### What changes were proposed in this pull request?
Push the rebase logic to the lower level of the parquet vectorized reader, to make the final code more vectorization-friendly.
### Why are the changes needed?
Parquet vectorized reader is carefully implemented, to make it more likely to be vectorized by the JVM. However, the newly added datetime rebase degrade the performance a lot, as it breaks vectorization, even if the datetime values don't need to rebase (this is very likely as dates before 1582 is rare).
### Does this PR introduce any user-facing change?
no
### How was this patch tested?
Run part of the `DateTimeRebaseBenchmark` locally. The results:
before this patch
```
[info] Load dates from parquet: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
[info] ------------------------------------------------------------------------------------------------------------------------
[info] after 1582, vec on, rebase off 2677 2838 142 37.4 26.8 1.0X
[info] after 1582, vec on, rebase on 3828 4331 805 26.1 38.3 0.7X
[info] before 1582, vec on, rebase off 2903 2926 34 34.4 29.0 0.9X
[info] before 1582, vec on, rebase on 4163 4197 38 24.0 41.6 0.6X
[info] Load timestamps from parquet: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
[info] ------------------------------------------------------------------------------------------------------------------------
[info] after 1900, vec on, rebase off 3537 3627 104 28.3 35.4 1.0X
[info] after 1900, vec on, rebase on 6891 7010 105 14.5 68.9 0.5X
[info] before 1900, vec on, rebase off 3692 3770 72 27.1 36.9 1.0X
[info] before 1900, vec on, rebase on 7588 7610 30 13.2 75.9 0.5X
```
After this patch
```
[info] Load dates from parquet: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
[info] ------------------------------------------------------------------------------------------------------------------------
[info] after 1582, vec on, rebase off 2758 2944 197 36.3 27.6 1.0X
[info] after 1582, vec on, rebase on 2908 2966 51 34.4 29.1 0.9X
[info] before 1582, vec on, rebase off 2840 2878 37 35.2 28.4 1.0X
[info] before 1582, vec on, rebase on 3407 3433 24 29.4 34.1 0.8X
[info] Load timestamps from parquet: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
[info] ------------------------------------------------------------------------------------------------------------------------
[info] after 1900, vec on, rebase off 3861 4003 139 25.9 38.6 1.0X
[info] after 1900, vec on, rebase on 4194 4283 77 23.8 41.9 0.9X
[info] before 1900, vec on, rebase off 3849 3937 79 26.0 38.5 1.0X
[info] before 1900, vec on, rebase on 7512 7546 55 13.3 75.1 0.5X
```
Date type is 30% faster if the values don't need to rebase, 20% faster if need to rebase.
Timestamp type is 60% faster if the values don't need to rebase, no difference if need to rebase.
Closes#28406 from cloud-fan/perf.
Lead-authored-by: Wenchen Fan <wenchen@databricks.com>
Co-authored-by: Maxim Gekk <max.gekk@gmail.com>
Signed-off-by: HyukjinKwon <gurwls223@apache.org>
### What changes were proposed in this pull request?
Reuse the `rebaseGregorianToJulianMicros()` and `rebaseJulianToGregorianMicros()` functions introduced by the PR #28119 in `DateTimeUtils`.`toJavaTimestamp()` and `fromJavaTimestamp()`. Actually, new implementation is derived from Spark 2.4 + rebasing via pre-calculated rebasing maps.
