## What changes were proposed in this pull request?
Some ML Models were using `defaultCopy` which expects a default constructor, and others were not setting the parent estimator. This change fixes these by creating a new instance of the model and explicitly setting values and parent.
## How was this patch tested?
Added `MLTestingUtils.checkCopy` to the offending models to tests to verify the copy is made and parent is set.
Author: Bryan Cutler <cutlerb@gmail.com>
Closes#17326 from BryanCutler/ml-model-copy-error-SPARK-19985.
## What changes were proposed in this pull request?
Add a set of catalog API in R
```
"currentDatabase",
"listColumns",
"listDatabases",
"listFunctions",
"listTables",
"recoverPartitions",
"refreshByPath",
"refreshTable",
"setCurrentDatabase",
```
https://github.com/apache/spark/pull/17483/files#diff-6929e6c5e59017ff954e110df20ed7ff
## How was this patch tested?
manual tests, unit tests
Author: Felix Cheung <felixcheung_m@hotmail.com>
Closes#17483 from felixcheung/rcatalog.
## What changes were proposed in this pull request?
If the shutdown hook called before the variable `uiTab` is set , it will throw a NullPointerException.
## How was this patch tested?
manual tests
Author: zuotingbing <zuo.tingbing9@zte.com.cn>
Closes#17496 from zuotingbing/SPARK-HiveThriftServer2.
JIRA Issue: https://issues.apache.org/jira/browse/SPARK-20123
## What changes were proposed in this pull request?
If $SPARK_HOME or $FWDIR variable contains spaces, then use "./dev/make-distribution.sh --name custom-spark --tgz -Psparkr -Phadoop-2.7 -Phive -Phive-thriftserver -Pmesos -Pyarn" build spark will failed.
## How was this patch tested?
manual tests
Author: zuotingbing <zuo.tingbing9@zte.com.cn>
Closes#17452 from zuotingbing/spark-bulid.
## What changes were proposed in this pull request?
Currently, `DataType.fromJson` throws `scala.MatchError` or `java.util.NoSuchElementException` in some cases when the JSON input is invalid as below:
```scala
DataType.fromJson(""""abcd"""")
```
```
java.util.NoSuchElementException: key not found: abcd
at ...
```
```scala
DataType.fromJson("""{"abcd":"a"}""")
```
```
scala.MatchError: JObject(List((abcd,JString(a)))) (of class org.json4s.JsonAST$JObject)
at ...
```
```scala
DataType.fromJson("""{"fields": [{"a":123}], "type": "struct"}""")
```
```
scala.MatchError: JObject(List((a,JInt(123)))) (of class org.json4s.JsonAST$JObject)
at ...
```
After this PR,
```scala
DataType.fromJson(""""abcd"""")
```
```
java.lang.IllegalArgumentException: Failed to convert the JSON string 'abcd' to a data type.
at ...
```
```scala
DataType.fromJson("""{"abcd":"a"}""")
```
```
java.lang.IllegalArgumentException: Failed to convert the JSON string '{"abcd":"a"}' to a data type.
at ...
```
```scala
DataType.fromJson("""{"fields": [{"a":123}], "type": "struct"}""")
at ...
```
```
java.lang.IllegalArgumentException: Failed to convert the JSON string '{"a":123}' to a field.
```
## How was this patch tested?
Unit test added in `DataTypeSuite`.
Author: hyukjinkwon <gurwls223@gmail.com>
Closes#17468 from HyukjinKwon/fromjson_exception.
## What changes were proposed in this pull request?
`BroadcastHint` should use child's statistics and set `isBroadcastable` to true.
## How was this patch tested?
Added a new stats estimation test for `BroadcastHint`.
Author: wangzhenhua <wangzhenhua@huawei.com>
Closes#17504 from wzhfy/broadcastHintEstimation.
### What changes were proposed in this pull request?
After we renames `Catalog`.`createExternalTable` to `createTable` in the PR: https://github.com/apache/spark/pull/16528, we also need to deprecate the corresponding functions in `SQLContext`.
### How was this patch tested?
N/A
Author: Xiao Li <gatorsmile@gmail.com>
Closes#17502 from gatorsmile/deprecateCreateExternalTable.
…anges.
## What changes were proposed in this pull request?
Document compression way little detail changes.
