The "getPersistentRDDs()" is a useful API of SparkContext to get cached RDDs. However, the JavaSparkContext does not have this API.
Add a simple getPersistentRDDs() to get java.util.Map<Integer, JavaRDD> for Java users.
Author: Junyang <fly.shenjy@gmail.com>
Closes#10978 from flyjy/master.
Remove spark.closure.serializer option and use JavaSerializer always
CC andrewor14 rxin I see there's a discussion in the JIRA but just thought I'd offer this for a look at what the change would be.
Author: Sean Owen <sowen@cloudera.com>
Closes#11150 from srowen/SPARK-12414.
This is the next iteration of tnachen's previous PR: https://github.com/apache/spark/pull/4027
In that PR, we resolved with andrewor14 and pwendell to implement the Mesos scheduler's support of `spark.executor.cores` to be consistent with YARN and Standalone. This PR implements that resolution.
This PR implements two high-level features. These two features are co-dependent, so they're implemented both here:
- Mesos support for spark.executor.cores
- Multiple executors per slave
We at Mesosphere have been working with Typesafe on a Spark/Mesos integration test suite: https://github.com/typesafehub/mesos-spark-integration-tests, which passes for this PR.
The contribution is my original work and I license the work to the project under the project's open source license.
Author: Michael Gummelt <mgummelt@mesosphere.io>
Closes#10993 from mgummelt/executor_sizing.
This PR improve the lookup of BytesToBytesMap by:
1. Generate code for calculate the hash code of grouping keys.
2. Do not use MemoryLocation, fetch the baseObject and offset for key and value directly (remove the indirection).
Author: Davies Liu <davies@databricks.com>
Closes#11010 from davies/gen_map.
Additional changes to #10835, mainly related to style and visibility. This patch also adds back a few deprecated methods for backward compatibility.
Author: Andrew Or <andrew@databricks.com>
Closes#10958 from andrewor14/task-metrics-to-accums-followups.
There is a bug when we try to grow the buffer, OOM is ignore wrongly (the assert also skipped by JVM), then we try grow the array again, this one will trigger spilling free the current page, the current record we inserted will be invalid.
The root cause is that JVM has less free memory than MemoryManager thought, it will OOM when allocate a page without trigger spilling. We should catch the OOM, and acquire memory again to trigger spilling.
And also, we could not grow the array in `insertRecord` of `InMemorySorter` (it was there just for easy testing).
Author: Davies Liu <davies@databricks.com>
Closes#11095 from davies/fix_expand.
Trivial search-and-replace to eliminate deprecation warnings in Scala 2.11.
Also works with 2.10
Author: Jakob Odersky <jakob@odersky.com>
Closes#11085 from jodersky/SPARK-13171.
These were ignored because they are incorrectly written; they don't actually trigger stage retries, which is what the tests are testing. These tests are now rewritten to induce stage retries through fetch failures.
Note: there were 2 tests before and now there's only 1. What happened? It turns out that the case where we only resubmit a subset of of the original missing partitions is very difficult to simulate in tests without potentially introducing flakiness. This is because the `DAGScheduler` removes all map outputs associated with a given executor when this happens, and we will need multiple executors to trigger this case, and sometimes the scheduler still removes map outputs from all executors.
Author: Andrew Or <andrew@databricks.com>
Closes#10969 from andrewor14/unignore-accum-test.
Currently the Master would always set an application's initial executor limit to infinity. If the user specified `spark.dynamicAllocation.initialExecutors`, the config would not take effect. This is similar to #11047 but for standalone mode.
Author: Andrew Or <andrew@databricks.com>
Closes#11054 from andrewor14/standalone-da-initial.
Building with scala 2.11 results in the warning trait SynchronizedBuffer in package mutable is deprecated: Synchronization via traits is deprecated as it is inherently unreliable. Consider java.util.concurrent.ConcurrentLinkedQueue as an alternative. Investigation shows we are already using ConcurrentLinkedQueue in other locations so switch our uses of SynchronizedBuffer to ConcurrentLinkedQueue.
Author: Holden Karau <holden@us.ibm.com>
Closes#11059 from holdenk/SPARK-13164-replace-deprecated-synchronized-buffer-in-core.
Add a local property to indicate if checkpointing all RDDs that are marked with the checkpoint flag, and enable it in Streaming
Author: Shixiong Zhu <shixiong@databricks.com>
Closes#10934 from zsxwing/recursive-checkpoint.
This is an existing issue uncovered recently by #10835. The reason for the exception was because the `SQLHistoryListener` gets all sorts of accumulators, not just the ones that represent SQL metrics. For example, the listener gets the `internal.metrics.shuffleRead.remoteBlocksFetched`, which is an Int, then it proceeds to cast the Int to a Long, which fails.
The fix is to mark accumulators representing SQL metrics using some internal metadata. Then we can identify which ones are SQL metrics and only process those in the `SQLHistoryListener`.
Author: Andrew Or <andrew@databricks.com>
Closes#10971 from andrewor14/fix-sql-history.
[SPARK-10873] Support column sort and search for History Server using jQuery DataTable and REST API. Before this commit, the history server was generated hard-coded html and can not support search, also, the sorting was disabled if there is any application that has more than one attempt. Supporting search and sort (over all applications rather than the 20 entries in the current page) in any case will greatly improve user experience.
1. Create the historypage-template.html for displaying application information in datables.
2. historypage.js uses jQuery to access the data from /api/v1/applications REST API, and use DataTable to display each application's information. For application that has more than one attempt, the RowsGroup is used to merge such entries while at the same time supporting sort and search.
3. "duration" and "lastUpdated" rest API are added to application's "attempts".
4. External javascirpt and css files for datatables, RowsGroup and jquery plugins are added with licenses clarified.
Snapshots for how it looks like now:
History page view:
![historypage](https://cloud.githubusercontent.com/assets/11683054/12184383/89bad774-b55a-11e5-84e4-b0276172976f.png)
Search:
![search](https://cloud.githubusercontent.com/assets/11683054/12184385/8d3b94b0-b55a-11e5-869a-cc0ef0a4242a.png)
Sort by started time:
![sort-by-started-time](https://cloud.githubusercontent.com/assets/11683054/12184387/8f757c3c-b55a-11e5-98c8-577936366566.png)
Author: zhuol <zhuol@yahoo-inc.com>
Closes#10648 from zhuoliu/10873.
Spark's `Partition` and `RDD.partitions` APIs have a contract which requires custom implementations of `RDD.partitions` to ensure that for all `x`, `rdd.partitions(x).index == x`; in other words, the `index` reported by a repartition needs to match its position in the partitions array.
If a custom RDD implementation violates this contract, then Spark has the potential to become stuck in an infinite recomputation loop when recomputing a subset of an RDD's partitions, since the tasks that are actually run will not correspond to the missing output partitions that triggered the recomputation. Here's a link to a notebook which demonstrates this problem: 5e8a5aa8d2/Violating%2520RDD.partitions%2520contract.html
In order to guard against this infinite loop behavior, this patch modifies Spark so that it fails fast and refuses to compute RDDs' whose `partitions` violate the API contract.
Author: Josh Rosen <joshrosen@databricks.com>
Closes#10932 from JoshRosen/SPARK-13021.
The high level idea is that instead of having the executors send both accumulator updates and TaskMetrics, we should have them send only accumulator updates. This eliminates the need to maintain both code paths since one can be implemented in terms of the other. This effort is split into two parts:
**SPARK-12895: Implement TaskMetrics using accumulators.** TaskMetrics is basically just a bunch of accumulable fields. This patch makes TaskMetrics a syntactic wrapper around a collection of accumulators so we don't need to send TaskMetrics from the executors to the driver.
**SPARK-12896: Send only accumulator updates to the driver.** Now that TaskMetrics are expressed in terms of accumulators, we can capture all TaskMetrics values if we just send accumulator updates from the executors to the driver. This completes the parent issue SPARK-10620.
While an effort has been made to preserve as much of the public API as possible, there were a few known breaking DeveloperApi changes that would be very awkward to maintain. I will gather the full list shortly and post it here.
Note: This was once part of #10717. This patch is split out into its own patch from there to make it easier for others to review. Other smaller pieces of already been merged into master.
Author: Andrew Or <andrew@databricks.com>
Closes#10835 from andrewor14/task-metrics-use-accums.
Fix Java function API methods for flatMap and mapPartitions to require producing only an Iterator, not Iterable. Also fix DStream.flatMap to require a function producing TraversableOnce only, not Traversable.
CC rxin pwendell for API change; tdas since it also touches streaming.
Author: Sean Owen <sowen@cloudera.com>
Closes#10413 from srowen/SPARK-3369.
Added color coding to the Executors page for Active Tasks, Failed Tasks, Completed Tasks and Task Time.
Active Tasks is shaded blue with it's range based on percentage of total cores used.
Failed Tasks is shaded red ranging over the first 10% of total tasks failed
Completed Tasks is shaded green ranging over 10% of total tasks including failed and active tasks, but only when there are active or failed tasks on that executor.
Task Time is shaded red when GC Time goes over 10% of total time with it's range directly corresponding to the percent of total time.
Author: Alex Bozarth <ajbozart@us.ibm.com>
Closes#10154 from ajbozarth/spark12149.
- Remove Akka dependency from core. Note: the streaming-akka project still uses Akka.
- Remove HttpFileServer
- Remove Akka configs from SparkConf and SSLOptions
- Rename `spark.akka.frameSize` to `spark.rpc.message.maxSize`. I think it's still worth to keep this config because using `DirectTaskResult` or `IndirectTaskResult` depends on it.
- Update comments and docs
Author: Shixiong Zhu <shixiong@databricks.com>
Closes#10854 from zsxwing/remove-akka.
This is a step in implementing SPARK-10620, which migrates TaskMetrics to accumulators.
TaskMetrics has a bunch of var's, some are fully public, some are `private[spark]`. This is bad coding style that makes it easy to accidentally overwrite previously set metrics. This has happened a few times in the past and caused bugs that were difficult to debug.
Instead, we should have get-or-create semantics, which are more readily understandable. This makes sense in the case of TaskMetrics because these are just aggregated metrics that we want to collect throughout the task, so it doesn't matter who's incrementing them.
Parent PR: #10717
Author: Andrew Or <andrew@databricks.com>
Author: Josh Rosen <joshrosen@databricks.com>
Author: andrewor14 <andrew@databricks.com>
Closes#10815 from andrewor14/get-or-create-metrics.
This is a small step in implementing SPARK-10620, which migrates TaskMetrics to accumulators. This patch is strictly a cleanup patch and introduces no change in functionality. It literally just renames 3 fields for consistency. Today we have:
```
inputMetrics.recordsRead
outputMetrics.bytesWritten
shuffleReadMetrics.localBlocksFetched
...
shuffleWriteMetrics.shuffleRecordsWritten
shuffleWriteMetrics.shuffleBytesWritten
shuffleWriteMetrics.shuffleWriteTime
```
The shuffle write ones are kind of redundant. We can drop the `shuffle` part in the method names. I added backward compatible (but deprecated) methods with the old names.
Parent PR: #10717
Author: Andrew Or <andrew@databricks.com>
Closes#10811 from andrewor14/rename-things.
This patch refactors portions of the BlockManager and CacheManager in order to avoid having to pass `evictedBlocks` lists throughout the code. It appears that these lists were only consumed by `TaskContext.taskMetrics`, so the new code now directly updates the metrics from the lower-level BlockManager methods.
Author: Josh Rosen <joshrosen@databricks.com>
Closes#10776 from JoshRosen/SPARK-10985.
This pull request removes the external block store API. This is rarely used, and the file system interface is actually a better, more standard way to interact with external storage systems.
There are some other things to remove also, as pointed out by JoshRosen. We will do those as follow-up pull requests.
Author: Reynold Xin <rxin@databricks.com>
Closes#10752 from rxin/remove-offheap.
This patch significantly speeds up the BlockManagerSuite's "SPARK-9591: getRemoteBytes from another location when Exception throw" test, reducing the test time from 45s to ~250ms. The key change was to set `spark.shuffle.io.maxRetries` to 0 (the code previously set `spark.network.timeout` to `2s`, but this didn't make a difference because the slowdown was not due to this timeout).
