## What changes were proposed in this pull request?
This happens in a schema evolution use case only when a user specifies the schema manually and use non-vectorized ORC deserializer code path.
There is a bug in `OrcDeserializer.scala` that results in `null`s being set at the wrong column position, and for state from previous records to remain uncleared in next records. There are more details for when exactly the bug gets triggered and what the outcome is in the [JIRA issue](https://jira.apache.org/jira/browse/SPARK-26859).
The high-level summary is that this bug results in severe data correctness issues, but fortunately the set of conditions to expose the bug are complicated and make the surface area somewhat small.
This change fixes the problem and adds a respective test.
## How was this patch tested?
Pass the Jenkins with the newly added test cases.
Closes#23766 from IvanVergiliev/fix-orc-deserializer.
Lead-authored-by: Ivan Vergiliev <ivan.vergiliev@gmail.com>
Co-authored-by: Dongjoon Hyun <dongjoon@apache.org>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
## What changes were proposed in this pull request?
This PR targets to support Arrow optimization for conversion from Spark DataFrame to R DataFrame.
Like PySpark side, it falls back to non-optimization code path when it's unable to use Arrow optimization.
This can be tested as below:
```bash
$ ./bin/sparkR --conf spark.sql.execution.arrow.enabled=true
```
```r
collect(createDataFrame(mtcars))
```
### Requirements
- R 3.5.x
- Arrow package 0.12+
```bash
Rscript -e 'remotes::install_github("apache/arrowapache-arrow-0.12.0", subdir = "r")'
```
**Note:** currently, Arrow R package is not in CRAN. Please take a look at ARROW-3204.
**Note:** currently, Arrow R package seems not supporting Windows. Please take a look at ARROW-3204.
### Benchmarks
**Shall**
```bash
sync && sudo purge
./bin/sparkR --conf spark.sql.execution.arrow.enabled=false --driver-memory 4g
```
```bash
sync && sudo purge
./bin/sparkR --conf spark.sql.execution.arrow.enabled=true --driver-memory 4g
```
**R code**
```r
df <- cache(createDataFrame(read.csv("500000.csv")))
count(df)
test <- function() {
options(digits.secs = 6) # milliseconds
start.time <- Sys.time()
collect(df)
end.time <- Sys.time()
time.taken <- end.time - start.time
print(time.taken)
}
test()
```
**Data (350 MB):**
```r
object.size(read.csv("500000.csv"))
350379504 bytes
```
"500000 Records" http://eforexcel.com/wp/downloads-16-sample-csv-files-data-sets-for-testing/
**Results**
```
Time difference of 221.32014 secs
```
```
Time difference of 15.51145 secs
```
The performance improvement was around **1426%**.
### Limitations:
- For now, Arrow optimization with R does not support when the data is `raw`, and when user explicitly gives float type in the schema. They produce corrupt values. In this case, we decide to fall back to non-optimization code path.
- Due to ARROW-4512, it cannot send and receive batch by batch. It has to send all batches in Arrow stream format at once. It needs improvement later.
## How was this patch tested?
Existing tests related with Arrow optimization cover this change. Also, manually tested.
Closes#23760 from HyukjinKwon/SPARK-26762.
Authored-by: Hyukjin Kwon <gurwls223@apache.org>
Signed-off-by: Hyukjin Kwon <gurwls223@apache.org>
## What changes were proposed in this pull request?
This is a followup PR for #21149.
New way uses unsafeRow.hashCode() as hash value in HashAggregate.
The unsafe row has [null bit set] etc., so the hash should be different from shuffle hash, and then we don't need a special seed.
## How was this patch tested?
UTs.
Closes#23821 from yucai/unsafe_hash.
Authored-by: yucai <yyu1@ebay.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
## What changes were proposed in this pull request?
This proposes adding a note on `QueryExecution.toRdd` regarding Spark's internal optimization callers would need to indicate.
## How was this patch tested?
This patch is a documentation change.
Closes#23822 from HeartSaVioR/MINOR-doc-add-note-query-execution-to-rdd.
Authored-by: Jungtaek Lim (HeartSaVioR) <kabhwan@gmail.com>
Signed-off-by: Hyukjin Kwon <gurwls223@apache.org>
## What changes were proposed in this pull request?
Continue the API refactor for streaming write, according to the [doc](https://docs.google.com/document/d/1vI26UEuDpVuOjWw4WPoH2T6y8WAekwtI7qoowhOFnI4/edit?usp=sharing).
The major changes:
1. rename `StreamingWriteSupport` to `StreamingWrite`
2. add `WriteBuilder.buildForStreaming`
3. update existing sinks, to move the creation of `StreamingWrite` to `Table`
## How was this patch tested?
existing tests
Closes#23702 from cloud-fan/stream-write.
Authored-by: Wenchen Fan <wenchen@databricks.com>
Signed-off-by: gatorsmile <gatorsmile@gmail.com>
## What changes were proposed in this pull request?
This PR partially revert SPARK-26744.
60caa92dea and 4dce45a599 were merged at similar time range independently. So the test failures were not caught.
- 60caa92dea happened to add a schema reading logic in writing path for overwrite mode as well.
- 4dce45a599 added some tests with overwrite modes with migrated ORC v2.
And the tests looks starting to fail.
I guess the discussion won't be short (see https://github.com/apache/spark/pull/23606#discussion_r257675083) and this PR proposes to disable the tests added at 4dce45a599 to unblock other PRs for now.