1.spark.eventLog.compress add 'Compression will use spark.io.compression.codec.'
2.spark.broadcast.compress add 'Compression will use spark.io.compression.codec.'
3,spark.rdd.compress add 'Compression will use spark.io.compression.codec.'
4.spark.io.compression.codec add 'event log describe'.
eg
Through the documents, I don't know what is compression mode about 'event log'.
## How was this patch tested?
manual tests
Please review http://spark.apache.org/contributing.html before opening a pull request.
Author: 郭小龙 10207633 <guo.xiaolong1@zte.com.cn>
Closes#17498 from guoxiaolongzte/SPARK-20177.
## What changes were proposed in this pull request?
- Encoder's deserializer must be resolved at the driver where the class is defined. Otherwise there are corner cases using nested classes where resolving at the executor can fail.
- Fixed flaky test related to processing time timeout. The flakiness is caused because the test thread (that adds data to memory source) has a race condition with the streaming query thread. When testing the manual clock, the goal is to add data and increment clock together atomically, such that a trigger sees new data AND updated clock simultaneously (both or none). This fix adds additional synchronization in when adding data; it makes sure that the streaming query thread is waiting on the manual clock to be incremented (so no batch is currently running) before adding data.
- Added`testQuietly` on some tests that generate a lot of error logs.
## How was this patch tested?
Multiple runs on existing unit tests
Author: Tathagata Das <tathagata.das1565@gmail.com>
Closes#17488 from tdas/SPARK-20165.
### What changes were proposed in this pull request?
`ParquetConversions` and `OrcConversions` should be treated as regular `Analyzer` rules. It is not reasonable to be part of `HiveSessionCatalog`. This PR also combines two rules `ParquetConversions` and `OrcConversions` to build a new rule `RelationConversions `.
After moving these two rules out of HiveSessionCatalog, the next step is to clean up, rename and move `HiveMetastoreCatalog` because it is not related to the hive package any more.
### How was this patch tested?
The existing test cases
Author: Xiao Li <gatorsmile@gmail.com>
Closes#17484 from gatorsmile/cleanup.
## What changes were proposed in this pull request?
Remove accumulator updates for internal.metrics.updatedBlockStatuses from SparkListenerTaskEnd entries in the history file. These can cause history files to grow to hundreds of GB because the value of the accumulator contains all tracked blocks.
## How was this patch tested?
Current History UI tests cover use of the history file.
Author: Ryan Blue <blue@apache.org>
Closes#17412 from rdblue/SPARK-20084-remove-block-accumulator-info.
## What changes were proposed in this pull request?
The query plan in an `AnalysisException` may be `null` when an `AnalysisException` object is serialized and then deserialized, since `plan` is marked `transient`. Or when someone throws an `AnalysisException` with a null query plan (which should not happen).
`def getMessage` is not tolerant of this and throws a `NullPointerException`, leading to loss of information about the original exception.
The fix is to add a `null` check in `getMessage`.
## How was this patch tested?
- Unit test
Author: Kunal Khamar <kkhamar@outlook.com>
Closes#17486 from kunalkhamar/spark-20164.
## What changes were proposed in this pull request?
After SPARK-20136, we report metadata timing metrics in scan operator. However, that timing metric doesn't include one of the most important part of metadata, which is partition pruning. This patch adds that time measurement to the scan metrics.
## How was this patch tested?
N/A - I tried adding a test in SQLMetricsSuite but it was extremely convoluted to the point that I'm not sure if this is worth it.
Author: Reynold Xin <rxin@databricks.com>
Closes#17476 from rxin/SPARK-20151.
## What changes were proposed in this pull request?
Instead of iterating all expressions that can return null for null inputs, we can just check `NullIntolerant`.
## How was this patch tested?
existing tests
Author: Wenchen Fan <wenchen@databricks.com>
Closes#17450 from cloud-fan/null.
## What changes were proposed in this pull request?
Few changes related to Intellij IDEA inspection.
## How was this patch tested?
Changes were tested by existing unit tests
Author: Denis Bolshakov <denis.bolshakov@onefactor.com>
Closes#17458 from dbolshak/SPARK-20127.
Fixed a few typos.
There is one more I'm not sure of:
```
Append mode uses watermark to drop old aggregation state. But the output of a
windowed aggregation is delayed the late threshold specified in `withWatermark()` as by
the modes semantics, rows can be added to the Result Table only once after they are
```
Not sure how to change `is delayed the late threshold`.
