As part of the initiative of preventing data loss on streaming driver failure, this sub-task implements a BlockRDD that is backed by HDFS. This BlockRDD can either read data from the Spark's BlockManager, or read the data from file-segments in write ahead log in HDFS.
Most of this code has been written by @harishreedharan
Author: Tathagata Das <tathagata.das1565@gmail.com>
Author: Hari Shreedharan <hshreedharan@apache.org>
Closes#2931 from tdas/driver-ha-rdd and squashes the following commits:
209e49c [Tathagata Das] Better fix to style issue.
4a5866f [Tathagata Das] Addressed one more comment.
ed5fbf0 [Tathagata Das] Minor updates.
b0a18b1 [Tathagata Das] Fixed import order.
20aa7c6 [Tathagata Das] Fixed more line length issues.
29aa099 [Tathagata Das] Fixed line length issues.
9e47b5b [Tathagata Das] Renamed class, simplified+added unit tests.
6e1bfb8 [Tathagata Das] Tweaks testuite to create spark contxt lazily to prevent contxt leaks.
9c86a61 [Tathagata Das] Merge pull request #22 from harishreedharan/driver-ha-rdd
2878c38 [Hari Shreedharan] Shutdown spark context after tests. Formatting/minor fixes
c709f2f [Tathagata Das] Merge pull request #21 from harishreedharan/driver-ha-rdd
5cce16f [Hari Shreedharan] Make sure getBlockLocations uses offset and length to find the blocks on HDFS
eadde56 [Tathagata Das] Transferred HDFSBackedBlockRDD for the driver-ha-working branch
This method is not implemented by the only subclass
(HashShuffleReader), nor is it ever called. While the
use of Scala's fancy "???" was pretty exciting, the method's
existence can only lead to confusion and it therefore should
be deleted.
mateiz was there a reason for adding this that I'm
missing?
Author: Kay Ousterhout <kayousterhout@gmail.com>
Closes#2966 from kayousterhout/SPARK-4102 and squashes the following commits:
532c564 [Kay Ousterhout] Added back commented-out method, as per Matei's request
904655e [Kay Ousterhout] [SPARK-4102] Remove unused ShuffleReader.stop() method.
This is part of a bigger effort to provide elastic scaling of executors within a Spark application ([SPARK-3174](https://issues.apache.org/jira/browse/SPARK-3174)). This PR does not provide any functionality by itself; it is a skeleton that is missing a mechanism to be added later in [SPARK-3822](https://issues.apache.org/jira/browse/SPARK-3822).
Comments and feedback are most welcome. For those of you reviewing this in detail, I highly recommend doing it through your favorite IDE instead of through the diff here.
Author: Andrew Or <andrewor14@gmail.com>
Author: Andrew Or <andrew@databricks.com>
Closes#2746 from andrewor14/scaling-heuristics and squashes the following commits:
8a4fdaa [Andrew Or] Merge branch 'master' of github.com:apache/spark into scaling-heuristics
e045df8 [Andrew Or] Add warning message (minor)
dfa31ec [Andrew Or] Fix tests
c0becc4 [Andrew Or] Merging with SPARK-3822
4784f93 [Andrew Or] Reword an awkward log message
181f27f [Andrew Or] Merge branch 'master' of github.com:apache/spark into scaling-heuristics
c79e907 [Andrew Or] Merge branch 'master' of github.com:apache/spark into scaling-heuristics
4672b90 [Andrew Or] It's nano time.
a6a30f2 [Andrew Or] Do not allow min/max executors of 0
c60ec33 [Andrew Or] Rewrite test logic with clocks
b00b680 [Andrew Or] Fix style
c3caa65 [Andrew Or] Merge branch 'master' of github.com:apache/spark into scaling-heuristics
7f9da14 [Andrew Or] Factor out logic to verify bounds on # executors (minor)
f279019 [Andrew Or] Add time mocking tests for polling loop
685e347 [Andrew Or] Factor out clock in polling loop to facilitate testing
3cea7f7 [Andrew Or] Use PrivateMethodTester to keep original class private
3156d81 [Andrew Or] Update comments and exception messages
92f36f9 [Andrew Or] Address minor review comments
abdea61 [Andrew Or] Merge branch 'master' of github.com:apache/spark into scaling-heuristics
2aefd09 [Andrew Or] Correct listener behavior
9fe6e44 [Andrew Or] Rename variables and configs + update comments and log messages
149cc32 [Andrew Or] Fix style
254c958 [Andrew Or] Merge branch 'master' of github.com:apache/spark into scaling-heuristics
5ff829b [Andrew Or] Add tests for ExecutorAllocationManager
19c6c4b [Andrew Or] Merge branch 'master' of github.com:apache/spark into scaling-heuristics
5896515 [Andrew Or] Move ExecutorAllocationManager out of scheduler package
9ca8945 [Andrew Or] Rewrite callbacks through the listener interface
5e336b9 [Andrew Or] Remove code from backend to avoid conflict with SPARK-3822
092d1fd [Andrew Or] Remove timeout logic for pending requests
1309fab [Andrew Or] Request executors by specifying the number pending
8bc0e9d [Andrew Or] Add logic to expire pending requests after timeouts
b750ee1 [Andrew Or] Express timers in terms of expiration times + remove retry logic
7f8dd47 [Andrew Or] Merge branch 'master' of github.com:apache/spark into scaling-heuristics
9d516cc [Andrew Or] Bug fix: Actually trigger the add timer / add retry timer
44f1832 [Andrew Or] Rename configs to include time units
eaae7ef [Andrew Or] Address various review comments
6f8be6c [Andrew Or] Beef up comments on what each of the timers mean
baaa403 [Andrew Or] Simplify variable names (minor)
42beec8 [Andrew Or] Reset whether the add threshold is crossed on cancellation
9bcc0bc [Andrew Or] ExecutorScalingManager -> ExecutorAllocationManager
2784398 [Andrew Or] Merge branch 'master' of github.com:apache/spark into scaling-heuristics
5a97d9e [Andrew Or] Log retry attempts in INFO + clean up logging
2f55c9f [Andrew Or] Do not keep requesting executors even after max attempts
0acd1cb [Andrew Or] Rewrite timer logic with polling
b3c7d44 [Andrew Or] Start the retry timer for adding executors at the right time
9b5f2ea [Andrew Or] Wording changes in comments and log messages
c2203a5 [Andrew Or] Simplify code to access the scheduler backend
e519d08 [Andrew Or] Simplify initialization code
2cc87a7 [Andrew Or] Add retry logic for removing executors
d0b34a6 [Andrew Or] Add retry logic for adding executors
9cc4649 [Andrew Or] Simplifying synchronization logic
67c03c7 [Andrew Or] Correct semantics of adding executors + update comments
6c48ab0 [Andrew Or] Update synchronization comment
8901900 [Andrew Or] Simplify remove policy + change the semantics of add policy
1cc8444 [Andrew Or] Minor wording change
ae5b64a [Andrew Or] Add synchronization
20ec6b9 [Andrew Or] First cut implementation of removing executors dynamically
4077ae2 [Andrew Or] Minor code re-organization
6f1fa66 [Andrew Or] First cut implementation of adding executors dynamically
b2e6dcc [Andrew Or] Add skeleton interface for requesting / killing executors
There is a chance that `thread` is null when calling `thread.interrupt()`.
```Scala
override def cancel(): Unit = this.synchronized {
_cancelled = true
if (thread != null) {
thread.interrupt()
}
}
```
Should put `thread = null` into a `synchronized` block to fix the race condition.