## How was this patch tested?
Existing tests.
Closes#23828 from HyukjinKwon/SPARK-26744.
Authored-by: Hyukjin Kwon <gurwls223@apache.org>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
## What changes were proposed in this pull request?
This adds two logical plans that implement the ReplaceData operation from the [logical plans SPIP](https://docs.google.com/document/d/1gYm5Ji2Mge3QBdOliFV5gSPTKlX4q1DCBXIkiyMv62A/edit?ts=5a987801#heading=h.m45webtwxf2d). These two plans will be used to implement Spark's `INSERT OVERWRITE` behavior for v2.
Specific changes:
* Add `SupportsTruncate`, `SupportsOverwrite`, and `SupportsDynamicOverwrite` to DSv2 write API
* Add `OverwriteByExpression` and `OverwritePartitionsDynamic` plans (logical and physical)
* Add new plans to DSv2 write validation rule `ResolveOutputRelation`
* Refactor `WriteToDataSourceV2Exec` into trait used by all DSv2 write exec nodes
## How was this patch tested?
* The v2 analysis suite has been updated to validate the new overwrite plans
* The analysis suite for `OverwriteByExpression` checks that the delete expression is resolved using the table's columns
* Existing tests validate that overwrite exec plan works
* Updated existing v2 test because schema is used to validate overwrite
Closes#23606 from rdblue/SPARK-26666-add-overwrite.
Authored-by: Ryan Blue <blue@apache.org>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
## What changes were proposed in this pull request?
The previous strategy for comparing Maps leveraged sorting (key, value) tuples by their _.toString. However, the _.toString representation of an arrays has nothing to do with it's content. If a map has array keys, it's (key, value) pairs would be compared with other maps essentially at random. This could results in false negatives in tests.
This changes first compares keys together to find the matching ones, and then compares associated values.
## How was this patch tested?
New unit test added.
Closes#23789 from ala/compare-map.
Authored-by: Ala Luszczak <ala@databricks.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
## What changes were proposed in this pull request?
The file source has a schema validation feature, which validates 2 schemas:
1. the user-specified schema when reading.
2. the schema of input data when writing.
If a file source doesn't support the schema, we can fail the query earlier.
This PR is to implement the same feature in the `FileDataSourceV2` framework. Comparing to `FileFormat`, `FileDataSourceV2` has multiple layers. The API is added in two places:
1. Read path: the table schema is determined in `TableProvider.getTable`. The actual read schema can be a subset of the table schema. This PR proposes to validate the actual read schema in `FileScan`.
2. Write path: validate the actual output schema in `FileWriteBuilder`.
## How was this patch tested?
Unit test
Closes#23714 from gengliangwang/schemaValidationV2.
Authored-by: Gengliang Wang <gengliang.wang@databricks.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
## What changes were proposed in this pull request?
This is a followup PR to fix two issues in #23601:
1. the class `FileWriterFactory` contains `conf: SerializableConfiguration` as a member, which is duplicated with `WriteJobDescription. serializableHadoopConf `. By removing it we can reduce the broadcast task binary size by around 70KB
2. The test suite `OrcV1QuerySuite`/`OrcV1QuerySuite`/`OrcV1PartitionDiscoverySuite` didn't change the configuration `SQLConf.USE_V1_SOURCE_WRITER_LIST` to `"orc"`. We should set the conf.
## How was this patch tested?
Unit test
Closes#23800 from gengliangwang/reduceWriteTaskSize.
Authored-by: Gengliang Wang <gengliang.wang@databricks.com>
Signed-off-by: Hyukjin Kwon <gurwls223@apache.org>
## What changes were proposed in this pull request?
In https://github.com/apache/spark/pull/23383, the file source V2 framework is implemented. In the PR, `FileIndex` is created as a member of `FileTable`, so that we can implement partition pruning like 0f9fcabb4a in the future(As data source V2 catalog is under development, partition pruning is removed from the PR)
However, after write path of file source V2 is implemented, I find that a simple write will create an unnecessary `FileIndex`, which is required by `FileTable`. This is a sort of regression. And we can see there is a warning message when writing to ORC files
```
WARN InMemoryFileIndex: The directory file:/tmp/foo was not found. Was it deleted very recently?
```
This PR is to make `FileIndex` as a lazy value in `FileTable`, so that we can avoid creating unnecessary `FileIndex` in the write path.
## How was this patch tested?
Existing unit test
Closes#23774 from gengliangwang/moveFileIndexInV2.
Authored-by: Gengliang Wang <gengliang.wang@databricks.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
## What changes were proposed in this pull request?
This is a fix for https://github.com/apache/spark/pull/23524, which did not stop cost-based join reorder when the CostBasedJoinReorder rule recurses down the tree and applies join reorder for nested joins with hints.
The issue had not been detected by the existing tests because CBO is disabled by default.
## How was this patch tested?
Enabled CBO for JoinHintSuite.
Closes#23759 from maryannxue/spark-26840.
Lead-authored-by: maryannxue <maryannxue@apache.org>
Co-authored-by: Dongjoon Hyun <dhyun@apple.com>
Signed-off-by: gatorsmile <gatorsmile@gmail.com>
## What changes were proposed in this pull request?
These builtin typed aggregate functions are not very useful:
1. users can just call the untyped ones and turn the resulting dataframe to a dataset. It has better performance.