Author: Seigneurin, Alexis (CONT) <Alexis.Seigneurin@capitalone.com>
Closes#17443 from aseigneurin/typos.
## What changes were proposed in this pull request?
while submit apps with -v or --verbose， we can print the right queue name, but if we set a queue name with `spark.yarn.queue` by --conf or in the spark-default.conf, we just got `null` for the queue in Parsed arguments.
```
bin/spark-shell -v --conf spark.yarn.queue=thequeue
Using properties file: /home/hadoop/spark-2.1.0-bin-apache-hdp2.7.3/conf/spark-defaults.conf
....
Adding default property: spark.yarn.queue=default
Parsed arguments:
master yarn
deployMode client
...
queue null
....
verbose true
Spark properties used, including those specified through
--conf and those from the properties file /home/hadoop/spark-2.1.0-bin-apache-hdp2.7.3/conf/spark-defaults.conf:
spark.yarn.queue -> thequeue
....
```
## How was this patch tested?
ut and local verify
Author: Kent Yao <yaooqinn@hotmail.com>
Closes#17430 from yaooqinn/SPARK-20096.
java.nio.Bits.unaligned() does not return true for the ppc64le arch.
see https://bugs.openjdk.java.net/browse/JDK-8165231
## What changes were proposed in this pull request?
check architecture
## How was this patch tested?
unit test
Author: samelamin <hussam.elamin@gmail.com>
Author: samelamin <sam_elamin@discovery.com>
Closes#17472 from samelamin/SPARK-19999.
…adoc
## What changes were proposed in this pull request?
Use recommended values for row boundaries in Window's scaladoc, i.e. `Window.unboundedPreceding`, `Window.unboundedFollowing`, and `Window.currentRow` (that were introduced in 2.1.0).
## How was this patch tested?
Local build
Author: Jacek Laskowski <jacek@japila.pl>
Closes#17417 from jaceklaskowski/window-expression-scaladoc.
## What changes were proposed in this pull request?
Implementations of strategies for resilient block replication for different resource managers that replicate the 3-replica strategy used by HDFS, where the first replica is on an executor, the second replica within the same rack as the executor and a third replica on a different rack.
The implementation involves providing two pluggable classes, one running in the driver that provides topology information for every host at cluster start and the second prioritizing a list of peer BlockManagerIds.
The prioritization itself can be thought of an optimization problem to find a minimal set of peers that satisfy certain objectives and replicating to these peers first. The objectives can be used to express richer constraints over and above HDFS like 3-replica strategy.
## How was this patch tested?
This patch was tested with unit tests for storage, along with new unit tests to verify prioritization behaviour.
Author: Shubham Chopra <schopra31@bloomberg.net>
Closes#13932 from shubhamchopra/PrioritizerStrategy.
## What changes were proposed in this pull request?
There are two examples in r folder missing the run commands.
In this PR, I just add the missing comment, which is consistent with other examples.
## How was this patch tested?
Manual test.
Author: wm624@hotmail.com <wm624@hotmail.com>
Closes#17474 from wangmiao1981/stat.
## What changes were proposed in this pull request?
The internal FileCommitProtocol interface returns all task commit messages in bulk to the implementation when a job finishes. However, it is sometimes useful to access those messages before the job completes, so that the driver gets incremental progress updates before the job finishes.
This adds an `onTaskCommit` listener to the internal api.
## How was this patch tested?
Unit tests.
cc rxin
Author: Eric Liang <ekl@databricks.com>
Closes#17475 from ericl/file-commit-api-ext.
## What changes were proposed in this pull request?
This patch adds explicit metadata operation timing and number of files in data source metrics. Those would be useful to include for performance profiling.
Screenshot of a UI with this change (num files and metadata time are new metrics):
<img width="321" alt="screen shot 2017-03-29 at 12 29 28 am" src="https://cloud.githubusercontent.com/assets/323388/24443272/d4ea58c0-1416-11e7-8940-ecb69375554a.png">
## How was this patch tested?
N/A
Author: Reynold Xin <rxin@databricks.com>
Closes#17465 from rxin/SPARK-20136.
## What changes were proposed in this pull request?
The column comment was missing while constructing the Hive TableSchema. This fix will preserve the original comment.