Author: zsxwing <zsxwing@gmail.com>
Closes#2957 from zsxwing/SPARK-4097 and squashes the following commits:
edf0aee [zsxwing] Add comments to explain the lock
c5cfeca [zsxwing] Fix the race condition of 'thread'
This is part of a broader effort to enable dynamic scaling of executors ([SPARK-3174](https://issues.apache.org/jira/browse/SPARK-3174)). This is intended to work alongside SPARK-3795 (#2746), SPARK-3796 and SPARK-3797, but is functionally independently of these other issues.
The logic is built on top of PraveenSeluka's changes at #2798. This is different from the changes there in a few major ways: (1) the mechanism is implemented within the existing scheduler backend framework rather than in new `Actor` classes. This also introduces a parent abstract class `YarnSchedulerBackend` to encapsulate common logic to communicate with the Yarn `ApplicationMaster`. (2) The interface of requesting executors exposed to the `SparkContext` is the same, but the communication between the scheduler backend and the AM uses total number executors desired instead of an incremental number. This is discussed in #2746 and explained in the comments in the code.
I have tested this significantly on a stable Yarn cluster.
------------
A remaining task for this issue is to tone down the error messages emitted when an executor is removed.
Currently, `SparkContext` and its components react as if the executor has failed, resulting in many scary error messages and eventual timeouts. While it's not strictly necessary to fix this as of the first-cut implementation of this mechanism, it would be good to add logic to distinguish this case. I prefer to address this in a separate PR. I have filed a separate JIRA for this task at SPARK-4134.
Author: Andrew Or <andrew@databricks.com>
Author: Andrew Or <andrewor14@gmail.com>
Closes#2840 from andrewor14/yarn-scaling-mechanism and squashes the following commits:
485863e [Andrew Or] Minor log message changes
4920be8 [Andrew Or] Clarify that public API is only for Yarn mode for now
1c57804 [Andrew Or] Reword a few comments + other review comments
6321140 [Andrew Or] Merge branch 'master' of github.com:apache/spark into yarn-scaling-mechanism
02836c0 [Andrew Or] Limit scope of synchronization
4e2ed7f [Andrew Or] Fix bug: keep track of removed executors properly
73ade46 [Andrew Or] Wording changes (minor)
2a7a6da [Andrew Or] Add `sc.killExecutor` as a shorthand (minor)
665f229 [Andrew Or] Mima excludes
79aa2df [Andrew Or] Simplify the request interface by asking for a total
04f625b [Andrew Or] Fix race condition that causes over-allocation of executors
f4783f8 [Andrew Or] Change the semantics of requesting executors
005a124 [Andrew Or] Fix tests
4628b16 [Andrew Or] Merge branch 'master' of github.com:apache/spark into yarn-scaling-mechanism
db4a679 [Andrew Or] Merge branch 'master' of github.com:apache/spark into yarn-scaling-mechanism
572f5c5 [Andrew Or] Unused import (minor)
f30261c [Andrew Or] Kill multiple executors rather than one at a time
de260d9 [Andrew Or] Simplify by skipping useless null check
9c52542 [Andrew Or] Simplify by skipping the TaskSchedulerImpl
97dd1a8 [Andrew Or] Merge branch 'master' of github.com:apache/spark into yarn-scaling-mechanism
d987b3e [Andrew Or] Move addWebUIFilters to Yarn scheduler backend
7b76d0a [Andrew Or] Expose mechanism in SparkContext as developer API
47466cd [Andrew Or] Refactor common Yarn scheduler backend logic
c4dfaac [Andrew Or] Avoid thrashing when removing executors
53e8145 [Andrew Or] Start yarn actor early to listen for AM registration message
bbee669 [Andrew Or] Add mechanism in yarn client mode
This PR encapsulates #2330, which is itself a continuation of #2240. The first goal of this PR is to provide an alternate, simpler implementation of the ConnectionManager which is based on Netty.
In addition to this goal, however, we want to resolve [SPARK-3796](https://issues.apache.org/jira/browse/SPARK-3796), which calls for a standalone shuffle service which can be integrated into the YARN NodeManager, Standalone Worker, or on its own. This PR makes the first step in this direction by ensuring that the actual Netty service is as small as possible and extracted from Spark core. Given this, we should be able to construct this standalone jar which can be included in other JVMs without incurring significant dependency or runtime issues. The actual work to ensure that such a standalone shuffle service would work in Spark will be left for a future PR, however.
In order to minimize dependencies and allow for the service to be long-running (possibly much longer-running than Spark, and possibly having to support multiple version of Spark simultaneously), the entire service has been ported to Java, where we have full control over the binary compatibility of the components and do not depend on the Scala runtime or version.
These issues: have been addressed by folding in #2330:
SPARK-3453: Refactor Netty module to use BlockTransferService interface
SPARK-3018: Release all buffers upon task completion/failure
SPARK-3002: Create a connection pool and reuse clients across different threads
SPARK-3017: Integration tests and unit tests for connection failures
SPARK-3049: Make sure client doesn't block when server/connection has error(s)
SPARK-3502: SO_RCVBUF and SO_SNDBUF should be bootstrap childOption, not option
SPARK-3503: Disable thread local cache in PooledByteBufAllocator
TODO before mergeable:
- [x] Implement uploadBlock()
- [x] Unit tests for RPC side of code
- [x] Performance testing (see comments [here](https://github.com/apache/spark/pull/2753#issuecomment-59475022))
- [x] Turn OFF by default (currently on for unit testing)
Author: Reynold Xin <rxin@apache.org>
Author: Aaron Davidson <aaron@databricks.com>
Author: cocoatomo <cocoatomo77@gmail.com>
Author: Patrick Wendell <pwendell@gmail.com>
Author: Prashant Sharma <prashant.s@imaginea.com>
Author: Davies Liu <davies.liu@gmail.com>
Author: Anand Avati <avati@redhat.com>
Closes#2753 from aarondav/netty and squashes the following commits:
cadfd28 [Aaron Davidson] Turn netty off by default
d7be11b [Aaron Davidson] Turn netty on by default
4a204b8 [Aaron Davidson] Fail block fetches if client connection fails
2b0d1c0 [Aaron Davidson] 100ch
0c5bca2 [Aaron Davidson] Merge branch 'master' of https://github.com/apache/spark into netty
14e37f7 [Aaron Davidson] Address Reynold's comments
8dfcceb [Aaron Davidson] Merge branch 'master' of https://github.com/apache/spark into netty
322dfc1 [Aaron Davidson] Address Reynold's comments, including major rename
e5675a4 [Aaron Davidson] Fail outstanding RPCs as well
ccd4959 [Aaron Davidson] Don't throw exception if client immediately fails
9da0bc1 [Aaron Davidson] Add RPC unit tests
d236dfd [Aaron Davidson] Remove no-op serializer :)
7b7a26c [Aaron Davidson] Fix Nio compile issue
dd420fd [Aaron Davidson] Merge branch 'master' of https://github.com/apache/spark into netty-test
939f276 [Aaron Davidson] Attempt to make comm. bidirectional
aa58f67 [cocoatomo] [SPARK-3909][PySpark][Doc] A corrupted format in Sphinx documents and building warnings
8dc1ded [cocoatomo] [SPARK-3867][PySpark] ./python/run-tests failed when it run with Python 2.6 and unittest2 is not installed
5b5dbe6 [Prashant Sharma] [SPARK-2924] Required by scala 2.11, only one fun/ctor amongst overriden alternatives, can have default argument(s).