2. the typed aggregate functions have subtle different behaviors regarding empty input.
I think we should get rid of these builtin typed agg functions and suggest users to use the untyped ones.
However, these functions are still useful as a demo of the `Aggregator` API, so I copied them to the example module.
## How was this patch tested?
N/A
Closes#23763 from cloud-fan/example.
Authored-by: Wenchen Fan <wenchen@databricks.com>
Signed-off-by: gatorsmile <gatorsmile@gmail.com>
## What changes were proposed in this pull request?
Updates FileFormatWriter to create a consistent Hadoop Job ID for a write.
## How was this patch tested?
Existing tests for regressions.
Closes#23777 from rdblue/SPARK-26873-fix-file-format-writer-job-ids.
Authored-by: Ryan Blue <blue@apache.org>
Signed-off-by: Marcelo Vanzin <vanzin@cloudera.com>
## What changes were proposed in this pull request?
This PR is a correctness fix in `HashAggregateExec` code generation. It forces evaluation of result expressions before calling `consume()` to avoid multiple executions.
This PR fixes a use case where an aggregate is nested into a broadcast join and appears on the "stream" side. The issue is that Broadcast join generates it's own loop. And without forcing evaluation of `resultExpressions` of `HashAggregateExec` before the join's loop these expressions can be executed multiple times giving incorrect results.
## How was this patch tested?
New UT was added.
Closes#23731 from peter-toth/SPARK-26572.
Authored-by: Peter Toth <peter.toth@gmail.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
## What changes were proposed in this pull request?
```java
public class SqlDemo {
public static void main(final String[] args) throws Exception {
SparkConf conf = new SparkConf().setAppName("spark-sql-demo");
JavaSparkContext sc = new JavaSparkContext(conf);
SparkSession ss = SparkSession.builder().enableHiveSupport().getOrCreate();
ss.sql("show databases").show();
}
}
```
Before https://issues.apache.org/jira/browse/SPARK-20946, the demo above point to the right hive metastore if the hive-site.xml is present. But now it can only point to the default in-memory one.
Catalog is now as a variable shared across SparkSessions, it is instantiated with SparkContext's conf. After https://issues.apache.org/jira/browse/SPARK-20946, Session level configs are not pass to SparkContext's conf anymore, so the enableHiveSupport API takes no affect on the catalog instance.
You can set spark.sql.catalogImplementation=hive application wide to solve the problem, or never create a sc before you call SparkSession.builder().enableHiveSupport().getOrCreate()
Here we respect the SparkSession level configuration at the first time to generate catalog within SharedState
## How was this patch tested?
1. add ut
2. manually
```scala
test("enableHiveSupport has right to determine the catalog while using an existing sc") {
val conf = new SparkConf().setMaster("local").setAppName("SharedState Test")
val sc = SparkContext.getOrCreate(conf)
val ss = SparkSession.builder().enableHiveSupport().getOrCreate()
assert(ss.sharedState.externalCatalog.unwrapped.isInstanceOf[HiveExternalCatalog],
"The catalog should be hive ")
val ss2 = SparkSession.builder().getOrCreate()
assert(ss2.sharedState.externalCatalog.unwrapped.isInstanceOf[HiveExternalCatalog],
"The catalog should be shared across sessions")
}
```
Without this fix, the above test will fail.
You can apply it to `org.apache.spark.sql.hive.HiveSharedStateSuite`,
and run,
```sbt
./build/sbt -Phadoop-2.7 -Phive "hive/testOnly org.apache.spark.sql.hive.HiveSharedStateSuite"
```
to verify.
Closes#23709 from yaooqinn/SPARK-26794.
Authored-by: Kent Yao <yaooqinn@hotmail.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
## What changes were proposed in this pull request?
According to Brian Goetz et al in Java Concurrency in Practice, the double checked locking pattern has worked since Java 5, but only if the resource is declared volatile:
> Subsequent changes in the JMM (Java 5.0 and later) have enabled DCL to work if resource is made volatile, and the performance impact of this is small since volatile reads are usually only slightly more expensive than nonvolatile reads.
CachedRDDBuilder. cachedColumnBuffers and CachedRDDBuilder.clearCache both use DCL to manage the resource ``_cachedColumnBuffers``. The missing ingredient is that ``_cachedColumnBuffers`` is not volatile.
Because of this, clearCache may see ``_cachedColumnBuffers`` as null, when in fact it is not, and therefore fail to un-cache the RDD. There may be other, more subtle bugs due to visibility issues.
To avoid these issues, this PR makes ``_cachedColumnBuffers`` volatile.
## How was this patch tested?
- Existing SQL unit tests
- Existing pyspark-sql tests
Closes#23768 from bersprockets/SPARK-26851.
Authored-by: Bruce Robbins <bersprockets@gmail.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
## What changes were proposed in this pull request?
In the PR, I propose to use `System.nanoTime()` instead of `System.currentTimeMillis()` in measurements of time intervals.
`System.currentTimeMillis()` returns current wallclock time and will follow changes to the system clock. Thus, negative wallclock adjustments can cause timeouts to "hang" for a long time (until wallclock time has caught up to its previous value again). This can happen when ntpd does a "step" after the network has been disconnected for some time. The most canonical example is during system bootup when DHCP takes longer than usual. This can lead to failures that are really hard to understand/reproduce. `System.nanoTime()` is guaranteed to be monotonically increasing irrespective of wallclock changes.