### Why are the changes needed?
The changes speed up conversions to/from java.sql.Timestamp, and as a consequence the PR improve performance of ORC datasource in loading/saving timestamps:
- Saving ~ **x2.8 faster** in master, and -11% against Spark 2.4.6
- Loading - **x3.2-4.5 faster** in master, -5% against Spark 2.4.6
Before:
```
Save timestamps to ORC: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
after 1582 59877 59877 0 1.7 598.8 0.0X
before 1582 61361 61361 0 1.6 613.6 0.0X
Load timestamps from ORC: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
after 1582, vec off 48197 48288 118 2.1 482.0 1.0X
after 1582, vec on 38247 38351 128 2.6 382.5 1.3X
before 1582, vec off 53179 53359 249 1.9 531.8 0.9X
before 1582, vec on 44076 44268 269 2.3 440.8 1.1X
```
After:
```
Save timestamps to ORC: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
after 1582 21250 21250 0 4.7 212.5 0.1X
before 1582 22105 22105 0 4.5 221.0 0.1X
Load timestamps from ORC: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
after 1582, vec off 14903 14933 40 6.7 149.0 1.0X
after 1582, vec on 8342 8426 73 12.0 83.4 1.8X
before 1582, vec off 15528 15575 76 6.4 155.3 1.0X
before 1582, vec on 9025 9075 61 11.1 90.2 1.7X
```
Spark 2.4.6-SNAPSHOT:
```
Save timestamps to ORC: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
after 1582 18858 18858 0 5.3 188.6 1.0X
before 1582 18508 18508 0 5.4 185.1 1.0X
Load timestamps from ORC: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
after 1582, vec off 14063 14177 143 7.1 140.6 1.0X
after 1582, vec on 5955 6029 100 16.8 59.5 2.4X
before 1582, vec off 14119 14126 7 7.1 141.2 1.0X
before 1582, vec on 5991 6007 25 16.7 59.9 2.3X
```
### Does this PR introduce any user-facing change?
Yes, the `to_utc_timestamp` function returns the later local timestamp in the case of overlapping local timestamps at daylight saving time. it's changed back to the 2.4 behavior.
### How was this patch tested?
- By existing test suite `DateTimeUtilsSuite`, `RebaseDateTimeSuite`, `DateFunctionsSuite`, `DateExpressionsSuites`, `ParquetIOSuite`, `OrcHadoopFsRelationSuite`.
- Re-generating results of the benchmarks `DateTimeBenchmark` and `DateTimeRebaseBenchmark` in the environment:
| Item | Description |
| ---- | ----|
| Region | us-west-2 (Oregon) |
| Instance | r3.xlarge |
| AMI | ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1 (ami-06f2f779464715dc5) |
| Java | OpenJDK 64-Bit Server VM 1.8.0_242 and OpenJDK 64-Bit Server VM 11.0.6+10 |
Closes#28189 from MaxGekk/optimize-to-from-java-timestamp.
Authored-by: Max Gekk <max.gekk@gmail.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
### What changes were proposed in this pull request?
In regular ORC reader when `spark.sql.orc.enableVectorizedReader` is set to `false`, I propose to use `DaysWritable` in reading DATE values from ORC files. Currently, days from ORC files are converted to java.sql.Date, and then to days in Proleptic Gregorian calendar. So, the conversion to Java type can be eliminated.
### Why are the changes needed?
- The PR fixes regressions in loading dates before the 1582 year from ORC files by when vectorised ORC reader is off.
- The changes improve performance of regular ORC reader for DATE columns.
- x3.6 faster comparing to the current master
- x1.9-x4.3 faster against Spark 2.4.6
Before (on JDK 8):
```
Load dates from ORC: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
after 1582, vec off 39651 39686 31 2.5 396.5 1.0X
after 1582, vec on 3647 3660 13 27.4 36.5 10.9X
before 1582, vec off 38155 38219 61 2.6 381.6 1.0X
before 1582, vec on 4041 4046 6 24.7 40.4 9.8X
```
After (on JDK 8):
```
Load dates from ORC: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
after 1582, vec off 10947 10971 28 9.1 109.5 1.0X
after 1582, vec on 3677 3702 36 27.2 36.8 3.0X
before 1582, vec off 11456 11472 21 8.7 114.6 1.0X
before 1582, vec on 4079 4103 21 24.5 40.8 2.7X
```
Spark 2.4.6:
