### What changes were proposed in this pull request?
This pull request introduces a helper class that simplifies usage of `Dataset.observe()` for batch datasets:
val observation = Observation("name")
val observed = ds.observe(observation, max($"id").as("max_id"))
observed.count()
val metrics = observation.get
### Why are the changes needed?
Currently, users are required to implement the `QueryExecutionListener` interface to retrieve the metrics, as well as apply some knowledge on threading and locking to pull the metrics over to the main thread. With the helper class, metrics can be retrieved from batch dataset processing with three lines of code (the action on the observed dataset does not count as a line of code here).
### Does this PR introduce _any_ user-facing change?
Yes, one new class and one `Dataset`` method.
### How was this patch tested?
Adds a unit test to `DataFrameSuite`, similar to `"get observable metrics by callback"` in `DataFrameCallbackSuite`.
Closes#31905 from EnricoMi/branch-observation.
Authored-by: Enrico Minack <github@enrico.minack.dev>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
### What changes were proposed in this pull request?
This PR fixes two issues highlighted in https://issues.apache.org/jira/browse/SPARK-36163:
- JDBC connection provider propagates incorrect connection properties.
- Ambiguity when more than one JDBC connection provider is available.
I updated `BasicConnectionProvider` to use `jdbcOptions.asConnectionProperties` to remove JDBC data source specific options.
I also added `connectionProvider` data source option that specifies the name of the provider, e.g. `db2`, `presto`, to allow enforcing this specific provider in case of ambiguity.
### Why are the changes needed?
Users can leverage `spark.sql.sources.disabledJdbcConnProviderList` but it is cumbersome and requires them to disable all other providers which could be problematic when using ambiguous providers in two or more different JDBC queries.
### Does this PR introduce _any_ user-facing change?
Yes
PROBLEM DESCRIPTION:
This introduces new JDBC data source option `connectionProvider` that allows users to select a specific JDBC connection provider based on the short name. I updated the SQL guide doc and README.
Before this change, the only way to resolve ambiguity was SQL conf to blacklist all of the other JDBC connection providers. After this change users will be able to specify the exact connection provider they need per data source.
### How was this patch tested?
I updated the existing `ConnectionProviderSuite` and added a new `BasicConnectionProviderSuite`.
Closes#33370 from sadikovi/fix-jdbc-conn-provider.
Authored-by: Ivan Sadikov <ivan.sadikov@databricks.com>
Signed-off-by: Hyukjin Kwon <gurwls223@apache.org>
### What changes were proposed in this pull request?
RemoveRedundantAliases may change DataWritingCommand's parameter's attribute name.
In the UT's case before RemoveRedundantAliases the partitionColumns is `CAL_DT`, and change by RemoveRedundantAliases and change to `cal_dt` then case the error case
### Why are the changes needed?
Fix bug
### Does this PR introduce _any_ user-facing change?
For below SQL case
```
sql("create table t1(cal_dt date) using parquet")
sql("insert into t1 values (date'2021-06-27'),(date'2021-06-28'),(date'2021-06-29'),(date'2021-06-30')")
sql("create view t1_v as select * from t1")
sql("CREATE TABLE t2 USING PARQUET PARTITIONED BY (CAL_DT) AS SELECT 1 AS FLAG,CAL_DT FROM t1_v WHERE CAL_DT BETWEEN '2021-06-27' AND '2021-06-28'")
sql("INSERT INTO t2 SELECT 2 AS FLAG,CAL_DT FROM t1_v WHERE CAL_DT BETWEEN '2021-06-29' AND '2021-06-30'")
```
Before this pr
```
sql("SELECT * FROM t2 WHERE CAL_DT BETWEEN '2021-06-29' AND '2021-06-30'").show
+----+------+
|FLAG|CAL_DT|
+----+------+
+----+------+
sql("SELECT * FROM t2 ").show
+----+----------+
|FLAG| CAL_DT|
+----+----------+
| 1|2021-06-27|
| 1|2021-06-28|
+----+----------+
```
After this pr
```
sql("SELECT * FROM t2 WHERE CAL_DT BETWEEN '2021-06-29' AND '2021-06-30'").show
+----+------+
|FLAG|CAL_DT|
+----+------+
| 2|2021-06-29|
| 2|2021-06-30|
+----+------+
sql("SELECT * FROM t2 ").show
+----+----------+
|FLAG| CAL_DT|
+----+----------+
| 1|2021-06-27|
| 1|2021-06-28|
| 2|2021-06-29|
| 2|2021-06-30|
+----+----------+
```
### How was this patch tested?
Added UT
Closes#33324 from AngersZhuuuu/SPARK-36093.
Authored-by: Angerszhuuuu <angers.zhu@gmail.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
### What changes were proposed in this pull request?
A hive partition can have different `PartitionDesc`s from `TableDesc` for describing Serde/InputFormatClass/OutputFormatClass, for a hive partitioned table, we shall respect those in `PartitionDesc`.
### Why are the changes needed?
in many cases, that Spark reads hive tables could result in surprise because of this issue.
### Does this PR introduce _any_ user-facing change?
yes, hive partition table that contains different serde/input/output could be recognized by Spark
### How was this patch tested?
new test added
Closes#33406 from yaooqinn/SPARK-36197.
Authored-by: Kent Yao <yao@apache.org>
Signed-off-by: Kent Yao <yao@apache.org>
### What changes were proposed in this pull request?
Currently, two AQE rules `OptimizeLocalShuffleReader` and `OptimizeSkewedJoin` run `EnsureRequirements` at the end to check if there are extra shuffles in the optimized plan and revert the optimization if extra shuffles are introduced.
This PR proposes to run `ValidateRequirements` instead, which is much simpler than `EnsureRequirements`. This PR also moves this check to `AdaptiveSparkPlanExec`, so that it's centralized instead of in each rule. After centralization, the batch name of optimizing the final stage is the same as normal stages, which makes more sense.
### Why are the changes needed?
`EnsureRequirements` is a big rule and even contains optimizations (remove unnecessary shuffles). `ValidateRequirements` is much faster to run and can avoid potential bugs as it has no optimization and is a pure check.
### Does this PR introduce _any_ user-facing change?
no
### How was this patch tested?
existing tests.
Closes#33396 from cloud-fan/aqe.
Authored-by: Wenchen Fan <wenchen@databricks.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
### What changes were proposed in this pull request?
This PR fixes typo for `spark.sql.legacy.allowCreatingManagedTableUsingNonemptyLocation` in `SQLConf.scala`.
### Why are the changes needed?
This is a [Broken windows theory](https://en.wikipedia.org/wiki/Broken_windows_theory) change.