2c5d9dc [Patrick Wendell] HOTFIX: Fix build issue with Akka 2.3.4 upgrade.
020691e [Davies Liu] [SPARK-3886] [PySpark] use AutoBatchedSerializer by default
ae4083a [Anand Avati] [SPARK-2805] Upgrade Akka to 2.3.4
29c6dcf [Aaron Davidson] [SPARK-3453] Netty-based BlockTransferService, extracted from Spark core
f7e7568 [Reynold Xin] Fixed spark.shuffle.io.receiveBuffer setting.
5d98ce3 [Reynold Xin] Flip buffer.
f6c220d [Reynold Xin] Merge with latest master.
407e59a [Reynold Xin] Fix style violation.
a0518c7 [Reynold Xin] Implemented block uploads.
4b18db2 [Reynold Xin] Copy the buffer in fetchBlockSync.
bec4ea2 [Reynold Xin] Removed OIO and added num threads settings.
1bdd7ee [Reynold Xin] Fixed tests.
d68f328 [Reynold Xin] Logging close() in case close() fails.
f63fb4c [Reynold Xin] Add more debug message.
6afc435 [Reynold Xin] Added logging.
c066309 [Reynold Xin] Implement java.io.Closeable interface.
519d64d [Reynold Xin] Mark private package visibility and MimaExcludes.
f0a16e9 [Reynold Xin] Fixed test hanging.
14323a5 [Reynold Xin] Removed BlockManager.getLocalShuffleFromDisk.
b2f3281 [Reynold Xin] Added connection pooling.
d23ed7b [Reynold Xin] Incorporated feedback from Norman: - use same pool for boss and worker - remove ioratio - disable caching of byte buf allocator - childoption sendbuf/receivebuf - fire exception through pipeline
9e0cb87 [Reynold Xin] Fixed BlockClientHandlerSuite
5cd33d7 [Reynold Xin] Fixed style violation.
cb589ec [Reynold Xin] Added more test cases covering cleanup when fault happens in ShuffleBlockFetcherIteratorSuite
1be4e8e [Reynold Xin] Shorten NioManagedBuffer and NettyManagedBuffer class names.
108c9ed [Reynold Xin] Forgot to add TestSerializer to the commit list.
b5c8d1f [Reynold Xin] Fixed ShuffleBlockFetcherIteratorSuite.
064747b [Reynold Xin] Reference count buffers and clean them up properly.
2b44cf1 [Reynold Xin] Added more documentation.
1760d32 [Reynold Xin] Use Epoll.isAvailable in BlockServer as well.
165eab1 [Reynold Xin] [SPARK-3453] Refactor Netty module to use BlockTransferService.
Sorter uses generic-typed key for sorting. When data is large, it creates lots of key objects, which is not efficient. We should reuse the key in Sorter for memory efficiency. This change is part of the petabyte sort implementation from rxin .
The `Sorter` class was written in Java and marked package private. So it is only available to `org.apache.spark.util.collection`. I renamed it to `TimSort` and add a simple wrapper of it, still called `Sorter`, in Scala, which is `private[spark]`.
The benchmark code is updated, which now resets the array before each run. Here is the result on sorting primitive Int arrays of size 25 million using Sorter:
~~~
[info] - Sorter benchmark for key-value pairs !!! IGNORED !!!
Java Arrays.sort() on non-primitive int array: Took 13237 ms
Java Arrays.sort() on non-primitive int array: Took 13320 ms
Java Arrays.sort() on non-primitive int array: Took 15718 ms
Java Arrays.sort() on non-primitive int array: Took 13283 ms
Java Arrays.sort() on non-primitive int array: Took 13267 ms
Java Arrays.sort() on non-primitive int array: Took 15122 ms
Java Arrays.sort() on non-primitive int array: Took 15495 ms
Java Arrays.sort() on non-primitive int array: Took 14877 ms
Java Arrays.sort() on non-primitive int array: Took 16429 ms
Java Arrays.sort() on non-primitive int array: Took 14250 ms
Java Arrays.sort() on non-primitive int array: (13878 ms first try, 14499 ms average)
Java Arrays.sort() on primitive int array: Took 2683 ms
Java Arrays.sort() on primitive int array: Took 2683 ms
Java Arrays.sort() on primitive int array: Took 2701 ms
Java Arrays.sort() on primitive int array: Took 2746 ms
Java Arrays.sort() on primitive int array: Took 2685 ms
Java Arrays.sort() on primitive int array: Took 2735 ms
Java Arrays.sort() on primitive int array: Took 2669 ms
Java Arrays.sort() on primitive int array: Took 2693 ms
Java Arrays.sort() on primitive int array: Took 2680 ms
Java Arrays.sort() on primitive int array: Took 2642 ms
Java Arrays.sort() on primitive int array: (2948 ms first try, 2691 ms average)
Sorter without key reuse on primitive int array: Took 10732 ms
Sorter without key reuse on primitive int array: Took 12482 ms
Sorter without key reuse on primitive int array: Took 10718 ms
Sorter without key reuse on primitive int array: Took 12650 ms
Sorter without key reuse on primitive int array: Took 10747 ms
Sorter without key reuse on primitive int array: Took 10783 ms
Sorter without key reuse on primitive int array: Took 12721 ms
Sorter without key reuse on primitive int array: Took 10604 ms
Sorter without key reuse on primitive int array: Took 10622 ms
Sorter without key reuse on primitive int array: Took 11843 ms
Sorter without key reuse on primitive int array: (11089 ms first try, 11390 ms average)
Sorter with key reuse on primitive int array: Took 5141 ms
Sorter with key reuse on primitive int array: Took 5298 ms
Sorter with key reuse on primitive int array: Took 5066 ms
Sorter with key reuse on primitive int array: Took 5164 ms
Sorter with key reuse on primitive int array: Took 5203 ms
Sorter with key reuse on primitive int array: Took 5274 ms
Sorter with key reuse on primitive int array: Took 5186 ms
Sorter with key reuse on primitive int array: Took 5159 ms
Sorter with key reuse on primitive int array: Took 5164 ms
Sorter with key reuse on primitive int array: Took 5078 ms
Sorter with key reuse on primitive int array: (5311 ms first try, 5173 ms average)
~~~
So with key reuse, it is faster and less likely to trigger GC.
Author: Xiangrui Meng <meng@databricks.com>
Author: Reynold Xin <rxin@apache.org>
Closes#2937 from mengxr/SPARK-4084 and squashes the following commits:
d73c3d0 [Xiangrui Meng] address comments
0b7b682 [Xiangrui Meng] fix mima
a72f53c [Xiangrui Meng] update timeIt
38ba50c [Xiangrui Meng] update timeIt
720f731 [Xiangrui Meng] add doc about JIT specialization
78f2879 [Xiangrui Meng] update tests
7de2efd [Xiangrui Meng] update the Sorter benchmark code to be correct
8626356 [Xiangrui Meng] add prepare to timeIt and update testsin SorterSuite
5f0d530 [Xiangrui Meng] update method modifiers of SortDataFormat
6ffbe66 [Xiangrui Meng] rename Sorter to TimSort and add a Scala wrapper that is private[spark]
b00db4d [Xiangrui Meng] doc and tests
cf94e8a [Xiangrui Meng] renaming
464ddce [Reynold Xin] cherry-pick rxin's commit
A global UTF8 constant is very helpful to handle encoding problems when converting between String and bytes. There are several solutions here:
1. Add `val UTF_8 = Charset.forName("UTF-8")` to Utils.scala
2. java.nio.charset.StandardCharsets.UTF_8 (require JDK7)
3. io.netty.util.CharsetUtil.UTF_8
4. com.google.common.base.Charsets.UTF_8
5. org.apache.commons.lang.CharEncoding.UTF_8
6. org.apache.commons.lang3.CharEncoding.UTF_8
IMO, I prefer option 1) because people can find it easily.
