### What changes were proposed in this pull request?
This PR aims to recover `IntervalBenchmark` and `DataTimeBenchmark` due to banning intervals as output.
### Why are the changes needed?
This PR recovers the benchmark suite.
### Does this PR introduce any user-facing change?
No.
### How was this patch tested?
Manually, re-run the benchmark.
Closes#27885 from yaooqinn/SPARK-31111-2.
Authored-by: Kent Yao <yaooqinn@hotmail.com>
Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
### What changes were proposed in this pull request?
fix the error caused by interval output in ExtractBenchmark
### Why are the changes needed?
fix a bug in the test
```scala
[info] Running case: cast to interval
[error] Exception in thread "main" org.apache.spark.sql.AnalysisException: Cannot use interval type in the table schema.;;
[error] OverwriteByExpression RelationV2[] noop-table, true, true
[error] +- Project [(subtractdates(cast(cast(id#0L as timestamp) as date), -719162) + subtracttimestamps(cast(id#0L as timestamp), -30610249419876544)) AS ((CAST(CAST(id AS TIMESTAMP) AS DATE) - DATE '0001-01-01') + (CAST(id AS TIMESTAMP) - TIMESTAMP '1000-01-01 01:02:03.123456'))#2]
[error] +- Range (1262304000, 1272304000, step=1, splits=Some(1))
[error]
[error] at org.apache.spark.sql.catalyst.util.TypeUtils$.failWithIntervalType(TypeUtils.scala:106)
[error] at org.apache.spark.sql.catalyst.analysis.CheckAnalysis.$anonfun$checkAnalysis$25(CheckAnalysis.scala:389)
[error] at org.a
```
### Does this PR introduce any user-facing change?
no
### How was this patch tested?
re-run benchmark
Closes#27867 from yaooqinn/SPARK-31111.
Authored-by: Kent Yao <yaooqinn@hotmail.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
### What changes were proposed in this pull request?
1. Rewrite DateTimeUtils methods `getHours()`, `getMinutes()`, `getSeconds()`, `getSecondsWithFraction()`, `getMilliseconds()` and `getMicroseconds()` using Java 8 time APIs. This will automatically switch the `Hour`, `Minute`, `Second` and `DatePart` expressions on Proleptic Gregorian calendar.
2. Remove unused methods and constant of DateTimeUtils - `to2001`, `YearZero `, `toYearZero` and `absoluteMicroSecond()`.
3. Remove unused value `timeZone` from `TimeZoneAwareExpression` since all expressions have been migrated to Java 8 time API, and legacy instance of `TimeZone` is not needed any more.
4. Change signatures of modified DateTimeUtils methods, and pass `ZoneId` instead of `TimeZone`. This will allow to avoid unnecessary conversions `TimeZone` -> `String` -> `ZoneId`.
5. Modify tests in `DateTimeUtilsSuite` and in `DateExpressionsSuite` to pass `ZoneId` instead of `TimeZone`. Correct the tests, to pass tested zone id instead of None.
### Why are the changes needed?
The changes fix the issue of wrong results returned by the `hour()`, `minute()`, `second()`, `date_part('millisecond', ...)` and `date_part('microsecond', ....)`, see example in [SPARK-30843](https://issues.apache.org/jira/browse/SPARK-30843).
### Does this PR introduce any user-facing change?
Yes. After the changes, the results of examples from SPARK-30843:
```sql
spark-sql> select hour(timestamp '0010-01-01 00:00:00');
0
spark-sql> select minute(timestamp '0010-01-01 00:00:00');
0
spark-sql> select second(timestamp '0010-01-01 00:00:00');
0
spark-sql> select date_part('milliseconds', timestamp '0010-01-01 00:00:00');
0.000
spark-sql> select date_part('microseconds', timestamp '0010-01-01 00:00:00');
0
```
### How was this patch tested?
- By existing test suites `DateTimeUtilsSuite`, `DateExpressionsSuite` and `DateFunctionsSuite`.
- Add new tests to `DateExpressionsSuite` and `DateTimeUtilsSuite` for 10 year, like:
```scala
input = date(10, 1, 1, 0, 0, 0, 0, zonePST)
assert(getHours(input, zonePST) === 0)
```
- Re-run `DateTimeBenchmark` using Amazon EC2.
| Item | Description |
| ---- | ----|
| Region | us-west-2 (Oregon) |
| Instance | r3.xlarge |
| AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) |
| Java | OpenJDK8/11 |
Closes#27596 from MaxGekk/localtimestamp-greg-cal.
Lead-authored-by: Maxim Gekk <max.gekk@gmail.com>
Co-authored-by: Max Gekk <max.gekk@gmail.com>
Co-authored-by: Ubuntu <ubuntu@ip-172-31-1-30.us-west-2.compute.internal>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
### What changes were proposed in this pull request?
In the PR, I propose to support pushed down filters in CSV datasource. The reason of pushing a filter up to `UnivocityParser` is to apply the filter as soon as all its attributes become available i.e. converted from CSV fields to desired values according to the schema. This allows to skip conversions of other values if the filter returns `false`. This can improve performance when pushed filters are highly selective and conversion of CSV string fields to desired values are comparably expensive ( for example, conversion to `TIMESTAMP` values).
Here are details of the implementation:
- `UnivocityParser.convert()` converts parsed CSV tokens one-by-one sequentially starting from index 0 up to `parsedSchema.length - 1`. At current index `i`, it applies filters that refer to attributes at row fields indexes `0..i`. If any filter returns `false`, it skips conversions of other input tokens.
- Pushed filters are converted to expressions. The expressions are bound to row positions according to `requiredSchema`. The expressions are compiled to predicates via generating Java code.
- To be able to apply predicates to partially initialized rows, the predicates are grouped, and combined via the `And` expression. Final predicate at index `N` can refer to row fields at the positions `0..N`, and can be applied to a row even if other fields at the positions `N+1..requiredSchema.lenght-1` are not set.
### Why are the changes needed?
The changes improve performance on synthetic benchmarks more **than 9 times** (on JDK 8 & 11):
```
OpenJDK 64-Bit Server VM 11.0.5+10 on Mac OS X 10.15.2
Intel(R) Core(TM) i7-4850HQ CPU 2.30GHz
Filters pushdown: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
w/o filters 11889 11945 52 0.0 118893.1 1.0X
pushdown disabled 11790 11860 115 0.0 117902.3 1.0X
w/ filters 1240 1278 33 0.1 12400.8 9.6X
```
### Does this PR introduce any user-facing change?
No
### How was this patch tested?
- Added new test suite `CSVFiltersSuite`
- Added tests to `CSVSuite` and `UnivocityParserSuite`
Closes#26973 from MaxGekk/csv-filters-pushdown.
Authored-by: Maxim Gekk <max.gekk@gmail.com>
Signed-off-by: HyukjinKwon <gurwls223@apache.org>
### What changes were proposed in this pull request?
This PR removes a dangling test result, `JSONBenchmark-jdk11-results.txt`.
This causes a case-sensitive issue on Mac.
```
$ git clone https://gitbox.apache.org/repos/asf/spark.git spark-gitbox
Cloning into 'spark-gitbox'...
remote: Counting objects: 671717, done.
remote: Compressing objects: 100% (258021/258021), done.
remote: Total 671717 (delta 329181), reused 560390 (delta 228097)
Receiving objects: 100% (671717/671717), 149.69 MiB | 950.00 KiB/s, done.
Resolving deltas: 100% (329181/329181), done.
Updating files: 100% (16090/16090), done.
warning: the following paths have collided (e.g. case-sensitive paths
on a case-insensitive filesystem) and only one from the same
colliding group is in the working tree:
'sql/core/benchmarks/JSONBenchmark-jdk11-results.txt'
'sql/core/benchmarks/JsonBenchmark-jdk11-results.txt'
```
### Why are the changes needed?
Previously, since the file name didn't match with `object JSONBenchmark`, it made a confusion when we ran the benchmark. So, 4e0e4e51c4 renamed `JSONBenchmark` to `JsonBenchmark`. However, at the same time frame, https://github.com/apache/spark/pull/26003 regenerated this file.