### Does this PR introduce _any_ user-facing change?
Yes, after merging this PR, the error message for commands such as
```python
spark.conf.set("spark.sql.legacy.allowCreatingManagedTableUsingNonemptyLocation", "true")
```
, users will get a typo-free exception.
### How was this patch tested?
This is a trivial change.
Closes#33389 from bessbd/patch-1.
Authored-by: Bessenyei Balázs Donát <9086834+bessbd@users.noreply.github.com>
Signed-off-by: Sean Owen <srowen@gmail.com>
### What changes were proposed in this pull request?
The current implement of `Sequence` accept `TimestampType`, `DateType` and `IntegralType`. This PR will let `Sequence` accepts `TimestampNTZType`.
### Why are the changes needed?
We can generate sequence for timestamp without time zone.
### Does this PR introduce _any_ user-facing change?
'Yes'.
This PR will let `Sequence` accepts `TimestampNTZType`.
### How was this patch tested?
New tests.
Closes#33360 from beliefer/SPARK-36090.
Lead-authored-by: gengjiaan <gengjiaan@360.cn>
Co-authored-by: Jiaan Geng <beliefer@163.com>
Signed-off-by: Max Gekk <max.gekk@gmail.com>
### What changes were proposed in this pull request?
This PR changes the behavior of the quoted interval literals like `SELECT INTERVAL '1 year 2 month'` to be converted to ANSI interval types.
### Why are the changes needed?
The tnit-to-unit interval literals and the unit list interval literals are converted to ANSI interval types but quoted interval literals are still converted to CalendarIntervalType.
```
-- Unit list interval literals
spark-sql> select interval 1 year 2 month;
1-2
-- Quoted interval literals
spark-sql> select interval '1 year 2 month';
1 years 2 months
```
### Does this PR introduce _any_ user-facing change?
Yes but the following sentence in `sql-migration-guide.md` seems to cover this change.
```
- In Spark 3.2, the unit list interval literals can not mix year-month fields (YEAR and MONTH) and day-time fields (WEEK, DAY, ..., MICROSECOND).
For example, `INTERVAL 1 day 1 hour` is invalid in Spark 3.2. In Spark 3.1 and earlier,
there is no such limitation and the literal returns value of `CalendarIntervalType`.
To restore the behavior before Spark 3.2, you can set `spark.sql.legacy.interval.enabled` to `true`.
```
### How was this patch tested?
Modified existing tests and add new tests.
Closes#33380 from sarutak/fix-interval-constructor.
Authored-by: Kousuke Saruta <sarutak@oss.nttdata.com>
Signed-off-by: Max Gekk <max.gekk@gmail.com>
### What changes were proposed in this pull request?
This patch removes an ignored test from `RocksDBSuite`.
### Why are the changes needed?
The removed test is now ignored. The test itself doesn't look making sense. For example, the condition for capturing exception is never matched. The test runs updates to RocksDB instances at same remote dir with same versions. This doesn't look like a case it will run through in practice.
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
Existing tests.
Closes#33401 from viirya/remove-ignore-test.
Authored-by: Liang-Chi Hsieh <viirya@gmail.com>
Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
### What changes were proposed in this pull request?
In `CatalogFileIndex.filterPartitions`, check the config `spark.sql.hive.metastorePartitionPruning` and don't pushdown predicates to remote HMS if it is false. Instead, fallback to the `listPartitions` API and do the filtering on the client side.
### Why are the changes needed?
Currently the config `spark.sql.hive.metastorePartitionPruning` is only effective for Hive tables, and for non-Hive tables we'd always use the `listPartitionsByFilter` API from HMS client. On the other hand, by default all data source tables also manage their partitions through HMS, when the config `spark.sql.hive.manageFilesourcePartitions` is turned on. Therefore, it seems reasonable to extend the above config for non-Hive tables as well.
In certain cases the remote HMS service could throw exceptions when using the `listPartitionsByFilter` API, which, on the Spark side, is unrecoverable at the current state. Therefore it would be better to allow users to disable the API by using the above config.
For instance, HMS only allow pushdown date column when direct SQL is used instead of JDO for interacting with the underlying RDBMS, and will throw exception otherwise. Even though the Spark Hive client will attempt to recover itself when the exception happens, it only does so when the config `hive.metastore.try.direct.sql` from remote HMS is `false`. There could be cases where the value of `hive.metastore.try.direct.sql` is true but remote HMS still throws exception.
### Does this PR introduce _any_ user-facing change?
Yes now the config `spark.sql.hive.metastorePartitionPruning` is extended for non-Hive tables which use HMS to manage their partition metadata.
### How was this patch tested?
Added a new unit test:
```
build/sbt "hive/testOnly *PruneFileSourcePartitionsSuite -- -z SPARK-36128"
```
Closes#33348 from sunchao/SPARK-36128-by-filter.
Authored-by: Chao Sun <sunchao@apple.com>
Signed-off-by: Liang-Chi Hsieh <viirya@gmail.com>
Introduction: this PR is the last part of SPARK-10816 (EventTime based sessionization (session window)). Please refer #31937 to see the overall view of the code change. (Note that code diff could be diverged a bit.)
### What changes were proposed in this pull request?
This PR proposes to support native session window. Please refer the comments/design doc in SPARK-10816 for more details on the rationalization and design (could be outdated a bit compared to the PR).
The definition of the boundary of "session window" is [the timestamp of start event ~ the timestamp of last event + gap duration). That said, unlike time window, session window is a dynamic window which can expand if new input row is added to the session. To handle expansion of session window, Spark defines session window per input row, and "merge" windows if they can be merged (boundaries are overlapped).
This PR leverages two different approaches on merging session windows:
1. merging session windows with Spark's aggregation logic (a variant of sort aggregation)
2. updating session window for all rows bound to the same session, and applying aggregation logic afterwards
First one is preferable as it outperforms compared to the second one, though it can be only used if merging session window can be applied altogether with aggregation. It is not applicable on all the cases, so second one is used to cover the remaining cases.
This PR also applies the optimization on merging input rows and existing sessions with retaining the order (group keys + start timestamp of session window), leveraging the fact the number of existing sessions per group key won't be huge.
The state format is versioned, so that we can bring a new state format if we find a better one.
### Why are the changes needed?
For now, to deal with sessionization, Spark requires end users to play with (flat)MapGroupsWithState directly which has a couple of major drawbacks:
1. (flat)MapGroupsWithState is lower level API and end users have to code everything in details for defining session window and merging windows
2. built-in aggregate functions cannot be used and end users have to deal with aggregation by themselves
3. (flat)MapGroupsWithState is only available in Scala/Java.