This is a PR for option 1) and only fixes Spark Core.
Author: zsxwing <zsxwing@gmail.com>
Closes#2781 from zsxwing/SPARK-3922 and squashes the following commits:
f974edd [zsxwing] Merge branch 'master' into SPARK-3922
2d27423 [zsxwing] Refactor spark-core to use Refactor spark-core to use Utils.UTF_8
This avoids reading torrent broadcast variables when they are referenced in the closure but not used in the closure. This is done by using a `lazy val` to read broadcast blocks
cc rxin JoshRosen for review
Author: Shivaram Venkataraman <shivaram@cs.berkeley.edu>
Closes#2871 from shivaram/broadcast-read-value and squashes the following commits:
1456d65 [Shivaram Venkataraman] Use getUsedTimeMs and remove readObject
d6c5ee9 [Shivaram Venkataraman] Use laxy val to implement readBroadcastBlock
0b34df7 [Shivaram Venkataraman] Merge branch 'master' of https://github.com/apache/spark into broadcast-read-value
9cec507 [Shivaram Venkataraman] Test if broadcast variables are read lazily
768b40b [Shivaram Venkataraman] Merge branch 'master' of https://github.com/apache/spark into broadcast-read-value
8792ed8 [Shivaram Venkataraman] Make torrent broadcast read blocks on use. This avoids reading broadcast variables when they are referenced in the closure but not used by the code.
`read()` may return fewer bytes than requested; when this occurred, the old code would silently return less data than requested, which might cause stream corruption errors. `skip()` faces similar issues, too.
This patch fixes several cases where we mis-handle these methods' return values.
Author: Josh Rosen <joshrosen@databricks.com>
Closes#2969 from JoshRosen/file-channel-read-fix and squashes the following commits:
e724a9f [Josh Rosen] Fix similar issue of not checking skip() return value.
cbc03ce [Josh Rosen] Update the other log message, too.
01e6015 [Josh Rosen] file.getName -> file.getAbsolutePath
d961d95 [Josh Rosen] Fix another issue in FileServerSuite.
b9265d2 [Josh Rosen] Fix a similar (minor) issue in TestUtils.
cd9d76f [Josh Rosen] Fix a similar error in Tachyon:
3db0008 [Josh Rosen] Fix a similar read() error in Utils.offsetBytes().
db985ed [Josh Rosen] Fix unsafe usage of FileChannel.read():
cc @rxin
Author: GuoQiang Li <witgo@qq.com>
Closes#2929 from witgo/SPARK-4064 and squashes the following commits:
20110f2 [GuoQiang Li] Modify the exception msg
3425225 [GuoQiang Li] review commits
2b07e49 [GuoQiang Li] If we create a lot of big broadcast variables, Spark may hang
This change replaces usages of colt with commons-math3 equivalents, and makes some minor necessary adjustments to related code and tests to match.
Author: Sean Owen <sowen@cloudera.com>
Closes#2928 from srowen/SPARK-4022 and squashes the following commits:
61a232f [Sean Owen] Fix failure due to different sampling in JavaAPISuite.sample()
16d66b8 [Sean Owen] Simplify seeding with call to reseedRandomGenerator
a1a78e0 [Sean Owen] Use Well19937c
31c7641 [Sean Owen] Fix Python Poisson test by choosing a different seed; about 88% of seeds should work but 1 didn't, it seems
5c9c67f [Sean Owen] Additional test fixes from review
d8f88e0 [Sean Owen] Replace colt with commons-math3. Some tests do not pass yet.
The patch takes advantage an API provided in Hadoop 2.5 that allows getting accurate data on Hadoop FileSystem bytes read. It eliminates the old method, which naively accepts the split size as the input bytes. An impact of this change will be that input metrics go away when using against Hadoop versions earlier thatn 2.5. I can add this back in, but my opinion is that no metrics are better than inaccurate metrics.
This is difficult to write a test for because we don't usually build against a version of Hadoop that contains the function we need. I've tested it manually on a pseudo-distributed cluster.
Author: Sandy Ryza <sandy@cloudera.com>
Closes#2087 from sryza/sandy-spark-2621 and squashes the following commits:
23010b8 [Sandy Ryza] Missing style fixes
74fc9bb [Sandy Ryza] Make getFSBytesReadOnThreadCallback private
1ab662d [Sandy Ryza] Clear things up a bit
984631f [Sandy Ryza] Switch from pull to push model and add test
7ef7b22 [Sandy Ryza] Add missing curly braces
219abc9 [Sandy Ryza] Fall back to split size
90dbc14 [Sandy Ryza] SPARK-2621. Update task InputMetrics incrementally
This change makes the destroy function public for broadcast variables. Motivation for the change is described in https://issues.apache.org/jira/browse/SPARK-4030.
This patch also logs where destroy was called from if a broadcast variable is used after destruction.
Author: Shivaram Venkataraman <shivaram@cs.berkeley.edu>
Closes#2922 from shivaram/broadcast-destroy and squashes the following commits:
a11abab [Shivaram Venkataraman] Fix scala style in Utils.scala
bed9c9d [Shivaram Venkataraman] Make destroy blocking by default
e80c1ab [Shivaram Venkataraman] Make destroy public for broadcast variables Also log where destroy was called from if a broadcast variable is used after destruction.
The shutdown hook of `DiskBlockManager` would remove localDirs. So do not need to register them with `Utils.registerShutdownDeleteDir`. It causes duplicate removal of these local dirs and corresponding exceptions.
Author: Liang-Chi Hsieh <viirya@gmail.com>
Closes#2826 from viirya/fix_duplicate_localdir_remove and squashes the following commits:
051d4b5 [Liang-Chi Hsieh] check dir existing and return empty List as default.
2b91a9c [Liang-Chi Hsieh] remove duplicate removal of local dirs.
This patch adds Selenium tests for Spark's web UI. To avoid adding extra
dependencies to the test environment, the tests use Selenium's HtmlUnitDriver,
which is pure-Java, instead of, say, ChromeDriver.
I added new tests to try to reproduce a few UI bugs reported on JIRA, namely
SPARK-3021, SPARK-2105, and SPARK-2527. I wasn't able to reproduce these bugs;
I suspect that the older ones might have been fixed by other patches.
In order to use HtmlUnitDriver, I added an explicit dependency on the
org.apache.httpcomponents version of httpclient in order to prevent jets3t's
older version from taking precedence on the classpath.
I also upgraded ScalaTest to 2.2.1.
Author: Josh Rosen <joshrosen@apache.org>
Author: Josh Rosen <joshrosen@databricks.com>
Closes#2474 from JoshRosen/webui-selenium-tests and squashes the following commits:
fcc9e83 [Josh Rosen] scalautils -> scalactic package rename
510e54a [Josh Rosen] [SPARK-3616] Add basic Selenium tests to WebUISuite.
This follows https://github.com/apache/spark/pull/2893 , but does not completely fix SPARK-3359 either. This fixes minor scaladoc/javadoc issues that Javadoc 8 will treat as errors.