## How was this patch tested?
By existing test suites.
Closes#23727 from MaxGekk/system-nanotime.
Lead-authored-by: Maxim Gekk <max.gekk@gmail.com>
Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com>
Signed-off-by: Sean Owen <sean.owen@databricks.com>
## What changes were proposed in this pull request?
There is a very old TODO in `HandleNullInputsForUDF`, saying that we can skip the null check if input is not nullable. We leverage the nullability info at many places, we can trust it here too.
## How was this patch tested?
re-enable an ignored test
Closes#23712 from cloud-fan/minor.
Lead-authored-by: Wenchen Fan <wenchen@databricks.com>
Co-authored-by: Xiao Li <gatorsmile@gmail.com>
Signed-off-by: Takeshi Yamamuro <yamamuro@apache.org>
## What changes were proposed in this pull request?
This PR targets to add vectorized `gapply()` in R, Arrow optimization.
This can be tested as below:
```bash
$ ./bin/sparkR --conf spark.sql.execution.arrow.enabled=true
```
```r
df <- createDataFrame(mtcars)
collect(gapply(df,
"gear",
function(key, group) {
data.frame(gear = key[[1]], disp = mean(group$disp) > group$disp)
},
structType("gear double, disp boolean")))
```
### Requirements
- R 3.5.x
- Arrow package 0.12+
```bash
Rscript -e 'remotes::install_github("apache/arrowapache-arrow-0.12.0", subdir = "r")'
```
**Note:** currently, Arrow R package is not in CRAN. Please take a look at ARROW-3204.
**Note:** currently, Arrow R package seems not supporting Windows. Please take a look at ARROW-3204.
### Benchmarks
**Shall**
```bash
sync && sudo purge
./bin/sparkR --conf spark.sql.execution.arrow.enabled=false
```
```bash
sync && sudo purge
./bin/sparkR --conf spark.sql.execution.arrow.enabled=true
```
**R code**
```r
rdf <- read.csv("500000.csv")
rdf <- rdf[, c("Month.of.Joining", "Weight.in.Kgs.")] # We're only interested in the key and values to calculate.
df <- cache(createDataFrame(rdf))
count(df)
test <- function() {
options(digits.secs = 6) # milliseconds
start.time <- Sys.time()
count(gapply(df,
"Month_of_Joining",
function(key, group) {
data.frame(Month_of_Joining = key[[1]], Weight_in_Kgs_ = mean(group$Weight_in_Kgs_) > group$Weight_in_Kgs_)
},
structType("Month_of_Joining integer, Weight_in_Kgs_ boolean")))
end.time <- Sys.time()
time.taken <- end.time - start.time
print(time.taken)
}
test()
```
**Data (350 MB):**
```r
object.size(read.csv("500000.csv"))
350379504 bytes
```
"500000 Records" http://eforexcel.com/wp/downloads-16-sample-csv-files-data-sets-for-testing/
**Results**
```
Time difference of 35.67459 secs
```
```
Time difference of 4.301399 secs
```
The performance improvement was around **829%**.
**Note that** I am 100% sure this PR improves more then 829% because I gave up testing it with non-Arrow optimization because it took super super super long when the data size becomes bigger.
### Limitations
- For now, Arrow optimization with R does not support when the data is `raw`, and when user explicitly gives float type in the schema. They produce corrupt values.
- Due to ARROW-4512, it cannot send and receive batch by batch. It has to send all batches in Arrow stream format at once. It needs improvement later.
## How was this patch tested?
Unit tests were added
**TODOs:**
- [x] Draft codes
- [x] make the tests passed
- [x] make the CRAN check pass
- [x] Performance measurement
- [x] Supportability investigation (for instance types)
Closes#23746 from HyukjinKwon/SPARK-26759.
Authored-by: Hyukjin Kwon <gurwls223@apache.org>
Signed-off-by: Hyukjin Kwon <gurwls223@apache.org>
## What changes were proposed in this pull request?
In https://github.com/apache/spark/issues/23569, the copy method of `ColumnarArray` is implemented.
To further improve it, we can return `UnsafeArrayData` for `date`/`timestamp` type in `ColumnarArray.copy()`.
## How was this patch tested?
Unit test
Closes#23761 from gengliangwang/copyDateAndTS.
Authored-by: Gengliang Wang <gengliang.wang@databricks.com>
Signed-off-by: Takeshi Yamamuro <yamamuro@apache.org>
## What changes were proposed in this pull request?
Implements the solution proposed in [SPARK-26696](https://issues.apache.org/jira/browse/SPARK-26696), a minor refactoring that allows frameworks to perform advanced type-preserving dataset transformations without carrying `Encoder` implicits from user code.
The change allows
```scala
def foo[A](ds: Dataset[A]): Dataset[A] =
ds.toDF().as[A](ds.encoder)
```
instead of
```scala
def foo[A: Encoder](ds: Dataset[A]): Dataset[A] =
ds.toDF().as[A](implicitly[Encoder[A]])
```
## How was this patch tested?
This patch was tested with an automated test that was later removed as it was deemed unnecessary per the discussion in this PR.
Closes#23620 from ssimeonov/ss_SPARK-26696.
Authored-by: Simeon Simeonov <sim@fastignite.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
## What changes were proposed in this pull request?
In the PR, I propose to covert underlying types of timestamp/date columns to strings, and store the converted values as column statistics. This makes statistics for timestamp/date columns independent from system time zone while saving and retrieving such statistics.