Along the way, I also cleaned up the way that we handle SparkConf in BlockManagerSuite: previously, each test would mutate a shared SparkConf instance, while now each test gets a fresh SparkConf.
Author: Josh Rosen <joshrosen@databricks.com>
Closes#10759 from JoshRosen/SPARK-12174.
When an Executor process is destroyed, the FileAppender that is asynchronously reading the stderr stream of the process can throw an IOException during read because the stream is closed. Before the ExecutorRunner destroys the process, the FileAppender thread is flagged to stop. This PR wraps the inputStream.read call of the FileAppender in a try/catch block so that if an IOException is thrown and the thread has been flagged to stop, it will safely ignore the exception. Additionally, the FileAppender thread was changed to use Utils.tryWithSafeFinally to better log any exception that do occur. Added unit tests to verify a IOException is thrown and logged if FileAppender is not flagged to stop, and that no IOException when the flag is set.
Author: Bryan Cutler <cutlerb@gmail.com>
Closes#10714 from BryanCutler/file-appender-read-ioexception-SPARK-9844.
We've already removed local execution but didn't deprecate `TaskContext.isRunningLocally()`; we should deprecate it for 2.0.
Author: Josh Rosen <joshrosen@databricks.com>
Closes#10751 from JoshRosen/remove-local-exec-from-taskcontext.
This problem lies in `BypassMergeSortShuffleWriter`, empty partition will also generate a temp shuffle file with several bytes. So here change to only create file when partition is not empty.
This problem only lies in here, no such issue in `HashShuffleWriter`.
Please help to review, thanks a lot.
Author: jerryshao <sshao@hortonworks.com>
Closes#10376 from jerryshao/SPARK-12400.
Fix the style violation (space before , and :).
This PR is a followup for #10643
Author: Kousuke Saruta <sarutak@oss.nttdata.co.jp>
Closes#10719 from sarutak/SPARK-12692-followup-core.
[SPARK-12582][Test] IndexShuffleBlockResolverSuite fails in windows
* IndexShuffleBlockResolverSuite fails in windows due to file is not closed.
* mv IndexShuffleBlockResolverSuite.scala from "test/java" to "test/scala".
https://issues.apache.org/jira/browse/SPARK-12582
Author: Yucai Yu <yucai.yu@intel.com>
Closes#10526 from yucai/master.
This patch deduplicates some test code in BlockManagerSuite. I'm splitting this change off from a larger PR in order to make things easier to review.
Author: Josh Rosen <joshrosen@databricks.com>
Closes#10667 from JoshRosen/block-mgr-tests-cleanup.
Replace Guava `Optional` with (an API clone of) Java 8 `java.util.Optional` (edit: and a clone of Guava `Optional`)
See also https://github.com/apache/spark/pull/10512
Author: Sean Owen <sowen@cloudera.com>
Closes#10513 from srowen/SPARK-4819.
Fix most build warnings: mostly deprecated API usages. I'll annotate some of the changes below. CC rxin who is leading the charge to remove the deprecated APIs.
Author: Sean Owen <sowen@cloudera.com>
Closes#10570 from srowen/SPARK-12618.
The default serializer in Kryo is FieldSerializer and it ignores transient fields and never calls `writeObject` or `readObject`. So we should register OpenHashMapBasedStateMap using `DefaultSerializer` to make it work with Kryo.
Author: Shixiong Zhu <shixiong@databricks.com>
Closes#10609 from zsxwing/SPARK-12591.
This PR removes `spark.cleaner.ttl` and the associated TTL-based metadata cleaning code.
Now that we have the `ContextCleaner` and a timer to trigger periodic GCs, I don't think that `spark.cleaner.ttl` is necessary anymore. The TTL-based cleaning isn't enabled by default, isn't included in our end-to-end tests, and has been a source of user confusion when it is misconfigured. If the TTL is set too low, data which is still being used may be evicted / deleted, leading to hard to diagnose bugs.
For all of these reasons, I think that we should remove this functionality in Spark 2.0. Additional benefits of doing this include marginally reduced memory usage, since we no longer need to store timetsamps in hashmaps, and a handful fewer threads.
Author: Josh Rosen <joshrosen@databricks.com>
Closes#10534 from JoshRosen/remove-ttl-based-cleaning.
Change Java countByKey, countApproxDistinctByKey return types to use Java Long, not Scala; update similar methods for consistency on java.long.Long.valueOf with no API change
Author: Sean Owen <sowen@cloudera.com>
Closes#10554 from srowen/SPARK-12604.
Whole code of Vector.scala, VectorSuite.scala and GraphKryoRegistrator.scala are no longer used so it's time to remove them in Spark 2.0.
Author: Kousuke Saruta <sarutak@oss.nttdata.co.jp>
Closes#10613 from sarutak/SPARK-12665.
Cartesian product use UnsafeExternalSorter without comparator to do spilling, it will NPE if spilling happens.
This bug also hitted by #10605
cc JoshRosen
Author: Davies Liu <davies@databricks.com>
Closes#10606 from davies/fix_spilling.
I looked at each case individually and it looks like they can all be removed. The only one that I had to think twice was toArray (I even thought about un-deprecating it, until I realized it was a problem in Java to have toArray returning java.util.List).
Author: Reynold Xin <rxin@databricks.com>
Closes#10569 from rxin/SPARK-12615.
This patch updates the ExecutorRunner's terminate path to use the new java 8 API
to terminate processes more forcefully if possible. If the executor is unhealthy,
it would previously ignore the destroy() call. Presumably, the new java API was
added to handle cases like this.
We could update the termination path in the future to use OS specific commands
for older java versions.
Author: Nong Li <nong@databricks.com>
Closes#10438 from nongli/spark-12486-executors.
### Remove AkkaRpcEnv
Keep `SparkEnv.actorSystem` because Streaming still uses it. Will remove it and AkkaUtils after refactoring Streaming actorStream API.
### Remove systemName
There are 2 places using `systemName`:
* `RpcEnvConfig.name`. Actually, although it's used as `systemName` in `AkkaRpcEnv`, `NettyRpcEnv` uses it as the service name to output the log `Successfully started service *** on port ***`. Since the service name in log is useful, I keep `RpcEnvConfig.name`.
* `def setupEndpointRef(systemName: String, address: RpcAddress, endpointName: String)`. Each `ActorSystem` has a `systemName`. Akka requires `systemName` in its URI and will refuse a connection if `systemName` is not matched. However, `NettyRpcEnv` doesn't use it. So we can remove `systemName` from `setupEndpointRef` since we are removing `AkkaRpcEnv`.
### Remove RpcEnv.uriOf
`uriOf` exists because Akka uses different URI formats for with and without authentication, e.g., `akka.ssl.tcp...` and `akka.tcp://...`. But `NettyRpcEnv` uses the same format. So it's not necessary after removing `AkkaRpcEnv`.
Author: Shixiong Zhu <shixiong@databricks.com>
Closes#10459 from zsxwing/remove-akka-rpc-env.
We switched to TorrentBroadcast in Spark 1.1, and HttpBroadcast has been undocumented since then. It's time to remove it in Spark 2.0.
Author: Reynold Xin <rxin@databricks.com>
Closes#10531 from rxin/SPARK-12588.
The web UI's paginated table uses Javascript to implement certain navigation controls, such as table sorting and the "go to page" form. This is unnecessary and should be simplified to use plain HTML form controls and links.
/cc zsxwing, who wrote this original code, and yhuai.
Author: Josh Rosen <joshrosen@databricks.com>
Closes#10441 from JoshRosen/simplify-paginated-table-sorting.
The feature was first added at commit: 7b877b2705 but was later removed (probably by mistake) at commit: fc8b58195a.
This change sets the default path of RDDs created via sc.textFile(...) to the path argument.
Here is the symptom:
* Using spark-1.5.2-bin-hadoop2.6:
scala> sc.textFile("/home/root/.bashrc").name
res5: String = null
scala> sc.binaryFiles("/home/root/.bashrc").name
res6: String = /home/root/.bashrc
* while using Spark 1.3.1:
scala> sc.textFile("/home/root/.bashrc").name
res0: String = /home/root/.bashrc
scala> sc.binaryFiles("/home/root/.bashrc").name
res1: String = /home/root/.bashrc
Author: Yaron Weinsberg <wyaron@gmail.com>
Author: yaron <yaron@il.ibm.com>
Closes#10456 from wyaron/master.
Restore the original value of os.arch property after each test
Since some of tests forced to set the specific value to os.arch property, we need to set the original value.
Author: Kazuaki Ishizaki <ishizaki@jp.ibm.com>
Closes#10289 from kiszk/SPARK-12311.
According the benchmark [1], LZ4-java could be 80% (or 30%) faster than Snappy.
After changing the compressor to LZ4, I saw 20% improvement on end-to-end time for a TPCDS query (Q4).
[1] https://github.com/ning/jvm-compressor-benchmark/wiki
cc rxin
Author: Davies Liu <davies@databricks.com>
Closes#10342 from davies/lz4.
When multiple workers exist in a host, we can bypass unnecessary remote access for broadcasts; block managers fetch broadcast blocks from the same host instead of remote hosts.
Author: Takeshi YAMAMURO <linguin.m.s@gmail.com>
Closes#10346 from maropu/OptimizeBlockLocationOrder.
Added `channelActive` to `RpcHandler` so that `NettyRpcHandler` doesn't need `clients` any more.
Author: Shixiong Zhu <shixiong@databricks.com>
Closes#10301 from zsxwing/network-events.
In discussion (SPARK-9552), we proposed a force kill in `killExecutors`. But if there is nothing to kill, it will return back with true (acknowledgement). And then, it causes the certain executor(s) (which is not eligible to kill) adding to pendingToRemove list for further actions.
In this patch, we'd like to change the return semantics. If there is nothing to kill, we will return "false". and therefore all those non-eligible executors won't be added to the pendingToRemove list.
vanzin andrewor14 As the follow up of PR#7888, please let me know your comments.
Author: Grace <jie.huang@intel.com>
Author: Jie Huang <hjie@fosun.com>
Author: Andrew Or <andrew@databricks.com>
Closes#9796 from GraceH/emptyPendingToRemove.
Not jira is created.
The original test is passed because the class cast is lazy (only when the object's method is invoked).
Author: Jeff Zhang <zjffdu@apache.org>
Closes#10371 from zjffdu/minor_fix.
`DAGSchedulerEventLoop` normally only logs errors (so it can continue to process more events, from other jobs). However, this is not desirable in the tests -- the tests should be able to easily detect any exception, and also shouldn't silently succeed if there is an exception.
This was suggested by mateiz on https://github.com/apache/spark/pull/7699. It may have already turned up an issue in "zero split job".
Author: Imran Rashid <irashid@cloudera.com>
Closes#8466 from squito/SPARK-10248.
These changes rework the implementations of `SimpleFutureAction`, `ComplexFutureAction`, `JobWaiter`, and `AsyncRDDActions` such that asynchronous callbacks on the generated `Futures` NEVER block waiting for a job to complete. A small amount of mutex synchronization is necessary to protect the internal fields that manage cancellation, but these locks are only held very briefly and in practice should almost never cause any blocking to occur. The existing blocking APIs of these classes are retained, but they simply delegate to the underlying non-blocking API and `Await` the results with indefinite timeouts.
Associated JIRA ticket: https://issues.apache.org/jira/browse/SPARK-9026
Also fixes: https://issues.apache.org/jira/browse/SPARK-4514
This pull request contains all my own original work, which I release to the Spark project under its open source license.
Author: Richard W. Eggert II <richard.eggert@gmail.com>
Closes#9264 from reggert/fix-futureaction.
Fix a minor typo (unbalanced bracket) in ResetSystemProperties.
Author: Holden Karau <holden@us.ibm.com>
Closes#10303 from holdenk/SPARK-12332-trivial-typo-in-ResetSystemProperties-comment.
**Problem.** In unified memory management, acquiring execution memory may lead to eviction of storage memory. However, the space freed from evicting cached blocks is distributed among all active tasks. Thus, an incorrect upper bound on the execution memory per task can cause the acquisition to fail, leading to OOM's and premature spills.