With native support of session window, end users simply use "session_window" like they use "window" for tumbling/sliding window, and leverage built-in aggregate functions as well as UDAFs to simply define aggregations.
Quoting the query example from test suite:
```
val inputData = MemoryStream[(String, Long)]
// Split the lines into words, treat words as sessionId of events
val events = inputData.toDF()
.select($"_1".as("value"), $"_2".as("timestamp"))
.withColumn("eventTime", $"timestamp".cast("timestamp"))
.selectExpr("explode(split(value, ' ')) AS sessionId", "eventTime")
.withWatermark("eventTime", "30 seconds")
val sessionUpdates = events
.groupBy(session_window($"eventTime", "10 seconds") as 'session, 'sessionId)
.agg(count("*").as("numEvents"))
.selectExpr("sessionId", "CAST(session.start AS LONG)", "CAST(session.end AS LONG)",
"CAST(session.end AS LONG) - CAST(session.start AS LONG) AS durationMs",
"numEvents")
```
which is same as StructuredSessionization (native session window is shorter and clearer even ignoring model classes).
39542bb81f/examples/src/main/scala/org/apache/spark/examples/sql/streaming/StructuredSessionization.scala (L66-L105)
(Worth noting that the code in StructuredSessionization only works with processing time. The code doesn't consider old event can update the start time of old session.)
### Does this PR introduce _any_ user-facing change?
Yes. This PR brings the new feature to support session window on both batch and streaming query, which adds a new function "session_window" which usage is similar with "window".
### How was this patch tested?
New test suites. Also tested with benchmark code.
Closes#33081 from HeartSaVioR/SPARK-34893-SPARK-10816-PR-31570-part-5.
Lead-authored-by: Jungtaek Lim <kabhwan.opensource@gmail.com>
Co-authored-by: Liang-Chi Hsieh <viirya@gmail.com>
Co-authored-by: Yuanjian Li <yuanjian.li@databricks.com>
Signed-off-by: Jungtaek Lim <kabhwan.opensource@gmail.com>
### What changes were proposed in this pull request?
This PR add the DPP + AQE support when spark can't reuse the broadcast but executing the DPP subquery is cheaper.
### Why are the changes needed?
Improve AQE + DPP
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
Adding new ut
Closes#32861 from JkSelf/supportDPP3.
Lead-authored-by: Ke Jia <ke.a.jia@intel.com>
Co-authored-by: Wenchen Fan <cloud0fan@gmail.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
this commit makes sure that for File Source V2 partition filters are
also taken into account when the readDataSchema is empty.
This is the case for queries like:
SELECT count(*) FROM tbl WHERE partition=foo
SELECT input_file_name() FROM tbl WHERE partition=foo
### What changes were proposed in this pull request?
As described in SPARK-35985 there is bug in the File Datasource V2 which prevents it to push down to the FileScanner for queries like the ones listed above.
### Why are the changes needed?
If partitions filters are not pushed down, the whole dataset will be scanned while only one partition is interesting.
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
An extra test was added which relies on the output of explain, as is done in other places.
Closes#33191 from steven-aerts/SPARK-35985.
Authored-by: Steven Aerts <steven.aerts@airties.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
### What changes were proposed in this pull request?
This PR proposes to move `spark.sql.sources.disabledJdbcConnProviderList` from SQLConf to StaticSQLConf which disallows to set in runtime.
### Why are the changes needed?
It's documented as a static configuration. we should make it as a static configuration properly.
### Does this PR introduce _any_ user-facing change?
Previously, the configuration can be set to different value but not effective.
Now it throws an exception if users try to set in runtime.
### How was this patch tested?
Existing unittest was fixed. That should verify the change.
Closes#33381 from HyukjinKwon/SPARK-36169.
Authored-by: Hyukjin Kwon <gurwls223@apache.org>
Signed-off-by: Hyukjin Kwon <gurwls223@apache.org>
### What changes were proposed in this pull request?
In the PR, I propose to propagate either the SQL config `spark.sql.parquet.datetimeRebaseModeInRead` or/and Parquet option `datetimeRebaseMode` to `ParquetFilters`. The `ParquetFilters` class uses the settings in conversions of dates/timestamps instances from datasource filters to values pushed via `FilterApi` to the `parquet-column` lib.
Before the changes, date/timestamp values expressed as days/microseconds/milliseconds are interpreted as offsets in Proleptic Gregorian calendar, and pushed to the parquet library as is. That works fine if timestamp/dates values in parquet files were saved in the `CORRECTED` mode but in the `LEGACY` mode, filter's values could not match to actual values.
After the changes, timestamp/dates values of filters pushed down to parquet libs such as `FilterApi.eq(col1, -719162)` are rebased according the rebase settings. For the example, if the rebase mode is `CORRECTED`, **-719162** is pushed down as is but if the current rebase mode is `LEGACY`, the number of days is rebased to **-719164**. For more context, the PR description https://github.com/apache/spark/pull/28067 shows the diffs between two calendars.
### Why are the changes needed?
The changes fix the bug portrayed by the following example from SPARK-36034:
```scala
In [27]: spark.conf.set("spark.sql.legacy.parquet.datetimeRebaseModeInWrite", "LEGACY")
>>> spark.sql("SELECT DATE '0001-01-01' AS date").write.mode("overwrite").parquet("date_written_by_spark3_legacy")
>>> spark.read.parquet("date_written_by_spark3_legacy").where("date = '0001-01-01'").show()
+----+
|date|
+----+
+----+
```
The result must have the date value `0001-01-01`.
### Does this PR introduce _any_ user-facing change?
In some sense, yes. Query results can be different in some cases. For the example above:
```scala
scala> spark.conf.set("spark.sql.parquet.datetimeRebaseModeInWrite", "LEGACY")
scala> spark.sql("SELECT DATE '0001-01-01' AS date").write.mode("overwrite").parquet("date_written_by_spark3_legacy")
scala> spark.read.parquet("date_written_by_spark3_legacy").where("date = '0001-01-01'").show(false)
+----------+
|date |
+----------+
|0001-01-01|
+----------+
```
### How was this patch tested?
By running the modified test suite `ParquetFilterSuite`:
```
$ build/sbt "test:testOnly *ParquetV1FilterSuite"
$ build/sbt "test:testOnly *ParquetV2FilterSuite"
```
Closes#33347 from MaxGekk/fix-parquet-ts-filter-pushdown.
Authored-by: Max Gekk <max.gekk@gmail.com>
Signed-off-by: Max Gekk <max.gekk@gmail.com>
### What changes were proposed in this pull request?
Support TimestampNTZ type in file partitioning
* When there is no provided schema and the default Timestamp type is TimestampNTZ , Spark should infer and parse the timestamp value partitions as TimestampNTZ.