Author: Sean Owen <sowen@cloudera.com>
Closes#2909 from srowen/SPARK-3359 and squashes the following commits:
f62c347 [Sean Owen] Fix some javadoc issues that javadoc 8 considers errors. This is not all of the errors turned up when javadoc 8 runs on output of genjavadoc.
In tests, we may want to have BlockManagers of size < 1MB (spark.storage.unrollMemoryThreshold). However, these BlockManagers are useless because we can't unroll anything in them ever. At the very least we need to log a warning.
tdas
Author: Andrew Or <andrew@databricks.com>
Closes#2917 from andrewor14/unroll-safely-logging and squashes the following commits:
38947e3 [Andrew Or] Warn against starting a block manager that's too small
fd621b4 [Andrew Or] Warn against failure to reserve initial memory threshold
This pull request is a first step towards the implementation of a stable, pull-based progress / status API for Spark (see [SPARK-2321](https://issues.apache.org/jira/browse/SPARK-2321)). For now, I'd like to discuss the basic implementation, API names, and overall interface design. Once we arrive at a good design, I'll go back and add additional methods to expose more information via these API.
#### Design goals:
- Pull-based API
- Usable from Java / Scala / Python (eventually, likely with a wrapper)
- Can be extended to expose more information without introducing binary incompatibilities.
- Returns immutable objects.
- Don't leak any implementation details, preserving our freedom to change the implementation.
#### Implementation:
- Add public methods (`getJobInfo`, `getStageInfo`) to SparkContext to allow status / progress information to be retrieved.
- Add public interfaces (`SparkJobInfo`, `SparkStageInfo`) for our API return values. These interfaces consist entirely of Java-style getter methods. The interfaces are currently implemented in Java. I decided to explicitly separate the interface from its implementation (`SparkJobInfoImpl`, `SparkStageInfoImpl`) in order to prevent users from constructing these responses themselves.
-Allow an existing JobProgressListener to be used when constructing a live SparkUI. This allows us to re-use this listeners in the implementation of this status API. There are a few reasons why this listener re-use makes sense:
- The status API and web UI are guaranteed to show consistent information.
- These listeners are already well-tested.
- The same garbage-collection / information retention configurations can apply to both this API and the web UI.
- Extend JobProgressListener to maintain `jobId -> Job` and `stageId -> Stage` mappings.
The progress API methods are implemented in a separate trait that's mixed into SparkContext. This helps to avoid SparkContext.scala from becoming larger and more difficult to read.
Author: Josh Rosen <joshrosen@databricks.com>
Author: Josh Rosen <joshrosen@apache.org>
Closes#2696 from JoshRosen/progress-reporting-api and squashes the following commits:
e6aa78d [Josh Rosen] Add tests.
b585c16 [Josh Rosen] Accept SparkListenerBus instead of more specific subclasses.
c96402d [Josh Rosen] Address review comments.
2707f98 [Josh Rosen] Expose current stage attempt id
c28ba76 [Josh Rosen] Update demo code:
646ff1d [Josh Rosen] Document spark.ui.retainedJobs.
7f47d6d [Josh Rosen] Clean up SparkUI constructors, per Andrew's feedback.
b77b3d8 [Josh Rosen] Merge remote-tracking branch 'origin/master' into progress-reporting-api
787444c [Josh Rosen] Move status API methods into trait that can be mixed into SparkContext.
f9a9a00 [Josh Rosen] More review comments:
3dc79af [Josh Rosen] Remove creation of unused listeners in SparkContext.
249ca16 [Josh Rosen] Address several review comments:
da5648e [Josh Rosen] Add example of basic progress reporting in Java.
7319ffd [Josh Rosen] Add getJobIdsForGroup() and num*Tasks() methods.
cc568e5 [Josh Rosen] Add note explaining that interfaces should not be implemented outside of Spark.
6e840d4 [Josh Rosen] Remove getter-style names and "consistent snapshot" semantics:
08cbec9 [Josh Rosen] Begin to sketch the interfaces for a stable, public status API.
ac2d13a [Josh Rosen] Add jobId->stage, stageId->stage mappings in JobProgressListener
24de263 [Josh Rosen] Create UI listeners in SparkContext instead of in Tabs:
If classes implementing Serializable or Externalizable interfaces throw
exceptions other than IOException or ClassNotFoundException from their
(de)serialization methods, then this results in an unhelpful
"IOException: unexpected exception type" rather than the actual exception that
produced the (de)serialization error.
This patch fixes this by adding a utility method that re-wraps any uncaught
exceptions in IOException (unless they are already instances of IOException).
Author: Josh Rosen <joshrosen@databricks.com>
Closes#2932 from JoshRosen/SPARK-4080 and squashes the following commits:
cd3a9be [Josh Rosen] [SPARK-4080] Only throw IOException from [write|read][Object|External].
https://issues.apache.org/jira/browse/SPARK-4067
currently , we call Utils.tryOrExit everywhere
AppClient
Executor
TaskSchedulerImpl
It makes the name of ExecutorUncaughtExceptionHandler unfit to the real case....
Author: Nan Zhu <nanzhu@Nans-MacBook-Pro.local>
Author: Nan Zhu <nanzhu@nans-mbp.home>
Closes#2913 from CodingCat/SPARK-4067 and squashes the following commits:
035ee3d [Nan Zhu] make RAT happy
e62e416 [Nan Zhu] add some general Exit code
a10b63f [Nan Zhu] refactor
In the existing code, each coarse-grained executor has two concurrently running actor systems. This causes many more error messages to be logged than necessary when the executor is lost or killed because we receive a disassociation event for each of these actor systems.
This is blocking #2840.
Author: Andrew Or <andrewor14@gmail.com>
Closes#2863 from andrewor14/executor-actor-system and squashes the following commits:
44ce2e0 [Andrew Or] Avoid starting two actor systems on each executor
In deploy.ClientArguments.isValidJarUrl, the url is checked as follows.
def isValidJarUrl(s: String): Boolean = s.matches("(.+):(.+)jar")
So, it allows like 'hdfs:file.jar' (no authority).
Author: Kousuke Saruta <sarutak@oss.nttdata.co.jp>
Closes#2925 from sarutak/uri-syntax-check-improvement and squashes the following commits:
cf06173 [Kousuke Saruta] Improved URI syntax checking
If Spark lunched multiple executors in one host for one application, every executor would download it dependent files and jars (if not using local: url) independently. It maybe result in huge latency. In my case, it result in 20 seconds latency to download dependent jars(size about 17M) when I lunched 32 executors in every host(total 4 hosts).
This patch will cache downloaded files and jars for executors to reduce network throughput and download latency. In my case, the latency was reduced from 20 seconds to less than 1 second.
Author: Li Zhihui <zhihui.li@intel.com>
Author: li-zhihui <zhihui.li@intel.com>
Closes#1616 from li-zhihui/cachefiles and squashes the following commits:
36940df [Li Zhihui] Close cache for local mode
935fed6 [Li Zhihui] Clean code.
f9330d4 [Li Zhihui] Clean code again
7050d46 [Li Zhihui] Clean code
074a422 [Li Zhihui] Fix: deal with spark.files.overwrite
03ed3a8 [li-zhihui] rename cache file name as XXXXXXXXX_cache
2766055 [li-zhihui] Use url.hashCode + timestamp as cachedFileName
76a7b66 [Li Zhihui] Clean code & use applcation work directory as cache directory
3510eb0 [Li Zhihui] Keep fetchFile private
2ffd742 [Li Zhihui] add comment for FileLock
e0ebd48 [Li Zhihui] Try and finally lock.release
7fb7c0b [Li Zhihui] Release lock before copy files
6b997bf [Li Zhihui] Executors of same application in same host should only download files & jars once
After take(), maybe there are some garbage left in the socket, then next task assigned to this worker will hang because of corrupted data.
