## What changes were proposed in this pull request?
This patch moves tags and unsafe modules into common directory to remove 2 top level non-user-facing directories.
## How was this patch tested?
Jenkins should suffice.
Author: Reynold Xin <rxin@databricks.com>
Closes#11426 from rxin/SPARK-13548.
## What changes were proposed in this pull request?
This PR defer the resolution from a id of dictionary to value until the column is actually accessed (inside getInt/getLong), this is very useful for those columns and rows that are filtered out. It's also useful for binary type, we will not need to copy all the byte arrays.
This PR also change the underlying type for small decimal that could be fit within a Int, in order to use getInt() to lookup the value from IntDictionary.
## How was this patch tested?
Manually test TPCDS Q7 with scale factor 10, saw about 30% improvements (after PR #11274).
Author: Davies Liu <davies@databricks.com>
Closes#11437 from davies/decode_dict.
Estimator for Generalized Linear Models(GLMs) which will be solved by IRLS.
cc mengxr
Author: Yanbo Liang <ybliang8@gmail.com>
Closes#11136 from yanboliang/spark-12811.
JIRA: https://issues.apache.org/jira/browse/SPARK-13511
## What changes were proposed in this pull request?
Current limit operator doesn't support wholestage codegen. This is open to add support for it.
In the `doConsume` of `GlobalLimit` and `LocalLimit`, we use a count term to count the processed rows. Once the row numbers catches the limit number, we set the variable `stopEarly` of `BufferedRowIterator` newly added in this pr to `true` that indicates we want to stop processing remaining rows. Then when the wholestage codegen framework checks `shouldStop()`, it will stop the processing of the row iterator.
Before this, the executed plan for a query `sqlContext.range(N).limit(100).groupBy().sum()` is:
TungstenAggregate(key=[], functions=[(sum(id#5L),mode=Final,isDistinct=false)], output=[sum(id)#6L])
+- TungstenAggregate(key=[], functions=[(sum(id#5L),mode=Partial,isDistinct=false)], output=[sum#9L])
+- GlobalLimit 100
+- Exchange SinglePartition, None
+- LocalLimit 100
+- Range 0, 1, 1, 524288000, [id#5L]
After add wholestage codegen support:
WholeStageCodegen
: +- TungstenAggregate(key=[], functions=[(sum(id#40L),mode=Final,isDistinct=false)], output=[sum(id)#41L])
: +- TungstenAggregate(key=[], functions=[(sum(id#40L),mode=Partial,isDistinct=false)], output=[sum#44L])
: +- GlobalLimit 100
: +- INPUT
+- Exchange SinglePartition, None
+- WholeStageCodegen
: +- LocalLimit 100
: +- Range 0, 1, 1, 524288000, [id#40L]
## How was this patch tested?
A test is added into BenchmarkWholeStageCodegen.
Author: Liang-Chi Hsieh <viirya@gmail.com>
Closes#11391 from viirya/wholestage-limit.
## What changes were proposed in this pull request?
This patch fixes the problem that pyspark fails on Windows because pyspark can't find ```spark-submit2.cmd```.
## How was this patch tested?
manual tests:
I ran ```bin\pyspark.cmd``` and checked if pyspark is launched correctly after this patch is applyed.
Author: Masayoshi TSUZUKI <tsudukim@oss.nttdata.co.jp>
Closes#11442 from tsudukim/feature/SPARK-13592.
JIRA: https://issues.apache.org/jira/browse/SPARK-13550
## What changes were proposed in this pull request?
Just add a java example for ml.clustering.BisectingKMeans
## How was this patch tested?
manual tests were done.
(If this patch involves UI changes, please attach a screenshot; otherwise, remove this)
Author: Zheng RuiFeng <ruifengz@foxmail.com>
Closes#11428 from zhengruifeng/ml_bkm_je.
JIRA: https://issues.apache.org/jira/browse/SPARK-13551
## What changes were proposed in this pull request?
Fix wrong comment and remove meanless lines in mllib.JavaBisectingKMeansExample
## How was this patch tested?
manual test
Author: Zheng RuiFeng <ruifengz@foxmail.com>
Closes#11429 from zhengruifeng/mllib_bkm_je.
