests (including `lint-scala`).
Author: Dongjoon Hyun <dongjoon@apache.org>
Closes#12221 from dongjoon-hyun/SPARK-14444.
## What changes were proposed in this pull request?
Added a new Executor Allocation Manager for the Streaming scheduler for doing Streaming Dynamic Allocation.
## How was this patch tested
Unit tests, and cluster tests.
Author: Tathagata Das <tathagata.das1565@gmail.com>
Closes#12154 from tdas/streaming-dynamic-allocation.
## What changes were proposed in this pull request?
As mentioned in the ticket this was because one get path in the refactored `BlockManager` did not check for remote storage.
## How was this patch tested?
Unit test, also verified manually with reproduction in the ticket.
cc JoshRosen
Author: Eric Liang <ekl@databricks.com>
Closes#12193 from ericl/spark-14252.
## What changes were proposed in this pull request?
While I was reviewing #12078, I found most of CoarseGrainedSchedulerBackend's mutable fields doesn't have any comments about the thread-safe assumptions and it's hard for people to figure out which part of codes should be protected by the lock. This PR just added comments/annotations for them and also added strict access modifiers for some fields.
## How was this patch tested?
Existing unit tests.
Author: Shixiong Zhu <shixiong@databricks.com>
Closes#12188 from zsxwing/comments.
Because SQL keeps track of all known configs, some customization was
needed in SQLConf to allow that, since the core API does not have that
feature.
Tested via existing (and slightly updated) unit tests.
Author: Marcelo Vanzin <vanzin@cloudera.com>
Closes#11570 from vanzin/SPARK-529-sql.
## What changes were proposed in this pull request?
In `LogPage`, the content to be rendered is defined as follows.
```
val content =
<html>
<body>
{linkToMaster}
<div>
<div style="float:left; margin-right:10px">{backButton}</div>
<div style="float:left;">{range}</div>
<div style="float:right; margin-left:10px">{nextButton}</div>
</div>
<br />
<div style="height:500px; overflow:auto; padding:5px;">
<pre>{logText}</pre>
</div>
</body>
</html>
UIUtils.basicSparkPage(content, logType + " log page for " + pageName)
```
As you can see, <html> and <body> tags will be rendered.
On the other hand, `UIUtils.basicSparkPage` will render those tags so those tags will be nested.
```
def basicSparkPage(
content: => Seq[Node],
title: String,
useDataTables: Boolean = false): Seq[Node] = {
<html>
<head>
{commonHeaderNodes}
{if (useDataTables) dataTablesHeaderNodes else Seq.empty}
<title>{title}</title>
</head>
<body>
<div class="container-fluid">
<div class="row-fluid">
<div class="span12">
<h3 style="vertical-align: middle; display: inline-block;">
<a style="text-decoration: none" href={prependBaseUri("/")}>
<img src={prependBaseUri("/static/spark-logo-77x50px-hd.png")} />
<span class="version"
style="margin-right: 15px;">{org.apache.spark.SPARK_VERSION}</span>
</a>
{title}
</h3>
</div>
</div>
{content}
</div>
</body>
</html>
}
```
These are the screen shots before this patch is applied.
And these are the ones after this patch is applied.
The appearance is not changed but the html source code is changed.
## How was this patch tested?
Manually run some jobs on my standalone-cluster and check the WebUI.
Author: Kousuke Saruta <sarutak@oss.nttdata.co.jp>
Closes#12170 from sarutak/SPARK-14397.
Use PartitionerAwareUnionRDD when possbile for optimizing shuffling and
preserving the partitioner.
Author: Guillaume Poulin <poulin.guillaume@gmail.com>
Closes#10382 from gpoulin/dstream_union_optimisation.
This change modifies the "assembly/" module to just copy needed
dependencies to its build directory, and modifies the packaging
script to pick those up (and remove duplicate jars packages in the
examples module).
I also made some minor adjustments to dependencies to remove some
test jars from the final packaging, and remove jars that conflict with each
other when packaged separately (e.g. servlet api).
Also note that this change restores guava in applications' classpaths, even
though it's still shaded inside Spark. This is now needed for the Hadoop
libraries that are packaged with Spark, which now are not processed by
the shade plugin.
Author: Marcelo Vanzin <vanzin@cloudera.com>
Closes#11796 from vanzin/SPARK-13579.