* When the provided Partition schema is TimestampNTZ, Spark should be able to parse the TimestampNTZ type partition column.
### Why are the changes needed?
File partitioning is an important feature and Spark should support TimestampNTZ type in it.
### Does this PR introduce _any_ user-facing change?
Yes, Spark supports TimestampNTZ type in file partitioning
### How was this patch tested?
Unit tests
Closes#33344 from gengliangwang/partition.
Authored-by: Gengliang Wang <gengliang@apache.org>
Signed-off-by: Gengliang Wang <gengliang@apache.org>
### What changes were proposed in this pull request?
This PR proposes to change the application of the operators for TimeWindow, from project -> filter, to filter -> project.
Currently Spark applies project, and filter, while filter is not dependent on project. That said, if the input rows are going to be filtered out via filter predicate, applying projection on these input rows are simply waste of time.
### Why are the changes needed?
This is a simple improvement requiring changes from a couple of lines.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
Existing tests.
Closes#33367 from HeartSaVioR/SPARK-36157.
Authored-by: Jungtaek Lim <kabhwan.opensource@gmail.com>
Signed-off-by: Liang-Chi Hsieh <viirya@gmail.com>
### What changes were proposed in this pull request?
This pr fix Parquet filter pushdown `NOT` `IN` predicate if its values exceeds `spark.sql.parquet.pushdown.inFilterThreshold`. For example: `Not(In(a, Array(2, 3, 7))`. We can not push down `not(and(gteq(a, 2), lteq(a, 7)))`.
### Why are the changes needed?
Fix bug.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
Unit test.
Closes#33365 from wangyum/SPARK-32792-3.
Authored-by: Yuming Wang <yumwang@ebay.com>
Signed-off-by: Max Gekk <max.gekk@gmail.com>
### What changes were proposed in this pull request?
Keep the output order is deterministic for `SHOW TBLPROPERTIES`
### Why are the changes needed?
[#33343](https://github.com/apache/spark/pull/33343#issue-689828187).
Keep the output order deterministic meaningful.
Since the properties are sorted and then compare result in the testcase for `SHOW TBLPROPERTIES`, it does not fail, but ideally, the output is ordered and deterministic.
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
existed ut test
Closes#33353 from Peng-Lei/order-ouput-properties.
Authored-by: PengLei <peng.8lei@gmail.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
### What changes were proposed in this pull request?
This PR fixes a behavior of `SHOW CREATE TABLE` added in `SPARK-33898` (#32931) to output deterministic result.
A test `SPARK-33898: SHOW CREATE TABLE` in `DataSourceV2SQLSuite` compares two `CREATE TABLE` statements. One is generated by `SHOW CREATE TABLE` against a created table and the other is expected `CREATE TABLE` statement.
The created table has options `from` and `to`, and they are declared in this order.
```
CREATE TABLE $t (
a bigint NOT NULL,
b bigint,
c bigint,
`extra col` ARRAY<INT>,
`<another>` STRUCT<x: INT, y: ARRAY<BOOLEAN>>
)
USING foo
OPTIONS (
from = 0,
to = 1)
COMMENT 'This is a comment'
TBLPROPERTIES ('prop1' = '1')
PARTITIONED BY (a)
LOCATION '/tmp'
```
And the expected `CREATE TABLE` in the test code is like as follows.
```
"CREATE TABLE testcat.ns1.ns2.tbl (",
"`a` BIGINT NOT NULL,",
"`b` BIGINT,",
"`c` BIGINT,",
"`extra col` ARRAY<INT>,",
"`<another>` STRUCT<`x`: INT, `y`: ARRAY<BOOLEAN>>)",
"USING foo",
"OPTIONS(",
"'from' = '0',",
"'to' = '1')",
"PARTITIONED BY (a)",
"COMMENT 'This is a comment'",
"LOCATION '/tmp'",
"TBLPROPERTIES(",
"'prop1' = '1')"
```
As you can see, the order of `from` and `to` is expected.
But options are implemented as `Map` so the order of key cannot be kept.
In fact, this test fails with Scala 2.13.
```
[info] - SPARK-33898: SHOW CREATE TABLE *** FAILED *** (515 milliseconds)
[info] Array("CREATE TABLE testcat.ns1.ns2.tbl (", "`a` BIGINT NOT NULL,", "`b` BIGINT,", "`c` BIGINT,", "`extra col` ARRAY<INT>,", "`<another>` STRUCT<`x`: INT, `y`: ARRAY<BOOLEAN>>)", "USING foo", "OPTIONS(", "'to' = '1',", "'from' = '0')", "PARTITIONED BY (a)", "COMMENT 'This is a comment'", "LOCATION '/tmp'", "TBLPROPERTIES(", "'prop1' = '1')") did not equal Array("CREATE TABLE testcat.ns1.ns2.tbl (", "`a` BIGINT NOT NULL,", "`b` BIGINT,", "`c` BIGINT,", "`extra col` ARRAY<INT>,", "`<another>` STRUCT<`x`: INT, `y`: ARRAY<BOOLEAN>>)", "USING foo", "OPTIONS(", "'from' = '0',", "'to' = '1')", "PARTITIONED BY (a)", "COMMENT 'This is a comment'", "LOCATION '/tmp'", "TBLPROPERTIES(", "'prop1' = '1')") (DataSourceV2SQLSuite.scala:1997)
```
In the current master, the test doesn't fail with Scala 2.12 but it's still non-deterministic.
### Why are the changes needed?
Bug fix.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
I confirmed that the modified test passed with both Scala 2.12 and Scala 2.13 with this change.
Closes#33343 from sarutak/fix-show-create-table-test.
Authored-by: Kousuke Saruta <sarutak@oss.nttdata.com>
Signed-off-by: Hyukjin Kwon <gurwls223@apache.org>
### What changes were proposed in this pull request?
`RegExpReplace` overrides `checkInputDataTypes` but doesn't do the basic type check.
This PR adds the type check so that the error message is more readable.
### Why are the changes needed?
bugfix
### Does this PR introduce _any_ user-facing change?
no
### How was this patch tested?
newly added test case
Closes#33357 from linhongliu-db/SPARK-36148-regexp-replace-check.
Authored-by: Linhong Liu <linhong.liu@databricks.com>
Signed-off-by: Max Gekk <max.gekk@gmail.com>
### What changes were proposed in this pull request?
Only use the zone ids that has no daylight saving for testing `localtimestamp`
### Why are the changes needed?
https://github.com/apache/spark/pull/33346#discussion_r670135296 MaxGekk suggests that we should avoid wrong results if possible.
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
Unit test
Closes#33354 from gengliangwang/FIxDST.