I bumped versions of stored statistics from 1 to 2 since the PR changes the format.
## How was this patch tested?
The changes were tested by `StatisticsCollectionSuite` and by `StatisticsSuite`.
Closes#23662 from MaxGekk/column-stats-time-date.
Lead-authored-by: Maxim Gekk <max.gekk@gmail.com>
Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com>
Co-authored-by: Hyukjin Kwon <gurwls223@apache.org>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
## What changes were proposed in this pull request?
Not all users wants to keep temporary checkpoint directories. Additionally hard to restore from it.
In this PR I've added a force delete flag which is default `false`. Additionally not clear for users when temporary checkpoint directory deleted so added log messages to explain this a bit more.
## How was this patch tested?
Existing + additional unit tests.
Closes#23732 from gaborgsomogyi/SPARK-26389.
Authored-by: Gabor Somogyi <gabor.g.somogyi@gmail.com>
Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
## What changes were proposed in this pull request?
This PR consists of the `test` components of #23665 only, minus the associated patch from that PR.
It adds a new unit test to `JsonSuite` which verifies that the `count()` returned from a `DataFrame` loaded from JSON containing empty lines does not include those empty lines in the record count. The test runs `count` prior to otherwise reading data from the `DataFrame`, so as to catch future cases where a pre-parsing optimization might result in `count` results inconsistent with existing behavior.
This PR is intended to be deployed alongside #23667; `master` currently causes the test to fail, as described in [SPARK-26745](https://issues.apache.org/jira/browse/SPARK-26745).
## How was this patch tested?
Manual testing, existing `JsonSuite` unit tests.
Closes#23674 from sumitsu/json_emptyline_count_test.
Authored-by: Branden Smith <branden.smith@publicismedia.com>
Signed-off-by: Hyukjin Kwon <gurwls223@apache.org>
## What changes were proposed in this pull request?
Update to Parquet Java 1.10.1.
## How was this patch tested?
Added a test from HyukjinKwon that validates the notEq case from SPARK-26677.
Closes#23704 from rdblue/SPARK-26677-fix-noteq-parquet-bug.
Lead-authored-by: Ryan Blue <blue@apache.org>
Co-authored-by: Hyukjin Kwon <gurwls223@apache.org>
Co-authored-by: Ryan Blue <rdblue@users.noreply.github.com>
Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
## What changes were proposed in this pull request?
Make .unpersist(), .destroy() non-blocking by default and adjust callers to request blocking only where important.
This also adds an optional blocking argument to Pyspark's RDD.unpersist(), which never had one.
## How was this patch tested?
Existing tests.
Closes#23685 from srowen/SPARK-26771.
Authored-by: Sean Owen <sean.owen@databricks.com>
Signed-off-by: Sean Owen <sean.owen@databricks.com>
## What changes were proposed in this pull request?
Right now, EventTimeStats.merge doesn't handle `zero.merge(zero)` correctly. This will make `avg` become `NaN`. And whatever gets merged with the result of `zero.merge(zero)`, `avg` will still be `NaN`. Then finally, we call `NaN.toLong` and get `0`, and the user will see the following incorrect report:
```
"eventTime" : {
"avg" : "1970-01-01T00:00:00.000Z",
"max" : "2019-01-31T12:57:00.000Z",
"min" : "2019-01-30T18:44:04.000Z",
"watermark" : "1970-01-01T00:00:00.000Z"
}
```
This issue was reported by liancheng .
This PR fixes the above issue.
## How was this patch tested?
The new unit tests.
Closes#23718 from zsxwing/merge-zero.
Authored-by: Shixiong Zhu <zsxwing@gmail.com>
Signed-off-by: Shixiong Zhu <zsxwing@gmail.com>
## What changes were proposed in this pull request?
Create a framework for write path of File Source V2.
Also, migrate write path of ORC to V2.
Supported:
* Write to file as Dataframe
Not Supported:
* Partitioning, which is still under development in the data source V2 project.
* Bucketing, which is still under development in the data source V2 project.
* Catalog.
## How was this patch tested?
Unit test
Closes#23601 from gengliangwang/orc_write.
Authored-by: Gengliang Wang <gengliang.wang@databricks.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
## What changes were proposed in this pull request?
A followup of https://github.com/apache/spark/pull/23388 .
`AggUtils.createAggregate` is not the right place to normalize the grouping expressions, as final aggregate is also created by it. The grouping expressions of final aggregate should be attributes which refer to the grouping expressions in partial aggregate.
This PR moves the normalization to the caller side of `AggUtils`.
## How was this patch tested?
existing tests
Closes#23692 from cloud-fan/follow.
Authored-by: Wenchen Fan <wenchen@databricks.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
## What changes were proposed in this pull request?
In #23639, the API `supportDataType` is refactored. We should also remove the method `verifyWriteSchema` and `verifyReadSchema` in `DataSourceUtils`.
Since the error message use `FileFormat.toString` to specify the data source naming, this PR also overriding the `toString` method in `AvroFileFormat`.
## How was this patch tested?
Unit test.
Closes#23699 from gengliangwang/SPARK-26716-followup.
Authored-by: Gengliang Wang <gengliang.wang@databricks.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
## What changes were proposed in this pull request?
This PR reverts JSON count optimization part of #21909.