**Example.** Suppose total memory is 1000B, cached blocks occupy 900B, `spark.memory.storageFraction` is 0.4, and there are two active tasks. In this case, the cap on task execution memory is 100B / 2 = 50B. If task A tries to acquire 200B, it will evict 100B of storage but can only acquire 50B because of the incorrect cap. For another example, see this [regression test](https://github.com/andrewor14/spark/blob/fix-oom/core/src/test/scala/org/apache/spark/memory/UnifiedMemoryManagerSuite.scala#L233) that I stole from JoshRosen.
**Solution.** Fix the cap on task execution memory. It should take into account the space that could have been freed by storage in addition to the current amount of memory available to execution. In the example above, the correct cap should have been 600B / 2 = 300B.
This patch also guards against the race condition (SPARK-12253):
(1) Existing tasks collectively occupy all execution memory
(2) New task comes in and blocks while existing tasks spill
(3) After tasks finish spilling, another task jumps in and puts in a large block, stealing the freed memory
(4) New task still cannot acquire memory and goes back to sleep
Author: Andrew Or <andrew@databricks.com>
Closes#10240 from andrewor14/fix-oom.
This patch adds documentation for Spark configurations that affect off-heap memory and makes some naming and validation improvements for those configs.
- Change `spark.memory.offHeapSize` to `spark.memory.offHeap.size`. This is fine because this configuration has not shipped in any Spark release yet (it's new in Spark 1.6).
- Deprecated `spark.unsafe.offHeap` in favor of a new `spark.memory.offHeap.enabled` configuration. The motivation behind this change is to gather all memory-related configurations under the same prefix.
- Add a check which prevents users from setting `spark.memory.offHeap.enabled=true` when `spark.memory.offHeap.size == 0`. After SPARK-11389 (#9344), which was committed in Spark 1.6, Spark enforces a hard limit on the amount of off-heap memory that it will allocate to tasks. As a result, enabling off-heap execution memory without setting `spark.memory.offHeap.size` will lead to immediate OOMs. The new configuration validation makes this scenario easier to diagnose, helping to avoid user confusion.
- Document these configurations on the configuration page.
Author: Josh Rosen <joshrosen@databricks.com>
Closes#10237 from JoshRosen/SPARK-12251.
This avoids bringing up yet another HTTP server on the driver, and
instead reuses the file server already managed by the driver's
RpcEnv. As a bonus, the repl now inherits the security features of
the network library.
There's also a small change to create the directory for storing classes
under the root temp dir for the application (instead of directly
under java.io.tmpdir).
Author: Marcelo Vanzin <vanzin@cloudera.com>
Closes#9923 from vanzin/SPARK-11563.
This patch fixes a bug in the eviction of storage memory by execution.
## The bug:
In general, execution should be able to evict storage memory when the total storage memory usage is greater than `maxMemory * spark.memory.storageFraction`. Due to a bug, however, Spark might wind up evicting no storage memory in certain cases where the storage memory usage was between `maxMemory * spark.memory.storageFraction` and `maxMemory`. For example, here is a regression test which illustrates the bug:
```scala
val maxMemory = 1000L
val taskAttemptId = 0L
val (mm, ms) = makeThings(maxMemory)
// Since we used the default storage fraction (0.5), we should be able to allocate 500 bytes
// of storage memory which are immune to eviction by execution memory pressure.
// Acquire enough storage memory to exceed the storage region size
assert(mm.acquireStorageMemory(dummyBlock, 750L, evictedBlocks))
assertEvictBlocksToFreeSpaceNotCalled(ms)
assert(mm.executionMemoryUsed === 0L)
assert(mm.storageMemoryUsed === 750L)
// At this point, storage is using 250 more bytes of memory than it is guaranteed, so execution
// should be able to reclaim up to 250 bytes of storage memory.
// Therefore, execution should now be able to require up to 500 bytes of memory:
assert(mm.acquireExecutionMemory(500L, taskAttemptId, MemoryMode.ON_HEAP) === 500L) // <--- fails by only returning 250L
assert(mm.storageMemoryUsed === 500L)
assert(mm.executionMemoryUsed === 500L)
assertEvictBlocksToFreeSpaceCalled(ms, 250L)
```
The problem relates to the control flow / interaction between `StorageMemoryPool.shrinkPoolToReclaimSpace()` and `MemoryStore.ensureFreeSpace()`. While trying to allocate the 500 bytes of execution memory, the `UnifiedMemoryManager` discovers that it will need to reclaim 250 bytes of memory from storage, so it calls `StorageMemoryPool.shrinkPoolToReclaimSpace(250L)`. This method, in turn, calls `MemoryStore.ensureFreeSpace(250L)`. However, `ensureFreeSpace()` first checks whether the requested space is less than `maxStorageMemory - storageMemoryUsed`, which will be true if there is any free execution memory because it turns out that `MemoryStore.maxStorageMemory = (maxMemory - onHeapExecutionMemoryPool.memoryUsed)` when the `UnifiedMemoryManager` is used.
The control flow here is somewhat confusing (it grew to be messy / confusing over time / as a result of the merging / refactoring of several components). In the pre-Spark 1.6 code, `ensureFreeSpace` was called directly by the `MemoryStore` itself, whereas in 1.6 it's involved in a confusing control flow where `MemoryStore` calls `MemoryManager.acquireStorageMemory`, which then calls back into `MemoryStore.ensureFreeSpace`, which, in turn, calls `MemoryManager.freeStorageMemory`.
## The solution:
The solution implemented in this patch is to remove the confusing circular control flow between `MemoryManager` and `MemoryStore`, making the storage memory acquisition process much more linear / straightforward. The key changes:
- Remove a layer of inheritance which made the memory manager code harder to understand (53841174760a24a0df3eb1562af1f33dbe340eb9).
- Move some bounds checks earlier in the call chain (13ba7ada77f87ef1ec362aec35c89a924e6987cb).
- Refactor `ensureFreeSpace()` so that the part which evicts blocks can be called independently from the part which checks whether there is enough free space to avoid eviction (7c68ca09cb1b12f157400866983f753ac863380e).
- Realize that this lets us remove a layer of overloads from `ensureFreeSpace` (eec4f6c87423d5e482b710e098486b3bbc4daf06).
- Realize that `ensureFreeSpace()` can simply be replaced with an `evictBlocksToFreeSpace()` method which is called [after we've already figured out](2dc842aea8/core/src/main/scala/org/apache/spark/memory/StorageMemoryPool.scala (L88)) how much memory needs to be reclaimed via eviction; (2dc842aea82c8895125d46a00aa43dfb0d121de9).
Along the way, I fixed some problems with the mocks in `MemoryManagerSuite`: the old mocks would [unconditionally](80a824d36e/core/src/test/scala/org/apache/spark/memory/MemoryManagerSuite.scala (L84)) report that a block had been evicted even if there was enough space in the storage pool such that eviction would be avoided.
I also fixed a problem where `StorageMemoryPool._memoryUsed` might become negative due to freed memory being double-counted when excution evicts storage. The problem was that `StorageMemoryPoolshrinkPoolToFreeSpace` would [decrement `_memoryUsed`](7c68ca09cb (diff-935c68a9803be144ed7bafdd2f756a0fL133)) even though `StorageMemoryPool.freeMemory` had already decremented it as each evicted block was freed. See SPARK-12189 for details.
Author: Josh Rosen <joshrosen@databricks.com>
Author: Andrew Or <andrew@databricks.com>
Closes#10170 from JoshRosen/SPARK-12165.
Because of AM failure, the target executor number between driver and AM will be different, which will lead to unexpected behavior in dynamic allocation. So when AM is re-registered with driver, state in `ExecutorAllocationManager` and `CoarseGrainedSchedulerBacked` should be reset.
This issue is originally addressed in #8737 , here re-opened again. Thanks a lot KaiXinXiaoLei for finding this issue.
andrewor14 and vanzin would you please help to review this, thanks a lot.
Author: jerryshao <sshao@hortonworks.com>
Closes#9963 from jerryshao/SPARK-10582.
The json endpoint for stages doesn't include information on the stage duration that is present in the UI. This looks like a simple oversight, they should be included. eg., the metrics should be included at api/v1/applications/<appId>/stages.
Metrics I've added are: submissionTime, firstTaskLaunchedTime and completionTime
Author: Xin Ren <iamshrek@126.com>
Closes#10107 from keypointt/SPARK-11155.
`ByteBuffer` doesn't guarantee all contents in `ByteBuffer.array` are valid. E.g, a ByteBuffer returned by `ByteBuffer.slice`. We should not use the whole content of `ByteBuffer` unless we know that's correct.
This patch fixed all places that use `ByteBuffer.array` incorrectly.
Author: Shixiong Zhu <shixiong@databricks.com>
Closes#10083 from zsxwing/bytebuffer-array.
We should upgrade to SBT 0.13.9, since this is a requirement in order to use SBT's new Maven-style resolution features (which will be done in a separate patch, because it's blocked by some binary compatibility issues in the POM reader plugin).
I also upgraded Scalastyle to version 0.8.0, which was necessary in order to fix a Scala 2.10.5 compatibility issue (see https://github.com/scalastyle/scalastyle/issues/156). The newer Scalastyle is slightly stricter about whitespace surrounding tokens, so I fixed the new style violations.
Author: Josh Rosen <joshrosen@databricks.com>
Closes#10112 from JoshRosen/upgrade-to-sbt-0.13.9.
This replaces https://github.com/apache/spark/pull/9696
Invoke Checkstyle and print any errors to the console, failing the step.
Use Google's style rules modified according to
https://cwiki.apache.org/confluence/display/SPARK/Spark+Code+Style+Guide
Some important checks are disabled (see TODOs in `checkstyle.xml`) due to
multiple violations being present in the codebase.
Suggest fixing those TODOs in a separate PR(s).
More on Checkstyle can be found on the [official website](http://checkstyle.sourceforge.net/).
Sample output (from [build 46345](https://amplab.cs.berkeley.edu/jenkins/job/SparkPullRequestBuilder/46345/consoleFull)) (duplicated because I run the build twice with different profiles):
> Checkstyle checks failed at following occurrences:
[ERROR] src/main/java/org/apache/spark/sql/execution/datasources/parquet/UnsafeRowParquetRecordReader.java:[217,7] (coding) MissingSwitchDefault: switch without "default" clause.
> [ERROR] src/main/java/org/apache/spark/sql/execution/datasources/parquet/SpecificParquetRecordReaderBase.java:[198,10] (modifier) ModifierOrder: 'protected' modifier out of order with the JLS suggestions.
> [ERROR] src/main/java/org/apache/spark/sql/execution/datasources/parquet/UnsafeRowParquetRecordReader.java:[217,7] (coding) MissingSwitchDefault: switch without "default" clause.
> [ERROR] src/main/java/org/apache/spark/sql/execution/datasources/parquet/SpecificParquetRecordReaderBase.java:[198,10] (modifier) ModifierOrder: 'protected' modifier out of order with the JLS suggestions.
> [error] running /home/jenkins/workspace/SparkPullRequestBuilder2/dev/lint-java ; received return code 1
Also fix some of the minor violations that didn't require sweeping changes.
Apologies for the previous botched PRs - I finally figured out the issue.
cr: JoshRosen, pwendell
> I state that the contribution is my original work, and I license the work to the project under the project's open source license.
Author: Dmitry Erastov <derastov@gmail.com>
Closes#9867 from dskrvk/master.
**Problem.** Event logs in 1.6 were much bigger than 1.5. I ran page rank and the event log size in 1.6 was almost 5x that in 1.5. I did a bisect to find that the RDD callsite added in #9398 is largely responsible for this.
**Solution.** This patch removes the long form of the callsite (which is not used!) from the event log. This reduces the size of the event log significantly.
*Note on compatibility*: if this patch is to be merged into 1.6.0, then it won't break any compatibility. Otherwise, if it is merged into 1.6.1, then we might need to add more backward compatibility handling logic (currently does not exist yet).
Author: Andrew Or <andrew@databricks.com>
Closes#10115 from andrewor14/smaller-event-logs.
We should try increasing a timeout in NettyBlockTransferSecuritySuite in order to reduce that suite's flakiness in Jenkins.
Author: Josh Rosen <joshrosen@databricks.com>
Closes#10113 from JoshRosen/SPARK-12082.
I have tried to address all the comments in pull request https://github.com/apache/spark/pull/2447.