## What changes were proposed in this pull request?
RDD.toLocalIterator() could be used to fetch one partition at a time to reduce the memory usage. Right now, for Dataset/Dataframe we have to use df.rdd.toLocalIterator, which is super slow also requires lots of memory (because of the Java serializer or even Kyro serializer).
This PR introduce an optimized toLocalIterator for Dataset/DataFrame, which is much faster and requires much less memory. For a partition with 5 millions rows, `df.rdd.toIterator` took about 100 seconds, but df.toIterator took less than 7 seconds. For 10 millions row, rdd.toIterator will crash (not enough memory) with 4G heap, but df.toLocalIterator could finished in 12 seconds.
The JDBC server has been updated to use DataFrame.toIterator.
## How was this patch tested?
Existing tests.
Author: Davies Liu <davies@databricks.com>
Closes#12114 from davies/local_iterator.
## What changes were proposed in this pull request?
Scala traits are difficult to maintain binary compatibility on, and as a result we had to introduce JavaSparkListener. In Spark 2.0 we can change SparkListener from a trait to an abstract class and then remove JavaSparkListener.
## How was this patch tested?
Updated related unit tests.
Author: Reynold Xin <rxin@databricks.com>
Closes#12142 from rxin/SPARK-14358.
## What changes were proposed in this pull request?
It's a mistake that HeartbeatReceiver object was made public in Spark 1.x.
## How was this patch tested?
N/A
Author: Reynold Xin <rxin@databricks.com>
Closes#12148 from rxin/SPARK-14364.
## What changes were proposed in this pull request?
This PR contains the following 5 types of maintenance fix over 59 files (+94 lines, -93 lines).
- Fix typos(exception/log strings, testcase name, comments) in 44 lines.
- Fix lint-java errors (MaxLineLength) in 6 lines. (New codes after SPARK-14011)
- Use diamond operators in 40 lines. (New codes after SPARK-13702)
- Fix redundant semicolon in 5 lines.
- Rename class `InferSchemaSuite` to `CSVInferSchemaSuite` in CSVInferSchemaSuite.scala.
## How was this patch tested?
Manual and pass the Jenkins tests.
Author: Dongjoon Hyun <dongjoon@apache.org>
Closes#12139 from dongjoon-hyun/SPARK-14355.
## What changes were proposed in this pull request?
This special cases 0 and 1 counts to avoid passing 0 degrees of freedom.
## How was this patch tested?
Tests run successfully. New test added.
## Note:
This recreates #11982 which was closed to due to non-updated diff. rxin srowen Commented there.
This also adds tests, reworks the code to perform the special casing (based on srowen's comments), and adds equality machinery for BoundedDouble, as well as changing how it is transformed to string.
Author: Marcin Tustin <mtustin@handybook.com>
Author: Marcin Tustin <mtustin@handy.com>
Closes#12016 from mtustin-handy/SPARK-14163.
## What changes were proposed in this pull request?
Appends s3 specific configurations and spark.hadoop configurations to hive configuration.
## How was this patch tested?
Tested by running a job on cluster.
…figurations to hive configuration.
Author: Sital Kedia <skedia@fb.com>
Closes#11876 from sitalkedia/hiveConf.
## What changes were proposed in this pull request?
Straggler references to Tachyon were removed:
- for docs, `tachyon` has been generalized as `off-heap memory`;
- for Mesos test suits, the key-value `tachyon:true`/`tachyon:false` has been changed to `os:centos`/`os:ubuntu`, since `os` is an example constrain used by the [Mesos official docs](http://mesos.apache.org/documentation/attributes-resources/).
## How was this patch tested?
Existing test suites.
Author: Liwei Lin <lwlin7@gmail.com>
Closes#12129 from lw-lin/tachyon-cleanup.
## What changes were proposed in this pull request?
This PR aims to fix all Scala-Style multiline comments into Java-Style multiline comments in Scala codes.
(All comment-only changes over 77 files: +786 lines, −747 lines)
## How was this patch tested?
Manual.
Author: Dongjoon Hyun <dongjoon@apache.org>
Closes#12130 from dongjoon-hyun/use_multiine_javadoc_comments.
## What changes were proposed in this pull request?
Adding long values for each Date in the ApplicationAttemptInfo API for easier use in code
## How was the this patch tested?
Tested with dev/run-tests
Author: Alex Bozarth <ajbozart@us.ibm.com>
Closes#11326 from ajbozarth/spark13241.