The Hive client library is not smart enough to notice that the current
user is a proxy user; so when using a proxy user, it fails to fetch
delegation tokens from the metastore because of a missing kerberos
TGT for the current user.
To fix it, just run the code that fetches the delegation token as the
real logged in user.
Tested on a kerberos cluster both submitting normally and with a proxy
user; Hive and HBase tokens are retrieved correctly in both cases.
Author: Marcelo Vanzin <vanzin@cloudera.com>
Closes#11358 from vanzin/SPARK-13478.
## What changes were proposed in this pull request?
This PR adds support for implementing whole state codegen for sort. Builds heaving on nongli 's PR: https://github.com/apache/spark/pull/11008 (which actually implements the feature), and adds the following changes on top:
- [x] Generated code updates peak execution memory metrics
- [x] Unit tests in `WholeStageCodegenSuite` and `SQLMetricsSuite`
## How was this patch tested?
New unit tests in `WholeStageCodegenSuite` and `SQLMetricsSuite`. Further, all existing sort tests should pass.
Author: Sameer Agarwal <sameer@databricks.com>
Author: Nong Li <nong@databricks.com>
Closes#11359 from sameeragarwal/sort-codegen.
## What changes were proposed in this pull request?
Just fixed the log place introduced by #11401
## How was this patch tested?
unit tests.
Author: Shixiong Zhu <shixiong@databricks.com>
Closes#11432 from zsxwing/SPARK-13522-follow-up.
## What changes were proposed in this pull request?
Sometimes, network disconnection event won't be triggered for other potential race conditions that we may not have thought of, then the executor will keep sending heartbeats to driver and won't exit.
This PR adds a new configuration `spark.executor.heartbeat.maxFailures` to kill Executor when it's unable to heartbeat to the driver more than `spark.executor.heartbeat.maxFailures` times.
## How was this patch tested?
unit tests
Author: Shixiong Zhu <shixiong@databricks.com>
Closes#11401 from zsxwing/SPARK-13522.
#### What changes were proposed in this pull request?
After analysis by Analyzer, two operators could have alias. They are `Project` and `Aggregate`. So far, we only rewrite and propagate constraints if `Alias` is defined in `Project`. This PR is to resolve this issue in `Aggregate`.
#### How was this patch tested?
Added a test case for `Aggregate` in `ConstraintPropagationSuite`.
marmbrus sameeragarwal
Author: gatorsmile <gatorsmile@gmail.com>
Closes#11422 from gatorsmile/validConstraintsInUnaryNodes.
https://issues.apache.org/jira/browse/SPARK-13507https://issues.apache.org/jira/browse/SPARK-13509
## What changes were proposed in this pull request?
This PR adds the support to write CSV data directly by a single call to the given path.
Several unitests were added for each functionality.
## How was this patch tested?
This was tested with unittests and with `dev/run_tests` for coding style
Author: hyukjinkwon <gurwls223@gmail.com>
Author: Hyukjin Kwon <gurwls223@gmail.com>
Closes#11389 from HyukjinKwon/SPARK-13507-13509.
## What changes were proposed in this pull request?
Nested classes defined within Scala objects are translated into Java static nested classes. Unlike inner classes, they don't need outer scopes. But the analyzer still thinks that an outer scope is required.
This PR fixes this issue simply by checking whether a nested class is static before looking up its outer scope.
## How was this patch tested?
A test case is added to `DatasetSuite`. It checks contents of a Dataset whose element type is a nested class declared in a Scala object.
Author: Cheng Lian <lian@databricks.com>
Closes#11421 from liancheng/spark-13540-object-as-outer-scope.
JIRA: https://issues.apache.org/jira/browse/SPARK-13506
## What changes were proposed in this pull request?
just chang R Snippet Comment in AssociationRulesSuite
## How was this patch tested?
unit test passsed
Author: Zheng RuiFeng <ruifengz@foxmail.com>
Closes#11387 from zhengruifeng/ars.
## What changes were proposed in this pull request?
Now by default, it shows as ascending order of appId. We might prefer to display as descending order by default, which will show the latest application at the top.
## How was this patch tested?
Manual tested. See screenshot below:
Author: zhuol <zhuol@yahoo-inc.com>
Closes#11357 from zhuoliu/13481.