Note that the second commit (using the new method in all internal code of all components) is quite intrusive and could be omitted.
Author: Jeroen Schot <jeroen.schot@surfsara.nl>
Closes#9767 from schot/master.
The existing `spark.memory.fraction` (default 0.75) gives the system 25% of the space to work with. For small heaps, this is not enough: e.g. default 1GB leaves only 250MB system memory. This is especially a problem in local mode, where the driver and executor are crammed in the same JVM. Members of the community have reported driver OOM's in such cases.
**New proposal.** We now reserve 300MB before taking the 75%. For 1GB JVMs, this leaves `(1024 - 300) * 0.75 = 543MB` for execution and storage. This is proposal (1) listed in the [JIRA](https://issues.apache.org/jira/browse/SPARK-12081).
Author: Andrew Or <andrew@databricks.com>
Closes#10081 from andrewor14/unified-memory-small-heaps.
The solution is the save the RDD partitioner in a separate file in the RDD checkpoint directory. That is, `<checkpoint dir>/_partitioner`. In most cases, whether the RDD partitioner was recovered or not, does not affect the correctness, only reduces performance. So this solution makes a best-effort attempt to save and recover the partitioner. If either fails, the checkpointing is not affected. This makes this patch safe and backward compatible.
Author: Tathagata Das <tathagata.das1565@gmail.com>
Closes#9983 from tdas/SPARK-12004.
This change seems large, but most of it is just replacing `byte[]`
with `ByteBuffer` and `new byte[]` with `ByteBuffer.allocate()`,
since it changes the network library's API.
The following are parts of the code that actually have meaningful
changes:
- The Message implementations were changed to inherit from a new
AbstractMessage that can optionally hold a reference to a body
(in the form of a ManagedBuffer); this is similar to how
ResponseWithBody worked before, except now it's not restricted
to just responses.
- The TransportFrameDecoder was pretty much rewritten to avoid
copies as much as possible; it doesn't rely on CompositeByteBuf
to accumulate incoming data anymore, since CompositeByteBuf
has issues when slices are retained. The code now is able to
create frames without having to resort to copying bytes except
for a few bytes (containing the frame length) in very rare cases.
- Some minor changes in the SASL layer to convert things back to
`byte[]` since the JDK SASL API operates on those.
Author: Marcelo Vanzin <vanzin@cloudera.com>
Closes#9987 from vanzin/SPARK-12007.
In the previous codes, `newDaemonCachedThreadPool` uses `SynchronousQueue`, which is wrong. `SynchronousQueue` is an empty queue that cannot cache any task. This patch uses `LinkedBlockingQueue` to fix it along with other fixes to make sure `newDaemonCachedThreadPool` can use at most `maxThreadNumber` threads, and after that, cache tasks to `LinkedBlockingQueue`.
Author: Shixiong Zhu <shixiong@databricks.com>
Closes#9978 from zsxwing/cached-threadpool.
This change does a couple of different things to make sure that the RpcEnv-level
code and the network library agree about the status of outstanding RPCs.
For RPCs that do not expect a reply ("RpcEnv.send"), support for one way
messages (hello CORBA!) was added to the network layer. This is a
"fire and forget" message that does not require any state to be kept
by the TransportClient; as a result, the RpcEnv 'Ack' message is not needed
anymore.
For RPCs that do expect a reply ("RpcEnv.ask"), the network library now
returns the internal RPC id; if the RpcEnv layer decides to time out the
RPC before the network layer does, it now asks the TransportClient to
forget about the RPC, so that if the network-level timeout occurs, the
client is not killed.
As part of implementing the above, I cleaned up some of the code in the
netty rpc backend, removing types that were not necessary and factoring
out some common code. Of interest is a slight change in the exceptions
when posting messages to a stopped RpcEnv; that's mostly to avoid nasty
error messages from the local-cluster backend when shutting down, which
pollutes the terminal output.
Author: Marcelo Vanzin <vanzin@cloudera.com>
Closes#9917 from vanzin/SPARK-11866.
- NettyRpcEnv::openStream() now correctly propagates errors to
the read side of the pipe.
- NettyStreamManager now throws if the file being transferred does
not exist.
- The network library now correctly handles zero-sized streams.
Author: Marcelo Vanzin <vanzin@cloudera.com>
Closes#9941 from vanzin/SPARK-11956.
This issue was addressed in https://github.com/apache/spark/pull/5494, but the fix in that PR, while safe in the sense that it will prevent the SparkContext from shutting down, misses the actual bug. The intent of `submitMissingTasks` should be understood as "submit the Tasks that are missing for the Stage, and run them as part of the ActiveJob identified by jobId". Because of a long-standing bug, the `jobId` parameter was never being used. Instead, we were trying to use the jobId with which the Stage was created -- which may no longer exist as an ActiveJob, hence the crash reported in SPARK-6880.
The correct fix is to use the ActiveJob specified by the supplied jobId parameter, which is guaranteed to exist at the call sites of submitMissingTasks.
This fix should be applied to all maintenance branches, since it has existed since 1.0.
kayousterhout pankajarora12
Author: Mark Hamstra <markhamstra@gmail.com>
Author: Imran Rashid <irashid@cloudera.com>
Closes#6291 from markhamstra/SPARK-6880.
This is continuation of SPARK-11761
Andrew suggested adding this protection. See tail of https://github.com/apache/spark/pull/9741
Author: tedyu <yuzhihong@gmail.com>
Closes#9852 from tedyu/master.
This change abstracts the code that serves jars / files to executors so that
each RpcEnv can have its own implementation; the akka version uses the existing
HTTP-based file serving mechanism, while the netty versions uses the new
stream support added to the network lib, which makes file transfers benefit
from the easier security configuration of the network library, and should also
reduce overhead overall.
The change includes a small fix to TransportChannelHandler so that it propagates
user events to downstream handlers.
Author: Marcelo Vanzin <vanzin@cloudera.com>
Closes#9530 from vanzin/SPARK-11140.
In PersistenceEngineSuite, we do not call `close()` on the PersistenceEngine at the end of the test. For the ZooKeeperPersistenceEngine, this causes us to leak a ZooKeeper client, causing the logs of unrelated tests to be periodically spammed with connection error messages from that client:
```
15/11/20 05:13:35.789 pool-1-thread-1-ScalaTest-running-PersistenceEngineSuite-SendThread(localhost:15741) INFO ClientCnxn: Opening socket connection to server localhost/127.0.0.1:15741. Will not attempt to authenticate using SASL (unknown error)
15/11/20 05:13:35.790 pool-1-thread-1-ScalaTest-running-PersistenceEngineSuite-SendThread(localhost:15741) WARN ClientCnxn: Session 0x15124ff48dd0000 for server null, unexpected error, closing socket connection and attempting reconnect
java.net.ConnectException: Connection refused
at sun.nio.ch.SocketChannelImpl.checkConnect(Native Method)
at sun.nio.ch.SocketChannelImpl.finishConnect(SocketChannelImpl.java:739)
at org.apache.zookeeper.ClientCnxnSocketNIO.doTransport(ClientCnxnSocketNIO.java:350)
at org.apache.zookeeper.ClientCnxn$SendThread.run(ClientCnxn.java:1068)
```
This patch fixes this by using a `finally` block.
Author: Josh Rosen <joshrosen@databricks.com>
Closes#9864 from JoshRosen/close-zookeeper-client-in-tests.
This patch reduces some RPC timeouts in order to speed up the slow "AkkaUtilsSuite.remote fetch ssl on - untrusted server", which used to take two minutes to run.
Author: Josh Rosen <joshrosen@databricks.com>
Closes#9869 from JoshRosen/SPARK-11650.
To make sure that all lineage is correctly truncated for TrackStateRDD when checkpointed.
Author: Tathagata Das <tathagata.das1565@gmail.com>
Closes#9831 from tdas/SPARK-11845.
SparkListenerSuite's _"onTaskGettingResult() called when result fetched remotely"_ test was extremely slow (1 to 4 minutes to run) and recently became extremely flaky, frequently failing with OutOfMemoryError.
The root cause was the fact that this was using `System.setProperty` to set the Akka frame size, which was not actually modifying the frame size. As a result, this test would allocate much more data than necessary. The fix here is to simply use SparkConf in order to configure the frame size.
Author: Josh Rosen <joshrosen@databricks.com>
Closes#9822 from JoshRosen/SPARK-11649.
[SPARK-6028](https://issues.apache.org/jira/browse/SPARK-6028) uses network module to implement RPC. However, there are some configurations named with `spark.shuffle` prefix in the network module.
This PR refactors them to make sure the user can control them in shuffle and RPC separately. The user can use `spark.rpc.*` to set the configuration for netty RPC.
Author: Shixiong Zhu <shixiong@databricks.com>
Closes#9481 from zsxwing/SPARK-10745.
https://issues.apache.org/jira/browse/SPARK-11792
The main changes include:
* Renaming `SizeEstimation` to `KnownSizeEstimation`. Hopefully this new name has more information.
* Making `estimatedSize` return `Long` instead of `Option[Long]`.
* In `UnsaveHashedRelation`, `estimatedSize` will delegate the work to `SizeEstimator` if we have not created a `BytesToBytesMap`.
Since we will put `UnsaveHashedRelation` to `BlockManager`, it is generally good to let it provide a more accurate size estimation. Also, if we do not put `BytesToBytesMap` directly into `BlockerManager`, I feel it is not really necessary to make `BytesToBytesMap` extends `KnownSizeEstimation`.
Author: Yin Huai <yhuai@databricks.com>
Closes#9813 from yhuai/SPARK-11792-followup.
Make sure we are using the context classloader when deserializing failed TaskResults instead of the Spark classloader.
The issue is that `enqueueFailedTask` was using the incorrect classloader which results in `ClassNotFoundException`.
Adds a test in TaskResultGetterSuite that compiles a custom exception, throws it on the executor, and asserts that Spark handles the TaskResult deserialization instead of returning `UnknownReason`.
See #9367 for previous comments
See SPARK-11195 for a full repro
Author: Hurshal Patel <hpatel516@gmail.com>
Closes#9779 from choochootrain/spark-11195-master.
This PR upgrade the version of RoaringBitmap to 0.5.10, to optimize the memory layout, will be much smaller when most of blocks are empty.
This PR is based on #9661 (fix conflicts), see all of the comments at https://github.com/apache/spark/pull/9661 .
Author: Kent Yao <yaooqinn@hotmail.com>
Author: Davies Liu <davies@databricks.com>
Author: Charles Allen <charles@allen-net.com>
Closes#9746 from davies/roaring_mapstatus.
By using the dynamic allocation, sometimes it occurs false killing for those busy executors. Some executors with assignments will be killed because of being idle for enough time (say 60 seconds). The root cause is that the Task-Launch listener event is asynchronized.
For example, some executors are under assigning tasks, but not sending out the listener notification yet. Meanwhile, the dynamic allocation's executor idle time is up (e.g., 60 seconds). It will trigger killExecutor event at the same time.
1. the timer expiration starts before the listener event arrives.
2. Then, the task is going to run on top of that killed/killing executor. It will lead to task failure finally.
Here is the proposal to fix it. We can add the force control for killExecutor. If the force control is not set (i.e., false), we'd better to check if the executor under killing is idle or busy. If the current executor has some assignment, we should not kill that executor and return back false (to indicate killing failure). In dynamic allocation, we'd better to turn off force killing (i.e., force = false), we will meet killing failure if tries to kill a busy executor. And then, the executor timer won't be invalid. Later on, the task assignment event arrives, we can remove the idle timer accordingly. So that we can avoid false killing for those busy executors in dynamic allocation.
For the rest of usages, the end users can decide if to use force killing or not by themselves. If to turn on that option, the killExecutor will do the action without any status checking.
Author: Grace <jie.huang@intel.com>
Author: Andrew Or <andrew@databricks.com>
Author: Jie Huang <jie.huang@intel.com>
Closes#7888 from GraceH/forcekill.
The basic idea is that:
The archive of the SparkR package itself, that is sparkr.zip, is created during build process and is contained in the Spark binary distribution. No change to it after the distribution is installed as the directory it resides ($SPARK_HOME/R/lib) may not be writable.