This patch adds support for caching blocks in the executor processes using direct / off-heap memory.
## User-facing changes
**Updated semantics of `OFF_HEAP` storage level**: In Spark 1.x, the `OFF_HEAP` storage level indicated that an RDD should be cached in Tachyon. Spark 2.x removed the external block store API that Tachyon caching was based on (see #10752 / SPARK-12667), so `OFF_HEAP` became an alias for `MEMORY_ONLY_SER`. As of this patch, `OFF_HEAP` means "serialized and cached in off-heap memory or on disk". Via the `StorageLevel` constructor, `useOffHeap` can be set if `serialized == true` and can be used to construct custom storage levels which support replication.
**Storage UI reporting**: the storage UI will now report whether in-memory blocks are stored on- or off-heap.
**Only supported by UnifiedMemoryManager**: for simplicity, this feature is only supported when the default UnifiedMemoryManager is used; applications which use the legacy memory manager (`spark.memory.useLegacyMode=true`) are not currently able to allocate off-heap storage memory, so using off-heap caching will fail with an error when legacy memory management is enabled. Given that we plan to eventually remove the legacy memory manager, this is not a significant restriction.
**Memory management policies:** the policies for dividing available memory between execution and storage are the same for both on- and off-heap memory. For off-heap memory, the total amount of memory available for use by Spark is controlled by `spark.memory.offHeap.size`, which is an absolute size. Off-heap storage memory obeys `spark.memory.storageFraction` in order to control the amount of unevictable storage memory. For example, if `spark.memory.offHeap.size` is 1 gigabyte and Spark uses the default `storageFraction` of 0.5, then up to 500 megabytes of off-heap cached blocks will be protected from eviction due to execution memory pressure. If necessary, we can split `spark.memory.storageFraction` into separate on- and off-heap configurations, but this doesn't seem necessary now and can be done later without any breaking changes.
**Use of off-heap memory does not imply use of off-heap execution (or vice-versa)**: for now, the settings controlling the use of off-heap execution memory (`spark.memory.offHeap.enabled`) and off-heap caching are completely independent, so Spark SQL can be configured to use off-heap memory for execution while continuing to cache blocks on-heap. If desired, we can change this in a followup patch so that `spark.memory.offHeap.enabled` affect the default storage level for cached SQL tables.
## Internal changes
- Rename `ByteArrayChunkOutputStream` to `ChunkedByteBufferOutputStream`
- It now returns a `ChunkedByteBuffer` instead of an array of byte arrays.
- Its constructor now accept an `allocator` function which is called to allocate `ByteBuffer`s. This allows us to control whether it allocates regular ByteBuffers or off-heap DirectByteBuffers.
- Because block serialization is now performed during the unroll process, a `ChunkedByteBufferOutputStream` which is configured with a `DirectByteBuffer` allocator will use off-heap memory for both unroll and storage memory.
- The `MemoryStore`'s MemoryEntries now tracks whether blocks are stored on- or off-heap.
- `evictBlocksToFreeSpace()` now accepts a `MemoryMode` parameter so that we don't try to evict off-heap blocks in response to on-heap memory pressure (or vice-versa).
- Make sure that off-heap buffers are properly de-allocated during MemoryStore eviction.
- The JVM limits the total size of allocated direct byte buffers using the `-XX:MaxDirectMemorySize` flag and the default tends to be fairly low (< 512 megabytes in some JVMs). To work around this limitation, this patch adds a custom DirectByteBuffer allocator which ignores this memory limit.
Author: Josh Rosen <joshrosen@databricks.com>
Closes#11805 from JoshRosen/off-heap-caching.
Currently, when max number of executor failures reached the `maxNumExecutorFailures`, `ApplicationMaster` will be killed and re-register another one.This time, `YarnAllocator` will be created a new instance.
But, the value of property `executorIdCounter` in `YarnAllocator` will reset to `0`. Then the Id of new executor will starting from `1`. This will confuse with the executor has already created before, which will cause FetchFailedException.
This situation is just in yarn client mode, so this is an issue in yarn client mode. For more details, [link to jira issues SPARK-12864](https://issues.apache.org/jira/browse/SPARK-12864)
This PR introduce a mechanism to initialize `executorIdCounter` after `ApplicationMaster` killed.
