This PR introduces `CatalystSchemaConverter` for converting Parquet schema to Spark SQL schema and vice versa. Original conversion code in `ParquetTypesConverter` is removed. Benefits of the new version are:
1. When converting Spark SQL schemas, it generates standard Parquet schemas conforming to [the most updated Parquet format spec] [1]. Converting to old style Parquet schemas is also supported via feature flag `spark.sql.parquet.followParquetFormatSpec` (which is set to `false` for now, and should be set to `true` after both read and write paths are fixed).
Note that although this version of Parquet format spec hasn't been officially release yet, Parquet MR 1.7.0 already sticks to it. So it should be safe to follow.
1. It implements backwards-compatibility rules described in the most updated Parquet format spec. Thus can recognize more schema patterns generated by other/legacy systems/tools.
1. Code organization follows convention used in [parquet-mr] [2], which is easier to follow. (Structure of `CatalystSchemaConverter` is similar to `AvroSchemaConverter`).
To fully implement backwards-compatibility rules in both read and write path, we also need to update `CatalystRowConverter` (which is responsible for converting Parquet records to `Row`s), `RowReadSupport`, and `RowWriteSupport`. These would be done in follow-up PRs.
TODO
- [x] More schema conversion test cases for legacy schema patterns.
[1]: ea09522659/LogicalTypes.md
[2]: https://github.com/apache/parquet-mr/
Author: Cheng Lian <lian@databricks.com>
Closes#6617 from liancheng/spark-6777 and squashes the following commits:
2a2062d [Cheng Lian] Don't convert decimals without precision information
b60979b [Cheng Lian] Adds a constructor which accepts a Configuration, and fixes default value of assumeBinaryIsString
743730f [Cheng Lian] Decimal scale shouldn't be larger than precision
a104a9e [Cheng Lian] Fixes Scala style issue
1f71d8d [Cheng Lian] Adds feature flag to allow falling back to old style Parquet schema conversion
ba84f4b [Cheng Lian] Fixes MapType schema conversion bug
13cb8d5 [Cheng Lian] Fixes MiMa failure
81de5b0 [Cheng Lian] Fixes UDT, workaround read path, and add tests
28ef95b [Cheng Lian] More AnalysisExceptions
b10c322 [Cheng Lian] Replaces require() with analysisRequire() which throws AnalysisException
cceaf3f [Cheng Lian] Implements backwards compatibility rules in CatalystSchemaConverter
make the `TakeOrdered` strategy and operator more general, such that it can optionally handle a projection when necessary
Author: Wenchen Fan <cloud0fan@outlook.com>
Closes#6780 from cloud-fan/limit and squashes the following commits:
34aa07b [Wenchen Fan] revert
07d5456 [Wenchen Fan] clean closure
20821ec [Wenchen Fan] fix
3676a82 [Wenchen Fan] address comments
b558549 [Wenchen Fan] address comments
214842b [Wenchen Fan] fix style
2d8be83 [Wenchen Fan] add LimitPushDown
948f740 [Wenchen Fan] fix existing
https://issues.apache.org/jira/browse/SPARK-8578
It is not very safe to use a custom output committer when append data to an existing dir. This changes adds the logic to check if we are appending data, and if so, we use the output committer associated with the file output format.
Author: Yin Huai <yhuai@databricks.com>
Closes#6964 from yhuai/SPARK-8578 and squashes the following commits:
43544c4 [Yin Huai] Do not use a custom output commiter when appendiing data.
This PR improves the error message shown when conflicting partition column names are detected. This can be particularly annoying and confusing when there are a large number of partitions while a handful of them happened to contain unexpected temporary file(s). Now all suspicious directories are listed as below:
```
java.lang.AssertionError: assertion failed: Conflicting partition column names detected:
Partition column name list #0: b, c, d
Partition column name list #1: b, c
Partition column name list #2: b
For partitioned table directories, data files should only live in leaf directories. Please check the following directories for unexpected files:
file:/tmp/foo/b=0
file:/tmp/foo/b=1
file:/tmp/foo/b=1/c=1
file:/tmp/foo/b=0/c=0
```
Author: Cheng Lian <lian@databricks.com>
Closes#6610 from liancheng/part-errmsg and squashes the following commits:
7d05f2c [Cheng Lian] Fixes Scala style issue
a149250 [Cheng Lian] Adds test case for the error message
6b74dd8 [Cheng Lian] Also lists suspicious non-leaf partition directories
a935eb8 [Cheng Lian] Improves error message when conflicting partition columns are found
Add `sampleBy` to DataFrame. rxin
Author: Xiangrui Meng <meng@databricks.com>
Closes#6769 from mengxr/SPARK-7157 and squashes the following commits:
991f26f [Xiangrui Meng] fix seed
4a14834 [Xiangrui Meng] move sampleBy to stat
832f7cc [Xiangrui Meng] add sampleBy to DataFrame
This PR only applies to master branch (1.5.0-SNAPSHOT) since it references `org.apache.parquet` classes which only appear in Parquet 1.7.0.