Recently, https://github.com/apache/spark/pull/27078 regenerates the results with the correct file name, `JsonBenchmark-jdk11-results.txt`. So, we can remove the old one.
### Does this PR introduce any user-facing change?
No. This is a test result.
### How was this patch tested?
Manually check the following correctly generated files in the master. And, check this PR removes the dangling one.
- https://github.com/apache/spark/blob/master/sql/core/benchmarks/JsonBenchmark-results.txt
- https://github.com/apache/spark/blob/master/sql/core/benchmarks/JsonBenchmark-jdk11-results.txtCloses#27180 from dongjoon-hyun/SPARK-REMOVE.
Authored-by: Dongjoon Hyun <dhyun@apple.com>
Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
### What changes were proposed in this pull request?
In the PR, I propose to replace `.collect()`, `.count()` and `.foreach(_ => ())` in SQL benchmarks and use the `NoOp` datasource. I added an implicit class to `SqlBasedBenchmark` with the `.noop()` method. It can be used in benchmark like: `ds.noop()`. The last one is unfolded to `ds.write.format("noop").mode(Overwrite).save()`.
### Why are the changes needed?
To avoid additional overhead that `collect()` (and other actions) has. For example, `.collect()` has to convert values according to external types and pull data to the driver. This can hide actual performance regressions or improvements of benchmarked operations.
### Does this PR introduce any user-facing change?
No
### How was this patch tested?
Re-run all modified benchmarks using Amazon EC2.
| Item | Description |
| ---- | ----|
| Region | us-west-2 (Oregon) |
| Instance | r3.xlarge (spot instance) |
| AMI | ami-06f2f779464715dc5 (ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1) |
| Java | OpenJDK8/10 |
- Run `TPCDSQueryBenchmark` using instructions from the PR #26049
```
# `spark-tpcds-datagen` needs this. (JDK8)
$ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4
$ export SPARK_HOME=$PWD
$ ./build/mvn clean package -DskipTests
# Generate data. (JDK8)
$ git clone gitgithub.com:maropu/spark-tpcds-datagen.git
$ cd spark-tpcds-datagen/
$ build/mvn clean package
$ mkdir -p /data/tpcds
$ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4`
```
- Other benchmarks ran by the script:
```
#!/usr/bin/env python3
import os
from sparktestsupport.shellutils import run_cmd
benchmarks = [
['sql/test', 'org.apache.spark.sql.execution.benchmark.AggregateBenchmark'],
['avro/test', 'org.apache.spark.sql.execution.benchmark.AvroReadBenchmark'],
['sql/test', 'org.apache.spark.sql.execution.benchmark.BloomFilterBenchmark'],
['sql/test', 'org.apache.spark.sql.execution.benchmark.DataSourceReadBenchmark'],
['sql/test', 'org.apache.spark.sql.execution.benchmark.DateTimeBenchmark'],
['sql/test', 'org.apache.spark.sql.execution.benchmark.ExtractBenchmark'],
['sql/test', 'org.apache.spark.sql.execution.benchmark.FilterPushdownBenchmark'],
['sql/test', 'org.apache.spark.sql.execution.benchmark.InExpressionBenchmark'],
['sql/test', 'org.apache.spark.sql.execution.benchmark.IntervalBenchmark'],
['sql/test', 'org.apache.spark.sql.execution.benchmark.JoinBenchmark'],
['sql/test', 'org.apache.spark.sql.execution.benchmark.MakeDateTimeBenchmark'],
['sql/test', 'org.apache.spark.sql.execution.benchmark.MiscBenchmark'],
['hive/test', 'org.apache.spark.sql.execution.benchmark.ObjectHashAggregateExecBenchmark'],
['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcNestedSchemaPruningBenchmark'],
['sql/test', 'org.apache.spark.sql.execution.benchmark.OrcV2NestedSchemaPruningBenchmark'],
['sql/test', 'org.apache.spark.sql.execution.benchmark.ParquetNestedSchemaPruningBenchmark'],
['sql/test', 'org.apache.spark.sql.execution.benchmark.RangeBenchmark'],
['sql/test', 'org.apache.spark.sql.execution.benchmark.UDFBenchmark'],
['sql/test', 'org.apache.spark.sql.execution.benchmark.WideSchemaBenchmark'],
['sql/test', 'org.apache.spark.sql.execution.benchmark.WideTableBenchmark'],
['hive/test', 'org.apache.spark.sql.hive.orc.OrcReadBenchmark'],
['sql/test', 'org.apache.spark.sql.execution.datasources.csv.CSVBenchmark'],
['sql/test', 'org.apache.spark.sql.execution.datasources.json.JsonBenchmark']
]
print('Set SPARK_GENERATE_BENCHMARK_FILES=1')
os.environ['SPARK_GENERATE_BENCHMARK_FILES'] = '1'
for b in benchmarks:
print("Run benchmark: %s" % b[1])
run_cmd(['build/sbt', '%s:runMain %s' % (b[0], b[1])])
```
Closes#27078 from MaxGekk/noop-in-benchmarks.
Lead-authored-by: Maxim Gekk <max.gekk@gmail.com>
Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com>
Co-authored-by: Dongjoon Hyun <dhyun@apple.com>
Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
### What changes were proposed in this pull request?
We are now able to handle whitespaces for integral and fractional types, and the leading or trailing whitespaces for interval, date, and timestamps. But the current interval parser is not able to identify whitespaces as separates as PostgreSQL can do.
This PR makes the whitespaces handling be consistent for nterval values.
Typed interval literal, multi-unit representation, and casting from strings are all supported.
```sql
postgres=# select interval E'1 \t day';
interval
----------
1 day
(1 row)
postgres=# select interval E'1\t' day;
interval
----------
1 day
(1 row)
```
### Why are the changes needed?
Whitespace handling should be consistent for interval value, and across different types in Spark.
PostgreSQL feature parity.
### Does this PR introduce any user-facing change?
Yes, the interval string of multi-units values which separated by whitespaces can be valid now.
### How was this patch tested?
add ut.
Closes#26662 from yaooqinn/SPARK-30026.
Authored-by: Kent Yao <yaooqinn@hotmail.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
### What changes were proposed in this pull request?
With the latest string to literal optimization https://github.com/apache/spark/pull/26256, some interval strings can not be cast when there are some spaces between signs and unit values. After state `PARSE_SIGN`, it directly goes to `PARSE_UNIT_VALUE` when takes a space character as the end. So when there are some white spaces come before the real unit value, it fails to parse, we should add a new state like `TRIM_VALUE` to trim all these spaces.
How to re-produce, which aim the revisions since https://github.com/apache/spark/pull/26256 is merged
```sql
select cast(v as interval) from values ('+ 1 second') t(v);
select cast(v as interval) from values ('- 1 second') t(v);
```
### Why are the changes needed?
bug fix
### Does this PR introduce any user-facing change?
no
### How was this patch tested?
1. ut
2. new benchmark test
Closes#26449 from yaooqinn/SPARK-29605.
Authored-by: Kent Yao <yaooqinn@hotmail.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
### What changes were proposed in this pull request?
In the PR, I propose new function `stringToInterval()` in `IntervalUtils` for converting `UTF8String` to `CalendarInterval`. The function is used in casting a `STRING` column to an `INTERVAL` column.
### Why are the changes needed?
The proposed implementation is ~10 times faster. For example, parsing 9 interval units on JDK 8:
Before:
```
9 units w/ interval 14004 14125 116 0.1 14003.6 0.0X
9 units w/o interval 13785 14056 290 0.1 13784.9 0.0X
```
After:
```
9 units w/ interval 1343 1344 1 0.7 1343.0 0.3X
9 units w/o interval 1345 1349 8 0.7 1344.6 0.3X
```
### Does this PR introduce any user-facing change?
No
### How was this patch tested?
- By new tests for `stringToInterval` in `IntervalUtilsSuite`
- By existing tests
Closes#26256 from MaxGekk/string-to-interval.
Authored-by: Maxim Gekk <max.gekk@gmail.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
First, a bit of background on the code being changed. The current code tracks
metric updates for each task, recording which metrics the task is monitoring
and the last update value.