## How was this patch tested?
I have added a new test case to test the column with/without comment.
Author: bomeng <bmeng@us.ibm.com>
Closes#17470 from bomeng/SPARK-20146.
## What changes were proposed in this pull request?
Commit 6c70a38 broke the build for scala 2.10. The commit uses some reflections which are not available in Scala 2.10. This PR fixes them.
## How was this patch tested?
Existing tests.
Author: Takuya UESHIN <ueshin@databricks.com>
Closes#17473 from ueshin/issues/SPARK-19088.
### What changes were proposed in this pull request?
`FalseLiteral` and `TrueLiteral` should have been eliminated by optimizer rule `BooleanSimplification`, but null literals might be added by optimizer rule `NullPropagation`. For safety, our filter estimation should handle all the eligible literal cases.
Our optimizer rule BooleanSimplification is unable to remove the null literal in many cases. For example, `a < 0 or null`. Thus, we need to handle null literal in filter estimation.
`Not` can be pushed down below `And` and `Or`. Then, we could see two consecutive `Not`, which need to be collapsed into one. Because of the limited expression support for filter estimation, we just need to handle the case `Not(null)` for avoiding incorrect error due to the boolean operation on null. For details, see below matrix.
```
not NULL = NULL
NULL or false = NULL
NULL or true = true
NULL or NULL = NULL
NULL and false = false
NULL and true = NULL
NULL and NULL = NULL
```
### How was this patch tested?
Added the test cases.
Author: Xiao Li <gatorsmile@gmail.com>
Closes#17446 from gatorsmile/constantFilterEstimation.
## What changes were proposed in this pull request?
This pr added `StructType.fromDDL` to convert a DDL format string into `StructType` for defining schemas in `functions.from_json`.
## How was this patch tested?
Added tests in `JsonFunctionsSuite`.
Author: Takeshi Yamamuro <yamamuro@apache.org>
Closes#17406 from maropu/SPARK-20009.
## What changes were proposed in this pull request?
Bugfix from [SPARK-19540.](https://github.com/apache/spark/pull/16826)
Cloning SessionState does not clone query execution listeners, so cloned session is unable to listen to events on queries.
## How was this patch tested?
- Unit test
Author: Kunal Khamar <kkhamar@outlook.com>
Closes#17379 from kunalkhamar/clone-bugfix.
## What changes were proposed in this pull request?
Allow Jenkins Python tests to use the installed conda to test Python 2.7 support & test pip installability.
## How was this patch tested?
Updated shell scripts, ran tests locally with installed conda, ran tests in Jenkins.
Author: Holden Karau <holden@us.ibm.com>
Closes#17355 from holdenk/SPARK-19955-support-python-tests-with-conda.
## What changes were proposed in this pull request?
Currently we use system classloader to find HBase jars, if it is specified by `--jars`, then it will be failed with ClassNotFound issue. So here changing to use child classloader.
Also putting added jars and main jar into classpath of submitted application in yarn cluster mode, otherwise HBase jars specified with `--jars` will never be honored in cluster mode, and fetching tokens in client side will always be failed.
## How was this patch tested?
Unit test and local verification.
Author: jerryshao <sshao@hortonworks.com>
Closes#17388 from jerryshao/SPARK-20059.
This change modifies the way block data is encrypted to make the more
common cases faster, while penalizing an edge case. As a side effect
of the change, all data that goes through the block manager is now
encrypted only when needed, including the previous path (broadcast
variables) where that did not happen.
The way the change works is by not encrypting data that is stored in
memory; so if a serialized block is in memory, it will only be encrypted
once it is evicted to disk.
The penalty comes when transferring that encrypted data from disk. If the
data ends up in memory again, it is as efficient as before; but if the
evicted block needs to be transferred directly to a remote executor, then
there's now a performance penalty, since the code now uses a custom
FileRegion implementation to decrypt the data before transferring.
This also means that block data transferred between executors now is
not encrypted (and thus relies on the network library encryption support
for secrecy). Shuffle blocks are still transferred in encrypted form,
since they're handled in a slightly different way by the code. This also
keeps compatibility with existing external shuffle services, which transfer
encrypted shuffle blocks, and avoids having to make the external service
aware of encryption at all.