We should make sure the socket is clean before reuse it, write END_OF_STREAM at the end, and check it after read out all result from python.
Author: Davies Liu <davies.liu@gmail.com>
Author: Davies Liu <davies@databricks.com>
Closes#2838 from davies/fix_reuse and squashes the following commits:
8872914 [Davies Liu] fix tests
660875b [Davies Liu] fix bug while reuse worker after take()
This commit fixes a bug in MapStatus that could cause jobs to wrongly return
empty results if those jobs contained stages with more than 2000 partitions
where most of those partitions were empty.
For jobs with > 2000 partitions, MapStatus uses HighlyCompressedMapStatus,
which only stores the average size of blocks. If the average block size is
zero, then this will cause all blocks to be reported as empty, causing
BlockFetcherIterator to mistakenly skip them.
For example, this would return an empty result:
sc.makeRDD(0 until 10, 1000).repartition(2001).collect()
This can also lead to deserialization errors (e.g. Snappy decoding errors)
for jobs with > 2000 partitions where the average block size is non-zero but
there is at least one empty block. In this case, the BlockFetcher attempts to
fetch empty blocks and fails when trying to deserialize them.
The root problem here is that MapStatus has a (previously undocumented)
correctness property that was violated by HighlyCompressedMapStatus:
If a block is non-empty, then getSizeForBlock must be non-zero.
I fixed this by modifying HighlyCompressedMapStatus to store the average size
of _non-empty_ blocks and to use a compressed bitmap to track which blocks are
empty.
I also removed a test which was broken as originally written: it attempted
to check that HighlyCompressedMapStatus's size estimation error was < 10%,
but this was broken because HighlyCompressedMapStatus is only used for map
statuses with > 2000 partitions, but the test only created 50.
Author: Josh Rosen <joshrosen@databricks.com>
Closes#2866 from JoshRosen/spark-4019 and squashes the following commits:
fc8b490 [Josh Rosen] Roll back hashset change, which didn't improve performance.
5faa0a4 [Josh Rosen] Incorporate review feedback
c8b8cae [Josh Rosen] Two performance fixes:
3b892dd [Josh Rosen] Address Reynold's review comments
ba2e71c [Josh Rosen] Add missing newline
609407d [Josh Rosen] Use Roaring Bitmap to track non-empty blocks.
c23897a [Josh Rosen] Use sets when comparing collect() results
91276a3 [Josh Rosen] [SPARK-4019] Fix MapStatus compression bug that could lead to empty results.
...r without a remove in between. The cause for that is unknown, and assumed a temp network issue.
However, since the second register is with a BlockManagerId on a different port, blockManagerInfo.contains() returns false, while blockManagerIdByExecutor returns Some. This inconsistency is caught in a conditional statement that does System.exit(1), which is a huge robustness issue for us.
The fix - simply remove the old id from both maps during register when this happens. We are mimicking the behavior of expireDeadHosts(), by doing local cleanup of the maps before trying to add new ones.
Also - added some logging for register and unregister.
This is just like https://github.com/apache/spark/pull/2854 except it's on master
Author: Tal Sliwowicz <tal.s@taboola.com>
Closes#2886 from tsliwowicz/master-block-mgr-removal and squashes the following commits:
094d508 [Tal Sliwowicz] some more white space change undone
41a2217 [Tal Sliwowicz] some more whitspaces change undone
7bcfc3d [Tal Sliwowicz] whitspaces fix
df9d98f [Tal Sliwowicz] Code review comments fixed
f48bce9 [Tal Sliwowicz] In long running contexts, we encountered the situation of double register without a remove in between. The cause for that is unknown, and assumed a temp network issue.
This PR fixes SPARK-3426, an issue where sort-based shuffle crashes if the
`spark.shuffle.spill.compress` and `spark.shuffle.compress` settings have
different values.
The problem is that sort-based shuffle's read and write paths use different
settings for determining whether to apply compression. ExternalSorter writes
runs to files using `TempBlockId` ids, which causes
`spark.shuffle.spill.compress` to be used for enabling compression, but these
spilled files end up being shuffled over the network and read as shuffle files
using `ShuffleBlockId` by BlockStoreShuffleFetcher, which causes
`spark.shuffle.compress` to be used for enabling decompression. As a result,
this leads to errors when these settings disagree.
Based on the discussions in #2247 and #2178, it sounds like we don't want to
remove the `spark.shuffle.spill.compress` setting. Therefore, I've tried to
come up with a fix where `spark.shuffle.spill.compress` is used to compress
data that's read and written locally and `spark.shuffle.compress` is used to
compress any data that will be fetched / read as shuffle blocks.
To do this, I split `TempBlockId` into two new id types, `TempLocalBlockId` and
`TempShuffleBlockId`, which map to `spark.shuffle.spill.compress` and
`spark.shuffle.compress`, respectively. ExternalAppendOnlyMap also used temp
blocks for spilling data. It looks like ExternalSorter was designed to be
a generic sorter but its configuration already happens to be tied to sort-based
shuffle, so I think it's fine if we use `spark.shuffle.compress` to compress
its spills; we can move the compression configuration to the constructor in
a later commit if we find that ExternalSorter is being used in other contexts
where we want different configuration options to control compression. To
summarize:
**Before:**
| | ExternalAppendOnlyMap | ExternalSorter |
|-------|------------------------------|------------------------------|
| Read | spark.shuffle.spill.compress | spark.shuffle.compress |
| Write | spark.shuffle.spill.compress | spark.shuffle.spill.compress |
**After:**
| | ExternalAppendOnlyMap | ExternalSorter |
|-------|------------------------------|------------------------|
| Read | spark.shuffle.spill.compress | spark.shuffle.compress |
| Write | spark.shuffle.spill.compress | spark.shuffle.compress |
Thanks to andrewor14 for debugging this with me!
Author: Josh Rosen <joshrosen@databricks.com>
Closes#2890 from JoshRosen/SPARK-3426 and squashes the following commits:
1921cf6 [Josh Rosen] Minor edit for clarity.
c8dd8f2 [Josh Rosen] Add comment explaining use of createTempShuffleBlock().
2c687b9 [Josh Rosen] Fix SPARK-3426.
91e7e40 [Josh Rosen] Combine tests into single test of all combinations
76ca65e [Josh Rosen] Add regression test for SPARK-3426.
runningLocally is deprecated now
Author: CrazyJvm <crazyjvm@gmail.com>
Closes#2879 from CrazyJvm/runningLocally and squashes the following commits:
bec0b3e [CrazyJvm] use isRunningLocally rather than runningLocally
Author: Sandy Ryza <sandy@cloudera.com>
Closes#789 from sryza/sandy-spark-1813 and squashes the following commits:
48b05e9 [Sandy Ryza] Simplify
b824932 [Sandy Ryza] Allow both spark.kryo.classesToRegister and spark.kryo.registrator at the same time
6a15bb7 [Sandy Ryza] Small fix
a2278c0 [Sandy Ryza] Respond to review comments
6ef592e [Sandy Ryza] SPARK-1813. Add a utility to SparkConf that makes using Kryo really easy
See [JIRA](https://issues.apache.org/jira/browse/SPARK-3994) for more information. Also adds
a note which warns against using these methods.