Once a SQL execution finishes, then the metrics for all the stages are
aggregated, by building a list with all (metric ID, value) pairs collected
for all tasks in the stages related to the execution, then grouping by metric
ID, and then calculating the values shown in the UI.
That is full of inefficiencies:
- in normal operation, all tasks will be tracking and updating the same
metrics. So recording the metric IDs per task is wasteful.
- tracking by task means we might be double-counting values if you have
speculative tasks (as a comment in the code mentions).
- creating a list of (metric ID, value) is extremely inefficient, because now
you have a huge map in memory storing boxed versions of the metric IDs and
values.
- same thing for the aggregation part, where now a Seq is built with the values
for each metric ID.
The end result is that for large queries, this code can become both really
slow, thus affecting the processing of events, and memory hungry.
The updated code changes the approach to the following:
- stages track metrics by their ID; this means the stage tracking code
naturally groups values, making aggregation later simpler.
- each metric ID being tracked uses a long array matching the number of
partitions of the stage; this means that it's cheap to update the value of
the metric once a task ends.
- when aggregating, custom code just concatenates the arrays corresponding to
the matching metric IDs; this is cheaper than the previous, boxing-heavy
approach.
The end result is that the listener uses about half as much memory as before
for tracking metrics, since it doesn't need to track metric IDs per task.
I captured heap dumps with the old and the new code during metric aggregation
in the listener, for an execution with 3 stages, 100k tasks per stage, 50
metrics updated per task. The dumps contained just reachable memory - so data
kept by the listener plus the variables in the aggregateMetrics() method.
With the old code, the thread doing aggregation references >1G of memory - and
that does not include temporary data created by the "groupBy" transformation
(for which the intermediate state is not referenced in the aggregation method).
The same thread with the new code references ~250M of memory. The old code uses
about ~250M to track all the metric values for that execution, while the new
code uses about ~130M. (Note the per-thread numbers include the amount used to
track the metrics - so, e.g., in the old case, aggregation was referencing
about ~750M of temporary data.)
I'm also including a small benchmark (based on the Benchmark class) so that we
can measure how much changes to this code affect performance. The benchmark
contains some extra code to measure things the normal Benchmark class does not,
given that the code under test does not really map that well to the
expectations of that class.
Running with the old code (I removed results that don't make much
sense for this benchmark):
```
[info] Java HotSpot(TM) 64-Bit Server VM 1.8.0_181-b13 on Linux 4.15.0-66-generic
[info] Intel(R) Core(TM) i7-6820HQ CPU 2.70GHz
[info] metrics aggregation (50 metrics, 100k tasks per stage): Best Time(ms) Avg Time(ms)
[info] --------------------------------------------------------------------------------------
[info] 1 stage(s) 2113 2118
[info] 2 stage(s) 4172 4392
[info] 3 stage(s) 7755 8460
[info]
[info] Stage Count Stage Proc. Time Aggreg. Time
[info] 1 614 1187
[info] 2 620 2480
[info] 3 718 5069
```
With the new code:
```
[info] Java HotSpot(TM) 64-Bit Server VM 1.8.0_181-b13 on Linux 4.15.0-66-generic
[info] Intel(R) Core(TM) i7-6820HQ CPU 2.70GHz
[info] metrics aggregation (50 metrics, 100k tasks per stage): Best Time(ms) Avg Time(ms)
[info] --------------------------------------------------------------------------------------
[info] 1 stage(s) 727 886
[info] 2 stage(s) 1722 1983
[info] 3 stage(s) 2752 3013
[info]
[info] Stage Count Stage Proc. Time Aggreg. Time
[info] 1 408 177
[info] 2 389 423
[info] 3 372 660
```
So the new code is faster than the old when processing task events, and about
an order of maginute faster when aggregating metrics.
Note this still leaves room for improvement; for example, using the above
measurements, 600ms is still a huge amount of time to spend in an event
handler. But I'll leave further enhancements for a separate change.
Tested with benchmarking code + existing unit tests.
Closes#26218 from vanzin/SPARK-29562.
Authored-by: Marcelo Vanzin <vanzin@cloudera.com>
Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
### What changes were proposed in this pull request?
Only use antlr4 to parse the interval string, and remove the duplicated parsing logic from `CalendarInterval`.
### Why are the changes needed?
Simplify the code and fix inconsistent behaviors.
### Does this PR introduce any user-facing change?
No
### How was this patch tested?
Pass the Jenkins with the updated test cases.
Closes#26190 from cloud-fan/parser.
Lead-authored-by: Wenchen Fan <wenchen@databricks.com>
Co-authored-by: Dongjoon Hyun <dongjoon@apache.org>
Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
### What changes were proposed in this pull request?
This is a follow-up of https://github.com/apache/spark/pull/26189 to regenerate the result on EC2.
### Why are the changes needed?
This will be used for the other PR reviews.
### Does this PR introduce any user-facing change?
No.
### How was this patch tested?
N/A.
Closes#26233 from dongjoon-hyun/SPARK-29533.
Authored-by: Dongjoon Hyun <dhyun@apple.com>
Signed-off-by: DB Tsai <d_tsai@apple.com>
### What changes were proposed in this pull request?
I extended `ExtractBenchmark` to support the `INTERVAL` type of the `source` parameter of the `date_part` function.
### Why are the changes needed?
- To detect performance issues while changing implementation of the `date_part` function in the future.
- To find out current performance bottlenecks in `date_part` for the `INTERVAL` type
### Does this PR introduce any user-facing change?
No
### How was this patch tested?
By running the benchmark and print out produced values per each `field` value.
Closes#26175 from MaxGekk/extract-interval-benchmark.
Authored-by: Maxim Gekk <max.gekk@gmail.com>
Signed-off-by: HyukjinKwon <gurwls223@apache.org>
### What changes were proposed in this pull request?
Added new benchmark `IntervalBenchmark` to measure performance of interval related functions. In the PR, I added benchmarks for casting strings to interval. In particular, interval strings with `interval` prefix and without it because there is special code for this da576a737c/common/unsafe/src/main/java/org/apache/spark/unsafe/types/CalendarInterval.java (L100-L103) . And also I added benchmarks for different number of units in interval strings, for example 1 unit is `interval 10 years`, 2 units w/o interval is `10 years 5 months`, and etc.
### Why are the changes needed?
- To find out current performance issues in casting to intervals
- The benchmark can be used while refactoring/re-implementing `CalendarInterval.fromString()` or `CalendarInterval.fromCaseInsensitiveString()`.
### Does this PR introduce any user-facing change?
No
### How was this patch tested?
By running the benchmark via the command:
```shell
SPARK_GENERATE_BENCHMARK_FILES=1 build/sbt "sql/test:runMain org.apache.spark.sql.execution.benchmark.IntervalBenchmark"
```
Closes#26189 from MaxGekk/interval-from-string-benchmark.
Authored-by: Maxim Gekk <max.gekk@gmail.com>
Signed-off-by: HyukjinKwon <gurwls223@apache.org>
### What changes were proposed in this pull request?
This PR aims the followings.
- Refactor `TPCDSQueryBenchmark` to use main method to improve the usability.
- Reduce the number of iteration from 5 to 2 because it takes too long. (2 is okay because we have `Stdev` field now. If there is an irregular run, we can notice easily with that).
- Generate one result file for TPCDS scale factor 1. (Note that this test suite can be used for the other scale factors, too.)
- AWS EC2 `r3.xlarge` with `ami-06f2f779464715dc5 (ubuntu-bionic-18.04-amd64-server-20190722.1)` is used.
This PR adds a JDK8 result based on the TPCDS ScaleFactor 1G data generated by the following.
```
# `spark-tpcds-datagen` needs this. (JDK8)
$ git clone https://github.com/apache/spark.git -b branch-2.4 --depth 1 spark-2.4
$ export SPARK_HOME=$PWD
$ ./build/mvn clean package -DskipTests
# Generate data. (JDK8)
$ git clone gitgithub.com:maropu/spark-tpcds-datagen.git
$ cd spark-tpcds-datagen/
$ build/mvn clean package
$ mkdir -p /data/tpcds
$ ./bin/dsdgen --output-location /data/tpcds/s1 // This need `Spark 2.4`
```
### Why are the changes needed?
Although the generated TPCDS data is random, we can keep the record.