We cannot distinguish the cases below without parsing:
```
[{...}, {...}]
```
```
[]
```
```
{...}
```
```bash
# empty string
```
when we `count()`. One line (input: IN) can be, 0 record, 1 record and multiple records and this is dependent on each input.
See also https://github.com/apache/spark/pull/23665#discussion_r251276720.
## How was this patch tested?
Manually tested.
Closes#23667 from HyukjinKwon/revert-SPARK-24959.
Authored-by: Hyukjin Kwon <gurwls223@apache.org>
Signed-off-by: Hyukjin Kwon <gurwls223@apache.org>
## What changes were proposed in this pull request?
This is a follow up of https://github.com/apache/spark/pull/23644/files , to make these methods less coupled with each other.
## How was this patch tested?
existing tests
Closes#23687 from cloud-fan/cache.
Authored-by: Wenchen Fan <wenchen@databricks.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
## What changes were proposed in this pull request?
After [recent changes](11e5f1bcd4) to CSV parsing to return partial results for bad CSV records, queries of wide CSV tables slowed considerably. That recent change resulted in every row being recreated, even when the associated input record had no parsing issues and the user specified no corrupt record field in his/her schema.
The change to FailureSafeParser.scala also impacted queries against wide JSON tables as well.
In this PR, I propose that a row should be recreated only if columns need to be shifted due to the existence of a corrupt column field in the user-supplied schema. Otherwise, the code should use the row as-is (For CSV input, it will have values for the columns that could be converted, and also null values for columns that could not be converted).
See benchmarks below. The CSV benchmark for 1000 columns went from 120144 ms to 89069 ms, a savings of 25% (this only brings the cost down to baseline levels. Again, see benchmarks below).
Similarly, the JSON benchmark for 1000 columns (added in this PR) went from 109621 ms to 80871 ms, also a savings of 25%.
Still, partial results functionality is preserved:
<pre>
bash-3.2$ cat test2.csv
"hello",1999-08-01,"last"
"there","bad date","field"
"again","2017-11-22","in file"
bash-3.2$ bin/spark-shell
...etc...
scala> val df = spark.read.schema("a string, b date, c string").csv("test2.csv")
df: org.apache.spark.sql.DataFrame = [a: string, b: date ... 1 more field]
scala> df.show
+-----+----------+-------+
| a| b| c|
+-----+----------+-------+
|hello|1999-08-01| last|
|there| null| field|
|again|2017-11-22|in file|
+-----+----------+-------+
scala> val df = spark.read.schema("badRecord string, a string, b date, c string").
| option("columnNameOfCorruptRecord", "badRecord").
| csv("test2.csv")
df: org.apache.spark.sql.DataFrame = [badRecord: string, a: string ... 2 more fields]
scala> df.show
+--------------------+-----+----------+-------+
| badRecord| a| b| c|
+--------------------+-----+----------+-------+
| null|hello|1999-08-01| last|
|"there","bad date...|there| null| field|
| null|again|2017-11-22|in file|
+--------------------+-----+----------+-------+
scala>
</pre>
### CSVBenchmark Benchmarks:
baseline = commit before partial results change
PR = this PR
master = master branch
[baseline_CSVBenchmark-results.txt](https://github.com/apache/spark/files/2697109/baseline_CSVBenchmark-results.txt)
[pr_CSVBenchmark-results.txt](https://github.com/apache/spark/files/2697110/pr_CSVBenchmark-results.txt)
[master_CSVBenchmark-results.txt](https://github.com/apache/spark/files/2697111/master_CSVBenchmark-results.txt)
### JSONBenchmark Benchmarks:
baseline = commit before partial results change
PR = this PR
master = master branch
[baseline_JSONBenchmark-results.txt](https://github.com/apache/spark/files/2711040/baseline_JSONBenchmark-results.txt)
[pr_JSONBenchmark-results.txt](https://github.com/apache/spark/files/2711041/pr_JSONBenchmark-results.txt)
[master_JSONBenchmark-results.txt](https://github.com/apache/spark/files/2711042/master_JSONBenchmark-results.txt)
## How was this patch tested?
- All SQL unit tests.
- Added 2 CSV benchmarks
- Python core and SQL tests
Closes#23336 from bersprockets/csv-wide-row-opt2.
Authored-by: Bruce Robbins <bersprockets@gmail.com>
Signed-off-by: Hyukjin Kwon <gurwls223@apache.org>
## What changes were proposed in this pull request?
For making test suite supporting both Parquet and Orc by reusing test cases, this patch extracts the methods for testing. For example, if we need to test a common feature shared by Parquet and Orc, we should be able to write test cases once and reuse them to test both formats.
This patch extracts the methods for testing and uses a variable `dataSourceName` to set up data format to test against with.
## How was this patch tested?
Existing tests.
Closes#23628 from viirya/datasource-test.
Authored-by: Liang-Chi Hsieh <viirya@gmail.com>
Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
## What changes were proposed in this pull request?
We should pass the existed `fileStatusCache` to `InMemoryFileIndex` even though there aren't partition columns.
## How was this patch tested?
Existed test. Extra tests can be added if there is a requirement.
Closes#23683 from ConeyLiu/filestatuscache.