When there is R source code contained in jars or Spark packages specified with "--jars" or "--packages" command line option, a temporary directory is created by calling Utils.createTempDir() where the R packages built from the R source code will be installed. The temporary directory is writable, and won't interfere with each other when there are multiple SparkR sessions, and will be deleted when this SparkR session ends. The R binary packages installed in the temporary directory then are packed into an archive named rpkg.zip.
sparkr.zip and rpkg.zip are distributed to the cluster in YARN modes.
The distribution of rpkg.zip in Standalone modes is not supported in this PR, and will be address in another PR.
Various R files are updated to accept multiple lib paths (one is for SparkR package, the other is for other R packages) so that these package can be accessed in R.
Author: Sun Rui <rui.sun@intel.com>
Closes#9390 from sun-rui/SPARK-10500.
Currently, all the shuffle writer will write to target path directly, the file could be corrupted by other attempt of the same partition on the same executor. They should write to temporary file then rename to target path, as what we do in output committer. In order to make the rename atomic, the temporary file should be created in the same local directory (FileSystem).
This PR is based on #9214 , thanks to squito . Closes#9214
Author: Davies Liu <davies@databricks.com>
Closes#9610 from davies/safe_shuffle.
This patch aims to reduce the test time and flakiness of HiveSparkSubmitSuite, SparkSubmitSuite, and CliSuite.
Key changes:
- Disable IO synchronization calls for Derby writes, since durability doesn't matter for tests. This was done for HiveCompatibilitySuite in #6651 and resulted in huge test speedups.
- Add a few missing `--conf`s to disable various Spark UIs. The CliSuite, in particular, never disabled these UIs, leaving it prone to port-contention-related flakiness.
- Fix two instances where tests defined `beforeAll()` methods which were never called because the appropriate traits were not mixed in. I updated these tests suites to extend `BeforeAndAfterEach` so that they play nicely with our `ResetSystemProperties` trait.
Author: Josh Rosen <joshrosen@databricks.com>
Closes#9623 from JoshRosen/SPARK-11647.
just trying to increase test coverage in the scheduler, this already works. It includes a regression test for SPARK-9809
copied some test utils from https://github.com/apache/spark/pull/5636, we can wait till that is merged first
Author: Imran Rashid <irashid@cloudera.com>
Closes#8402 from squito/test_retry_in_shared_shuffle_dep.
Changed AppClient to be non-blocking in `receiveAndReply` by using a separate thread to wait for response and reply to the context. The threads are managed by a thread pool. Also added unit tests for the AppClient interface.
Author: Bryan Cutler <bjcutler@us.ibm.com>
Closes#9317 from BryanCutler/appClient-receiveAndReply-SPARK-10827.
In order to lay the groundwork for proper off-heap memory support in SQL / Tungsten, we need to extend our MemoryManager to perform bookkeeping for off-heap memory.
## User-facing changes
This PR introduces a new configuration, `spark.memory.offHeapSize` (name subject to change), which specifies the absolute amount of off-heap memory that Spark and Spark SQL can use. If Tungsten is configured to use off-heap execution memory for allocating data pages, then all data page allocations must fit within this size limit.
## Internals changes
This PR contains a lot of internal refactoring of the MemoryManager. The key change at the heart of this patch is the introduction of a `MemoryPool` class (name subject to change) to manage the bookkeeping for a particular category of memory (storage, on-heap execution, and off-heap execution). These MemoryPools are not fixed-size; they can be dynamically grown and shrunk according to the MemoryManager's policies. In StaticMemoryManager, these pools have fixed sizes, proportional to the legacy `[storage|shuffle].memoryFraction`. In the new UnifiedMemoryManager, the sizes of these pools are dynamically adjusted according to its policies.
There are two subclasses of `MemoryPool`: `StorageMemoryPool` manages storage memory and `ExecutionMemoryPool` manages execution memory. The MemoryManager creates two execution pools, one for on-heap memory and one for off-heap. Instances of `ExecutionMemoryPool` manage the logic for fair sharing of their pooled memory across running tasks (in other words, the ShuffleMemoryManager-like logic has been moved out of MemoryManager and pushed into these ExecutionMemoryPool instances).
I think that this design is substantially easier to understand and reason about than the previous design, where most of these responsibilities were handled by MemoryManager and its subclasses. To see this, take at look at how simple the logic in `UnifiedMemoryManager` has become: it's now very easy to see when memory is dynamically shifted between storage and execution.
## TODOs
- [x] Fix handful of test failures in the MemoryManagerSuites.
- [x] Fix remaining TODO comments in code.
- [ ] Document new configuration.
- [x] Fix commented-out tests / asserts:
- [x] UnifiedMemoryManagerSuite.
- [x] Write tests that exercise the new off-heap memory management policies.
Author: Josh Rosen <joshrosen@databricks.com>
Closes#9344 from JoshRosen/offheap-memory-accounting.
https://issues.apache.org/jira/browse/SPARK-10116
This is really trivial, just happened to notice it -- if `XORShiftRandom.hashSeed` is really supposed to have random bits throughout (as the comment implies), it needs to do something for the conversion to `long`.
mengxr mkolod
Author: Imran Rashid <irashid@cloudera.com>
Closes#8314 from squito/SPARK-10116.
This brings the support of off-heap memory for array inside BytesToBytesMap and InMemorySorter, then we could allocate all the memory from off-heap for execution.
Closes#8068
Author: Davies Liu <davies@databricks.com>
Closes#9477 from davies/unsafe_timsort.
OutputCommitCoordinator uses a map in a place where an array would suffice, increasing its memory consumption for result stages with millions of tasks.
This patch replaces that map with an array. The only tricky part of this is reasoning about the range of possible array indexes in order to make sure that we never index out of bounds.
Author: Josh Rosen <joshrosen@databricks.com>
Closes#9274 from JoshRosen/SPARK-11307.
This is an updated version of #8995 by a-roberts. Original description follows:
Snappy now supports concatenation of serialized streams, this patch contains a version number change and the "does not support" test is now a "supports" test.
Snappy 1.1.2 changelog mentions:
> snappy-java-1.1.2 (22 September 2015)
> This is a backward compatible release for 1.1.x.
> Add AIX (32-bit) support.
> There is no upgrade for the native libraries of the other platforms.
> A major change since 1.1.1 is a support for reading concatenated results of SnappyOutputStream(s)
> snappy-java-1.1.2-RC2 (18 May 2015)
> Fix#107: SnappyOutputStream.close() is not idempotent
> snappy-java-1.1.2-RC1 (13 May 2015)
> SnappyInputStream now supports reading concatenated compressed results of SnappyOutputStream
> There has been no compressed format change since 1.0.5.x. So You can read the compressed results > interchangeablly between these versions.
> Fixes a problem when java.io.tmpdir does not exist.
Closes#8995.
Author: Adam Roberts <aroberts@uk.ibm.com>
Author: Josh Rosen <joshrosen@databricks.com>
Closes#9439 from JoshRosen/update-snappy.
In YARN mode, when preemption is enabled, we may leave executors in a
zombie state while we wait to retrieve the reason for which the executor
exited. This is so that we don't account for failed tasks that were
running on a preempted executor.
The issue is that while we wait for this information, the scheduler
might decide to schedule tasks on the executor, which will never be
able to run them. Other side effects include the block manager still
considering the executor available to cache blocks, for example.
So, when we know that an executor went down but we don't know why,
stop everything related to the executor, except its running tasks.
Only when we know the reason for the exit (or give up waiting for
it) we do update the running tasks.
This is achieved by a new `disableExecutor()` method in the
`Schedulable` interface. For managers that do not behave like this
(i.e. every one but YARN), the existing `executorLost()` method
will behave the same way it did before.
On top of that change, a few minor changes that made debugging easier,
and fixed some other minor issues:
- The cluster-mode AM was printing a misleading log message every
time an executor disconnected from the driver (because the akka
actor system was shared between driver and AM).
- Avoid sending unnecessary requests for an executor's exit reason
when we already know it was explicitly disabled / killed. This
avoids both multiple requests, and unnecessary requests that would
just cause warning messages on the AM (in the explicit kill case).
- Tone down a log message about the executor being lost when it
exited normally (e.g. preemption)
- Wake up the AM monitor thread when requests for executor loss
reasons arrive too, so that we can more quickly remove executors
from this zombie state.
Author: Marcelo Vanzin <vanzin@cloudera.com>
Closes#8887 from vanzin/SPARK-10622.
The test functionality should be the same, but without using mockito; logs don't
really say anything useful but I suspect it may be the cause of the flakiness,
since updating mocks when multiple threads may be using it doesn't work very
well. It also allows some other cleanup (= less test code in FsHistoryProvider).
Author: Marcelo Vanzin <vanzin@cloudera.com>
Closes#9425 from vanzin/SPARK-11466.
DriverDescription refactored to case class because it included no mutable fields.
ApplicationDescription had one mutable field, which was appUiUrl. This field was set by the driver to point to the driver web UI. Master was modifying this field when the application was removed to redirect requests to history server. This was wrong because objects which are sent over the wire should be immutable. Now appUiUrl is immutable in ApplicationDescription and always points to the driver UI even if it is already shutdown. The UI url which master exposes to the user and modifies dynamically is now included into ApplicationInfo - a data object which describes the application state internally in master. That URL in ApplicationInfo is initialised with the value from ApplicationDescription.
ApplicationDescription also included value user, which is now a part of case class fields.
Author: Jacek Lewandowski <lewandowski.jacek@gmail.com>
Closes#9299 from jacek-lewandowski/SPARK-11344.
"Client mode" means the RPC env will not listen for incoming connections.
This allows certain processes in the Spark stack (such as Executors or
tha YARN client-mode AM) to act as pure clients when using the netty-based
RPC backend, reducing the number of sockets needed by the app and also the
number of open ports.
Client connections are also preferred when endpoints that actually have
a listening socket are involved; so, for example, if a Worker connects
to a Master and the Master needs to send a message to a Worker endpoint,
that client connection will be used, even though the Worker is also
listening for incoming connections.
With this change, the workaround for SPARK-10987 isn't necessary anymore, and
is removed. The AM connects to the driver in "client mode", and that connection
is used for all driver <-> AM communication, and so the AM is properly notified
when the connection goes down.
Author: Marcelo Vanzin <vanzin@cloudera.com>
Closes#9210 from vanzin/SPARK-10997.
JIRA: https://issues.apache.org/jira/browse/SPARK-11271
As reported in the JIRA ticket, when there are too many tasks, the memory usage of MapStatus will cause problem. Use BitSet instead of RoaringBitMap should be more efficient in memory usage.
Author: Liang-Chi Hsieh <viirya@appier.com>
Closes#9243 from viirya/mapstatus-bitset.
Use standard JDK APIs for that (with a little help from Guava). Most of the
changes here are in test code, since there were no tests specific to that
part of the code.
Author: Marcelo Vanzin <vanzin@cloudera.com>
Closes#9257 from vanzin/SPARK-11073.
Large HDFS clusters may take a while to leave safe mode when starting; this change
makes the HS wait for that before doing checks about its configuraton. This means
the HS won't stop right away if HDFS is in safe mode and the configuration is not
correct, but that should be a very uncommon situation.
Author: Marcelo Vanzin <vanzin@cloudera.com>
Closes#9043 from vanzin/SPARK-11020.
[SPARK-11338: HistoryPage not multi-tenancy enabled ...](https://issues.apache.org/jira/browse/SPARK-11338)
- `HistoryPage.scala` ...prepending all page links with the web proxy (`uiRoot`) path
- `HistoryServerSuite.scala` ...adding a test case to verify all site-relative links are prefixed when the environment variable `APPLICATION_WEB_PROXY_BASE` (or System property `spark.ui.proxyBase`) is set
Author: Christian Kadner <ckadner@us.ibm.com>
Closes#9291 from ckadner/SPARK-11338 and squashes the following commits:
01d2f35 [Christian Kadner] [SPARK-11338][WebUI] nit fixes
d054bd7 [Christian Kadner] [SPARK-11338][WebUI] prependBaseUri in method makePageLink
8bcb3dc [Christian Kadner] [SPARK-11338][WebUI] Prepend application links on HistoryPage with uiRoot path
Since we do not need to preserve a page before calling compute(), MapPartitionsWithPreparationRDD is not needed anymore.