Author: Aaron Davidson <aaron@databricks.com>
Closes#2839 from aarondav/countByKey and squashes the following commits:
d6fdb2a [Aaron Davidson] Respond to comments
e1f06d3 [Aaron Davidson] [SPARK-3994] Use standard Aggregator code path for countByKey and countByValue
If an executor fails without being scheduled to run any tasks, then `DAGScheduler` won't notify `BlockManagerMasterActor` that the associated block manager should be removed. Instead, the associated block manager will be expired only after a few rounds of heartbeat timeouts. In terms of removal treatment, there should really be no distinction between executors that have been scheduled tasks and those that have not.
The fix, then, is to add all known executors to `TaskSchedulerImpl`'s `activeExecutorIds` whether or not it has been scheduled a task. In fact, the existing comment above `activeExecutorIds` is
```
// Which executor IDs we have executors on
val activeExecutorIds = new HashSet[String]
```
not "Which executors have been scheduled tasks thus far."
Author: Andrew Or <andrewor14@gmail.com>
Closes#2865 from andrewor14/active-executors and squashes the following commits:
ff3172b [Andrew Or] Add all known executors to `activeExecutorIds`
Just found a typo. Should not use "%f" for Long.
Author: zsxwing <zsxwing@gmail.com>
Closes#2875 from zsxwing/SPARK-4035 and squashes the following commits:
ce347e2 [zsxwing] Fix a wrong format specifier
This PR makes several changes to TorrentBroadcast in order to make
it easier to reason about, which should help when debugging SPARK-3958.
The key changes:
- Remove all state from the global TorrentBroadcast object. This state
consisted mainly of configuration options, like the block size and
compression codec, and was read by the blockify / unblockify methods.
Unfortunately, the use of `lazy val` for `BLOCK_SIZE` meant that the block
size was always determined by the first SparkConf that TorrentBroadast was
initialized with; as a result, unit tests could not properly test
TorrentBroadcast with different block sizes.
Instead, blockifyObject and unBlockifyObject now accept compression codecs
and blockSizes as arguments. These arguments are supplied at the call sites
inside of TorrentBroadcast instances. Each TorrentBroadcast instance
determines these values from SparkEnv's SparkConf. I was careful to ensure
that we do not accidentally serialize CompressionCodec or SparkConf objects
as part of the TorrentBroadcast object.
- Remove special-case handling of local-mode in TorrentBroadcast. I don't
think that broadcast implementations should know about whether we're running
in local mode. If we want to optimize the performance of broadcast in local
mode, then we should detect this at a higher level and use a dummy
LocalBroadcastFactory implementation instead.
Removing this code fixes a subtle error condition: in the old local mode
code, a failure to find the broadcast in the local BlockManager would lead
to an attempt to deblockify zero blocks, which could lead to confusing
deserialization or decompression errors when we attempted to decompress
an empty byte array. This should never have happened, though: a failure to
find the block in local mode is evidence of some other error. The changes
here will make it easier to debug those errors if they ever happen.
- Add a check that throws an exception when attempting to deblockify an
empty array.
- Use ScalaCheck to add a test to check that TorrentBroadcast's
blockifyObject and unBlockifyObject methods are inverses.
- Misc. cleanup and logging improvements.
Author: Josh Rosen <joshrosen@databricks.com>
Closes#2844 from JoshRosen/torrentbroadcast-bugfix and squashes the following commits:
1e8268d [Josh Rosen] Address Reynold's review comments
2a9fdfd [Josh Rosen] Address Reynold's review comments.
c3b08f9 [Josh Rosen] Update TorrentBroadcast tests to reflect removal of special local-mode optimizations.
5c22782 [Josh Rosen] Store broadcast variable's value in the driver.
33fc754 [Josh Rosen] Change blockify/unblockifyObject to accept serializer as argument.
618a872 [Josh Rosen] [SPARK-3958] TorrentBroadcast cleanup / debugging improvements.
Before, if the master node is killed and restarted, the worker nodes
would not attempt to reconnect to the Master. Therefore, when the Master
node was restarted, the worker nodes needed to be restarted as well.
Now, when the Master node is disconnected, the worker nodes will
continuously ping the master node in attempts to reconnect to it. Once
the master node restarts, it will detect one of the registration
requests from its former workers. The result is that the cluster
re-enters a healthy state.
In addition, when the master does not receive a heartbeat from the
worker, the worker was removed; however, when the worker sent a
heartbeat to the master, the master used to ignore the heartbeat. Now,
a master that receives a heartbeat from a worker that had been
disconnected will request the worker to re-attempt the registration
process, at which point the worker will send a RegisterWorker request
and be re-connected accordingly.
Re-connection attempts per worker are submitted every N seconds, where N
is configured by the property spark.worker.reconnect.interval - this has
a default of 60 seconds right now.
Author: mcheah <mcheah@palantir.com>
Closes#2828 from mccheah/reconnect-dead-workers and squashes the following commits:
83f8bc9 [mcheah] [SPARK-3736] More informative log message, and fixing some indentation.
fe0e02f [mcheah] [SPARK-3736] Moving reconnection logic to registerWithMaster().
94ddeca [mcheah] [SPARK-3736] Changing a log warning to a log info.
a698e35 [mcheah] [SPARK-3736] Addressing PR comment to make some defs private.
b9a3077 [mcheah] [SPARK-3736] Addressing PR comments related to reconnection.
2ad5ed5 [mcheah] [SPARK-3736] Cancel attempts to reconnect if the master changes.
b5b34af [mcheah] [SPARK-3736] Workers reconnect when disassociated from the master.
The problem caused by #1966
CC YanTangZhai andrewor14
Author: GuoQiang Li <witgo@qq.com>
Closes#2858 from witgo/SPARK-4010 and squashes the following commits:
9866fbf [GuoQiang Li] Spark UI returns 500 in yarn-client mode
Kernel 2.6.32 bug will lead to unexpected behavior of transferTo in copyStream, and this will corrupt the shuffle output file in sort-based shuffle, which will somehow introduce PARSING_ERROR(2), deserialization error or offset out of range. Here fix this by adding append flag, also add some position checking code. Details can be seen in [SPARK-3948](https://issues.apache.org/jira/browse/SPARK-3948).
Author: jerryshao <saisai.shao@intel.com>
Closes#2824 from jerryshao/SPARK-3948 and squashes the following commits:
be0533a [jerryshao] Address the comments
a82b184 [jerryshao] add configuration to control the NIO way of copying stream
e17ada2 [jerryshao] Fix kernel 2.6.32 bug led unexpected behavior of transferTo
This PR adds a Java API for AsyncRDDActions and promotes the API from `Experimental` to stable.
Author: Josh Rosen <joshrosen@apache.org>
Author: Josh Rosen <joshrosen@databricks.com>
Closes#2760 from JoshRosen/async-rdd-actions-in-java and squashes the following commits:
0d45fbc [Josh Rosen] Whitespace fix.
ad3ae53 [Josh Rosen] Merge remote-tracking branch 'origin/master' into async-rdd-actions-in-java
c0153a5 [Josh Rosen] Remove unused variable.
e8e2867 [Josh Rosen] Updates based on Marcelo's review feedback
7a1417f [Josh Rosen] Removed unnecessary java.util import.
6f8f6ac [Josh Rosen] Fix import ordering.
ff28e49 [Josh Rosen] Add MiMa excludes and fix a scalastyle error.