Author: zhonghaihua <793507405@qq.com>
Closes#10794 from zhonghaihua/initExecutorIdCounterAfterAMKilled.
JIRA: https://issues.apache.org/jira/browse/SPARK-13674
## What changes were proposed in this pull request?
Sample operator doesn't support wholestage codegen now. This pr is to add support to it.
## How was this patch tested?
A test is added into `BenchmarkWholeStageCodegen`. Besides, all tests should be passed.
Author: Liang-Chi Hsieh <simonh@tw.ibm.com>
Author: Liang-Chi Hsieh <viirya@gmail.com>
Closes#11517 from viirya/add-wholestage-sample.
## What changes were proposed in this pull request?
Currently in Spark on YARN, configurations can be passed through SparkConf, env and command arguments, some parts are duplicated, like client argument and SparkConf. So here propose to simplify the command arguments.
## How was this patch tested?
This patch is tested manually with unit test.
CC vanzin tgravescs , please help to suggest this proposal. The original purpose of this JIRA is to remove `ClientArguments`, through refactoring some arguments like `--class`, `--arg` are not so easy to replace, so here I remove the most part of command line arguments, only keep the minimal set.
Author: jerryshao <sshao@hortonworks.com>
Closes#11603 from jerryshao/SPARK-12343.
## What changes were proposed in this pull request?
This PR support multiple Python UDFs within single batch, also improve the performance.
```python
>>> from pyspark.sql.types import IntegerType
>>> sqlContext.registerFunction("double", lambda x: x * 2, IntegerType())
>>> sqlContext.registerFunction("add", lambda x, y: x + y, IntegerType())
>>> sqlContext.sql("SELECT double(add(1, 2)), add(double(2), 1)").explain(True)
== Parsed Logical Plan ==
'Project [unresolvedalias('double('add(1, 2)), None),unresolvedalias('add('double(2), 1), None)]
+- OneRowRelation$
== Analyzed Logical Plan ==
double(add(1, 2)): int, add(double(2), 1): int
Project [double(add(1, 2))#14,add(double(2), 1)#15]
+- Project [double(add(1, 2))#14,add(double(2), 1)#15]
+- Project [pythonUDF0#16 AS double(add(1, 2))#14,pythonUDF0#18 AS add(double(2), 1)#15]
+- EvaluatePython [add(pythonUDF1#17, 1)], [pythonUDF0#18]
+- EvaluatePython [double(add(1, 2)),double(2)], [pythonUDF0#16,pythonUDF1#17]
+- OneRowRelation$
== Optimized Logical Plan ==
Project [pythonUDF0#16 AS double(add(1, 2))#14,pythonUDF0#18 AS add(double(2), 1)#15]
+- EvaluatePython [add(pythonUDF1#17, 1)], [pythonUDF0#18]
+- EvaluatePython [double(add(1, 2)),double(2)], [pythonUDF0#16,pythonUDF1#17]
+- OneRowRelation$
== Physical Plan ==
WholeStageCodegen
: +- Project [pythonUDF0#16 AS double(add(1, 2))#14,pythonUDF0#18 AS add(double(2), 1)#15]
: +- INPUT
+- !BatchPythonEvaluation [add(pythonUDF1#17, 1)], [pythonUDF0#16,pythonUDF1#17,pythonUDF0#18]
+- !BatchPythonEvaluation [double(add(1, 2)),double(2)], [pythonUDF0#16,pythonUDF1#17]
+- Scan OneRowRelation[]
```
## How was this patch tested?
Added new tests.
Using the following script to benchmark 1, 2 and 3 udfs,
```
df = sqlContext.range(1, 1 << 23, 1, 4)
double = F.udf(lambda x: x * 2, LongType())
print df.select(double(df.id)).count()
print df.select(double(df.id), double(df.id + 1)).count()
print df.select(double(df.id), double(df.id + 1), double(df.id + 2)).count()
```
Here is the results:
N | Before | After | speed up
---- |------------ | -------------|------
1 | 22 s | 7 s | 3.1X
2 | 38 s | 13 s | 2.9X
3 | 58 s | 16 s | 3.6X
This benchmark ran locally with 4 CPUs. For 3 UDFs, it launched 12 Python before before this patch, 4 process after this patch. After this patch, it will use less memory for multiple UDFs than before (less buffering).