```
Load dates from ORC: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
after 1582, vec off 48169 48276 96 2.1 481.7 1.0X
after 1582, vec on 5375 5410 41 18.6 53.7 9.0X
before 1582, vec off 22353 22482 198 4.5 223.5 2.2X
before 1582, vec on 5474 5475 1 18.3 54.7 8.8X
```
### Does this PR introduce any user-facing change?
No
### How was this patch tested?
- By existing tests suites like `DateTimeUtilsSuite`
- Checked for `hive-1.2` by:
```
./build/sbt -Phive-1.2 "test:testOnly *OrcHadoopFsRelationSuite"
```
- Re-run `DateTimeRebaseBenchmark` in the environment:
| Item | Description |
| ---- | ----|
| Region | us-west-2 (Oregon) |
| Instance | r3.xlarge |
| AMI | ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1 (ami-06f2f779464715dc5) |
| Java | OpenJDK 64-Bit Server VM 1.8.0_242 and OpenJDK 64-Bit Server VM 11.0.6+10 |
Closes#28169 from MaxGekk/orc-optimize-dates.
Authored-by: Max Gekk <max.gekk@gmail.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
### What changes were proposed in this pull request?
In the PR, I propose to optimise the `DateTimeUtils`.`rebaseJulianToGregorianMicros()` and `rebaseGregorianToJulianMicros()` functions, and make them faster by using pre-calculated rebasing tables. This approach allows to avoid expensive conversions via local timestamps. For example, the `America/Los_Angeles` time zone has just a few time points when difference between Proleptic Gregorian calendar and the hybrid calendar (Julian + Gregorian since 1582-10-15) is changed in the time interval 0001-01-01 .. 2100-01-01:
| i | local timestamp | Proleptic Greg. seconds | Hybrid (Julian+Greg) seconds | difference in minutes|
| -- | ------- |----|----| ---- |
|0|0001-01-01 00:00|-62135568422|-62135740800|-2872|
|1|0100-03-01 00:00|-59006333222|-59006419200|-1432|
|...|...|...|...|...|
|13|1582-10-15 00:00|-12219264422|-12219264000|7|
|14|1883-11-18 12:00|-2717640000|-2717640000|0|
The difference in microseconds between Proleptic and hybrid calendars for any local timestamp in time intervals `[local timestamp(i), local timestamp(i+1))`, and for any microseconds in the time interval `[Gregorian micros(i), Gregorian micros(i+1))` is the same. In this way, we can rebase an input micros by following the steps:
1. Look at the table, and find the time interval where the micros falls to
2. Take the difference between 2 calendars for this time interval
3. Add the difference to the input micros. The result is rebased microseconds that has the same local timestamp representation.
Here are details of the implementation:
- Pre-calculated tables are stored to JSON files `gregorian-julian-rebase-micros.json` and `julian-gregorian-rebase-micros.json` in the resource folder of `sql/catalyst`. The diffs and switch time points are stored as seconds, for example:
```json
[
{
"tz" : "America/Los_Angeles",
"switches" : [ -62135740800, -59006419200, ... , -2717640000 ],
"diffs" : [ 172378, 85978, ..., 0 ]
}
]
```
The JSON files are generated by 2 tests in `RebaseDateTimeSuite` - `generate 'gregorian-julian-rebase-micros.json'` and `generate 'julian-gregorian-rebase-micros.json'`. Both tests are disabled by default.
The `switches` time points are ordered from old to recent timestamps. This condition is checked by the test `validate rebase records in JSON files` in `RebaseDateTimeSuite`. Also sizes of the `switches` and `diffs` arrays are the same (this is checked by the same test).
- The **_Asia/Tehran, Iran, Africa/Casablanca and Africa/El_Aaiun_** time zones weren't added to the JSON files, see [SPARK-31385](https://issues.apache.org/jira/browse/SPARK-31385)
- The rebase info from the JSON files is placed to hash tables - `gregJulianRebaseMap` and `julianGregRebaseMap`. I use `AnyRefMap` because it is almost 2 times faster than Scala's immutable Map. Also I tried `java.util.HashMap` but it has worse lookup time than `AnyRefMap` in our case.