Authored-by: Gengliang Wang <gengliang@apache.org>
Signed-off-by: Max Gekk <max.gekk@gmail.com>
### What changes were proposed in this pull request?
The threshold of the test case "datetime function localtimestamp" is small, which leads to flaky test results
https://github.com/gengliangwang/spark/runs/3067396143?check_suite_focus=true
This PR is to increase the threshold for checking two the different current local datetimes from 5ms to 1 second. (The test case of current_timestamp uses 5 seconds)
### Why are the changes needed?
Fix flaky test
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
Unit test
Closes#33346 from gengliangwang/fixFlaky.
Authored-by: Gengliang Wang <gengliang@apache.org>
Signed-off-by: Gengliang Wang <gengliang@apache.org>
### What changes were proposed in this pull request?
Adds error classes to some of the exceptions in QueryCompilationErrors.
### Why are the changes needed?
Improves auditing for developers and adds useful fields for users (error class and SQLSTATE).
### Does this PR introduce _any_ user-facing change?
Yes, fills in missing error class and SQLSTATE fields.
### How was this patch tested?
Existing tests and new unit tests.
Closes#33309 from karenfeng/group-compilation-errors-1.
Authored-by: Karen Feng <karen.feng@databricks.com>
Signed-off-by: Hyukjin Kwon <gurwls223@apache.org>
### What changes were proposed in this pull request?
spark function from_json output field name, field type and field value when FAILFAST mode throw exception.
### Why are the changes needed?
This infoormation is very important for devlops to find where error input data is located.
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
org/apache/spark/sql/JsonFunctionsSuite.scala:598
test("[SPARK-36069] from_json invalid json schema - check field name and field value")
Closes#33297 from geekyouth/feature/FAILFAST_output_fidelaName_fieldValue_dataType.
Lead-authored-by: Geek <forsupergeeker@gmail.com>
Co-authored-by: 极客青年 <forsupergeeker@gmail.com>
Signed-off-by: Max Gekk <max.gekk@gmail.com>
### What changes were proposed in this pull request?
Add `CoalescedPartitionSpec(0, 0, _)` check if a `CoalescedPartitionSpec` is coalesced.
### Why are the changes needed?
Fix corner case.
### Does this PR introduce _any_ user-facing change?
yes, UI may be changed
### How was this patch tested?
Add test
Closes#33342 from ulysses-you/SPARK-35639-FOLLOW.
Authored-by: ulysses-you <ulyssesyou18@gmail.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
### What changes were proposed in this pull request?
Fix the skipping values logic in Parquet vectorized reader when column index is effective, by considering nulls and only call `ParquetVectorUpdater.skipValues` when the values are non-null.
### Why are the changes needed?
Currently, the Parquet vectorized reader may not work correctly if column index filtering is effective, and the data page contains null values. For instance, let's say we have two columns `c1: BIGINT` and `c2: STRING`, and the following pages:
```
* c1 500 500 500 500
* |---------|---------|---------|---------|
* |-------|-----|-----|---|---|---|---|---|
* c2 400 300 300 200 200 200 200 200
```
and suppose we have a query like the following:
```sql
SELECT * FROM t WHERE c1 = 500
```
this will create a Parquet row range `[500, 1000)` which, when applied to `c2`, will require us to skip all the rows in `[400,500)`. However the current logic for skipping rows is via `updater.skipValues(n, valueReader)` which is incorrect since this skips the next `n` non-null values. In the case when nulls are present, this will not work correctly.
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
Added a new test in `ParquetColumnIndexSuite`.
Closes#33330 from sunchao/SPARK-36123-skip-nulls.
Authored-by: Chao Sun <sunchao@apple.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
### What changes were proposed in this pull request?
DATE/TIMESTAMP literals support years 0000 to 9999. However, internally we support a range that is much larger.
We can add or subtract large intervals from a date/timestamp and the system will happily process and display large negative and positive dates.
Since we obviously cannot put this genie back into the bottle the only thing we can do is allow matching DATE/TIMESTAMP literals.
### Why are the changes needed?
make spark more usable and bug fix
### Does this PR introduce _any_ user-facing change?
Yes, after this PR, below SQL will have different results
```sql
select cast('-10000-1-2' as date) as date_col
-- before PR: NULL
-- after PR: -10000-1-2
```
```sql
select cast('2021-4294967297-11' as date) as date_col
-- before PR: 2021-01-11
-- after PR: NULL
```
### How was this patch tested?
newly added test cases
Closes#32959 from linhongliu-db/SPARK-35780-full-range-datetime.
Lead-authored-by: Linhong Liu <linhong.liu@databricks.com>
Co-authored-by: Linhong Liu <67896261+linhongliu-db@users.noreply.github.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
Introduction: this PR is a part of SPARK-10816 (EventTime based sessionization (session window)). Please refer #31937 to see the overall view of the code change. (Note that code diff could be diverged a bit.)
### What changes were proposed in this pull request?
This PR introduces MergingSortWithSessionWindowStateIterator, which does "merge sort" between input rows and sessions in state based on group key and session's start time.
Note that the iterator does merge sort among input rows and sessions grouped by grouping key. The iterator doesn't provide sessions in state which keys don't exist in input rows. For input rows, the iterator will provide all rows regardless of the existence of matching sessions in state.
MergingSortWithSessionWindowStateIterator works on the precondition that given iterator is sorted by "group keys + start time of session window", and the iterator still retains the characteristic of the sort.
### Why are the changes needed?
This part is a one of required on implementing SPARK-10816.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
New UT added.
Closes#33077 from HeartSaVioR/SPARK-34892-SPARK-10816-PR-31570-part-4.
Authored-by: Jungtaek Lim <kabhwan.opensource@gmail.com>
Signed-off-by: Jungtaek Lim <kabhwan.opensource@gmail.com>
### What changes were proposed in this pull request?
Fix [comment](https://github.com/apache/spark/pull/32488#issuecomment-879315179)
This PR fix rule `UnwrapCastInBinaryComparison` bug. Rule UnwrapCastInBinaryComparison should skip In expression when in.list contains an expression that is not literal.
- In
Before this pr, the following example will throw an exception.
```scala
withTable("tbl") {
sql("CREATE TABLE tbl (d decimal(33, 27)) USING PARQUET")
sql("SELECT d FROM tbl WHERE d NOT IN (d + 1)")
}
```
- InSet
As the analyzer guarantee that all the elements in the `inSet.hset` are literal, so this is not an issue for `InSet`.
fbf53dee37/sql/catalyst/src/main/scala/org/apache/spark/sql/catalyst/optimizer/expressions.scala (L264-L279)
### Does this PR introduce _any_ user-facing change?
No, only bug fix.