Part of task for [SPARK-11219](https://issues.apache.org/jira/browse/SPARK-11219) to make PySpark MLlib parameter description formatting consistent. This is for the regression module. Also, updated 2 params in classification to read as `Supported values:` to be consistent.
closes#10600
Author: vijaykiran <mail@vijaykiran.com>
Author: Bryan Cutler <cutlerb@gmail.com>
Closes#11404 from BryanCutler/param-desc-consistent-regression-SPARK-12633.
## What changes were proposed in this pull request?
* The default value of ```regParam``` of PySpark MLlib ```LogisticRegressionWithLBFGS``` should be consistent with Scala which is ```0.0```. (This is also consistent with ML ```LogisticRegression```.)
* BTW, if we use a known updater(L1 or L2) for binary classification, ```LogisticRegressionWithLBFGS``` will call the ML implementation. We should update the API doc to clarifying ```numCorrections``` will have no effect if we fall into that route.
* Make a pass for all parameters of ```LogisticRegressionWithLBFGS```, others are set properly.
cc mengxr dbtsai
## How was this patch tested?
No new tests, it should pass all current tests.
Author: Yanbo Liang <ybliang8@gmail.com>
Closes#11424 from yanboliang/spark-13545.
JIRA: https://issues.apache.org/jira/browse/SPARK-13537
## What changes were proposed in this pull request?
In readBytes of VectorizedPlainValuesReader, we use buffer[offset] to access bytes in buffer. It is incorrect because offset is added with Platform.BYTE_ARRAY_OFFSET when initialization. We should fix it.
## How was this patch tested?
`ParquetHadoopFsRelationSuite` sometimes (depending on the randomly generated data) will be [failed](https://amplab.cs.berkeley.edu/jenkins/job/SparkPullRequestBuilder/52136/consoleFull) by this bug. After applying this, the test can be passed.
I added a test to `ParquetHadoopFsRelationSuite` with the data which will fail without this patch.
The error exception:
[info] ParquetHadoopFsRelationSuite:
[info] - test all data types - StringType (440 milliseconds)
[info] - test all data types - BinaryType (434 milliseconds)
[info] - test all data types - BooleanType (406 milliseconds)
20:59:38.618 ERROR org.apache.spark.executor.Executor: Exception in task 0.0 in stage 2597.0 (TID 67966)
java.lang.ArrayIndexOutOfBoundsException: 46
at org.apache.spark.sql.execution.datasources.parquet.VectorizedPlainValuesReader.readBytes(VectorizedPlainValuesReader.java:88)
Author: Liang-Chi Hsieh <viirya@gmail.com>
Closes#11418 from viirya/fix-readbytes.
## What changes were proposed in this pull request?
As the title says, this moves the three modules currently in network/ into common/network-*. This removes one top level, non-user-facing folder.
## How was this patch tested?
Compilation and existing tests. We should run both SBT and Maven.
Author: Reynold Xin <rxin@databricks.com>
Closes#11409 from rxin/SPARK-13529.
## What changes were proposed in this pull request?
This creates a `SessionState`, which groups a few fields that existed in `SQLContext`. Because `HiveContext` extends `SQLContext` we also need to make changes there. This is mainly a cleanup task that will soon pave the way for merging the two contexts.
## How was this patch tested?
Existing unit tests; this patch introduces no change in behavior.
Author: Andrew Or <andrew@databricks.com>
Closes#11405 from andrewor14/refactor-session.
## What changes were proposed in this pull request?
Fix readBytes in VectorizedPlainValuesReader. This fixes a copy and paste issue.
## How was this patch tested?
Ran ParquetHadoopFsRelationSuite which failed before this.
Author: Nong Li <nong@databricks.com>
Closes#11414 from nongli/spark-13533.
JIRA: https://issues.apache.org/jira/browse/SPARK-13530
## What changes were proposed in this pull request?
By enabling vectorized parquet scanner by default, the unit test `ParquetHadoopFsRelationSuite` based on `HadoopFsRelationTest` will be failed due to the lack of short type support in `UnsafeRowParquetRecordReader`. We should fix it.