The hash maps store the switch time points and diffs in microseconds precision to avoid conversions from microseconds to seconds in the runtime.
- I moved the code related to days and microseconds rebasing to the separate object `RebaseDateTime` to do not pollute `DateTimeUtils`. Tests related to date-time rebasing are moved to `RebaseDateTimeSuite` for the same reason.
- I placed rebasing via local timestamp to separate methods that require zone id as the first parameter assuming that the caller has zone id already. This allows to void unnecessary retrieving the default time zone. The methods are marked as `private[sql]` because they are used in `RebaseDateTimeSuite` as reference implementation.
- Modified the `rebaseGregorianToJulianMicros()` and `rebaseJulianToGregorianMicros()` methods in `RebaseDateTime` to look up the rebase tables first of all. If hash maps don't contain rebasing info for the given time zone id, the methods falls back to the implementation via local timestamps. This allows to support time zones specified as zone offsets like '-08:00'.
### Why are the changes needed?
To make timestamps rebasing faster:
- Saving timestamps to parquet files is ~ **x3.8 faster**
- Loading timestamps from parquet files is ~**x2.8 faster**.
- Loading timestamps by Vectorized reader ~**x4.6 faster**.
### Does this PR introduce any user-facing change?
No
### How was this patch tested?
- Added the test `validate rebase records in JSON files` to `RebaseDateTimeSuite`. The test validates 2 json files from the resource folder - `gregorian-julian-rebase-micros.json` and `julian-gregorian-rebase-micros.json`, and it checks per each time zone records that
- the number of switch points is equal to the number of diffs between calendars. If the numbers are different, this will violate the assumption made in `RebaseDateTime.rebaseMicros`.
- swith points are ordered from old to recent timestamps. This pre-condition is required for linear search in the `rebaseMicros` function.
- Added the test `optimization of micros rebasing - Gregorian to Julian` to `RebaseDateTimeSuite` which iterates over timestamps from 0001-01-01 to 2100-01-01 with the steps 1 ± 0.5 months, and checks that optimised function `RebaseDateTime`.`rebaseGregorianToJulianMicros()` returns the same result as non-optimised one. The check is performed for the UTC, PST, CET, Africa/Dakar, America/Los_Angeles, Antarctica/Vostok, Asia/Hong_Kong, Europe/Amsterdam time zones.
- Added the test `optimization of micros rebasing - Julian to Gregorian` to `RebaseDateTimeSuite` which does similar checks as the test above but for rebasing from the hybrid calendar (Julian + Gregorian) to Proleptic Gregorian calendar.
- The tests for days rebasing are moved from `DateTimeUtilsSuite` to `RebaseDateTimeSuite` because the rebasing related code is moved from `DateTimeUtils` to the separate object `RebaseDateTime`.
- Re-run `DateTimeRebaseBenchmark` at the America/Los_Angeles time zone (it is set explicitly in the PR #28127):
| Item | Description |
| ---- | ----|
| Region | us-west-2 (Oregon) |
| Instance | r3.xlarge |
| AMI | ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1 (ami-06f2f779464715dc5) |
| Java | OpenJDK 64-Bit Server VM 1.8.0_242 and OpenJDK 64-Bit Server VM 11.0.6+10 |
Closes#28119 from MaxGekk/optimize-rebase-micros.
Authored-by: Max Gekk <max.gekk@gmail.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
### What changes were proposed in this pull request?
In the PR, I propose to set the `America/Los_Angeles` time zone in the date-time benchmarks `DateTimeBenchmark` and `DateTimeRebaseBenchmark` via `withDefaultTimeZone(LA)` and `withSQLConf(SQLConf.SESSION_LOCAL_TIMEZONE.key -> LA.getId)`.