Author: Davies Liu <davies@databricks.com>
Closes#12057 from davies/multi_udfs.
## What changes were proposed in this pull request?
Track executor information like host and port, cache size, running tasks.
TODO: tests
## How was this patch tested?
N/A
Author: Wenchen Fan <wenchen@databricks.com>
Closes#11888 from cloud-fan/status-tracker.
## What changes were proposed in this pull request?
This PR try to use `incUpdatedBlockStatuses ` to update the `updatedBlockStatuses ` when removing blocks, making sure `BlockManager` correctly updates `updatedBlockStatuses`
## How was this patch tested?
test("updated block statuses") in BlockManagerSuite.scala
Author: jeanlyn <jeanlyn92@gmail.com>
Closes#12091 from jeanlyn/updateBlock.
## What changes were proposed in this pull request?
Redirect error message to logWarning
## How was this patch tested?
Unit tests, manual tests
JoshRosen
Author: Nishkam Ravi <nishkamravi@gmail.com>
Closes#12052 from nishkamravi2/master_warning.
## What changes were proposed in this pull request?
For MemoryMode.OFF_HEAP, Unsafe.getInt etc. are used with no restriction.
However, the Oracle implementation uses these methods only if the class variable unaligned (commented as "Cached unaligned-access capability") is true, which seems to be calculated whether the architecture is i386, x86, amd64, or x86_64.
I think we should perform similar check for the use of Unsafe.
Reference: https://github.com/netty/netty/blob/4.1/common/src/main/java/io/netty/util/internal/PlatformDependent0.java#L112
## How was this patch tested?
Unit test suite
Author: tedyu <yuzhihong@gmail.com>
Closes#11943 from tedyu/master.
## What changes were proposed in this pull request?
This PR brings the support for chained Python UDFs, for example
```sql
select udf1(udf2(a))
select udf1(udf2(a) + 3)
select udf1(udf2(a) + udf3(b))
```
Also directly chained unary Python UDFs are put in single batch of Python UDFs, others may require multiple batches.
For example,
```python
>>> sqlContext.sql("select double(double(1))").explain()
== Physical Plan ==
WholeStageCodegen
: +- Project [pythonUDF#10 AS double(double(1))#9]
: +- INPUT
+- !BatchPythonEvaluation double(double(1)), [pythonUDF#10]
+- Scan OneRowRelation[]
>>> sqlContext.sql("select double(double(1) + double(2))").explain()
== Physical Plan ==
WholeStageCodegen
: +- Project [pythonUDF#19 AS double((double(1) + double(2)))#16]
: +- INPUT
+- !BatchPythonEvaluation double((pythonUDF#17 + pythonUDF#18)), [pythonUDF#17,pythonUDF#18,pythonUDF#19]
+- !BatchPythonEvaluation double(2), [pythonUDF#17,pythonUDF#18]
+- !BatchPythonEvaluation double(1), [pythonUDF#17]
+- Scan OneRowRelation[]
```
TODO: will support multiple unrelated Python UDFs in one batch (another PR).
## How was this patch tested?
Added new unit tests for chained UDFs.
Author: Davies Liu <davies@databricks.com>
Closes#12014 from davies/py_udfs.
Add a new api endpoint `/api/v1/version` to retrieve various version info. This PR only adds support for finding the current spark version, however other version info such as jvm or scala versions can easily be added.
Author: Jakob Odersky <jodersky@gmail.com>
Closes#10760 from jodersky/SPARK-10570.
## What changes were proposed in this pull request?
The event timeline doesn't show on job page if an executor is removed with a multiple line reason. This PR replaces all new line characters in the reason string with spaces.
## How was this patch tested?
Verified on the Web UI.
Author: Carson Wang <carson.wang@intel.com>
Closes#12029 from carsonwang/eventTimeline.
## What changes were proposed in this pull request?
Refactor RRDD by separating the common logic interacting with the R worker to a new class RRunner, which can be used to evaluate R UDFs.
Now RRDD relies on RRuner for RDD computation and RRDD could be reomved if we want to remove RDD API in SparkR later.
## How was this patch tested?
dev/lint-r
SparkR unit tests
Author: Sun Rui <rui.sun@intel.com>
Closes#12024 from sun-rui/SPARK-12792_new.
## What changes were proposed in this pull request?
This is a follow-up fix of #9963, in #9963 we handle this stale states clean-up work only for dynamic allocation enabled scenario. Here we should also clean the states in `CoarseGrainedSchedulerBackend` for dynamic allocation disabled scenario.