Author: Cheng Lian <lian@databricks.com>
Closes#6683 from liancheng/output-committer-docs and squashes the following commits:
b4648b8 [Cheng Lian] Removes spark.sql.sources.outputCommitterClass as it's not a public option
ee63923 [Cheng Lian] Updates docs and comments of data sources and Parquet output committer options
The logical plan `Expand` takes the `output` as constructor argument, which break the references chain. We need to refactor the code, as well as the column pruning.
Author: Cheng Hao <hao.cheng@intel.com>
Closes#5780 from chenghao-intel/expand and squashes the following commits:
76e4aa4 [Cheng Hao] revert the change for case insenstive
7c10a83 [Cheng Hao] refactor the grouping sets
the syntax was incorrect in the example in explode
Author: lockwobr <lockwobr@gmail.com>
Closes#6943 from lockwobr/master and squashes the following commits:
3d864d1 [lockwobr] updated the documentation for explode
Users can now do
```scala
left.join(broadcast(right), "joinKey")
```
to give the query planner a hint that "right" DataFrame is small and should be broadcasted.
Author: Reynold Xin <rxin@databricks.com>
Closes#6751 from rxin/broadcastjoin-hint and squashes the following commits:
953eec2 [Reynold Xin] Code review feedback.
88752d8 [Reynold Xin] Fixed import.
8187b88 [Reynold Xin] [SPARK-8300] DataFrame hint for broadcast join.
Deprecates ```callUdf``` in favor of ```callUDF```.
Author: BenFradet <benjamin.fradet@gmail.com>
Closes#6902 from BenFradet/SPARK-8356 and squashes the following commits:
ef4e9d8 [BenFradet] deprecated callUDF, use udf instead
9b1de4d [BenFradet] reinstated unit test for the deprecated callUdf
cbd80a5 [BenFradet] deprecated callUdf in favor of callUDF
Currently we auto alias expression in parser. However, during parser phase we don't have enough information to do the right alias. For example, Generator that has more than 1 kind of element need MultiAlias, ExtractValue don't need Alias if it's in middle of a ExtractValue chain.
Author: Wenchen Fan <cloud0fan@outlook.com>
Closes#6647 from cloud-fan/alias and squashes the following commits:
552eba4 [Wenchen Fan] fix python
5b5786d [Wenchen Fan] fix agg
73a90cb [Wenchen Fan] fix case-preserve of ExtractValue
4cfd23c [Wenchen Fan] fix order by
d18f401 [Wenchen Fan] refine
9f07359 [Wenchen Fan] address comments
39c1aef [Wenchen Fan] small fix
33640ec [Wenchen Fan] auto alias expressions in analyzer
This PR fixes a Parquet output file name collision bug which may cause data loss. Changes made:
1. Identify each write job issued by `InsertIntoHadoopFsRelation` with a UUID
All concrete data sources which extend `HadoopFsRelation` (Parquet and ORC for now) must use this UUID to generate task output file path to avoid name collision.
2. Make `TestHive` use a local mode `SparkContext` with 32 threads to increase parallelism
The major reason for this is that, the original parallelism of 2 is too low to reproduce the data loss issue. Also, higher concurrency may potentially caught more concurrency bugs during testing phase. (It did help us spotted SPARK-8501.)
3. `OrcSourceSuite` was updated to workaround SPARK-8501, which we detected along the way.
NOTE: This PR is made a little bit more complicated than expected because we hit two other bugs on the way and have to work them around. See [SPARK-8501] [1] and [SPARK-8513] [2].
[1]: https://github.com/liancheng/spark/tree/spark-8501
[2]: https://github.com/liancheng/spark/tree/spark-8513
----
Some background and a summary of offline discussion with yhuai about this issue for better understanding:
In 1.4.0, we added `HadoopFsRelation` to abstract partition support of all data sources that are based on Hadoop `FileSystem` interface. Specifically, this makes partition discovery, partition pruning, and writing dynamic partitions for data sources much easier.
To support appending, the Parquet data source tries to find out the max part number of part-files in the destination directory (i.e., `<id>` in output file name `part-r-<id>.gz.parquet`) at the beginning of the write job. In 1.3.0, this step happens on driver side before any files are written. However, in 1.4.0, this is moved to task side. Unfortunately, for tasks scheduled later, they may see wrong max part number generated of files newly written by other finished tasks within the same job. This actually causes a race condition. In most cases, this only causes nonconsecutive part numbers in output file names. But when the DataFrame contains thousands of RDD partitions, it's likely that two tasks may choose the same part number, then one of them gets overwritten by the other.