### Does this PR introduce any user-facing change?
No. (This is dev-only test benchmark).
### How was this patch tested?
Manually run the benchmark. Please note that you need to have TPCDS data.
```
SPARK_GENERATE_BENCHMARK_FILES=1 build/sbt "sql/test:runMain org.apache.spark.sql.execution.benchmark.TPCDSQueryBenchmark --data-location /data/tpcds/s1"
```
Closes#26049 from dongjoon-hyun/SPARK-25668.
Authored-by: Dongjoon Hyun <dhyun@apple.com>
Signed-off-by: HyukjinKwon <gurwls223@apache.org>
### What changes were proposed in this pull request?
This PR renames `object JSONBenchmark` to `object JsonBenchmark` and the benchmark result file `JSONBenchmark-results.txt` to `JsonBenchmark-results.txt`.
### Why are the changes needed?
Since the file name doesn't match with `object JSONBenchmark`, it makes a confusion when we run the benchmark. In addition, this makes the automation difficult.
```
$ find . -name JsonBenchmark.scala
./sql/core/src/test/scala/org/apache/spark/sql/execution/datasources/json/JsonBenchmark.scala
```
```
$ build/sbt "sql/test:runMain org.apache.spark.sql.execution.datasources.json.JsonBenchmark"
[info] Running org.apache.spark.sql.execution.datasources.json.JsonBenchmark
[error] Error: Could not find or load main class org.apache.spark.sql.execution.datasources.json.JsonBenchmark
```
### Does this PR introduce any user-facing change?
No.
### How was this patch tested?
This is just renaming.
Closes#26008 from dongjoon-hyun/SPARK-RENAME-JSON.
Authored-by: Dongjoon Hyun <dhyun@apple.com>
Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
### What changes were proposed in this pull request?
This PR regenerates the `sql/core` benchmarks in JDK8/11 to compare the result. In general, we compare the ratio instead of the time. However, in this PR, the average time is compared. This PR should be considered as a rough comparison.
**A. EXPECTED CASES(JDK11 is faster in general)**
- [x] BloomFilterBenchmark (JDK11 is faster except one case)
- [x] BuiltInDataSourceWriteBenchmark (JDK11 is faster at CSV/ORC)
- [x] CSVBenchmark (JDK11 is faster except five cases)
- [x] ColumnarBatchBenchmark (JDK11 is faster at `boolean`/`string` and some cases in `int`/`array`)
- [x] DatasetBenchmark (JDK11 is faster with `string`, but is slower for `long` type)
- [x] ExternalAppendOnlyUnsafeRowArrayBenchmark (JDK11 is faster except two cases)
- [x] ExtractBenchmark (JDK11 is faster except HOUR/MINUTE/SECOND/MILLISECONDS/MICROSECONDS)
- [x] HashedRelationMetricsBenchmark (JDK11 is faster)
- [x] JSONBenchmark (JDK11 is much faster except eight cases)
- [x] JoinBenchmark (JDK11 is faster except five cases)
- [x] OrcNestedSchemaPruningBenchmark (JDK11 is faster in nine cases)
- [x] PrimitiveArrayBenchmark (JDK11 is faster)
- [x] SortBenchmark (JDK11 is faster except `Arrays.sort` case)
- [x] UDFBenchmark (N/A, values are too small)
- [x] UnsafeArrayDataBenchmark (JDK11 is faster except one case)
- [x] WideTableBenchmark (JDK11 is faster except two cases)
**B. CASES WE NEED TO INVESTIGATE MORE LATER**
- [x] AggregateBenchmark (JDK11 is slower in general)
- [x] CompressionSchemeBenchmark (JDK11 is slower in general except `string`)
- [x] DataSourceReadBenchmark (JDK11 is slower in general)
- [x] DateTimeBenchmark (JDK11 is slightly slower in general except `parsing`)
- [x] MakeDateTimeBenchmark (JDK11 is slower except two cases)
- [x] MiscBenchmark (JDK11 is slower except ten cases)
- [x] OrcV2NestedSchemaPruningBenchmark (JDK11 is slower)
- [x] ParquetNestedSchemaPruningBenchmark (JDK11 is slower except six cases)
- [x] RangeBenchmark (JDK11 is slower except one case)
`FilterPushdownBenchmark/InExpressionBenchmark/WideSchemaBenchmark` will be compared later because it took long timer.
### Why are the changes needed?
According to the result, there are some difference between JDK8/JDK11.
This will be a baseline for the future improvement and comparison. Also, as a reproducible environment, the following environment is used.
- Instance: `r3.xlarge`
- OS: `CentOS Linux release 7.5.1804 (Core)`
- JDK:
- `OpenJDK Runtime Environment (build 1.8.0_222-b10)`
- `OpenJDK Runtime Environment 18.9 (build 11.0.4+11-LTS)`
### Does this PR introduce any user-facing change?
No.
### How was this patch tested?
This is a test-only PR. We need to run benchmark.
Closes#26003 from dongjoon-hyun/SPARK-29320.
Authored-by: Dongjoon Hyun <dhyun@apple.com>
Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
### What changes were proposed in this pull request?
Added new expression `SecondWithFraction` which produces the `seconds` part of timestamps/dates with fractional part containing microseconds. This expression is used only in the `DatePart` expression. As the result, `date_part()` and `extract` return seconds and microseconds as the fractional part of the seconds part when `field` is `SECOND` (or synonyms).
### Why are the changes needed?
The `date_part()` and `extract` were added to maintain feature parity with PostgreSQL which has different behavior for the `SECOND` value of the `field` parameter. The fix is needed to behave in the same way. Here is PostgreSQL's output:
```sql
# SELECT date_part('SECONDS', timestamp'2019-10-01 00:00:01.000001');
date_part
-----------
1.000001
(1 row)
```
### Does this PR introduce any user-facing change?
Yes, type of `date_part('SECOND', ...)` is changed from `INT` to `DECIMAL(8, 6)`.
Before:
```sql
spark-sql> SELECT date_part('SECONDS', '2019-10-01 00:00:01.000001');
1
```
After:
```sql
spark-sql> SELECT date_part('SECONDS', '2019-10-01 00:00:01.000001');
1.000001
```
### How was this patch tested?
- Added new tests to `DateExpressionSuite` for the `SecondWithFraction` expression
- Regenerated results of `date_part.sql`, `extract.sql` and `timestamp.sql`
- Updated results of `ExtractBenchmark`
Closes#25986 from MaxGekk/extract-seconds-from-timestamp.
Authored-by: Maxim Gekk <max.gekk@gmail.com>
Signed-off-by: HyukjinKwon <gurwls223@apache.org>
### What changes were proposed in this pull request?
Refactoring of the `DateTimeUtils.getEpoch()` function by avoiding decimal operations that are pretty expensive, and converting the final result to the decimal type at the end.
### Why are the changes needed?
The changes improve performance of the `getEpoch()` method at least up to **20 times**.
Before:
```
Invoke extract for timestamp: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
cast to timestamp 256 277 33 39.0 25.6 1.0X
EPOCH of timestamp 23455 23550 131 0.4 2345.5 0.0X
```
After:
```
Invoke extract for timestamp: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
cast to timestamp 255 294 34 39.2 25.5 1.0X
EPOCH of timestamp 1049 1054 9 9.5 104.9 0.2X
```
### Does this PR introduce any user-facing change?
No
### How was this patch tested?
By existing test from `DateExpressionSuite`.
Closes#25881 from MaxGekk/optimize-extract-epoch.
Authored-by: Maxim Gekk <max.gekk@gmail.com>
Signed-off-by: HyukjinKwon <gurwls223@apache.org>
### What changes were proposed in this pull request?
Changed the `DateTimeUtils.getMilliseconds()` by avoiding the decimal division, and replacing it by setting scale and precision while converting microseconds to the decimal type.