The error exception:
[info] ParquetHadoopFsRelationSuite:
[info] - test all data types - StringType (499 milliseconds)
[info] - test all data types - BinaryType (447 milliseconds)
[info] - test all data types - BooleanType (520 milliseconds)
[info] - test all data types - ByteType (418 milliseconds)
00:22:58.920 ERROR org.apache.spark.executor.Executor: Exception in task 0.0 in stage 124.0 (TID 1949)
org.apache.commons.lang.NotImplementedException: Unimplemented type: ShortType
at org.apache.spark.sql.execution.datasources.parquet.UnsafeRowParquetRecordReader$ColumnReader.readIntBatch(UnsafeRowParquetRecordReader.java:769)
at org.apache.spark.sql.execution.datasources.parquet.UnsafeRowParquetRecordReader$ColumnReader.readBatch(UnsafeRowParquetRecordReader.java:640)
at org.apache.spark.sql.execution.datasources.parquet.UnsafeRowParquetRecordReader$ColumnReader.access$000(UnsafeRowParquetRecordReader.java:461)
at org.apache.spark.sql.execution.datasources.parquet.UnsafeRowParquetRecordReader.nextBatch(UnsafeRowParquetRecordReader.java:224)
## How was this patch tested?
The unit test `ParquetHadoopFsRelationSuite` based on `HadoopFsRelationTest` will be [failed](https://amplab.cs.berkeley.edu/jenkins/job/SparkPullRequestBuilder/52110/consoleFull) due to the lack of short type support in UnsafeRowParquetRecordReader. By adding this support, the test can be passed.
Author: Liang-Chi Hsieh <viirya@gmail.com>
Closes#11412 from viirya/add-shorttype-support.
It registers more Scala classes, including ListBuffer to support Kryo with FPGrowth.
See https://github.com/twitter/chill/releases for Chill's change log.
Author: mark800 <yky800@126.com>
Closes#11041 from mark800/master.
## What changes were proposed in this pull request?
Change the default of the flag to enable this feature now that the implementation is complete.
## How was this patch tested?
The new parquet reader should be a drop in, so will be exercised by the existing tests.
Author: Nong Li <nong@databricks.com>
Closes#11397 from nongli/spark-13518.
## What changes were proposed in this pull request?
We provide a very limited set of cluster management script in Spark for Tachyon, although Tachyon itself provides a much better version of it. Given now Spark users can simply use Tachyon as a normal file system and does not require extensive configurations, we can remove this management capabilities to simplify Spark bash scripts.
Note that this also reduces coupling between a 3rd party external system and Spark's release scripts, and would eliminate possibility for failures such as Tachyon being renamed or the tar balls being relocated.
## How was this patch tested?
N/A
Author: Reynold Xin <rxin@databricks.com>
Closes#11400 from rxin/release-script.
Due to the people.apache.org -> home.apache.org migration, we need to update our packaging scripts to publish artifacts to the new server. Because the new server only supports sftp instead of ssh, we need to update the scripts to use lftp instead of ssh + rsync.
Author: Josh Rosen <joshrosen@databricks.com>
Closes#11350 from JoshRosen/update-release-scripts-for-apache-home.
## What changes were proposed in this pull request?
When the driver removes an executor's state, the connection between the driver and the executor may be still alive so that the executor cannot exit automatically (E.g., Master will send RemoveExecutor when a work is lost but the executor is still alive), so the driver should try to tell the executor to stop itself. Otherwise, we will leak an executor.
This PR modified the driver to send `StopExecutor` to the executor when it's removed.
## How was this patch tested?
manual test: increase the worker heartbeat interval to force it's always timeout and the leak executors are gone.
Author: Shixiong Zhu <shixiong@databricks.com>
Closes#11399 from zsxwing/SPARK-13519.
## What changes were proposed in this pull request?
After SPARK-13028, we should add Python API for MaxAbsScaler.
## How was this patch tested?
unit test
Author: zlpmichelle <zlpmichelle@gmail.com>
Closes#11393 from zlpmichelle/master.
## What changes were proposed in this pull request?
TaskContext supports task completion callback, which gets called regardless of task failures. However, there is no way for the listener to know if there is an error. This patch adds a new listener that gets called when a task fails.
## How was the this patch tested?
New unit test case and integration test case covering the code path
Author: Reynold Xin <rxin@databricks.com>
Closes#11340 from rxin/SPARK-13465.