This PR basically revert #8543, #8511, #8038, #8011
Author: Davies Liu <davies@databricks.com>
Closes#9381 from davies/remove_prepare2.
This PR introduce a mechanism to call spill() on those SQL operators that support spilling (for example, BytesToBytesMap, UnsafeExternalSorter and ShuffleExternalSorter) if there is not enough memory for execution. The preserved first page is needed anymore, so removed.
Other Spillable objects in Spark core (ExternalSorter and AppendOnlyMap) are not included in this PR, but those could benefit from this (trigger others' spilling).
The PrepareRDD may be not needed anymore, could be removed in follow up PR.
The following script will fail with OOM before this PR, finished in 150 seconds with 2G heap (also works in 1.5 branch, with similar duration).
```python
sqlContext.setConf("spark.sql.shuffle.partitions", "1")
df = sqlContext.range(1<<25).selectExpr("id", "repeat(id, 2) as s")
df2 = df.select(df.id.alias('id2'), df.s.alias('s2'))
j = df.join(df2, df.id==df2.id2).groupBy(df.id).max("id", "id2")
j.explain()
print j.count()
```
For thread-safety, here what I'm got:
1) Without calling spill(), the operators should only be used by single thread, no safety problems.
2) spill() could be triggered in two cases, triggered by itself, or by other operators. we can check trigger == this in spill(), so it's still in the same thread, so safety problems.
3) if it's triggered by other operators (right now cache will not trigger spill()), we only spill the data into disk when it's in scanning stage (building is finished), so the in-memory sorter or memory pages are read-only, we only need to synchronize the iterator and change it.
4) During scanning, the iterator will only use one record in one page, we can't free this page, because the downstream is currently using it (used by UnsafeRow or other objects). In BytesToBytesMap, we just skip the current page, and dump all others into disk. In UnsafeExternalSorter, we keep the page that is used by current record (having the same baseObject), free it when loading the next record. In ShuffleExternalSorter, the spill() will not trigger during scanning.
5) In order to avoid deadlock, we didn't call acquireMemory during spill (so we reused the pointer array in InMemorySorter).
Author: Davies Liu <davies@databricks.com>
Closes#9241 from davies/force_spill.
Commit af3bc59d1f introduced new
functionality so that if an executor dies for a reason that's not
caused by one of the tasks running on the executor (e.g., due to
pre-emption), Spark doesn't count the failure towards the maximum
number of failures for the task. That commit introduced some vague
naming that this commit attempts to fix; in particular:
(1) The variable "isNormalExit", which was used to refer to cases where
the executor died for a reason unrelated to the tasks running on the
machine, has been renamed (and reversed) to "exitCausedByApp". The problem
with the existing name is that it's not clear (at least to me!) what it
means for an exit to be "normal"; the new name is intended to make the
purpose of this variable more clear.
(2) The variable "shouldEventuallyFailJob" has been renamed to
"countTowardsTaskFailures". This variable is used to determine whether
a task's failure should be counted towards the maximum number of failures
allowed for a task before the associated Stage is aborted. The problem
with the existing name is that it can be confused with implying that
the task's failure should immediately cause the stage to fail because it
is somehow fatal (this is the case for a fetch failure, for example: if
a task fails because of a fetch failure, there's no point in retrying,
and the whole stage should be failed).
Author: Kay Ousterhout <kayousterhout@gmail.com>
Closes#9164 from kayousterhout/SPARK-11178.
… ReceiverTracker and ReceiverSchedulingPolicy to use it
This PR includes the following changes:
1. Add a new preferred location format, `executor_<host>_<executorID>` (e.g., "executor_localhost_2"), to support specifying the executor locations for RDD.
2. Use the new preferred location format in `ReceiverTracker` to optimize the starting time of Receivers when there are multiple executors in a host.
The goal of this PR is to enable the streaming scheduler to place receivers (which run as tasks) in specific executors. Basically, I want to have more control on the placement of the receivers such that they are evenly distributed among the executors. We tried to do this without changing the core scheduling logic. But it does not allow specifying particular executor as preferred location, only at the host level. So if there are two executors in the same host, and I want two receivers to run on them (one on each executor), I cannot specify that. Current code only specifies the host as preference, which may end up launching both receivers on the same executor. We try to work around it but restarting a receiver when it does not launch in the desired executor and hope that next time it will be started in the right one. But that cause lots of restarts, and delays in correctly launching the receiver.
So this change, would allow the streaming scheduler to specify the exact executor as the preferred location. Also this is not exposed to the user, only the streaming scheduler uses this.
Author: zsxwing <zsxwing@gmail.com>
Closes#9181 from zsxwing/executor-location.
This patch refactors the MemoryManager class structure. After #9000, Spark had the following classes:
- MemoryManager
- StaticMemoryManager
- ExecutorMemoryManager
- TaskMemoryManager
- ShuffleMemoryManager
This is fairly confusing. To simplify things, this patch consolidates several of these classes:
- ShuffleMemoryManager and ExecutorMemoryManager were merged into MemoryManager.
- TaskMemoryManager is moved into Spark Core.
**Key changes and tasks**:
- [x] Merge ExecutorMemoryManager into MemoryManager.
- [x] Move pooling logic into Allocator.
- [x] Move TaskMemoryManager from `spark-unsafe` to `spark-core`.
- [x] Refactor the existing Tungsten TaskMemoryManager interactions so Tungsten code use only this and not both this and ShuffleMemoryManager.
- [x] Refactor non-Tungsten code to use the TaskMemoryManager instead of ShuffleMemoryManager.
- [x] Merge ShuffleMemoryManager into MemoryManager.
- [x] Move code
- [x] ~~Simplify 1/n calculation.~~ **Will defer to followup, since this needs more work.**
- [x] Port ShuffleMemoryManagerSuite tests.
- [x] Move classes from `unsafe` package to `memory` package.
- [ ] Figure out how to handle the hacky use of the memory managers in HashedRelation's broadcast variable construction.
- [x] Test porting and cleanup: several tests relied on mock functionality (such as `TestShuffleMemoryManager.markAsOutOfMemory`) which has been changed or broken during the memory manager consolidation
- [x] AbstractBytesToBytesMapSuite
- [x] UnsafeExternalSorterSuite
- [x] UnsafeFixedWidthAggregationMapSuite
- [x] UnsafeKVExternalSorterSuite
**Compatiblity notes**:
- This patch introduces breaking changes in `ExternalAppendOnlyMap`, which is marked as `DevloperAPI` (likely for legacy reasons): this class now cannot be used outside of a task.
Author: Josh Rosen <joshrosen@databricks.com>
Closes#9127 from JoshRosen/SPARK-10984.
This test can take a little while to finish on slow / loaded machines.
Author: Marcelo Vanzin <vanzin@cloudera.com>
Closes#9235 from vanzin/SPARK-11134.
```
// My machine only has 8 cores
$ bin/spark-shell --master local[32]
scala> val df = sc.parallelize(Seq((1, 1), (2, 2))).toDF("a", "b")
scala> df.as("x").join(df.as("y"), $"x.a" === $"y.a").count()
Caused by: java.io.IOException: Unable to acquire 2097152 bytes of memory
at org.apache.spark.util.collection.unsafe.sort.UnsafeExternalSorter.acquireNewPage(UnsafeExternalSorter.java:351)
```
Author: Andrew Or <andrew@databricks.com>
Closes#9209 from andrewor14/fix-local-page-size.
There's a lot of duplication between SortShuffleManager and UnsafeShuffleManager. Given that these now provide the same set of functionality, now that UnsafeShuffleManager supports large records, I think that we should replace SortShuffleManager's serialized shuffle implementation with UnsafeShuffleManager's and should merge the two managers together.
Author: Josh Rosen <joshrosen@databricks.com>
Closes#8829 from JoshRosen/consolidate-sort-shuffle-implementations.
Correct the logic to return `HDFSCacheTaskLocation` instance when the input `str` is a in memory location.
Author: zhichao.li <zhichao.li@intel.com>
Closes#9096 from zhichao-li/uselessBranch.
JIRA: https://issues.apache.org/jira/browse/SPARK-11051
When a `RDD` is materialized and checkpointed, its partitions and dependencies are cleared. If we allow local checkpointing on it and assign `LocalRDDCheckpointData` to its `checkpointData`. Next time when the RDD is materialized again, the error will be thrown.
Author: Liang-Chi Hsieh <viirya@appier.com>
Closes#9072 from viirya/no-localcheckpoint-after-checkpoint.
Because the registration RPC was not really an RPC, but a bunch of
disconnected messages, it was possible for other messages to be
sent before the reply to the registration arrived, and that would
confuse the Worker. Especially in local-cluster mode, the worker was
succeptible to receiving an executor request before it received a
message from the master saying registration succeeded.
On top of the above, the change also fixes a ClassCastException when
the registration fails, which also affects the executor registration
protocol. Because the `ask` is issued with a specific return type,
if the error message (of a different type) was returned instead, the
code would just die with an exception. This is fixed by having a common
base trait for these reply messages.
Author: Marcelo Vanzin <vanzin@cloudera.com>
Closes#9138 from vanzin/SPARK-11131.
#9084 uncovered that many tests that test spilling don't actually spill. This is a follow-up patch to fix that to ensure our unit tests actually catch potential bugs in spilling. The size of this patch is inflated by the refactoring of `ExternalSorterSuite`, which had a lot of duplicate code and logic.
Author: Andrew Or <andrew@databricks.com>
Closes#9124 from andrewor14/spilling-tests.
If the heartbeat receiver kills executors (and new ones are not registered to replace them), the idle timeout for the old executors will be lost (and then change a total number of executors requested by Driver), So new ones will be not to asked to replace them.
For example, executorsPendingToRemove=Set(1), and executor 2 is idle timeout before a new executor is asked to replace executor 1. Then driver kill executor 2, and sending RequestExecutors to AM. But executorsPendingToRemove=Set(1,2), So AM doesn't allocate a executor to replace 1.
see: https://github.com/apache/spark/pull/8668
Author: KaiXinXiaoLei <huleilei1@huawei.com>
Author: huleilei <huleilei1@huawei.com>
Closes#8945 from KaiXinXiaoLei/pendingexecutor.
Internal accumulators don't write the internal flag to event log. So on the history server Web UI, all accumulators are not internal. This causes incorrect peak execution memory and unwanted accumulator table displayed on the stage page.
To fix it, I add the "internal" property of AccumulableInfo when writing the event log.
Author: Carson Wang <carson.wang@intel.com>
Closes#9061 from carsonwang/accumulableBug.
Restrict tasks (of job) to only 1 to ensure that the causing Exception asserted for job failure is the deliberately thrown DAGSchedulerSuiteDummyException intended, not an UnsupportedOperationException from any second/subsequent tasks that can propagate from a race condition during code execution.
Author: shellberg <sah@zepler.org>
Closes#9076 from shellberg/shellberg-DAGSchedulerSuite-misbehavedResultHandlerTest-patch-1.
A few more changes:
1. Renamed IDVerifier -> RpcEndpointVerifier
2. Renamed NettyRpcAddress -> RpcEndpointAddress
3. Simplified NettyRpcHandler a bit by removing the connection count tracking. This is OK because I now force spark.shuffle.io.numConnectionsPerPeer to 1
4. Reduced spark.rpc.connect.threads to 64. It would be great to eventually remove this extra thread pool.
5. Minor cleanup & documentation.
Author: Reynold Xin <rxin@databricks.com>
Closes#9112 from rxin/SPARK-11096.
This patch unifies the memory management of the storage and execution regions such that either side can borrow memory from each other. When memory pressure arises, storage will be evicted in favor of execution. To avoid regressions in cases where storage is crucial, we dynamically allocate a fraction of space for storage that execution cannot evict. Several configurations are introduced:
- **spark.memory.fraction (default 0.75)**: fraction of the heap space used for execution and storage. The lower this is, the more frequently spills and cached data eviction occur. The purpose of this config is to set aside memory for internal metadata, user data structures, and imprecise size estimation in the case of sparse, unusually large records.
- **spark.memory.storageFraction (default 0.5)**: size of the storage region within the space set aside by `spark.memory.fraction`. Cached data may only be evicted if total storage exceeds this region.