### Why are the changes needed?
This improves performance of `extract` and `date_part()` by more than **50 times**:
Before:
```
Invoke extract for timestamp: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative Invoke extract for timestamp: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
cast to timestamp 397 428 45 25.2 39.7 1.0X
MILLISECONDS of timestamp 36723 36761 63 0.3 3672.3 0.0X
```
After:
```
Invoke extract for timestamp: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
cast to timestamp 278 284 6 36.0 27.8 1.0X
MILLISECONDS of timestamp 592 606 13 16.9 59.2 0.5X
```
### Does this PR introduce any user-facing change?
No
### How was this patch tested?
By existing test suite - `DateExpressionsSuite`
Closes#25871 from MaxGekk/optimize-epoch-millis.
Lead-authored-by: Maxim Gekk <max.gekk@gmail.com>
Co-authored-by: Dongjoon Hyun <dhyun@apple.com>
Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
### What changes were proposed in this pull request?
Refactored SQL-related benchmark and made them depend on `SqlBasedBenchmark`. In particular, creation of Spark session are moved into `override def getSparkSession: SparkSession`.
### Why are the changes needed?
This should simplify maintenance of SQL-based benchmarks by reducing the number of dependencies. In the future, it should be easier to refactor & extend all SQL benchmarks by changing only one trait. Finally, all SQL-based benchmarks will look uniformly.
### Does this PR introduce any user-facing change?
No
### How was this patch tested?
By running the modified benchmarks.
Closes#25828 from MaxGekk/sql-benchmarks-refactoring.
Lead-authored-by: Maxim Gekk <max.gekk@gmail.com>
Co-authored-by: Dongjoon Hyun <dongjoon@apache.org>
Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
### What changes were proposed in this pull request?
Added new benchmarks for `make_date()` and `make_timestamp()` to detect performance issues, and figure out functions speed on foldable arguments.
- `make_date()` is benchmarked on fully foldable arguments.
- `make_timestamp()` is benchmarked on corner case `60.0`, foldable time fields and foldable date.
### Why are the changes needed?
To find out inputs where `make_date()` and `make_timestamp()` have performance problems. This should be useful in the future optimizations of the functions and users apps.
### Does this PR introduce any user-facing change?
No
### How was this patch tested?
By running the benchmark and manually checking generated dates/timestamps.
Closes#25813 from MaxGekk/make_datetime-benchmark.
Authored-by: Maxim Gekk <max.gekk@gmail.com>
Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
### What changes were proposed in this pull request?
In the PR, I propose to create an instance of `TimestampFormatter` only once at the initialization, and reuse it inside of `nullSafeEval()` and `doGenCode()` in the case when the `fmt` parameter is foldable.
### Why are the changes needed?
The changes improve performance of the `date_format()` function.
Before:
```
format date: Best/Avg Time(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------
format date wholestage off 7180 / 7181 1.4 718.0 1.0X
format date wholestage on 7051 / 7194 1.4 705.1 1.0X
```
After:
```
format date: Best/Avg Time(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------
format date wholestage off 4787 / 4839 2.1 478.7 1.0X
format date wholestage on 4736 / 4802 2.1 473.6 1.0X
```
### Does this PR introduce any user-facing change?
No.
### How was this patch tested?
By existing test suites `DateExpressionsSuite` and `DateFunctionsSuite`.
Closes#25782 from MaxGekk/date_format-foldable.
Authored-by: Maxim Gekk <max.gekk@gmail.com>
Signed-off-by: HyukjinKwon <gurwls223@apache.org>
### What changes were proposed in this pull request?
In the PR, I propose to extend `ExtractBenchmark` and add new ones for:
- `EXTRACT` and `DATE` as input column
- the `DATE_PART` function and `DATE`/`TIMESTAMP` input column
### Why are the changes needed?
The `EXTRACT` expression is rebased on the `DATE_PART` expression by the PR https://github.com/apache/spark/pull/25410 where some of sub-expressions take `DATE` column as the input (`Millennium`, `Year` and etc.) but others require `TIMESTAMP` column (`Hour`, `Minute`). Separate benchmarks for `DATE` should exclude overhead of implicit conversions `DATE` <-> `TIMESTAMP`.
### Does this PR introduce any user-facing change?
No, it doesn't.
### How was this patch tested?
- Regenerated results of `ExtractBenchmark`
Closes#25772 from MaxGekk/date_part-benchmark.
Authored-by: Maxim Gekk <max.gekk@gmail.com>
Signed-off-by: Takeshi Yamamuro <yamamuro@apache.org>
## What changes were proposed in this pull request?
Added new benchmark `ExtractBenchmark` for the `EXTRACT(field FROM source)` function. It was executed on all currently supported values of the `field` argument: `MILLENNIUM`, `CENTURY`, `DECADE`, `YEAR`, `ISOYEAR`, `QUARTER`, `MONTH`, `WEEK`, `DAY`, `DAYOFWEEK`, `HOUR`, `MINUTE`, `SECOND`, `MILLISECONDS`, `MICROSECONDS`, `EPOCH`. The `cast(id as timestamp)` was taken as the `source` argument.
## How was this patch tested?
By running the benchmark via:
```
$ SPARK_GENERATE_BENCHMARK_FILES=1 build/sbt "sql/test:runMain org.apache.spark.sql.execution.benchmark.ExtractBenchmark"
```
Closes#25462 from MaxGekk/extract-benchmark.
Lead-authored-by: Maxim Gekk <max.gekk@gmail.com>
Co-authored-by: Dongjoon Hyun <dhyun@apple.com>
Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
## What changes were proposed in this pull request?
Performance issue using explode was found when a complex field contains huge array is to get duplicated as the number of exploded array elements. Given example:
```scala
val df = spark.sparkContext.parallelize(Seq(("1",
Array.fill(M)({
val i = math.random
(i.toString, (i + 1).toString, (i + 2).toString, (i + 3).toString)
})))).toDF("col", "arr")
.selectExpr("col", "struct(col, arr) as st")
.selectExpr("col", "st.col as col1", "explode(st.arr) as arr_col")
```
The explode causes `st` to be duplicated as many as the exploded elements.
Benchmarks it:
```
[info] Java HotSpot(TM) 64-Bit Server VM 1.8.0_202-b08 on Mac OS X 10.14.4
[info] Intel(R) Core(TM) i7-8750H CPU 2.20GHz
[info] generate big nested struct array: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
[info] ------------------------------------------------------------------------------------------------------------------------
[info] generate big nested struct array wholestage off 52668 53162 699 0.0 877803.4 1.0X
[info] generate big nested struct array wholestage on 47261 49093 1125 0.0 787690.2 1.1X
[info]
```
The query plan:
```
== Physical Plan ==
Project [col#508, st#512.col AS col1#515, arr_col#519]
+- Generate explode(st#512.arr), [col#508, st#512], false, [arr_col#519]
+- Project [_1#503 AS col#508, named_struct(col, _1#503, arr, _2#504) AS st#512]
+- SerializeFromObject [staticinvoke(class org.apache.spark.unsafe.types.UTF8String, StringType, fromString, knownnotnull(assertnotnull(input[0, scala.Tuple2, true]))._1, true, false) AS _1#503, mapobjects(MapObjects_loopValue84, MapObjects_loopIsNull84, ObjectType(class scala.Tuple4), if (isnull(lambdavariable(MapObjects_loopValue84, MapObjects_loopIsNull84, ObjectType(class scala.Tuple4), true))) null else named_struct(_1, staticinvoke(class org.apache.spark.unsafe.types.UTF8String, StringType, fromString, knownnotnull(lambdavariable(MapObjects_loopValue84, MapObjects_loopIsNull84, ObjectType(class scala.Tuple4), true))._1, true, false), _2, staticinvoke(class org.apache.spark.unsafe.types.UTF8String, StringType, fromString, knownnotnull(lambdavariable(MapObjects_loopValue84, MapObjects_loopIsNull84, ObjectType(class scala.Tuple4), true))._2, true, false), _3, staticinvoke(class org.apache.spark.unsafe.types.UTF8String, StringType, fromString, knownnotnull(lambdavariable(MapObjects_loopValue84, MapObjects_loopIsNull84, ObjectType(class scala.Tuple4), true))._3, true, false), _4, staticinvoke(class org.apache.spark.unsafe.types.UTF8String, StringType, fromString, knownnotnull(lambdavariable(MapObjects_loopValue84, MapObjects_loopIsNull84, ObjectType(class scala.Tuple4), true))._4, true, false)), knownnotnull(assertnotnull(input[0, scala.Tuple2, true]))._2, None) AS _2#504]
+- Scan[obj#534]
```
This patch takes nested column pruning approach to prune unnecessary nested fields. It adds a projection of the needed nested fields as aliases on the child of `Generate`, and substitutes them by alias attributes on the projection on top of `Generate`.