## What changes were proposed in this pull request?
This patch includes these performance fixes:
- Remove unnecessary setNotNull() calls. The NULL bits are cleared already.
- Speed up RLE group decoding
- Speed up dictionary decoding by decoding NULLs directly into the result.
## How was this patch tested?
(Please explain how this patch was tested. E.g. unit tests, integration tests, manual tests)
In addition to the updated benchmarks, on TPCDS, the result of these changes
running Q55 (sf40) is:
```
TPCDS: Best/Avg Time(ms) Rate(M/s) Per Row(ns)
---------------------------------------------------------------------------------
q55 (Before) 6398 / 6616 18.0 55.5
q55 (After) 4983 / 5189 23.1 43.3
```
Author: Nong Li <nong@databricks.com>
Closes#11375 from nongli/spark-13499.
## What changes were proposed in this pull request?
Currently, BroadcastNestedLoopJoin is implemented for worst case, it's too slow, very easy to hang forever. This PR will create fast path for some joinType and buildSide, also improve the worst case (will use much less memory than before).
Before this PR, one task requires O(N*K) + O(K) in worst cases, N is number of rows from one partition of streamed table, it could hang the job (because of GC).
In order to workaround this for InnerJoin, we have to disable auto-broadcast, switch to CartesianProduct: This could be workaround for InnerJoin, see https://forums.databricks.com/questions/6747/how-do-i-get-a-cartesian-product-of-a-huge-dataset.html
In this PR, we will have fast path for these joins :
InnerJoin with BuildLeft or BuildRight
LeftOuterJoin with BuildRight
RightOuterJoin with BuildLeft
LeftSemi with BuildRight
These fast paths are all stream based (take one pass on streamed table), required O(1) memory.
All other join types and build types will take two pass on streamed table, one pass to find the matched rows that includes streamed part, which require O(1) memory, another pass to find the rows from build table that does not have a matched row from streamed table, which required O(K) memory, K is the number rows from build side, one bit per row, should be much smaller than the memory for broadcast. The following join types work in this way:
LeftOuterJoin with BuildLeft
RightOuterJoin with BuildRight
FullOuterJoin with BuildLeft or BuildRight
LeftSemi with BuildLeft
This PR also added tests for all the join types for BroadcastNestedLoopJoin.
After this PR, for InnerJoin with one small table, BroadcastNestedLoopJoin should be faster than CartesianProduct, we don't need that workaround anymore.
## How was the this patch tested?
Added unit tests.
Author: Davies Liu <davies@databricks.com>
Closes#11328 from davies/nested_loop.
## What changes were proposed in this pull request?
This PR fixes the order of modifier from `abstract public` into `public abstract`.
Currently, when we run `./dev/lint-java`, it shows the error.
```
Checkstyle checks failed at following occurrences:
[ERROR] src/main/java/org/apache/spark/util/sketch/CountMinSketch.java:[53,10] (modifier) ModifierOrder: 'public' modifier out of order with the JLS suggestions.
```
## How was this patch tested?
```
$ ./dev/lint-java
Checkstyle checks passed.
```
Author: Dongjoon Hyun <dongjoon@apache.org>
Closes#11390 from dongjoon-hyun/fix_modifier_order.
## What changes were proposed in this pull request?
This PR replaces example codes in `mllib-linear-methods.md` using `include_example`
by doing the followings:
* Extracts the example codes(Scala,Java,Python) as files in `example` module.
* Merges some dialog-style examples into a single file.
* Hide redundant codes in HTML for the consistency with other docs.
## How was the this patch tested?
manual test.
This PR can be tested by document generations, `SKIP_API=1 jekyll build`.
Author: Dongjoon Hyun <dongjoon@apache.org>
Closes#11320 from dongjoon-hyun/SPARK-11381.
Part of task for [SPARK-11219](https://issues.apache.org/jira/browse/SPARK-11219) to make PySpark MLlib parameter description formatting consistent. This is for the tree module.
closes#10601
Author: Bryan Cutler <cutlerb@gmail.com>
Author: vijaykiran <mail@vijaykiran.com>
Closes#11353 from BryanCutler/param-desc-consistent-tree-SPARK-12634.