- **spark.memory.useLegacyMode (default false)**: whether to use the memory management that existed in Spark 1.5 and before. This is mainly for backward compatibility.
For a detailed description of the design, see [SPARK-10000](https://issues.apache.org/jira/browse/SPARK-10000). This patch builds on top of the `MemoryManager` interface introduced in #9000.
Author: Andrew Or <andrew@databricks.com>
Closes#9084 from andrewor14/unified-memory-manager.
I'm going through the implementation right now for post-doc review. Adding more comments and renaming things as I go through them.
I also want to write higher level documentation about how the whole thing works -- but those will come in other pull requests.
Author: Reynold Xin <rxin@databricks.com>
Closes#9091 from rxin/rpc-review.
https://issues.apache.org/jira/browse/SPARK-10858
The issue here is that in resolveURI we default to calling new File(path).getAbsoluteFile().toURI(). But if the path passed in already has a # in it then File(path) will think that is supposed to be part of the actual file path and not a fragment so it changes # to %23. Then when we try to parse that later in Client as a URI it doesn't recognize there is a fragment.
so to fix we just check if there is a fragment, still create the File like we did before and then add the fragment back on.
Author: Tom Graves <tgraves@yahoo-inc.com>
Closes#9035 from tgravescs/SPARK-10858.
This change adds an API that encapsulates information about an app
launched using the library. It also creates a socket-based communication
layer for apps that are launched as child processes; the launching
application listens for connections from launched apps, and once
communication is established, the channel can be used to send updates
to the launching app, or to send commands to the child app.
The change also includes hooks for local, standalone/client and yarn
masters.
Author: Marcelo Vanzin <vanzin@cloudera.com>
Closes#7052 from vanzin/SPARK-8673.
This patch introduces a `MemoryManager` that is the central arbiter of how much memory to grant to storage and execution. This patch is primarily concerned only with refactoring while preserving the existing behavior as much as possible.
This is the first step away from the existing rigid separation of storage and execution memory, which has several major drawbacks discussed on the [issue](https://issues.apache.org/jira/browse/SPARK-10956). It is the precursor of a series of patches that will attempt to address those drawbacks.
Author: Andrew Or <andrew@databricks.com>
Author: Josh Rosen <joshrosen@databricks.com>
Author: andrewor14 <andrew@databricks.com>
Closes#9000 from andrewor14/memory-manager.
This PR just reverted 02144d6745 to remerge #6457 and also included the commits in #8905.
Author: zsxwing <zsxwing@gmail.com>
Closes#8944 from zsxwing/SPARK-6028.
Compatibility between history server script and functionality
The history server has its argument parsing class in HistoryServerArguments. However, this doesn't get involved in the start-history-server.sh codepath where the $0 arg is assigned to spark.history.fs.logDirectory and all other arguments discarded (e.g --property-file.)
This stops the other options being usable from this script
Author: Joshi <rekhajoshm@gmail.com>
Author: Rekha Joshi <rekhajoshm@gmail.com>
Closes#8758 from rekhajoshm/SPARK-10317.
Fix the following issues in StandaloneDynamicAllocationSuite:
1. It should not assume master and workers start in order
2. It should not assume master and workers get ready at once
3. It should not assume the application is already registered with master after creating SparkContext
4. It should not access Master.app and idToApp which are not thread safe
The changes includes:
* Use `eventually` to wait until master and workers are ready to fix 1 and 2
* Use `eventually` to wait until the application is registered with master to fix 3
* Use `askWithRetry[MasterStateResponse](RequestMasterState)` to get the application info to fix 4
Author: zsxwing <zsxwing@gmail.com>
Closes#8914 from zsxwing/fix-StandaloneDynamicAllocationSuite.
This makes two changes:
- Allow reduce tasks to fetch multiple map output partitions -- this is a pretty small change to HashShuffleFetcher
- Move shuffle locality computation out of DAGScheduler and into ShuffledRDD / MapOutputTracker; this was needed because the code in DAGScheduler wouldn't work for RDDs that fetch multiple map output partitions from each reduce task
I also added an AdaptiveSchedulingSuite that creates RDDs depending on multiple map output partitions.
Author: Matei Zaharia <matei@databricks.com>
Closes#8844 from mateiz/spark-9852.
The DiskBlockObjectWriter constructor took a BlockId parameter but never used it. As part of some general cleanup in these interfaces, this patch refactors its constructor to eliminate this parameter.
Author: Josh Rosen <joshrosen@databricks.com>
Closes#8871 from JoshRosen/disk-block-object-writer-blockid-cleanup.
The current shuffle code has an interface named ShuffleReader with only one implementation, HashShuffleReader. This naming is confusing, since the same read path code is used for both sort- and hash-based shuffle. This patch addresses this by renaming HashShuffleReader to BlockStoreShuffleReader.
Author: Josh Rosen <joshrosen@databricks.com>
Closes#8825 from JoshRosen/shuffle-reader-cleanup.
The job group, and job descriptions information is passed through thread local properties, and get inherited by child threads. In case of spark streaming, the streaming jobs inherit these properties from the thread that called streamingContext.start(). This may not make sense.
1. Job group: This is mainly used for cancelling a group of jobs together. It does not make sense to cancel streaming jobs like this, as the effect will be unpredictable. And its not a valid usecase any way, to cancel a streaming context, call streamingContext.stop()
2. Job description: This is used to pass on nice text descriptions for jobs to show up in the UI. The job description of the thread that calls streamingContext.start() is not useful for all the streaming jobs, as it does not make sense for all of the streaming jobs to have the same description, and the description may or may not be related to streaming.
The solution in this PR is meant for the Spark master branch, where local properties are inherited by cloning the properties. The job group and job description in the thread that starts the streaming scheduler are explicitly removed, so that all the subsequent child threads does not inherit them. Also, the starting is done in a new child thread, so that setting the job group and description for streaming, does not change those properties in the thread that called streamingContext.start().
Author: Tathagata Das <tathagata.das1565@gmail.com>
Closes#8781 from tdas/SPARK-10649.
Track pending tasks by partition ID instead of Task objects.
Before this change, failure & retry could result in a case where a stage got submitted before the map output from its dependencies get registered. This was due to an error in the condition for registering map outputs.
Author: hushan[胡珊] <hushan@xiaomi.com>
Author: Imran Rashid <irashid@cloudera.com>
Closes#7699 from squito/SPARK-5259.
I noticed only one block manager registered with master in an unsuccessful build (https://amplab.cs.berkeley.edu/jenkins/job/Spark-Master-SBT/AMPLAB_JENKINS_BUILD_PROFILE=hadoop2.2,label=spark-test/3534/)
```
15/09/16 13:02:30.981 pool-1-thread-1-ScalaTest-running-BroadcastSuite INFO SparkContext: Running Spark version 1.6.0-SNAPSHOT
...
15/09/16 13:02:38.133 sparkDriver-akka.actor.default-dispatcher-19 INFO BlockManagerMasterEndpoint: Registering block manager localhost:48196 with 530.3 MB RAM, BlockManagerId(0, localhost, 48196)
```
In addition, the first block manager needed 7+ seconds to start. But the test expected 2 block managers so it failed.
However, there was no exception in this log file. So I checked a successful build (https://amplab.cs.berkeley.edu/jenkins/job/Spark-Master-SBT/3536/AMPLAB_JENKINS_BUILD_PROFILE=hadoop2.2,label=spark-test/) and it needed 4-5 seconds to set up the local cluster:
```
15/09/16 18:11:27.738 sparkWorker1-akka.actor.default-dispatcher-5 INFO Worker: Running Spark version 1.6.0-SNAPSHOT
...
15/09/16 18:11:30.838 sparkDriver-akka.actor.default-dispatcher-20 INFO BlockManagerMasterEndpoint: Registering block manager localhost:54202 with 530.3 MB RAM, BlockManagerId(1, localhost, 54202)
15/09/16 18:11:32.112 sparkDriver-akka.actor.default-dispatcher-20 INFO BlockManagerMasterEndpoint: Registering block manager localhost:32955 with 530.3 MB RAM, BlockManagerId(0, localhost, 32955)
```
In this build, the first block manager needed only 3+ seconds to start.
Comparing these two builds, I guess it's possible that the local cluster in `BroadcastSuite` cannot be ready in 10 seconds if the Jenkins worker is busy. So I just increased the timeout to 60 seconds to see if this can fix the issue.
Author: zsxwing <zsxwing@gmail.com>
Closes#8813 from zsxwing/fix-BroadcastSuite.
It does not make much sense to set `spark.shuffle.spill` or `spark.sql.planner.externalSort` to false: I believe that these configurations were initially added as "escape hatches" to guard against bugs in the external operators, but these operators are now mature and well-tested. In addition, these configurations are not handled in a consistent way anymore: SQL's Tungsten codepath ignores these configurations and will continue to use spilling operators. Similarly, Spark Core's `tungsten-sort` shuffle manager does not respect `spark.shuffle.spill=false`.
This pull request removes these configurations, adds warnings at the appropriate places, and deletes a large amount of code which was only used in code paths that did not support spilling.
Author: Josh Rosen <joshrosen@databricks.com>
Closes#8831 from JoshRosen/remove-ability-to-disable-spilling.
When speculative execution is enabled, consider a scenario where the authorized committer of a particular output partition fails during the OutputCommitter.commitTask() call. In this case, the OutputCommitCoordinator is supposed to release that committer's exclusive lock on committing once that task fails. However, due to a unit mismatch (we used task attempt number in one place and task attempt id in another) the lock will not be released, causing Spark to go into an infinite retry loop.
This bug was masked by the fact that the OutputCommitCoordinator does not have enough end-to-end tests (the current tests use many mocks). Other factors contributing to this bug are the fact that we have many similarly-named identifiers that have different semantics but the same data types (e.g. attemptNumber and taskAttemptId, with inconsistent variable naming which makes them difficult to distinguish).
This patch adds a regression test and fixes this bug by always using task attempt numbers throughout this code.
Author: Josh Rosen <joshrosen@databricks.com>
Closes#8544 from JoshRosen/SPARK-10381.
*Note: this is for master branch only.* The fix for branch-1.5 is at #8721.
The query execution ID is currently passed from a thread to its children, which is not the intended behavior. This led to `IllegalArgumentException: spark.sql.execution.id is already set` when running queries in parallel, e.g.:
```
(1 to 100).par.foreach { _ =>
sc.parallelize(1 to 5).map { i => (i, i) }.toDF("a", "b").count()
}
```
The cause is `SparkContext`'s local properties are inherited by default. This patch adds a way to exclude keys we don't want to be inherited, and makes SQL go through that code path.
Author: Andrew Or <andrew@databricks.com>
Closes#8710 from andrewor14/concurrent-sql-executions.
This patch adds support for submitting map stages in a DAG individually so that we can make downstream decisions after seeing statistics about their output, as part of SPARK-9850. I also added more comments to many of the key classes in DAGScheduler. By itself, the patch is not super useful except maybe to switch between a shuffle and broadcast join, but with the other subtasks of SPARK-9850 we'll be able to do more interesting decisions.
The main entry point is SparkContext.submitMapStage, which lets you run a map stage and see stats about the map output sizes. Other stats could also be collected through accumulators. See AdaptiveSchedulingSuite for a short example.
Author: Matei Zaharia <matei@databricks.com>
Closes#8180 from mateiz/spark-9851.
This is a follow-up patch to #8723. I missed one case there.
Author: Andrew Or <andrew@databricks.com>
Closes#8727 from andrewor14/fix-threading-suite.
This is a followup to #8499 which adds a Scalastyle rule to mandate the use of SparkHadoopUtil's JobContext accessor methods and fixes the existing violations.
Author: Josh Rosen <joshrosen@databricks.com>
Closes#8521 from JoshRosen/SPARK-10330-part2.
Fix a few Java API test style issues: unused generic types, exceptions, wrong assert argument order
Author: Sean Owen <sowen@cloudera.com>
Closes#8706 from srowen/SPARK-10547.
This commit ensures if an assertion fails within a thread, it will ultimately fail the test. Otherwise we end up potentially masking real bugs by not propagating assertion failures properly.
Author: Andrew Or <andrew@databricks.com>
Closes#8723 from andrewor14/fix-threading-suite.