### How was this patch tested?
New test.
Closes#33335 from cfmcgrady/SPARK-36130.
Authored-by: Fu Chen <cfmcgrady@gmail.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
### What changes were proposed in this pull request?
Fix `CustomShuffleReaderExec.hasCoalescedPartition` so that it returns true only if some original partitions got combined
### Why are the changes needed?
W/o this change `CustomShuffleReaderExec` description can report `coalesced` even though partitions are unchanged
### Does this PR introduce _any_ user-facing change?
Yes, the `Arguments` in the node description is now accurate:
```
(16) CustomShuffleReader
Input [3]: [registration#4, sum#85, count#86L]
Arguments: coalesced
```
### How was this patch tested?
Existing tests
Closes#32872 from ekoifman/PRISM-77023-fix-hasCoalescedPartition.
Authored-by: Eugene Koifman <eugene.koifman@workday.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
### What changes were proposed in this pull request?
`LOCALTIMESTAMP()` is a datetime value function from ANSI SQL.
The syntax show below:
```
<datetime value function> ::=
<current date value function>
| <current time value function>
| <current timestamp value function>
| <current local time value function>
| <current local timestamp value function>
<current date value function> ::=
CURRENT_DATE
<current time value function> ::=
CURRENT_TIME [ <left paren> <time precision> <right paren> ]
<current local time value function> ::=
LOCALTIME [ <left paren> <time precision> <right paren> ]
<current timestamp value function> ::=
CURRENT_TIMESTAMP [ <left paren> <timestamp precision> <right paren> ]
<current local timestamp value function> ::=
LOCALTIMESTAMP [ <left paren> <timestamp precision> <right paren> ]
```
`LOCALTIMESTAMP()` returns the current timestamp at the start of query evaluation as TIMESTAMP WITH OUT TIME ZONE. This is similar to `CURRENT_TIMESTAMP()`.
Note we need to update the optimization rule `ComputeCurrentTime` so that Spark returns the same result in a single query if the function is called multiple times.
### Why are the changes needed?
`CURRENT_TIMESTAMP()` returns the current timestamp at the start of query evaluation.
`LOCALTIMESTAMP()` returns the current timestamp without time zone at the start of query evaluation.
The `LOCALTIMESTAMP` function is an ANSI SQL.
The `LOCALTIMESTAMP` function is very useful.
### Does this PR introduce _any_ user-facing change?
'Yes'. Support new function `LOCALTIMESTAMP()`.
### How was this patch tested?
New tests.
Closes#33258 from beliefer/SPARK-36037.
Lead-authored-by: gengjiaan <gengjiaan@360.cn>
Co-authored-by: Jiaan Geng <beliefer@163.com>
Co-authored-by: Wenchen Fan <cloud0fan@gmail.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
### What changes were proposed in this pull request?
Refactor `ParquetColumnIndexSuite` and allow better code reuse.
### Why are the changes needed?
A few methods in the test suite can share the same utility method `checkUnalignedPages` so it's better to do that and remove code duplication.
Additionally, `parquet.enable.dictionary` is tested for both `true` and `false` combination.
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
Existing tests.
Closes#33334 from sunchao/SPARK-35743-test-refactoring.
Authored-by: Chao Sun <sunchao@apple.com>
Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
Introduction: this PR is a part of SPARK-10816 (`EventTime based sessionization (session window)`). Please refer #31937 to see the overall view of the code change. (Note that code diff could be diverged a bit.)
### What changes were proposed in this pull request?
This PR introduces state store manager for session window in streaming query. Session window in batch query wouldn't need to leverage state store manager.
This PR ensures versioning on state format for state store manager, so that we can apply further optimization after releasing Spark version. StreamingSessionWindowStateManager is a trait defining the available methods in session window state store manager. Its subclasses are classes implementing the trait with versioning.
The format of version 1 leverages the new feature of "prefix match scan" to represent the session windows:
* full key : [ group keys, start time in session window ]
* prefix key [ group keys ]
### Why are the changes needed?
This part is a one of required on implementing SPARK-10816.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
New test suite added
Closes#31989 from HeartSaVioR/SPARK-34891-SPARK-10816-PR-31570-part-3.
Authored-by: Jungtaek Lim <kabhwan.opensource@gmail.com>
Signed-off-by: Liang-Chi Hsieh <viirya@gmail.com>
### What changes were proposed in this pull request?
Support TimestampNTZ type column in SQL command Cache table
### Why are the changes needed?
Cache table should support the new timestamp type.
### Does this PR introduce _any_ user-facing change?
Yes, the TimemstampNTZ type column can used in `CACHE TABLE`
### How was this patch tested?
Unit test
Closes#33322 from gengliangwang/cacheTable.
Authored-by: Gengliang Wang <gengliang@apache.org>
Signed-off-by: Max Gekk <max.gekk@gmail.com>
### What changes were proposed in this pull request?
Make sure all physical plans of TPCDS queries are valid (satisfy the partitioning requirement).
### Why are the changes needed?
improve test coverage
### Does this PR introduce _any_ user-facing change?
no
### How was this patch tested?
N/A
Closes#33248 from cloud-fan/aqe2.
Lead-authored-by: Wenchen Fan <cloud0fan@gmail.com>
Co-authored-by: Wenchen Fan <wenchen@databricks.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
### What changes were proposed in this pull request?
This PR adds an INVALID_FIELD_NAME error class for the errors in `StructType.findNestedField`. It also cleans up the code there and adds UT for this method.
### Why are the changes needed?
follow the new error message framework
### Does this PR introduce _any_ user-facing change?
no
### How was this patch tested?
existing tests
Closes#33282 from cloud-fan/error.
Authored-by: Wenchen Fan <wenchen@databricks.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
### What changes were proposed in this pull request?
Remove the private method `checkIntervalStringDataType()` from `IntervalUtils` since it hasn't been used anymore after https://github.com/apache/spark/pull/33242.
### Why are the changes needed?
To improve code maintenance.
### Does this PR introduce _any_ user-facing change?
No. The method is private, and it existing in code base for short time.
### How was this patch tested?
By existing GAs/tests.
Closes#33321 from MaxGekk/SPARK-35735-remove-unused-method.
Authored-by: Max Gekk <max.gekk@gmail.com>
Signed-off-by: Max Gekk <max.gekk@gmail.com>
### What changes were proposed in this pull request?
This PR allow the parser to parse unit list interval literals like `'3' day '10' hours '3' seconds` or `'8' years '3' months` as `YearMonthIntervalType` or `DayTimeIntervalType`.
### Why are the changes needed?
For ANSI compliance.
### Does this PR introduce _any_ user-facing change?
Yes. I noted the following things in the `sql-migration-guide.md`.