The serialization and deserialization APIs in the SerializerManager now
do not do encryption automatically; callers need to explicitly wrap their
streams with an appropriate crypto stream before using those.
As a result of these changes, some of the workarounds added in SPARK-19520
are removed here.
Testing: a new trait ("EncryptionFunSuite") was added that provides an easy
way to run a test twice, with encryption on and off; broadcast, block manager
and caching tests were modified to use this new trait so that the existing
tests exercise both encrypted and non-encrypted paths. I also ran some
applications with encryption turned on to verify that they still work,
including streaming tests that failed without the fix for SPARK-19520.
Author: Marcelo Vanzin <vanzin@cloudera.com>
Closes#17295 from vanzin/SPARK-19556.
## What changes were proposed in this pull request?
It is not super intuitive how to update SQLMetric on the driver side. This patch introduces a new SQLMetrics.postDriverMetricUpdates function to do that, and adds documentation to make it more obvious.
## How was this patch tested?
Updated a test case to use this method.
Author: Reynold Xin <rxin@databricks.com>
Closes#17464 from rxin/SPARK-20134.
## What changes were proposed in this pull request?
A pyspark wrapper for spark.ml.stat.ChiSquareTest
## How was this patch tested?
unit tests
doctests
Author: Bago Amirbekian <bago@databricks.com>
Closes#17421 from MrBago/chiSquareTestWrapper.
## What changes were proposed in this pull request?
We must set the taskset to zombie before the DAGScheduler handles the taskEnded event. It's possible the taskEnded event will cause the DAGScheduler to launch a new stage attempt (this happens when map output data was lost), and if this happens before the taskSet has been set to zombie, it will appear that we have conflicting task sets.
Author: liujianhui <liujianhui@didichuxing>
Closes#17208 from liujianhuiouc/spark-19868.
## What changes were proposed in this pull request?
When we build the deserializer expression for map type, we will use `StaticInvoke` to call `ArrayBasedMapData.toScalaMap`, and declare the return type as `scala.collection.immutable.Map`. If the map is inside an Option, we will wrap this `StaticInvoke` with `WrapOption`, which requires the input to be `scala.collect.Map`. Ideally this should be fine, as `scala.collection.immutable.Map` extends `scala.collect.Map`, but our `ObjectType` is too strict about this, this PR fixes it.
## How was this patch tested?
new regression test
Author: Wenchen Fan <wenchen@databricks.com>
Closes#17454 from cloud-fan/map.
## What changes were proposed in this pull request?
In the current Spark on YARN code, we will obtain tokens from provided services, but we're not going to add these tokens to the current user's credentials. This will make all the following operations to these services still require TGT rather than delegation tokens. This is unnecessary since we already got the tokens, also this will lead to failure in user impersonation scenario, because the TGT is granted by real user, not proxy user.
So here changing to put all the tokens to the current UGI, so that following operations to these services will honor tokens rather than TGT, and this will further handle the proxy user issue mentioned above.
## How was this patch tested?
Local verified in secure cluster.
vanzin tgravescs mridulm dongjoon-hyun please help to review, thanks a lot.
Author: jerryshao <sshao@hortonworks.com>
Closes#17335 from jerryshao/SPARK-19995.
## What changes were proposed in this pull request?
Commit ea361165e1 moved most of the logic from the SessionState classes into an accompanying builder. This makes the existence of the `HiveSessionState` redundant. This PR removes the `HiveSessionState`.
## How was this patch tested?
Existing tests.
Author: Herman van Hovell <hvanhovell@databricks.com>
Closes#17457 from hvanhovell/SPARK-20126.
## What changes were proposed in this pull request?
Join reorder algorithm should keep exactly the same order of output attributes in the top project.
For example, if user want to select a, b, c, after reordering, we should output a, b, c in the same order as specified by user, instead of b, a, c or other orders.
## How was this patch tested?
A new test case is added in `JoinReorderSuite`.
Author: wangzhenhua <wangzhenhua@huawei.com>
Closes#17453 from wzhfy/keepOrderInProject.
## What changes were proposed in this pull request?
TPCDS q45 fails becuase:
`ReorderJoin` collects all predicates and try to put them into join condition when creating ordered join. If a predicate with an IN subquery (`ListQuery`) is in a join condition instead of a filter condition, `RewritePredicateSubquery.rewriteExistentialExpr` would fail to convert the subquery to an `ExistenceJoin`, and thus result in error.