Authored-by: Xianyang Liu <xianyang.liu@intel.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
## What changes were proposed in this pull request?
Following https://github.com/apache/spark/pull/23430, this PR does the API refactor for continuous read, w.r.t. the [doc](https://docs.google.com/document/d/1uUmKCpWLdh9vHxP7AWJ9EgbwB_U6T3EJYNjhISGmiQg/edit?usp=sharing)
The major changes:
1. rename `XXXContinuousReadSupport` to `XXXContinuousStream`
2. at the beginning of continuous streaming execution, convert `StreamingRelationV2` to `StreamingDataSourceV2Relation` directly, instead of `StreamingExecutionRelation`.
3. remove all the hacks as we have finished all the read side API refactor
## How was this patch tested?
existing tests
Closes#23619 from cloud-fan/continuous.
Authored-by: Wenchen Fan <wenchen@databricks.com>
Signed-off-by: gatorsmile <gatorsmile@gmail.com>
## What changes were proposed in this pull request?
Added the following benchmarks:
- Extract components from timestamp like year, month, day and etc.
- Current date and time
- Date arithmetic like date_add, date_sub
- Format dates and timestamps
- Convert timestamps from/to UTC
Closes#23661 from MaxGekk/datetime-benchmark.
Authored-by: Maxim Gekk <max.gekk@gmail.com>
Signed-off-by: Herman van Hovell <hvanhovell@databricks.com>
## What changes were proposed in this pull request?
Adjust mem settings in UnifiedMemoryManager used in test suites to ha…ve execution memory > 0
Ref: https://github.com/apache/spark/pull/23457#issuecomment-457409976
## How was this patch tested?
Existing tests
Closes#23645 from srowen/SPARK-26725.
Authored-by: Sean Owen <sean.owen@databricks.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
## What changes were proposed in this pull request?
When performing non-cascading cache invalidation, `recache` is called on the other cache entries which are dependent on the cache being invalidated. It leads to the the physical plans of those cache entries being re-compiled. For those cache entries, if the cache RDD has already been persisted, chances are there will be inconsistency between the data and the new plan. It can cause a correctness issue if the new plan's `outputPartitioning` or `outputOrdering` is different from the that of the actual data, and meanwhile the cache is used by another query that asks for specific `outputPartitioning` or `outputOrdering` which happens to match the new plan but not the actual data.
The fix is to keep the cache entry as it is if the data has been loaded, otherwise re-build the cache entry, with a new plan and an empty cache buffer.
## How was this patch tested?
Added UT.
Closes#23644 from maryannxue/spark-26708.
Lead-authored-by: maryannxue <maryannxue@apache.org>
Co-authored-by: Xiao Li <gatorsmile@gmail.com>
Signed-off-by: gatorsmile <gatorsmile@gmail.com>
## What changes were proposed in this pull request?
1. Remove parameter `isReadPath`. The supported types of read/write should be the same.
2. Disallow reading `NullType` for ORC data source. In #21667 and #21389, it was supposed that ORC supports reading `NullType`, but can't write it. This doesn't make sense. I read docs and did some tests. ORC doesn't support `NullType`.
## How was this patch tested?
Unit tset
Closes#23639 from gengliangwang/supportDataType.
Authored-by: Gengliang Wang <gengliang.wang@databricks.com>
Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
## What changes were proposed in this pull request?
Spark replaces `CurrentTimestamp` with `CurrentBatchTimestamp`.
However, `CurrentBatchTimestamp` is `TimeZoneAwareExpression` while `CurrentTimestamp` isn't.
Without TimeZoneId, `CurrentBatchTimestamp` becomes unresolved and raises `UnresolvedException`.
Since `CurrentDate` is `TimeZoneAwareExpression`, there is no problem with `CurrentDate`.
This PR reverts the [previous patch](https://github.com/apache/spark/pull/23609) on `MicroBatchExecution` and fixes the root cause.
## How was this patch tested?
Pass the Jenkins with the updated test cases.
Closes#23660 from dongjoon-hyun/SPARK-26379.
Authored-by: Dongjoon Hyun <dongjoon@apache.org>
Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
## What changes were proposed in this pull request?
Add verification of plan integrity with regards to special expressions being hosted only in supported operators. Specifically:
- `AggregateExpression`: should only be hosted in `Aggregate`, or indirectly in `Window`
- `WindowExpression`: should only be hosted in `Window`
- `Generator`: should only be hosted in `Generate`
This will help us catch errors in future optimizer rules that incorrectly hoist special expression out of their supported operator.
TODO: This PR actually caught a bug in the analyzer in the test case `SPARK-23957 Remove redundant sort from subquery plan(scalar subquery)` in `SubquerySuite`, where a `max()` aggregate function is hosted in a `Sort` operator in the analyzed plan, which is invalid. That test case is disabled in this PR.
SPARK-26741 has been opened to track the fix in the analyzer.
## How was this patch tested?
Added new test case in `OptimizerStructuralIntegrityCheckerSuite`
Closes#23658 from rednaxelafx/plan-integrity.
Authored-by: Kris Mok <kris.mok@databricks.com>
Signed-off-by: gatorsmile <gatorsmile@gmail.com>
## What changes were proposed in this pull request?
This PR targets to support Arrow optimization for conversion from R DataFrame to Spark DataFrame.
Like PySpark side, it falls back to non-optimization code path when it's unable to use Arrow optimization.