346e46e [Josh Rosen] [SPARK-3902] Stabilize AsyncRDDActions; add Java API.
This patch attempts to fix SPARK-2546 in `branch-1.0` and `branch-1.1`. The underlying problem is that thread-safety issues in Hadoop Configuration objects may cause Spark tasks to get stuck in infinite loops. The approach taken here is to clone a new copy of the JobConf for each task rather than sharing a single copy between tasks. Note that there are still Configuration thread-safety issues that may affect the driver, but these seem much less likely to occur in practice and will be more complex to fix (see discussion on the SPARK-2546 ticket).
This cloning is guarded by a new configuration option (`spark.hadoop.cloneConf`) and is disabled by default in order to avoid unexpected performance regressions for workloads that are unaffected by the Configuration thread-safety issues.
Author: Josh Rosen <joshrosen@apache.org>
Closes#2684 from JoshRosen/jobconf-fix-backport and squashes the following commits:
f14f259 [Josh Rosen] Add configuration option to control cloning of Hadoop JobConf.
b562451 [Josh Rosen] Remove unused jobConfCacheKey field.
dd25697 [Josh Rosen] [SPARK-2546] [1.0 / 1.1 backport] Clone JobConf for each task.
(cherry picked from commit 2cd40db2b3)
Signed-off-by: Josh Rosen <joshrosen@databricks.com>
Conflicts:
core/src/main/scala/org/apache/spark/rdd/HadoopRDD.scala
Make JavaPairRDD.collectAsMap result Serializable since Java Maps generally are
Author: Sean Owen <sowen@cloudera.com>
Closes#2805 from srowen/SPARK-3926 and squashes the following commits:
ecb78ee [Sean Owen] Fix conflict between java.io.Serializable and use of Scala's Serializable
f4717f9 [Sean Owen] Oops, fix compile problem
ae1b36f [Sean Owen] Expand to cover Maps returned from other Java API methods as well
51c26c2 [Sean Owen] Make JavaPairRDD.collectAsMap result Serializable since Java Maps generally are
There is a unused variable(count) in saveAsHadoopDataset in PairRDDFunctions.scala. The initial idea of this variable seems to count the number of records, so I am adding a log statement to log the number of records that has been written to the writer.
Author: likun <jacky.likun@huawei.com>
Author: jackylk <jacky.likun@huawei.com>
Closes#2791 from jackylk/SPARK-3935 and squashes the following commits:
a874047 [jackylk] removing the unused variable in PairRddFunctions.scala
3bf43c7 [likun] log the number of records has been written
Its hard to debug which broadcast variables refer to what in a big codebase. Printing call site information helps in debugging.
Author: Shivaram Venkataraman <shivaram@cs.berkeley.edu>
Closes#2829 from shivaram/spark-broadcast-print and squashes the following commits:
cd6dbdf [Shivaram Venkataraman] Print call site information for broadcasts
JobProgressPage could not show Fair Scheduler Pools section sometimes.
SparkContext starts webui and then postEnvironmentUpdate. Sometimes JobProgressPage is accessed between webui starting and postEnvironmentUpdate, then the lazy val isFairScheduler will be false. The Fair Scheduler Pools section will not display any more.
Author: yantangzhai <tyz0303@163.com>
Author: YanTangZhai <hakeemzhai@tencent.com>
Closes#1966 from YanTangZhai/SPARK-3067 and squashes the following commits:
d4323f8 [yantangzhai] update [SPARK-3067] JobProgressPage could not show Fair Scheduler Pools section sometimes
8a00106 [YanTangZhai] Merge pull request #6 from apache/master
b6391cc [yantangzhai] revert [SPARK-3067] JobProgressPage could not show Fair Scheduler Pools section sometimes
d2226cd [yantangzhai] [SPARK-3067] JobProgressPage could not show Fair Scheduler Pools section sometimes
cbcba66 [YanTangZhai] Merge pull request #3 from apache/master
aac7f7b [yantangzhai] [SPARK-3067] JobProgressPage could not show Fair Scheduler Pools section sometimes
cdef539 [YanTangZhai] Merge pull request #1 from apache/master
Sorry. I found that I forgot to add `afterExecute` for `handleConnectExecutor` in #2593.
Author: zsxwing <zsxwing@gmail.com>
Closes#2794 from zsxwing/SPARK-3741 and squashes the following commits:
a0bc4dd [zsxwing] Add afterExecute for handleConnectExecutor
In BlockManagermasterActor, _remainingMem would increase memSize for twice when updateBlockInfo if new storageLevel is invalid and old storageLevel is "useMemory". Also, _remainingMem should increase with original memory size instead of new memSize.
Author: Zhang, Liye <liye.zhang@intel.com>
Closes#2792 from liyezhang556520/spark-3941-remainMem and squashes the following commits:
3d487cc [Zhang, Liye] make the code concise
0380a32 [Zhang, Liye] [SPARK-3941][CORE] _remainingmem should not increase twice when updateBlockInfo
Something about the 2.3.4 upgrade seems to have made the issue manifest where all the services disconnect from each other after exactly 1000 seconds (which is the heartbeat interval). [This post](https://groups.google.com/forum/#!topic/akka-user/X3xzpTCbEFs) suggests that heartbeat pause should be greater than heartbeat interval, and increasing the pause from 600s to 6000s seems to have rectified the issue. My current cluster has now exceeded 1400s of uptime without failure!
I do not know why this fixed it, because the threshold we have set for the failure detector is the exponent of a timeout, and 300 is extremely large. Perhaps the default failure detector changed in 2.3.4 and now ignores threshold.
Author: Aaron Davidson <aaron@databricks.com>
Closes#2784 from aarondav/fix-timeout and squashes the following commits:
bd1151a [Aaron Davidson] Increase pause, don't decrease interval
9cb0372 [Aaron Davidson] [SPARK-3923] Decrease Akka heartbeat interval below heartbeat pause
This is a small number of clean-up changes on top of #2782. Closes#2782.
Author: Prashant Sharma <prashant.s@imaginea.com>
Author: Patrick Wendell <pwendell@gmail.com>
Closes#2803 from pwendell/pr-2782 and squashes the following commits:
56d5b7a [Patrick Wendell] Minor clean-up
44089ec [Patrick Wendell] Clean-up the TaskContext API.
ed551ce [Prashant Sharma] Fixed a typo
df261d0 [Prashant Sharma] Josh's suggestion
facf3b1 [Prashant Sharma] Fixed the mima issue.
7ecc2fe [Prashant Sharma] CR, Moved implementations to TaskContextImpl
bbd9e05 [Prashant Sharma] adding missed out files to git.
ef633f5 [Prashant Sharma] SPARK-3874, Provide stable TaskContext API
Customized pickler should be registered before unpickling, but in executor, there is no way to register the picklers before run the tasks.
So, we need to register the picklers in the tasks itself, duplicate the javaToPython() and pythonToJava() in MLlib, call SerDe.initialize() before pickling or unpickling.
Author: Davies Liu <davies.liu@gmail.com>
Closes#2830 from davies/fix_pickle and squashes the following commits:
0c85fb9 [Davies Liu] revert the privacy change
6b94e15 [Davies Liu] use JavaConverters instead of JavaConversions
0f02050 [Davies Liu] hotfix: Customized pickler does not work in cluster
Author: Shiti <ssaxena.ece@gmail.com>
Closes#2810 from Shiti/master and squashes the following commits:
051d82f [Shiti] setting the default value of uri scheme to "file" where matching "file" or None yields the same result