The results of affected benchmarks was given on an Amazon EC2 instance w/ the configuration:
| Item | Description |
| ---- | ----|
| Region | us-west-2 (Oregon) |
| Instance | r3.xlarge |
| AMI | ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1 (ami-06f2f779464715dc5) |
| Java | OpenJDK8/11 |
### Why are the changes needed?
Performance of date-time functions can depend on the system JVM time zone or SQL config `spark.sql.session.timeZone`. The changes allow to avoid any fluctuations of benchmarks results related to time zones, and set a reliable baseline for future optimization.
### Does this PR introduce any user-facing change?
No
### How was this patch tested?
By regenerating results of DateTimeBenchmark and DateTimeRebaseBenchmark.
Closes#28127 from MaxGekk/set-timezone-in-benchmarks.
Authored-by: Max Gekk <max.gekk@gmail.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
### What changes were proposed in this pull request?
1. Fix the `rebaseGregorianToJulianMicros()` function in `DateTimeUtils` by passing the daylight saving offset associated with the input `micros` to the constructed instance of `GregorianCalendar`. The problem is in `cal.getTimeInMillis` which returns earliest instant in the case of local date-time overlaps, see https://github.com/AdoptOpenJDK/openjdk-jdk8u/blob/master/jdk/src/share/classes/java/util/GregorianCalendar.java#L2783-L2786 . I fixed the issue by keeping the standard zone offset as is, and set the DST offset only. I don't set `ZONE_OFFSET` because time zone resolution works differently in Java 8 and Java 7 time APIs. So, if I would set the standard zone offsets too, this could change the behavior, and rebasing won't give the same result as Spark 2.4.
2. Fix `rebaseJulianToGregorianMicros()` by changing resulted zoned date-time if `DST_OFFSET` is zero which means the input date-time has passed an autumn daylight savings cutover. So, I take the latest local timestamp out of 2 overlapped timestamps. Otherwise I return a zoned date-time w/o any modification because it is equal to calling the `withEarlierOffsetAtOverlap()` method, so, we can optimize the case.
### Why are the changes needed?
This fixes the bug of loosing of DST offset info in rebasing timestamps via local date-time. For example, there are 2 different timestamps in the `America/Los_Angeles` time zone: `2019-11-03T01:00:00-07:00` and `2019-11-03T01:00:00-08:00`, though they are mapped to the same local date-time `2019-11-03T01:00`, see
<img width="456" alt="Screen Shot 2020-04-02 at 10 19 24" src="https://user-images.githubusercontent.com/1580697/78245697-95a7da00-74f0-11ea-9eba-c08138851cb3.png">
Currently, the UTC timestamp `2019-11-03T09:00:00Z` is converted to `2019-11-03T01:00:00-08:00`, and then to `2019-11-03T01:00:00` (in the original calendar, for instance Proleptic Gregorian calendar) and back to the UTC timestamp `2019-11-03T08:00:00Z` (in the hybrid calendar - Gregorian for the timestamp). That's wrong because the local timestamp must be converted to the original timestamp `2019-11-03T09:00:00Z`.
### Does this PR introduce any user-facing change?
Yes
### How was this patch tested?
- Added a test to `DateTimeUtilsSuite` which checks that rebased micros are the same as the input during DST. The result must be the same if Java 8 and 7 time API functions return the same time zone offsets.