This can be tested as below:
```bash
$ ./bin/sparkR --conf spark.sql.execution.arrow.enabled=true
```
```r
collect(createDataFrame(mtcars))
```
### Requirements
- R 3.5.x
- Arrow package 0.12+
```bash
Rscript -e 'remotes::install_github("apache/arrowapache-arrow-0.12.0", subdir = "r")'
```
**Note:** currently, Arrow R package is not in CRAN. Please take a look at ARROW-3204.
**Note:** currently, Arrow R package seems not supporting Windows. Please take a look at ARROW-3204.
### Benchmarks
**Shall**
```bash
sync && sudo purge
./bin/sparkR --conf spark.sql.execution.arrow.enabled=false
```
```bash
sync && sudo purge
./bin/sparkR --conf spark.sql.execution.arrow.enabled=true
```
**R code**
```r
createDataFrame(mtcars) # Initializes
rdf <- read.csv("500000.csv")
test <- function() {
options(digits.secs = 6) # milliseconds
start.time <- Sys.time()
createDataFrame(rdf)
end.time <- Sys.time()
time.taken <- end.time - start.time
print(time.taken)
}
test()
```
**Data (350 MB):**
```r
object.size(read.csv("500000.csv"))
350379504 bytes
```
"500000 Records" http://eforexcel.com/wp/downloads-16-sample-csv-files-data-sets-for-testing/
**Results**
```
Time difference of 29.9468 secs
```
```
Time difference of 3.222129 secs
```
The performance improvement was around **950%**.
Actually, this PR improves around **1200%**+ because this PR includes a small optimization about regular R DataFrame -> Spark DatFrame. See https://github.com/apache/spark/pull/22954#discussion_r231847272
### Limitations:
For now, Arrow optimization with R does not support when the data is `raw`, and when user explicitly gives float type in the schema. They produce corrupt values.
In this case, we decide to fall back to non-optimization code path.
## How was this patch tested?
Small test was added.
I manually forced to set this optimization `true` for _all_ R tests and they were _all_ passed (with few of fallback warnings).
**TODOs:**
- [x] Draft codes
- [x] make the tests passed
- [x] make the CRAN check pass
- [x] Performance measurement
- [x] Supportability investigation (for instance types)
- [x] Wait for Arrow 0.12.0 release
- [x] Fix and match it to Arrow 0.12.0
Closes#22954 from HyukjinKwon/r-arrow-createdataframe.
Lead-authored-by: hyukjinkwon <gurwls223@apache.org>
Co-authored-by: Hyukjin Kwon <gurwls223@apache.org>
Signed-off-by: Hyukjin Kwon <gurwls223@apache.org>
## What changes were proposed in this pull request?
This PR makes hardcoded configs about spark memory and storage to use `ConfigEntry` and put them in the config package.
## How was this patch tested?
Existing unit tests.
Closes#23623 from SongYadong/configEntry_for_mem_storage.
Authored-by: SongYadong <song.yadong1@zte.com.cn>
Signed-off-by: Sean Owen <sean.owen@databricks.com>
## What changes were proposed in this pull request?
This patch proposes to fix issue on adding `current_timestamp` / `current_date` with streaming query.
The root reason is that Spark transforms `CurrentTimestamp`/`CurrentDate` to `CurrentBatchTimestamp` in MicroBatchExecution which makes transformed attributes not-yet-resolved. They will be resolved by IncrementalExecution.
(In ContinuousExecution, Spark doesn't allow using `current_timestamp` and `current_date` so it has been OK.)
It's OK for DataSource V1 sink because it simply leverages transformed logical plan and don't evaluate until they're resolved, but for DataSource V2 sink, Spark tries to extract the schema of transformed logical plan in prior to IncrementalExecution, and unresolved attributes will raise errors.
This patch fixes the issue via having separate pre-resolved logical plan to pass the schema to StreamingWriteSupport safely.
## How was this patch tested?
Added UT.
Closes#23609 from HeartSaVioR/SPARK-26379.
Authored-by: Jungtaek Lim (HeartSaVioR) <kabhwan@gmail.com>
Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
## What changes were proposed in this pull request?
Noop data source for batch was added in [#23471](https://github.com/apache/spark/pull/23471).
In this PR I've added the streaming part.
## How was this patch tested?
Additional unit tests.
Closes#23631 from gaborgsomogyi/SPARK-26649.
Authored-by: Gabor Somogyi <gabor.g.somogyi@gmail.com>
Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
## What changes were proposed in this pull request?
When reading from empty tables, the optimization `OptimizeMetadataOnlyQuery` may return wrong results:
```
sql("CREATE TABLE t (col1 INT, p1 INT) USING PARQUET PARTITIONED BY (p1)")
sql("INSERT INTO TABLE t PARTITION (p1 = 5) SELECT ID FROM range(1, 1)")
sql("SELECT MAX(p1) FROM t")
```
The result is supposed to be `null`. However, with the optimization the result is `5`.
The rule is originally ported from https://issues.apache.org/jira/browse/HIVE-1003 in #13494. In Hive, the rule is disabled by default in a later release(https://issues.apache.org/jira/browse/HIVE-15397), due to the same problem.
It is hard to completely avoid the correctness issue. Because data sources like Parquet can be metadata-only. Spark can't tell whether it is empty or not without actually reading it. This PR disable the optimization by default.
## How was this patch tested?
Unit test
Closes#23635 from gengliangwang/optimizeMetadata.
Lead-authored-by: Gengliang Wang <gengliang.wang@databricks.com>
Co-authored-by: Xiao Li <gatorsmile@gmail.com>
Signed-off-by: gatorsmile <gatorsmile@gmail.com>