Please review, CC andrewor14 lianhuiwang , thanks a lot.
## How was this patch tested?
Run the unit test locally, also with integration test manually.
Author: jerryshao <sshao@hortonworks.com>
Closes#11366 from jerryshao/SPARK-13447.
## What changes were proposed in this pull request?
We have a streaming job using `FlumePollInputStream` always driver OOM after few days, here is some driver heap dump before OOM
```
num #instances #bytes class name
----------------------------------------------
1: 13845916 553836640 org.apache.spark.storage.BlockStatus
2: 14020324 336487776 org.apache.spark.storage.StreamBlockId
3: 13883881 333213144 scala.collection.mutable.DefaultEntry
4: 8907 89043952 [Lscala.collection.mutable.HashEntry;
5: 62360 65107352 [B
6: 163368 24453904 [Ljava.lang.Object;
7: 293651 20342664 [C
...
```
`BlockStatus` and `StreamBlockId` keep on growing, and the driver OOM in the end.
After investigated, i found the `executorIdToStorageStatus` in `StorageStatusListener` seems never remove the blocks from `StorageStatus`.
In order to fix the issue, i try to use `onBlockUpdated` replace `onTaskEnd ` , so we can update the block informations(add blocks, drop the block from memory to disk and delete the blocks) in time.
## How was this patch tested?
Existing unit tests and manual tests
Author: jeanlyn <jeanlyn92@gmail.com>
Closes#11779 from jeanlyn/fix_driver_oom.
## What changes were proposed in this pull request?
Extract the workaround for HADOOP-10622 introduced by #11940 into UninterruptibleThread so that we can test and reuse it.
## How was this patch tested?
Unit tests
Author: Shixiong Zhu <shixiong@databricks.com>
Closes#11971 from zsxwing/uninterrupt.
## What changes were proposed in this pull request?
Call `executor.stop` in a new thread to eliminate deadlock.
## How was this patch tested?
Existing unit tests
Author: Shixiong Zhu <shixiong@databricks.com>
Closes#12012 from zsxwing/SPARK-14180.
## What changes were proposed in this pull request?
Currently, for the key that can not fit within a long, we build a hash map for UnsafeHashedRelation, it's converted to BytesToBytesMap after serialization and deserialization. We should build a BytesToBytesMap directly to have better memory efficiency.
In order to do that, BytesToBytesMap should support multiple (K,V) pair with the same K, Location.putNewKey() is renamed to Location.append(), which could append multiple values for the same key (same Location). `Location.newValue()` is added to find the next value for the same key.
## How was this patch tested?
Existing tests. Added benchmark for broadcast hash join with duplicated keys.
Author: Davies Liu <davies@databricks.com>
Closes#11870 from davies/map2.
Refactor RRDD by separating the common logic interacting with the R worker to a new class RRunner, which can be used to evaluate R UDFs.
Now RRDD relies on RRuner for RDD computation and RRDD could be reomved if we want to remove RDD API in SparkR later.
Author: Sun Rui <rui.sun@intel.com>
Closes#10947 from sun-rui/SPARK-12792.
JIRA: https://issues.apache.org/jira/browse/SPARK-13742
## What changes were proposed in this pull request?
`RandomSampler.sample` currently accepts iterator as input and output another iterator. This makes it inappropriate to use in wholestage codegen of `Sampler` operator #11517. This change is to add non-iterator interface to `RandomSampler`.
This change adds a new method `def sample(): Int` to the trait `RandomSampler`. As we don't need to know the actual values of the sampling items, so this new method takes no arguments.
This method will decide whether to sample the next item or not. It returns how many times the next item will be sampled.
For `BernoulliSampler` and `BernoulliCellSampler`, the returned sampling times can only be 0 or 1. It simply means whether to sample the next item or not.
For `PoissonSampler`, the returned value can be more than 1, meaning the next item will be sampled multiple times.
## How was this patch tested?
Tests are added into `RandomSamplerSuite`.
Author: Liang-Chi Hsieh <simonh@tw.ibm.com>
Author: Liang-Chi Hsieh <viirya@appier.com>
Author: Liang-Chi Hsieh <viirya@gmail.com>
Closes#11578 from viirya/random-sampler-no-iterator.