We should prevent push down of IN subquery to Join operator.
## How was this patch tested?
Add a new test case in `FilterPushdownSuite`.
Author: wangzhenhua <wangzhenhua@huawei.com>
Closes#17428 from wzhfy/noSubqueryInJoinCond.
### What changes were proposed in this pull request?
Changed the pattern to match the first n characters in the location field so that the string truncation does not affect it.
### How was this patch tested?
N/A
Author: Xiao Li <gatorsmile@gmail.com>
Closes#17448 from gatorsmile/fixTestCAse.
## What changes were proposed in this pull request?
The current SessionState initialization code path is quite complex. A part of the creation is done in the SessionState companion objects, a part of the creation is one inside the SessionState class, and a part is done by passing functions.
This PR refactors this code path, and consolidates SessionState initialization into a builder class. This SessionState will not do any initialization and just becomes a place holder for the various Spark SQL internals. This also lays the ground work for two future improvements:
1. This provides us with a start for removing the `HiveSessionState`. Removing the `HiveSessionState` would also require us to move resource loading into a separate class, and to (re)move metadata hive.
2. This makes it easier to customize the Spark Session. Currently you will need to create a custom version of the builder. I have added hooks to facilitate this. A future step will be to create a semi stable API on top of this.
## How was this patch tested?
Existing tests.
Author: Herman van Hovell <hvanhovell@databricks.com>
Closes#17433 from hvanhovell/SPARK-20100.
## What changes were proposed in this pull request?
Existing logic seemingly writes null to the BatchCommitLog, even though it does additional checks to write '{}' (valid json) to the log. This PR simplifies the logic by disallowing use of `log.add(batchId, metadata)` and instead using `log.add(batchId)`. No question of specifying metadata, so no confusion related to null.
## How was this patch tested?
Existing tests pass.
Author: Tathagata Das <tathagata.das1565@gmail.com>
Closes#17444 from tdas/SPARK-19876-1.
## What changes were proposed in this pull request?
Executors cache a list of their peers that is refreshed by default every minute. The cached stale references were randomly being used for replication. Since those executors were removed from the master, they did not occur in the block locations as reported by the master. This was fixed by
1. Refreshing peer cache in the block manager before trying to pro-actively replicate. This way the probability of replicating to a failed executor is eliminated.
2. Explicitly stopping the block manager in the tests. This shuts down the RPC endpoint use by the block manager. This way, even if a block manager tries to replicate using a stale reference, the replication logic should take care of refreshing the list of peers after failure.
## How was this patch tested?
Tested manually
Author: Shubham Chopra <schopra31@bloomberg.net>
Author: Kay Ousterhout <kayousterhout@gmail.com>
Author: Shubham Chopra <shubhamchopra@users.noreply.github.com>
Closes#17325 from shubhamchopra/SPARK-19803.
## What changes were proposed in this pull request?
It seems `checkType` and the type string in `structField` are not being tested closely. This string format currently seems SparkR-specific (see d1f6c64c4b/sql/core/src/main/scala/org/apache/spark/sql/api/r/SQLUtils.scala (L93-L131)) but resembles SQL type definition.
Therefore, it seems nicer if we test positive/negative cases in R side.
## How was this patch tested?
Unit tests in `test_sparkSQL.R`.
Author: hyukjinkwon <gurwls223@gmail.com>
Closes#17439 from HyukjinKwon/r-typestring-tests.
## What changes were proposed in this pull request?
The master snapshot publisher builds are currently broken due to two minor build issues:
1. For unknown reasons, the LFTP `mkdir -p` command began throwing errors when the remote directory already exists. This change of behavior might have been caused by configuration changes in the ASF's SFTP server, but I'm not entirely sure of that. To work around this problem, this patch updates the script to ignore errors from the `lftp mkdir -p` commands.
2. The PySpark `setup.py` file references a non-existent `pyspark.ml.stat` module, causing Python packaging to fail by complaining about a missing directory. The fix is to simply drop that line from the setup script.
## How was this patch tested?
The LFTP fix was tested by manually running the failing commands on AMPLab Jenkins against the ASF SFTP server. The PySpark fix was tested locally.
Author: Josh Rosen <joshrosen@databricks.com>
Closes#17437 from JoshRosen/spark-20102.