Benchmarks it after the change:
```
[info] Java HotSpot(TM) 64-Bit Server VM 1.8.0_202-b08 on Mac OS X 10.14.4
[info] Intel(R) Core(TM) i7-8750H CPU 2.20GHz
[info] generate big nested struct array: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
[info] ------------------------------------------------------------------------------------------------------------------------
[info] generate big nested struct array wholestage off 311 331 28 0.2 5188.6 1.0X
[info] generate big nested struct array wholestage on 297 312 15 0.2 4947.3 1.0X
[info]
```
The query plan:
```
== Physical Plan ==
Project [col#592, _gen_alias_608#608 AS col1#599, arr_col#603]
+- Generate explode(st#596.arr), [col#592, _gen_alias_608#608], false, [arr_col#603]
+- Project [_1#587 AS col#592, named_struct(col, _1#587, arr, _2#588) AS st#596, _1#587 AS _gen_alias_608#608]
+- SerializeFromObject [staticinvoke(class org.apache.spark.unsafe.types.UTF8String, StringType, fromString, knownnotnull(assertnotnull(in
put[0, scala.Tuple2, true]))._1, true, false) AS _1#587, mapobjects(MapObjects_loopValue102, MapObjects_loopIsNull102, ObjectType(class scala.Tuple4),
if (isnull(lambdavariable(MapObjects_loopValue102, MapObjects_loopIsNull102, ObjectType(class scala.Tuple4), true))) null else named_struct(_1, staticinvoke(class org.apache.spark.unsafe.types.UTF8String, StringType, fromString, knownnotnull(lambdavariable(MapObjects_loopValue102, MapObjects_loopIsNull102, ObjectType(class scala.Tuple4), true))._1, true, false), _2, staticinvoke(class org.apache.spark.unsafe.types.UTF8String, StringType, fromString, knownnotnull(lambdavariable(MapObjects_loopValue102, MapObjects_loopIsNull102, ObjectType(class scala.Tuple4), true))._2, true, false), _3, staticinvoke(class org.apache.spark.unsafe.types.UTF8String, StringType, fromString, knownnotnull(lambdavariable(MapObjects_loopValue102, MapObjects_loopIsNull102, ObjectType(class scala.Tuple4), true))._3, true, false), _4, staticinvoke(class org.apache.spark.unsafe.types.UTF8String, StringType, fromString, knownnotnull(lambdavariable(MapObjects_loopValue102, MapObjects_loopIsNull102, ObjectType(class scala.Tuple4), true))._4, true, false)), knownnotnull(assertnotnull(input[0, scala.Tuple2, true]))._2, None) AS _2#588]
+- Scan[obj#586]
```
This behavior is controlled by a SQL config `spark.sql.optimizer.expression.nestedPruning.enabled`.
## How was this patch tested?
Added benchmark.
Closes#24637 from viirya/SPARK-27707.
Lead-authored-by: Liang-Chi Hsieh <viirya@gmail.com>
Co-authored-by: Dongjoon Hyun <dhyun@apple.com>
Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
## What changes were proposed in this pull request?
`NestedColumnAliasing` rule covers `GetStructField` only, currently. It means that some nested field extraction expressions aren't pruned. For example, if only accessing a nested field in an array of struct (`GetArrayStructFields`), this column isn't pruned.
This patch extends the rule to cover general nested field cases, including `GetArrayStructFields`.
## How was this patch tested?
Added tests.
Closes#24599 from viirya/nested-pruning-extract-value.
Lead-authored-by: Liang-Chi Hsieh <viirya@gmail.com>
Co-authored-by: Dongjoon Hyun <dhyun@apple.com>
Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
## What changes were proposed in this pull request?
As outlined in the JIRA by JoshRosen, our conversion mechanism from catalyst types to scala ones is pretty inefficient for primitive data types. Indeed, in these cases, most of the times we are adding useless calls to `identity` function or anyway to functions which return the same value. Using the information we have when we generate the code, we can avoid most of these overheads.
## How was this patch tested?
Here is a simple test which shows the benefit that this PR can bring:
```
test("SPARK-27684: perf evaluation") {
val intLongUdf = ScalaUDF(
(a: Int, b: Long) => a + b, LongType,
Literal(1) :: Literal(1L) :: Nil,
true :: true :: Nil,
nullable = false)
val plan = generateProject(
MutableProjection.create(Alias(intLongUdf, s"udf")() :: Nil),
intLongUdf)
plan.initialize(0)
var i = 0
val N = 100000000
val t0 = System.nanoTime()
while(i < N) {
plan(EmptyRow).get(0, intLongUdf.dataType)
plan(EmptyRow).get(0, intLongUdf.dataType)
plan(EmptyRow).get(0, intLongUdf.dataType)
plan(EmptyRow).get(0, intLongUdf.dataType)
plan(EmptyRow).get(0, intLongUdf.dataType)
plan(EmptyRow).get(0, intLongUdf.dataType)
plan(EmptyRow).get(0, intLongUdf.dataType)
plan(EmptyRow).get(0, intLongUdf.dataType)
plan(EmptyRow).get(0, intLongUdf.dataType)
plan(EmptyRow).get(0, intLongUdf.dataType)
i += 1
}
val t1 = System.nanoTime()
println(s"Avg time: ${(t1 - t0).toDouble / N} ns")
}
```
The output before the patch is:
```
Avg time: 51.27083294 ns
```
after, we get:
```
Avg time: 11.85874227 ns
```
which is ~5X faster.
Moreover a benchmark has been added for Scala UDF. The output after the patch can be seen in this PR, before the patch, the output was:
```
================================================================================================
UDF with mixed input types
================================================================================================
Java HotSpot(TM) 64-Bit Server VM 1.8.0_152-b16 on Mac OS X 10.13.6
Intel(R) Core(TM) i7-4558U CPU 2.80GHz
long/nullable int/string to string: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
long/nullable int/string to string wholestage off 257 287 42 0,4 2569,5 1,0X
long/nullable int/string to string wholestage on 158 172 18 0,6 1579,0 1,6X
Java HotSpot(TM) 64-Bit Server VM 1.8.0_152-b16 on Mac OS X 10.13.6
Intel(R) Core(TM) i7-4558U CPU 2.80GHz
long/nullable int/string to option: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
long/nullable int/string to option wholestage off 104 107 5 1,0 1037,9 1,0X
long/nullable int/string to option wholestage on 80 92 12 1,2 804,0 1,3X
Java HotSpot(TM) 64-Bit Server VM 1.8.0_152-b16 on Mac OS X 10.13.6
Intel(R) Core(TM) i7-4558U CPU 2.80GHz
long/nullable int to primitive: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
long/nullable int to primitive wholestage off 71 76 7 1,4 712,1 1,0X
long/nullable int to primitive wholestage on 64 71 6 1,6 636,2 1,1X
================================================================================================
UDF with primitive types
================================================================================================
Java HotSpot(TM) 64-Bit Server VM 1.8.0_152-b16 on Mac OS X 10.13.6
Intel(R) Core(TM) i7-4558U CPU 2.80GHz
long/nullable int to string: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
long/nullable int to string wholestage off 60 60 0 1,7 600,3 1,0X
long/nullable int to string wholestage on 55 64 8 1,8 551,2 1,1X
Java HotSpot(TM) 64-Bit Server VM 1.8.0_152-b16 on Mac OS X 10.13.6
Intel(R) Core(TM) i7-4558U CPU 2.80GHz
long/nullable int to option: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
long/nullable int to option wholestage off 66 73 9 1,5 663,0 1,0X
long/nullable int to option wholestage on 30 32 2 3,3 300,7 2,2X
Java HotSpot(TM) 64-Bit Server VM 1.8.0_152-b16 on Mac OS X 10.13.6
Intel(R) Core(TM) i7-4558U CPU 2.80GHz
long/nullable int/string to primitive: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
long/nullable int/string to primitive wholestage off 32 35 5 3,2 316,7 1,0X
long/nullable int/string to primitive wholestage on 41 68 17 2,4 414,0 0,8X
```
The improvements are particularly visible in the second case, ie. when only primitive types are used as inputs.