ShuffleManager implementations are currently not given type information for
the key, value and combiner classes. Serialization of shuffle objects relies
on objects being JavaSerializable, with methods defined for reading/writing
the object or, alternatively, serialization via Kryo which uses reflection.
Serialization systems like Avro, Thrift and Protobuf generate classes with
zero argument constructors and explicit schema information
(e.g. IndexedRecords in Avro have get, put and getSchema methods).
By serializing the key, value and combiner class names in ShuffleDependency,
shuffle implementations will have access to schema information when
registerShuffle() is called.
Author: Matt Massie <massie@cs.berkeley.edu>
Closes#7403 from massie/shuffle-classtags.
This is a regression introduced in #4960, this commit fixes it and adds a test.
tnachen andrewor14 please review, this should be an easy one.
Author: Iulian Dragos <jaguarul@gmail.com>
Closes#8653 from dragos/issue/mesos/fine-grained-maxExecutorCores.
The architecture is that, in YARN mode, if the driver detects that an executor has disconnected, it asks the ApplicationMaster why the executor died. If the ApplicationMaster is aware that the executor died because of preemption, all tasks associated with that executor are not marked as failed. The executor
is still removed from the driver's list of available executors, however.
There's a few open questions:
1. Should standalone mode have a similar "get executor loss reason" as well? I localized this change as much as possible to affect only YARN, but there could be a valid case to differentiate executor losses in standalone mode as well.
2. I make a pretty strong assumption in YarnAllocator that getExecutorLossReason(executorId) will only be called once per executor id; I do this so that I can remove the metadata from the in-memory map to avoid object accumulation. It's not clear if I'm being overly zealous to save space, however.
cc vanzin specifically for review because it collided with some earlier YARN scheduling work.
cc JoshRosen because it's similar to output commit coordination we did in the past
cc andrewor14 for our discussion on how to get executor exit codes and loss reasons
Author: mcheah <mcheah@palantir.com>
Closes#8007 from mccheah/feature/preemption-handling.
We introduced the Netty network module for shuffle in Spark 1.2, and has turned it on by default for 3 releases. The old ConnectionManager is difficult to maintain. If we merge the patch now, by the time it is released, it would be 1 yr for which ConnectionManager is off by default. It's time to remove it.
Author: Reynold Xin <rxin@databricks.com>
Closes#8161 from rxin/SPARK-9767.
[SPARK-9591](https://issues.apache.org/jira/browse/SPARK-9591)
When we getting the broadcast variable, we can fetch the block form several location,but now when connecting the lost blockmanager(idle for enough time removed by driver when using dynamic resource allocate and so on) will cause task fail,and the worse case will cause the job fail.
Author: jeanlyn <jeanlyn92@gmail.com>
Closes#7927 from jeanlyn/catch_exception.
This contribution is my original work and I license the work to the project under the project's open source license.
Author: Pat Shields <yeoldefortran@gmail.com>
Closes#7979 from pashields/env-loading-on-driver.
This is pretty minor, just trying to improve the readability of `DAGSchedulerSuite`, I figure every bit helps. Before whenever I read this test, I never knew what "should work" and "should be ignored" really meant -- this adds some asserts & updates comments to make it more clear. Also some reformatting per a suggestion from markhamstra on https://github.com/apache/spark/pull/7699
Author: Imran Rashid <irashid@cloudera.com>
Closes#8434 from squito/SPARK-10247.
The ```Stage``` class now tracks whether there were a sufficient number of consecutive failures of that stage to trigger an abort.
To avoid an infinite loop of stage retries, we abort the job completely after 4 consecutive stage failures for one stage. We still allow more than 4 consecutive stage failures if there is an intervening successful attempt for the stage, so that in very long-lived applications, where a stage may get reused many times, we don't abort the job after failures that have been recovered from successfully.
I've added test cases to exercise the most obvious scenarios.
Author: Ilya Ganelin <ilya.ganelin@capitalone.com>
Closes#5636 from ilganeli/SPARK-5945.
SPARK-4223.
Currently we support setting view and modify acls but you have to specify a list of users. It would be nice to support * meaning all users have access.
Manual tests to verify that: "*" works for any user in:
a. Spark ui: view and kill stage. Done.
b. Spark history server. Done.
c. Yarn application killing. Done.
Author: zhuol <zhuol@yahoo-inc.com>
Closes#8398 from zhuoliu/4223.
In SMJ, the first ExternalSorter could consume all the memory before spilling, then the second can not even acquire the first page.
Before we have a better memory allocator, SMJ should call prepare() before call any compute() of it's children.
cc rxin JoshRosen
Author: Davies Liu <davies@databricks.com>
Closes#8511 from davies/smj_memory.
This change aims at speeding up the dev cycle a little bit, by making
sure that all tests behave the same w.r.t. where the code to be tested
is loaded from. Namely, that means that tests don't rely on the assembly
anymore, rather loading all needed classes from the build directories.
The main change is to make sure all build directories (classes and test-classes)
are added to the classpath of child processes when running tests.
YarnClusterSuite required some custom code since the executors are run
differently (i.e. not through the launcher library, like standalone and
Mesos do).
I also found a couple of tests that could leak a SparkContext on failure,
and added code to handle those.
With this patch, it's possible to run the following command from a clean
source directory and have all tests pass:
mvn -Pyarn -Phadoop-2.4 -Phive-thriftserver install
Author: Marcelo Vanzin <vanzin@cloudera.com>
Closes#7629 from vanzin/SPARK-9284.
Replace `JavaConversions` implicits with `JavaConverters`
Most occurrences I've seen so far are necessary conversions; a few have been avoidable. None are in critical code as far as I see, yet.
Author: Sean Owen <sowen@cloudera.com>
Closes#8033 from srowen/SPARK-9613.
The peak execution memory metric was introduced in SPARK-8735. That was before Tungsten was enabled by default, so it assumed that `spark.sql.unsafe.enabled` must be explicitly set to true. The result is that the memory is not displayed by default.
Author: Andrew Or <andrew@databricks.com>
Closes#8345 from andrewor14/show-memory-default.
https://issues.apache.org/jira/browse/SPARK-9439
In general, Yarn apps should be robust to NodeManager restarts. However, if you run spark with the external shuffle service on, after a NM restart all shuffles fail, b/c the shuffle service has lost some state with info on each executor. (Note the shuffle data is perfectly fine on disk across a NM restart, the problem is we've lost the small bit of state that lets us *find* those files.)
The solution proposed here is that the external shuffle service can write out its state to leveldb (backed by a local file) every time an executor is added. When running with yarn, that file is in the NM's local dir. Whenever the service is started, it looks for that file, and if it exists, it reads the file and re-registers all executors there.
Nothing is changed in non-yarn modes with this patch. The service is not given a place to save the state to, so it operates the same as before. This should make it easy to update other cluster managers as well, by just supplying the right file & the equivalent of yarn's `initializeApplication` -- I'm not familiar enough with those modes to know how to do that.
Author: Imran Rashid <irashid@cloudera.com>
Closes#7943 from squito/leveldb_external_shuffle_service_NM_restart and squashes the following commits:
0d285d3 [Imran Rashid] review feedback
70951d6 [Imran Rashid] Merge branch 'master' into leveldb_external_shuffle_service_NM_restart
5c71c8c [Imran Rashid] save executor to db before registering; style
2499c8c [Imran Rashid] explicit dependency on jackson-annotations
795d28f [Imran Rashid] review feedback
81f80e2 [Imran Rashid] Merge branch 'master' into leveldb_external_shuffle_service_NM_restart
594d520 [Imran Rashid] use json to serialize application executor info
1a7980b [Imran Rashid] version
8267d2a [Imran Rashid] style
e9f99e8 [Imran Rashid] cleanup the handling of bad dbs a little
9378ba3 [Imran Rashid] fail gracefully on corrupt leveldb files
acedb62 [Imran Rashid] switch to writing out one record per executor
79922b7 [Imran Rashid] rely on yarn to call stopApplication; assorted cleanup
12b6a35 [Imran Rashid] save registered executors when apps are removed; add tests
c878fbe [Imran Rashid] better explanation of shuffle service port handling
694934c [Imran Rashid] only open leveldb connection once per service
d596410 [Imran Rashid] store executor data in leveldb
59800b7 [Imran Rashid] Files.move in case renaming is unsupported
32fe5ae [Imran Rashid] Merge branch 'master' into external_shuffle_service_NM_restart
d7450f0 [Imran Rashid] style
f729e2b [Imran Rashid] debugging
4492835 [Imran Rashid] lol, dont use a PrintWriter b/c of scalastyle checks
0a39b98 [Imran Rashid] Merge branch 'master' into external_shuffle_service_NM_restart
55f49fc [Imran Rashid] make sure the service doesnt die if the registered executor file is corrupt; add tests
245db19 [Imran Rashid] style
62586a6 [Imran Rashid] just serialize the whole executors map
bdbbf0d [Imran Rashid] comments, remove some unnecessary changes
857331a [Imran Rashid] better tests & comments
bb9d1e6 [Imran Rashid] formatting
bdc4b32 [Imran Rashid] rename
86e0cb9 [Imran Rashid] for tests, shuffle service finds an open port
23994ff [Imran Rashid] style
7504de8 [Imran Rashid] style
a36729c [Imran Rashid] cleanup
efb6195 [Imran Rashid] proper unit test, and no longer leak if apps stop during NM restart
dd93dc0 [Imran Rashid] test for shuffle service w/ NM restarts
d596969 [Imran Rashid] cleanup imports
0e9d69b [Imran Rashid] better names
9eae119 [Imran Rashid] cleanup lots of duplication
1136f44 [Imran Rashid] test needs to have an actual shuffle
0b588bd [Imran Rashid] more fixes ...
ad122ef [Imran Rashid] more fixes
5e5a7c3 [Imran Rashid] fix build
c69f46b [Imran Rashid] maybe working version, needs tests & cleanup ...
bb3ba49 [Imran Rashid] minor cleanup
36127d3 [Imran Rashid] wip
b9d2ced [Imran Rashid] incomplete setup for external shuffle service tests
In Scala, `Seq.fill` always seems to return a List. Accessing a list by index is an O(N) operation. Thus, the following code will be really slow (~10 seconds on my machine):
```scala
val numItems = 100000
val s = Seq.fill(numItems)(1)
for (i <- 0 until numItems) s(i)
```
It turns out that we had a loop like this in DAGScheduler code, although it's a little tricky to spot. In `getPreferredLocsInternal`, there's a call to `getCacheLocs(rdd)(partition)`. The `getCacheLocs` call returns a Seq. If this Seq is a List and the RDD contains many partitions, then indexing into this list will cost O(partitions). Thus, when we loop over our tasks to compute their individual preferred locations we implicitly perform an N^2 loop, reducing scheduling throughput.
This patch fixes this by replacing `Seq` with `Array`.
Author: Josh Rosen <joshrosen@databricks.com>
Closes#8178 from JoshRosen/dagscheduler-perf.
The shuffle locality patch made the DAGScheduler aware of shuffle data,
but for RDDs that have both narrow and shuffle dependencies, it can
cause them to place tasks based on the shuffle dependency instead of the
narrow one. This case is common in iterative join-based algorithms like
PageRank and ALS, where one RDD is hash-partitioned and one isn't.
Author: Matei Zaharia <matei@databricks.com>
Closes#8220 from mateiz/shuffle-loc-fix.
In these tests, we use a custom listener and we assert on fields in the stage / task completion events. However, these events are posted in a separate thread so they're not guaranteed to be posted in time. This commit fixes this flakiness through a job end registration callback.
Author: Andrew Or <andrew@databricks.com>
Closes#8176 from andrewor14/fix-accumulator-suite.
I think that we should pass additional configuration flags to disable the driver UI and Master REST server in SparkSubmitSuite and HiveSparkSubmitSuite. This might cut down on port-contention-related flakiness in Jenkins.
Author: Josh Rosen <joshrosen@databricks.com>
Closes#8124 from JoshRosen/disable-ui-in-sparksubmitsuite.
… allocation are set. Now, dynamic allocation is set to false when num-executors is explicitly specified as an argument. Consequently, executorAllocationManager in not initialized in the SparkContext.
Author: Niranjan Padmanabhan <niranjan.padmanabhan@cloudera.com>
Closes#7657 from neurons/SPARK-9092.