- Run the following code to check that there is no difference between rebased and original micros for modern timestamps:
```scala
test("rebasing differences") {
withDefaultTimeZone(getZoneId("America/Los_Angeles")) {
val start = instantToMicros(LocalDateTime.of(1, 1, 1, 0, 0, 0)
.atZone(getZoneId("America/Los_Angeles"))
.toInstant)
val end = instantToMicros(LocalDateTime.of(2030, 1, 1, 0, 0, 0)
.atZone(getZoneId("America/Los_Angeles"))
.toInstant)
var micros = start
var diff = Long.MaxValue
var counter = 0
while (micros < end) {
val rebased = rebaseGregorianToJulianMicros(micros)
val curDiff = rebased - micros
if (curDiff != diff) {
counter += 1
diff = curDiff
val ldt = microsToInstant(micros).atZone(getZoneId("America/Los_Angeles")).toLocalDateTime
println(s"local date-time = $ldt diff = ${diff / MICROS_PER_MINUTE} minutes")
}
micros += 30 * MICROS_PER_MINUTE
}
println(s"counter = $counter")
}
}
```
```
local date-time = 0001-01-01T00:00 diff = -2872 minutes
local date-time = 0100-03-01T00:00 diff = -1432 minutes
local date-time = 0200-03-01T00:00 diff = 7 minutes
local date-time = 0300-03-01T00:00 diff = 1447 minutes
local date-time = 0500-03-01T00:00 diff = 2887 minutes
local date-time = 0600-03-01T00:00 diff = 4327 minutes
local date-time = 0700-03-01T00:00 diff = 5767 minutes
local date-time = 0900-03-01T00:00 diff = 7207 minutes
local date-time = 1000-03-01T00:00 diff = 8647 minutes
local date-time = 1100-03-01T00:00 diff = 10087 minutes
local date-time = 1300-03-01T00:00 diff = 11527 minutes
local date-time = 1400-03-01T00:00 diff = 12967 minutes
local date-time = 1500-03-01T00:00 diff = 14407 minutes
local date-time = 1582-10-15T00:00 diff = 7 minutes
local date-time = 1883-11-18T12:22:58 diff = 0 minutes
counter = 15
```
The code is not added to `DateTimeUtilsSuite` because it takes > 30 seconds.
- By running the updated benchmark `DateTimeRebaseBenchmark` via the command:
```
SPARK_GENERATE_BENCHMARK_FILES=1 build/sbt "sql/test:runMain org.apache.spark.sql.execution.benchmark.DateTimeRebaseBenchmark"
```
in the environment:
| Item | Description |
| ---- | ----|
| Region | us-west-2 (Oregon) |
| Instance | r3.xlarge |
| AMI | ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1 (ami-06f2f779464715dc5) |
| Java | OpenJDK 1.8.0_242-8u242/11.0.6+10 |
Closes#28101 from MaxGekk/fix-local-date-overlap.
Lead-authored-by: Maxim Gekk <max.gekk@gmail.com>
Co-authored-by: Max Gekk <max.gekk@gmail.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
### What changes were proposed in this pull request?
In the PR, I propose to add new benchmarks to `DateTimeRebaseBenchmark` for saving and loading dates/timestamps to/from ORC files. I extracted common code from the benchmark for Parquet datasource and place it to the methods `caseName()` and `getPath()`. Added benchmarks for ORC save/load dates before and after 1582-10-15 because an implementation may have different performance for dates before the Julian calendar cutover day, see #28067 as an example.
### Why are the changes needed?
To have the base line for future optimizations of `fromJavaDate()`/`toJavaDate()` and `toJavaTimestamp()`/`fromJavaTimestamp()` in `DateTimeUtils`. The methods are used while saving/loading dates/timestamps by ORC datasource.
### Does this PR introduce any user-facing change?
No
### How was this patch tested?
By running the updated benchmark `DateTimeRebaseBenchmark` via the command:
```
SPARK_GENERATE_BENCHMARK_FILES=1 build/sbt "sql/test:runMain org.apache.spark.sql.execution.benchmark.DateTimeRebaseBenchmark"
```
in the environment:
| Item | Description |
| ---- | ----|
| Region | us-west-2 (Oregon) |
| Instance | r3.xlarge |
| AMI | ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1 (ami-06f2f779464715dc5) |
| Java | OpenJDK 1.8.0_242-8u242/11.0.6+10 |
Closes#28076 from MaxGekk/rebase-benchmark-orc.