Closes#24636 from mgaido91/SPARK-27684.
Authored-by: Marco Gaido <marcogaido91@gmail.com>
Signed-off-by: Josh Rosen <rosenville@gmail.com>
## What changes were proposed in this pull request?
Added new JSON benchmarks related to date and timestamps operations:
- Write date/timestamp to JSON files
- `to_json()` and `from_json()` for dates and timestamps
- Read date/timestamps from JSON files, and infer schemas
- Parse and infer schemas from `Dataset[String]`
Also existing JSON benchmarks are ported on `NoOp` datasource.
Closes#24430 from MaxGekk/json-datetime-benchmark.
Authored-by: Maxim Gekk <max.gekk@gmail.com>
Signed-off-by: HyukjinKwon <gurwls223@apache.org>
## What changes were proposed in this pull request?
Added new CSV benchmarks related to date and timestamps operations:
- Write date/timestamp to CSV files
- `to_csv()` and `from_csv()` for dates and timestamps
- Read date/timestamps from CSV files, and infer schemas
- Parse and infer schemas from `Dataset[String]`
Also existing CSV benchmarks are ported on `NoOp` datasource.
Closes#24429 from MaxGekk/csv-timestamp-benchmark.
Authored-by: Maxim Gekk <max.gekk@gmail.com>
Signed-off-by: HyukjinKwon <gurwls223@apache.org>
## What changes were proposed in this pull request?
In the PR, I propose to parse strings to timestamps in microsecond precision by the ` to_timestamp()` function if the specified pattern contains a sub-pattern for seconds fractions.
Closes#24342
## How was this patch tested?
By `DateFunctionsSuite` and `DateExpressionsSuite`
Closes#24420 from MaxGekk/to_timestamp-microseconds3.
Authored-by: Maxim Gekk <max.gekk@gmail.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
## What changes were proposed in this pull request?
We added nested schema pruning support to Orc V2 recently. The benchmark result should be updated. The benchmark numbers are obtained by running benchmark on r3.xlarge machine.
## How was this patch tested?
Test only change.
Closes#24399 from viirya/update-orcv2-benchmark.
Authored-by: Liang-Chi Hsieh <viirya@gmail.com>
Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
## What changes were proposed in this pull request?
In the PR, I propose simpler implementation of `toJavaTimestamp()`/`fromJavaTimestamp()` by reusing existing functions of `DateTimeUtils`. This will allow to:
- Simply implementation of `toJavaTimestamp()`, and handle properly negative inputs.
- Detect `Long` overflow in conversion of milliseconds (`java.sql.Timestamp`) to microseconds (Catalyst's Timestamp).
## How was this patch tested?
By existing test suites `DateTimeUtilsSuite`, `DateFunctionsSuite`, `DateExpressionsSuite` and `CastSuite`. And by new benchmark for export/import timestamps added to `DateTimeBenchmark`:
Before:
```
To/from java.sql.Timestamp: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
From java.sql.Timestamp 290 335 49 17.2 58.0 1.0X
Collect longs 1234 1681 487 4.1 246.8 0.2X
Collect timestamps 1718 1755 63 2.9 343.7 0.2X
```
After:
```
To/from java.sql.Timestamp: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
From java.sql.Timestamp 283 301 19 17.7 56.6 1.0X
Collect longs 1048 1087 36 4.8 209.6 0.3X
Collect timestamps 1425 1479 56 3.5 285.1 0.2X
```
Closes#24311 from MaxGekk/conv-java-sql-date-timestamp.
Authored-by: Maxim Gekk <max.gekk@gmail.com>
Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
## What changes were proposed in this pull request?
Added new benchmarks for:
1. JSON functions: `from_json`, `json_tuple` and `get_json_object`
2. Parsing `Dataset[String]` with JSON records
3. Comparing just splitting input text by lines with schema inferring, per-line parsing when encoding is set and not set.
Also existing benchmarks were refactored to use the `NoOp` datasource to eliminate overhead of triggers like `.filter((_: Row) => true).count()`.
## How was this patch tested?
By running `JSONBenchmark` locally.
Closes#24252 from MaxGekk/json-benchmark-func.
Authored-by: Maxim Gekk <max.gekk@gmail.com>
Signed-off-by: Hyukjin Kwon <gurwls223@apache.org>
## What changes were proposed in this pull request?
This is a follow up PR for #23943 in order to update the benchmark result with EC2 `r3.xlarge` instance.
## How was this patch tested?
N/A. (Manually compare the diff)
Closes#24078 from dongjoon-hyun/SPARK-27034.
Authored-by: Dongjoon Hyun <dhyun@apple.com>
Signed-off-by: DB Tsai <d_tsai@apple.com>
## What changes were proposed in this pull request?
We only supported nested schema pruning for Parquet previously. This proposes to support nested schema pruning for ORC too.
Note: This only covers ORC v1. For ORC v2, the necessary change is at the schema pruning rule. We should deal with ORC v2 as a TODO item, in order to reduce review burden.
## How was this patch tested?
Added tests.
Closes#23943 from viirya/nested-schema-pruning-orc.
Authored-by: Liang-Chi Hsieh <viirya@gmail.com>
Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
## What changes were proposed in this pull request?
We have benchmark of nested schema pruning, but only for Parquet. This adds similar benchmark for ORC. This is used with nested schema pruning of ORC.
## How was this patch tested?
Added test.
Closes#23955 from viirya/orc-nested-schema-pruning-benchmark.
Authored-by: Liang-Chi Hsieh <viirya@gmail.com>
Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
## What changes were proposed in this pull request?
This PR optimizes `InSet` expressions for byte, short, integer, date types. It is a follow-up on PR #21442 from dbtsai.
`In` expressions are compiled into a sequence of if-else statements, which results in O\(n\) time complexity. `InSet` is an optimized version of `In`, which is supposed to improve the performance if all values are literals and the number of elements is big enough. However, `InSet` actually worsens the performance in many cases due to various reasons.
The main idea of this PR is to use Java `switch` statements to significantly improve the performance of `InSet` expressions for bytes, shorts, ints, dates. All `switch` statements are compiled into `tableswitch` and `lookupswitch` bytecode instructions. We will have O\(1\) time complexity if our case values are compact and `tableswitch` can be used. Otherwise, `lookupswitch` will give us O\(log n\).
Locally, I tried Spark `OpenHashSet` and primitive collections from `fastutils` in order to solve the boxing issue in `InSet`. Both options significantly decreased the memory consumption and `fastutils` improved the time compared to `HashSet` from Scala. However, the switch-based approach was still more than two times faster even on 500+ non-compact elements.
I also noticed that applying the switch-based approach on less than 10 elements gives a relatively minor improvement compared to the if-else approach. Therefore, I placed the switch-based logic into `InSet` and added a new config to track when it is applied. Even if we migrate to primitive collections at some point, the switch logic will be still faster unless the number of elements is really big. Another option is to have a separate `InSwitch` expression. However, this would mean we need to modify other places (e.g., `DataSourceStrategy`).
See [here](https://docs.oracle.com/javase/specs/jvms/se7/html/jvms-3.html#jvms-3.10) and [here](https://stackoverflow.com/questions/10287700/difference-between-jvms-lookupswitch-and-tableswitch) for more information.
This PR does not cover long values as Java `switch` statements cannot be used on them. However, we can have a follow-up PR with an approach similar to binary search.