## What changes were proposed in this pull request?
This is another try of PR #11323.
This PR removes DataFrame RDD operations except for `foreach` and `foreachPartitions` (they are actions rather than transformations). Original calls are now replaced by calls to methods of `DataFrame.rdd`.
PR #11323 was reverted because it introduced a regression: both `DataFrame.foreach` and `DataFrame.foreachPartitions` wrap underlying RDD operations with `withNewExecutionId` to track Spark jobs. But they are removed in #11323.
## How was the this patch tested?
No extra tests are added. Existing tests should do the work.
Author: Cheng Lian <lian@databricks.com>
Closes#11388 from liancheng/remove-df-rdd-ops.
Obtain the hive metastore and hbase token as well as hdfs token in DelegationToeknRenewer to supoort long-running application of spark on hbase or thriftserver.
Author: huangzhaowei <carlmartinmax@gmail.com>
Closes#10645 from SaintBacchus/SPARK-12523.
## What changes were proposed in this pull request?
This patch creates the public API for runtime configuration and an implementation for it. The public runtime configuration includes configs for existing SQL, as well as Hadoop Configuration.
This new interface is currently dead code. It will be added to SQLContext and a session entry point to Spark when we add that.
## How was this patch tested?
a new unit test suite
Author: Reynold Xin <rxin@databricks.com>
Closes#11378 from rxin/SPARK-13487.
## What changes were proposed in this pull request?
This Pull request is used for the fix SPARK-12941, creating a data type mapping to Oracle for the corresponding data type"Stringtype" from dataframe. This PR is for the master branch fix, where as another PR is already tested with the branch 1.4
## How was the this patch tested?
(Please explain how this patch was tested. E.g. unit tests, integration tests, manual tests)
This patch was tested using the Oracle docker .Created a new integration suite for the same.The oracle.jdbc jar was to be downloaded from the maven repository.Since there was no jdbc jar available in the maven repository, the jar was downloaded from oracle site manually and installed in the local; thus tested. So, for SparkQA test case run, the ojdbc jar might be manually placed in the local maven repository(com/oracle/ojdbc6/11.2.0.2.0) while Spark QA test run.
Author: thomastechs <thomas.sebastian@tcs.com>
Closes#11306 from thomastechs/master.
## What changes were proposed in this pull request?
Add ```approxQuantile``` for SparkR.
## How was this patch tested?
unit tests
Author: Yanbo Liang <ybliang8@gmail.com>
Closes#11383 from yanboliang/spark-13504 and squashes the following commits:
4f17adb [Yanbo Liang] Add approxQuantile for SparkR
Add export/import for all estimators and transformers(which have Scala implementation) under pyspark/ml/regression.py.
yanboliang Please help to review.
For doctest, I though it's enough to add one since it's common usage. But I can add to all if we want it.
Author: Tommy YU <tummyyu@163.com>
Closes#11000 from Wenpei/spark-13033-ml.regression-exprot-import and squashes the following commits:
3646b36 [Tommy YU] address review comments
9cddc98 [Tommy YU] change base on review and pr 11197
cc61d9d [Tommy YU] remove default parameter set
19535d4 [Tommy YU] add export/import to regression
44a9dc2 [Tommy YU] add import/export for ml.regression
jira: https://issues.apache.org/jira/browse/SPARK-13028
MaxAbsScaler works in a very similar way as MinMaxScaler, but scales in a way that the training data lies within the range [-1, 1] by dividing through the largest maximum value in each feature. The motivation to use this scaling includes robustness to very small standard deviations of features and preserving zero entries in sparse data.
Unlike StandardScaler and MinMaxScaler, MaxAbsScaler does not shift/center the data, and thus does not destroy any sparsity.
Something similar from sklearn:
http://scikit-learn.org/stable/modules/generated/sklearn.preprocessing.MaxAbsScaler.html#sklearn.preprocessing.MaxAbsScaler
Author: Yuhao Yang <hhbyyh@gmail.com>
Closes#10939 from hhbyyh/maxabs and squashes the following commits:
fd8bdcd [Yuhao Yang] add tag and some optimization on fit
648fced [Yuhao Yang] Merge remote-tracking branch 'upstream/master' into maxabs
75bebc2 [Yuhao Yang] Merge remote-tracking branch 'upstream/master' into maxabs
cb10bb6 [Yuhao Yang] remove minmax
91ef8f3 [Yuhao Yang] ut added
8ab0747 [Yuhao Yang] Merge remote-tracking branch 'upstream/master' into maxabs
a9215b5 [Yuhao Yang] max abs scaler
This pr added benchmark codes for Encoder#compress().