* Unit list interval literals are parsed as `YearMonthIntervaType` or `DayTimeIntervalType` instead of `CalendarIntervalType`.
* `WEEK`, `MILLISECONS`, `MICROSECOND` and `NANOSECOND` are not valid units for unit list interval literals.
* Units of year-month and day-time cannot be mixed like `1 YEAR 2 MINUTES`.
### How was this patch tested?
New tests and modified tests.
Closes#32949 from sarutak/day-time-multi-units.
Authored-by: Kousuke Saruta <sarutak@oss.nttdata.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
### What changes were proposed in this pull request?
Add new SQL function `to_timestamp_ltz`
syntax:
```
to_timestamp_ltz(timestamp_str_column[, fmt])
to_timestamp_ltz(timestamp_column)
to_timestamp_ltz(date_column)
```
### Why are the changes needed?
As the result of to_timestamp become consistent with the SQL configuration spark.sql.timestmapType and there is already a SQL function to_timestmap_ntz, we need new function to_timestamp_ltz to construct timestamp with local time zone values.
### Does this PR introduce _any_ user-facing change?
Yes, a new function for constructing timestamp with local time zone values
### How was this patch tested?
Unit test
Closes#33318 from gengliangwang/to_timestamp_ltz.
Authored-by: Gengliang Wang <gengliang@apache.org>
Signed-off-by: Gengliang Wang <gengliang@apache.org>
### What changes were proposed in this pull request?
This PR modifies `DecorrelateInnerQuery` to handle the COUNT bug for lateral subqueries. Similar to SPARK-15370, rewriting lateral subqueries as joins can change the semantics of the subquery and lead to incorrect answers.
However we can't reuse the existing code to handle the count bug for correlated scalar subqueries because it assumes the subquery to have a specific shape (either with Filter + Aggregate or Aggregate as the root node). Instead, this PR proposes a more generic way to handle the COUNT bug. If an Aggregate is subject to the COUNT bug, we insert a left outer domain join between the outer query and the aggregate with a `alwaysTrue` marker and rewrite the final result conditioning on the marker. For example:
```sql
-- t1: [(0, 1), (1, 2)]
-- t2: [(0, 2), (0, 3)]
select * from t1 left outer join lateral (select count(*) from t2 where t2.c1 = t1.c1)
```
Without count bug handling, the query plan is
```
Project [c1#44, c2#45, count(1)#53L]
+- Join LeftOuter, (c1#48 = c1#44)
:- LocalRelation [c1#44, c2#45]
+- Aggregate [c1#48], [count(1) AS count(1)#53L, c1#48]
+- LocalRelation [c1#48]
```
and the answer is wrong:
```
+---+---+--------+
|c1 |c2 |count(1)|
+---+---+--------+
|0 |1 |2 |
|1 |2 |null |
+---+---+--------+
```
With the count bug handling:
```
Project [c1#1, c2#2, count(1)#10L]
+- Join LeftOuter, (c1#34 <=> c1#1)
:- LocalRelation [c1#1, c2#2]
+- Project [if (isnull(alwaysTrue#32)) 0 else count(1)#33L AS count(1)#10L, c1#34]
+- Join LeftOuter, (c1#5 = c1#34)
:- Aggregate [c1#1], [c1#1 AS c1#34]
: +- LocalRelation [c1#1]
+- Aggregate [c1#5], [count(1) AS count(1)#33L, c1#5, true AS alwaysTrue#32]
+- LocalRelation [c1#5]
```
and we have the correct answer:
```
+---+---+--------+
|c1 |c2 |count(1)|
+---+---+--------+
|0 |1 |2 |
|1 |2 |0 |
+---+---+--------+
```
### Why are the changes needed?
Fix a correctness bug with lateral join rewrite.
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
Added SQL query tests. The results are consistent with Postgres' results.
Closes#33070 from allisonwang-db/spark-35551-lateral-count-bug.
Authored-by: allisonwang-db <allison.wang@databricks.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
### What changes were proposed in this pull request?
This patch proposes to check data after adding data to topic in `KafkaSourceStressSuite`.
### Why are the changes needed?
The test logic in `KafkaSourceStressSuite` is not stable. For example, https://github.com/apache/spark/runs/3049244904.
Once we add data to a topic and then delete the topic before checking data, the expected answer is different to retrieved data from the sink.
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
Existing tests.
Closes#33311 from viirya/stream-assert.
Authored-by: Liang-Chi Hsieh <viirya@gmail.com>
Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
### What changes were proposed in this pull request?
This PR reverts https://github.com/apache/spark/pull/32455 and its followup https://github.com/apache/spark/pull/32536 , because the new janino version has a bug that is not fixed yet: https://github.com/janino-compiler/janino/pull/148
### Why are the changes needed?
avoid regressions
### Does this PR introduce _any_ user-facing change?
no
### How was this patch tested?
existing tests
Closes#33302 from cloud-fan/revert.
Authored-by: Wenchen Fan <wenchen@databricks.com>
Signed-off-by: Hyukjin Kwon <gurwls223@apache.org>
### What changes were proposed in this pull request?
A followup of #32980. We should use `subExprCode` to avoid duplicate call of `addNewFunction`.
### Why are the changes needed?
Avoid duplicate all of `addNewFunction`.
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
Existing test.
Closes#33305 from viirya/fix-minor.
Authored-by: Liang-Chi Hsieh <viirya@gmail.com>
Signed-off-by: Liang-Chi Hsieh <viirya@gmail.com>
### What changes were proposed in this pull request?
Support new functions make_timestamp_ntz and make_timestamp_ltz
Syntax:
* `make_timestamp_ntz(year, month, day, hour, min, sec)`: Create local date-time from year, month, day, hour, min, sec fields
* `make_timestamp_ltz(year, month, day, hour, min, sec[, timezone])`: Create current timestamp with local time zone from year, month, day, hour, min, sec and timezone fields
### Why are the changes needed?
As the result of `make_timestamp` become consistent with the SQL configuration `spark.sql.timestmapType`, we need these two new functions to construct timestamp literals. They align to the functions [`make_timestamp` and `make_timestamptz`](https://www.postgresql.org/docs/9.4/functions-datetime.html) in PostgreSQL
### Does this PR introduce _any_ user-facing change?
Yes, two new datetime functions: make_timestamp_ntz and make_timestamp_ltz.
### How was this patch tested?
End-to-end tests.
Closes#33299 from gengliangwang/make_timestamp_ntz_ltz.
Authored-by: Gengliang Wang <gengliang@apache.org>
Signed-off-by: Max Gekk <max.gekk@gmail.com>
### What changes were proposed in this pull request?
This PR group exception messages in sql/core/src/main/scala/org/apache/spark/sql/execution/command
### Why are the changes needed?
It will largely help with standardization of error messages and its maintenance.