## How was this patch tested?
There are new tests that verify the logic of the proposed optimization.
The performance was evaluated using existing benchmarks. This PR was also tested on an EC2 instance (OpenJDK 64-Bit Server VM 1.8.0_191-b12 on Linux 4.14.77-70.59.amzn1.x86_64, Intel(R) Xeon(R) CPU E5-2686 v4 2.30GHz).
## Notes
- [This link](http://hg.openjdk.java.net/jdk8/jdk8/langtools/file/30db5e0aaf83/src/share/classes/com/sun/tools/javac/jvm/Gen.java#l1153) contains source code that decides between `tableswitch` and `lookupswitch`. The logic was re-used in the benchmarks. See the `isLookupSwitch` method.
Closes#23171 from aokolnychyi/spark-26205.
Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com>
Co-authored-by: Dongjoon Hyun <dhyun@apple.com>
Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
## What changes were proposed in this pull request?
In the PR, I propose to convert time zone string to `TimeZone` by converting it to `ZoneId` which uses `ZoneOffset` internally. The `ZoneOffset` class of JDK 8 has a cache already: http://hg.openjdk.java.net/jdk8/jdk8/jdk/file/687fd7c7986d/src/share/classes/java/time/ZoneOffset.java#l205 . In this way, there is no need to support cache of time zones in Spark.
The PR removes `computedTimeZones` from `DateTimeUtils`, and uses `ZoneId.of` to convert time zone id string to `ZoneId` and to `TimeZone` at the end.
## How was this patch tested?
The changes were tested by
Closes#23812 from MaxGekk/timezone-cache.
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?
This adds `NestedSchemaPruningBenchmark` to show the nested schema pruning performance clearly and to verify new PR's performance benefit and to prevent the future performance degradation.
## How was this patch tested?
Manually run the benchmark.
```
SPARK_GENERATE_BENCHMARK_FILES=1 build/sbt "sql/test:runMain org.apache.spark.sql.execution.benchmark.NestedSchemaPruningBenchmark"
```
Closes#23862 from dongjoon-hyun/SPARK-NESTED-SCHEMA-PRUNING-BM.
Lead-authored-by: Dongjoon Hyun <dhyun@apple.com>
Co-authored-by: DB Tsai <d_tsai@apple.com>
Signed-off-by: DB Tsai <d_tsai@apple.com>
## What changes were proposed in this pull request?
In the PR, I propose to simplify timestamp truncation to quarter of year by using *java.time* API directly. The `LocalDate` instance can be truncation to quarter timestamp via adjusting by chrono field `IsoFields.DAY_OF_QUARTER`.
## How was this patch tested?
This was checked by existing test suite - `DateTimeUtilsSuite`.
Closes#23808 from MaxGekk/date-quarter-of-year.
Authored-by: Maxim Gekk <max.gekk@gmail.com>
Signed-off-by: Sean Owen <sean.owen@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?
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?
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?
This PR contains benchmarks for `In` and `InSet` expressions. They cover literals of different data types and will help us to decide where to integrate the switch-based logic for bytes/shorts/ints.
As discussed in [PR-23171](https://github.com/apache/spark/pull/23171), one potential approach is to convert `In` to `InSet` if all elements are literals independently of data types and the number of elements. According to the results of this PR, we might want to keep the threshold for the number of elements. The if-else approach approach might be faster for some data types on a small number of elements (structs? arrays? small decimals?).
### byte / short / int / long
Unless the number of items is really big, `InSet` is slower than `In` because of autoboxing .
Interestingly, `In` scales worse on bytes/shorts than on ints/longs. For example, `InSet` starts to match the performance on around 50 bytes/shorts while this does not happen on the same number of ints/longs. This is a bit strange as shorts/bytes (e.g., `(byte) 1`, `(short) 2`) are represented as ints in the bytecode.
### float / double
Use cases on floats/doubles also suffer from autoboxing. Therefore, `In` outperforms `InSet` on 10 elements.
Similarly to shorts/bytes, `In` scales worse on floats/doubles than on ints/longs because the equality condition is more complicated (e.g., `java.lang.Float.isNaN(filter_valueArg_0) && java.lang.Float.isNaN(9.0F)) || filter_valueArg_0 == 9.0F`).
### decimal
The reason why we have separate benchmarks for small and large decimals is that Spark might use longs to represent decimals in some cases.
If this optimization happens, then `equals` will be nothing else as comparing longs. If this does not happen, Spark will create an instance of `scala.BigDecimal` and use it for comparisons. The latter is more expensive.
`Decimal$hashCode` will always use `scala.BigDecimal$hashCode` even if the number is small enough to fit into a long variable. As a consequence, we see that use cases on small decimals are faster with `In` as they are using long comparisons under the hood. Large decimal values are always faster with `InSet`.
### string
`UTF8String$equals` is not cheap. Therefore, `In` does not really outperform `InSet` as in previous use cases.
### timestamp / date
Under the hood, timestamp/date values will be represented as long/int values. So, `In` allows us to avoid autoboxing.
### array
Arrays are working as expected. `In` is faster on 5 elements while `InSet` is faster on 15 elements. The benchmarks are using `UnsafeArrayData`.
### struct
`InSet` is always faster than `In` for structs. These benchmarks use `GenericInternalRow`.
Closes#23291 from aokolnychyi/spark-26203.
Lead-authored-by: Anton Okolnychyi <aokolnychyi@apple.com>
Co-authored-by: Dongjoon Hyun <dongjoon@apache.org>
Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
## What changes were proposed in this pull request?
This PR aims to remove internal ORC configuration to simplify the code path for Spark 3.0.0. This removes the configuration `spark.sql.orc.copyBatchToSpark` and related ORC codes including tests and benchmarks.
## How was this patch tested?
Pass the Jenkins with the reduced test coverage.
Closes#23503 from dongjoon-hyun/SPARK-26584.
Authored-by: Dongjoon Hyun <dongjoon@apache.org>
Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
## What changes were proposed in this pull request?
Refactor ExternalAppendOnlyUnsafeRowArrayBenchmark to use main method.
## How was this patch tested?
Manually tested and regenerated results.
Please note that `spark.memory.debugFill` setting has a huge impact on this benchmark. Since it is set to true by default when running the benchmark from SBT, we need to disable it:
```
SPARK_GENERATE_BENCHMARK_FILES=1 build/sbt ";project sql;set javaOptions in Test += \"-Dspark.memory.debugFill=false\";test:runMain org.apache.spark.sql.execution.ExternalAppendOnlyUnsafeRowArrayBenchmark"
```
Closes#22617 from peter-toth/SPARK-25484.
Lead-authored-by: Peter Toth <peter.toth@gmail.com>
Co-authored-by: Peter Toth <ptoth@hortonworks.com>
Co-authored-by: Dongjoon Hyun <dongjoon@apache.org>
Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
## What changes were proposed in this pull request?
Regarding the performance issue of SPARK-26155, it reports the issue on TPC-DS. I think it is better to add a benchmark for `LongToUnsafeRowMap` which is the root cause of performance regression.
It can be easier to show performance difference between different metric implementations in `LongToUnsafeRowMap`.
## How was this patch tested?
Manually run added benchmark.
Closes#23284 from viirya/SPARK-26337.
Authored-by: Liang-Chi Hsieh <viirya@gmail.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
## What changes were proposed in this pull request?
1. OrcReadBenchmark is under hive module, so the way to run it should be
```
build/sbt "hive/test:runMain <this class>"
```
2. The benchmark "String with Nulls Scan" should be with case "String with Nulls Scan(5%/50%/95%)", not "(0.05%/0.5%/0.95%)"
3. Add the null value percentages in the test case names of DataSourceReadBenchmark, for the benchmark "String with Nulls Scan" .
## How was this patch tested?
Re-run benchmarks
Closes#22965 from gengliangwang/fixHiveOrcReadBenchmark.
Lead-authored-by: Gengliang Wang <gengliang.wang@databricks.com>
Co-authored-by: Gengliang Wang <ltnwgl@gmail.com>
Co-authored-by: Dongjoon Hyun <dongjoon@apache.org>
Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