Also, it replaced the benchmark results with new ones because the output format of `Benchmark` changed.
Author: Takeshi YAMAMURO <linguin.m.s@gmail.com>
Closes#11236 from maropu/CompressionSpike.
## Motivation
As a pre-requisite to off-heap caching of blocks, we need a mechanism to prevent pages / blocks from being evicted while they are being read. With on-heap objects, evicting a block while it is being read merely leads to memory-accounting problems (because we assume that an evicted block is a candidate for garbage-collection, which will not be true during a read), but with off-heap memory this will lead to either data corruption or segmentation faults.
## Changes
### BlockInfoManager and reader/writer locks
This patch adds block-level read/write locks to the BlockManager. It introduces a new `BlockInfoManager` component, which is contained within the `BlockManager`, holds the `BlockInfo` objects that the `BlockManager` uses for tracking block metadata, and exposes APIs for locking blocks in either shared read or exclusive write modes.
`BlockManager`'s `get*()` and `put*()` methods now implicitly acquire the necessary locks. After a `get()` call successfully retrieves a block, that block is locked in a shared read mode. A `put()` call will block until it acquires an exclusive write lock. If the write succeeds, the write lock will be downgraded to a shared read lock before returning to the caller. This `put()` locking behavior allows us store a block and then immediately turn around and read it without having to worry about it having been evicted between the write and the read, which will allow us to significantly simplify `CacheManager` in the future (see #10748).
See `BlockInfoManagerSuite`'s test cases for a more detailed specification of the locking semantics.
### Auto-release of locks at the end of tasks
Our locking APIs support explicit release of locks (by calling `unlock()`), but it's not always possible to guarantee that locks will be released prior to the end of the task. One reason for this is our iterator interface: since our iterators don't support an explicit `close()` operator to signal that no more records will be consumed, operations like `take()` or `limit()` don't have a good means to release locks on their input iterators' blocks. Another example is broadcast variables, whose block locks can only be released at the end of the task.
To address this, `BlockInfoManager` uses a pair of maps to track the set of locks acquired by each task. Lock acquisitions automatically record the current task attempt id by obtaining it from `TaskContext`. When a task finishes, code in `Executor` calls `BlockInfoManager.unlockAllLocksForTask(taskAttemptId)` to free locks.
### Locking and the MemoryStore
In order to prevent in-memory blocks from being evicted while they are being read, the `MemoryStore`'s `evictBlocksToFreeSpace()` method acquires write locks on blocks which it is considering as candidates for eviction. These lock acquisitions are non-blocking, so a block which is being read will not be evicted. By holding write locks until the eviction is performed or skipped (in case evicting the blocks would not free enough memory), we avoid a race where a new reader starts to read a block after the block has been marked as an eviction candidate but before it has been removed.
### Locking and remote block transfer
This patch makes small changes to to block transfer and network layer code so that locks acquired by the BlockTransferService are released as soon as block transfer messages are consumed and released by Netty. This builds on top of #11193, a bug fix related to freeing of network layer ManagedBuffers.
## FAQ
- **Why not use Java's built-in [`ReadWriteLock`](https://docs.oracle.com/javase/7/docs/api/java/util/concurrent/locks/ReadWriteLock.html)?**
Our locks operate on a per-task rather than per-thread level. Under certain circumstances a task may consist of multiple threads, so using `ReadWriteLock` would mean that we might call `unlock()` from a thread which didn't hold the lock in question, an operation which has undefined semantics. If we could rely on Java 8 classes, we might be able to use [`StampedLock`](https://docs.oracle.com/javase/8/docs/api/java/util/concurrent/locks/StampedLock.html) to work around this issue.
- **Why not detect "leaked" locks in tests?**:
See above notes about `take()` and `limit`.
Author: Josh Rosen <joshrosen@databricks.com>
Closes#10705 from JoshRosen/pin-pages.