### Does this PR introduce any user-facing change?
No. Error messages remain unchanged.
### How was this patch tested?
No new tests - pass all original tests to make sure it doesn't break any existing behavior.
Closes#32951 from dgd-contributor/SPARK-33603_grouping_execution/command.
Authored-by: dgd-contributor <dgd_contributor@viettel.com.vn>
Signed-off-by: Gengliang Wang <gengliang@apache.org>
### What changes were proposed in this pull request?
This PR proposes to introduce a new feature "prefix match scan" on state store, which enables users of state store (mostly stateful operators) to group the keys into logical groups, and scan the keys in the same group efficiently.
For example, if the schema of the key of state store is `[ sessionId | session.start ]`, we can scan with prefix key which schema is `[ sessionId ]` (leftmost 1 column) and retrieve all key-value pairs in state store which keys are matched with given prefix key.
This PR will bring the API changes, though the changes are done in the developer API.
* Registering the prefix key
We propose to make an explicit change to the init() method of StateStoreProvider, as below:
```
def init(
stateStoreId: StateStoreId,
keySchema: StructType,
valueSchema: StructType,
numColsPrefixKey: Int,
storeConfs: StateStoreConf,
hadoopConf: Configuration): Unit
```
Please note that we remove an unused parameter “keyIndexOrdinal” as well. The parameter is coupled with getRange() which we will remove as well. See below for rationalization.
Here we provide the number of columns we take to project the prefix key from the full key. If the operator doesn’t leverage prefix match scan, the value can (and should) be 0, because the state store provider may optimize the underlying storage format which may bring extra overhead.
We would like to apply some restrictions on prefix key to simplify the functionality:
* Prefix key is a part of the full key. It can’t be the same as the full key.
* That said, the full key will be the (prefix key + remaining parts), and both prefix key and remaining parts should have at least one column.
* We always take the columns from the leftmost sequentially, like “seq.take(nums)”.
* We don’t allow reordering of the columns.
* We only guarantee “equality” comparison against prefix keys, and don’t support the prefix “range” scan.
* We only support scanning on the keys which match with the prefix key.
* E.g. We don’t support the range scan from user A to user B due to technical complexity. That’s the reason we can’t leverage the existing getRange API.
As we mentioned, we want to make an explicit change to the init() method of StateStoreProvider which would break backward compatibility, assuming that 3rd party state store providers need to update their code in any way to support prefix match scan. Given RocksDB state store provider is being donated to the OSS and plan to be available in Spark 3.2, the majority of the users would migrate to the built-in state store providers, which would remedy the concerns.
* Scanning key-value pairs matched to the prefix key
We propose to add a new method to the ReadStateStore (and StateStore by inheritance), as below:
```
def prefixScan(prefixKey: UnsafeRow): Iterator[UnsafeRowPair]
```
We require callers to pass the `prefixKey` which would have the same schema with the registered prefix key schema. In other words, the schema of the parameter `prefixKey` should match to the projection of the prefix key on the full key based on the number of columns for the prefix key.
The method contract is clear - the method will return the iterator which will give the key-value pairs whose prefix key is matched with the given prefix key. Callers should only rely on the contract and should not expect any other characteristics based on specific details on the state store provider.
In the caller’s point of view, the prefix key is only used for retrieving key-value pairs via prefix match scan. Callers should keep using the full key to do CRUD.
Note that this PR also proposes to make a breaking change, removal of getRange(), which is never be implemented properly and hence never be called properly.
### Why are the changes needed?
* Introducing prefix match scan feature
Currently, the API in state store is only based on key-value data structure. This lacks on advanced data structures like list-like one, which required us to implement the data structure on our own whenever we need it. We had one in stream-stream join, and we were about to have another one in native session window. The custom implementation of data structure based on the state store API tends to be complicated and has to deal with multiple state stores.
We decided to enhance the state store API a bit to remove the requirement for native session window to implement its own. From the operator of native session window, it will just need to do prefix scan on group key to retrieve all sessions belonging to the group key.
Thanks to adding the feature to the part of state store API, this would enable state store providers to optimize the implementation based on the characteristic. (e.g. We will implement this in RocksDB state store provider via leveraging the characteristic that RocksDB sorts the key by natural order of binary format.)
* Removal of getRange API
Before introducing this we sought the way to leverage getRange, but it's quite hard to implement efficiently, with respecting its method contract. Spark always calls the method with (None, None) parameter and all the state store providers (including built-in) implement it as just calling iterator(), which is not respecting the method contract. That said, we can replace all getRange() usages to iterator(), and remove the API to remove any confusions/concerns.
### Does this PR introduce _any_ user-facing change?
Yes for the end users & maintainers of 3rd party state store provider. They will need to upgrade their state store provider implementations to adopt this change.
### How was this patch tested?
Added UT, and also existing UTs to make sure it doesn't break anything.
Closes#33038 from HeartSaVioR/SPARK-35861.
Authored-by: Jungtaek Lim <kabhwan.opensource@gmail.com>
Signed-off-by: Liang-Chi Hsieh <viirya@gmail.com>
### What changes were proposed in this pull request?
Combine `readBatch` and `readIntegers` in `VectorizedRleValuesReader` by having them share the same `readBatchInternal` method.
### Why are the changes needed?
`readBatch` and `readIntegers` share similar code path and this Jira aims to combine them into one method for easier maintenance.
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
Existing tests as this is just a refactoring.
Closes#33271 from sunchao/SPARK-35743-read-integers.
Authored-by: Chao Sun <sunchao@apple.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
### What changes were proposed in this pull request?
The SQL function TO_TIMESTAMP should return different results based on the default timestamp type:
* when "spark.sql.timestampType" is TIMESTAMP_NTZ, return TimestampNTZType literal
* when "spark.sql.timestampType" is TIMESTAMP_LTZ, return TimestampType literal
This PR also refactor the class GetTimestamp and GetTimestampNTZ to reduce duplicated code.
### Why are the changes needed?
As "spark.sql.timestampType" sets the default timestamp type, the to_timestamp function should behave consistently with it.
### Does this PR introduce _any_ user-facing change?
Yes, when the value of "spark.sql.timestampType" is TIMESTAMP_NTZ, the result type of `TO_TIMESTAMP` is of TIMESTAMP_NTZ type.
### How was this patch tested?
Unit test
Closes#33280 from gengliangwang/to_timestamp.
Authored-by: Gengliang Wang <gengliang@apache.org>
Signed-off-by: Max Gekk <max.gekk@gmail.com>