Lead-authored-by: Max Gekk <max.gekk@gmail.com>
Co-authored-by: Maxim Gekk <max.gekk@gmail.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
### What changes were proposed in this pull request?
In the PR, I propose to replace current implementation of the `rebaseGregorianToJulianDays()` and `rebaseJulianToGregorianDays()` functions in `DateTimeUtils` by new one which is based on the fact that difference between Proleptic Gregorian and the hybrid (Julian+Gregorian) calendars was changed only 14 times for entire supported range of valid dates `[0001-01-01, 9999-12-31]`:
| date | Proleptic Greg. days | Hybrid (Julian+Greg) days | diff|
| ---- | ----|----|----|
|0001-01-01|-719162|-719164|-2|
|0100-03-01|-682944|-682945|-1|
|0200-03-01|-646420|-646420|0|
|0300-03-01|-609896|-609895|1|
|0500-03-01|-536847|-536845|2|
|0600-03-01|-500323|-500320|3|
|0700-03-01|-463799|-463795|4|
|0900-03-01|-390750|-390745|5|
|1000-03-01|-354226|-354220|6|
|1100-03-01|-317702|-317695|7|
|1300-03-01|-244653|-244645|8|
|1400-03-01|-208129|-208120|9|
|1500-03-01|-171605|-171595|10|
|1582-10-15|-141427|-141427|0|
For the given days since the epoch, the proposed implementation finds the range of days which the input days belongs to, and adds the diff in days between calendars to the input. The result is rebased days since the epoch in the target calendar.
For example, if need to rebase -650000 days from Proleptic Gregorian calendar to the hybrid calendar. In that case, the input falls to the bucket [-682944, -646420), the diff associated with the range is -1. To get the rebased days in Julian calendar, we should add -1 to -650000, and the result is -650001.
### Why are the changes needed?
To make dates rebasing faster.
### Does this PR introduce any user-facing change?
No, the results should be the same for valid range of the `DATE` type `[0001-01-01, 9999-12-31]`.
### How was this patch tested?
- Added 2 tests to `DateTimeUtilsSuite` for the `rebaseGregorianToJulianDays()` and `rebaseJulianToGregorianDays()` functions. The tests check that results of old and new implementation (optimized version) are the same for all supported dates.
- Re-run `DateTimeRebaseBenchmark` on:
| Item | Description |
| ---- | ----|
| Region | us-west-2 (Oregon) |
| Instance | r3.xlarge |
| AMI | ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1 (ami-06f2f779464715dc5) |
| Java | OpenJDK8/11 |
Closes#28067 from MaxGekk/optimize-rebasing.
Lead-authored-by: Maxim Gekk <max.gekk@gmail.com>
Co-authored-by: Max Gekk <max.gekk@gmail.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
### What changes were proposed in this pull request?
In the PR, I propose to add new benchmark `DateTimeRebaseBenchmark` which should measure the performance of rebasing of dates/timestamps from/to to the hybrid calendar (Julian+Gregorian) to/from Proleptic Gregorian calendar:
1. In write, it saves separately dates and timestamps before and after 1582 year w/ and w/o rebasing.
2. In read, it loads previously saved parquet files by vectorized reader and by regular reader.
Here is the summary of benchmarking:
- Saving timestamps is **~6 times slower**
- Loading timestamps w/ vectorized **off** is **~4 times slower**
- Loading timestamps w/ vectorized **on** is **~10 times slower**
### Why are the changes needed?
To know the impact of date-time rebasing introduced by #27915, #27953, #27807.
### Does this PR introduce any user-facing change?
No
### How was this patch tested?
Run the `DateTimeRebaseBenchmark` benchmark using Amazon EC2:
| Item | Description |
| ---- | ----|
| Region | us-west-2 (Oregon) |
| Instance | r3.xlarge |
| AMI | ubuntu/images/hvm-ssd/ubuntu-bionic-18.04-amd64-server-20190722.1 (ami-06f2f779464715dc5) |
| Java | OpenJDK8/11 |
Closes#28057 from MaxGekk/rebase-bechmark.
Lead-authored-by: Maxim Gekk <max.gekk@gmail.com>
Co-authored-by: Max Gekk <max.gekk@gmail.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
Wenchen Fan
Max Gekk