Before `HadoopFsRelation`, Spark SQL already supports appending data to Hive tables. From a user's perspective, these two look similar. However, they differ a lot internally. When data are inserted into Hive tables via Spark SQL, `InsertIntoHiveTable` simulates Hive's behaviors:
1. Write data to a temporary location
2. Move data in the temporary location to the final destination location using
- `Hive.loadTable()` for non-partitioned table
- `Hive.loadPartition()` for static partitions
- `Hive.loadDynamicPartitions()` for dynamic partitions
The important part is that, `Hive.copyFiles()` is invoked in step 2 to move the data to the destination directory (I found the name is kinda confusing since no "copying" occurs here, we are just moving and renaming stuff). If a file in the source directory and another file in the destination directory happen to have the same name, say `part-r-00001.parquet`, the former is moved to the destination directory and renamed with a `_copy_N` postfix (`part-r-00001_copy_1.parquet`). That's how Hive handles appending and avoids name collision between different write jobs.
Some alternatives fixes considered for this issue:
1. Use a similar approach as Hive
This approach is not preferred in Spark 1.4.0 mainly because file metadata operations in S3 tend to be slow, especially for tables with lots of file and/or partitions. That's why `InsertIntoHadoopFsRelation` just inserts to destination directory directly, and is often used together with `DirectParquetOutputCommitter` to reduce latency when working with S3. This means, we don't have the chance to do renaming, and must avoid name collision from the very beginning.
2. Same as 1.3, just move max part number detection back to driver side
This isn't doable because unlike 1.3, 1.4 also takes dynamic partitioning into account. When inserting into dynamic partitions, we don't know which partition directories will be touched on driver side before issuing the write job. Checking all partition directories is simply too expensive for tables with thousands of partitions.
3. Add extra component to output file names to avoid name collision
This seems to be the only reasonable solution for now. To be more specific, we need a JOB level unique identifier to identify all write jobs issued by `InsertIntoHadoopFile`. Notice that TASK level unique identifiers can NOT be used. Because in this way a speculative task will write to a different output file from the original task. If both tasks succeed, duplicate output will be left behind. Currently, the ORC data source adds `System.currentTimeMillis` to the output file name for uniqueness. This doesn't work because of exactly the same reason.
That's why this PR adds a job level random UUID in `BaseWriterContainer` (which is used by `InsertIntoHadoopFsRelation` to issue write jobs). The drawback is that record order is not preserved any more (output files of a later job may be listed before those of a earlier job). However, we never promise to preserve record order when writing data, and Hive doesn't promise this either because the `_copy_N` trick breaks the order.
Author: Cheng Lian <lian@databricks.com>
Closes#6864 from liancheng/spark-8406 and squashes the following commits:
db7a46a [Cheng Lian] More comments
f5c1133 [Cheng Lian] Addresses comments
85c478e [Cheng Lian] Workarounds SPARK-8513
088c76c [Cheng Lian] Adds comment about SPARK-8501
99a5e7e [Cheng Lian] Uses job level UUID in SimpleTextRelation and avoids double task abortion
4088226 [Cheng Lian] Works around SPARK-8501
1d7d206 [Cheng Lian] Adds more logs
8966bbb [Cheng Lian] Fixes Scala style issue
18b7003 [Cheng Lian] Uses job level UUID to take speculative tasks into account
3806190 [Cheng Lian] Lets TestHive use all cores by default
748dbd7 [Cheng Lian] Adding UUID to output file name to avoid accidental overwriting
In earlier versions of Spark SQL we casted `TimestampType` and `DataType` to `StringType` when it was involved in a binary comparison with a `StringType`. This allowed comparing a timestamp with a partial date as a user would expect.
- `time > "2014-06-10"`
- `time > "2014"`
In 1.4.0 we tried to cast the String instead into a Timestamp. However, since partial dates are not a valid complete timestamp this results in `null` which results in the tuple being filtered.
This PR restores the earlier behavior. Note that we still special case equality so that these comparisons are not affected by not printing zeros for subsecond precision.
Author: Michael Armbrust <michael@databricks.com>
Closes#6888 from marmbrus/timeCompareString and squashes the following commits:
bdef29c [Michael Armbrust] test partial date
1f09adf [Michael Armbrust] special handling of equality
1172c60 [Michael Armbrust] more test fixing
4dfc412 [Michael Armbrust] fix tests
aaa9508 [Michael Armbrust] newline
04d908f [Michael Armbrust] [SPARK-8420][SQL] Fix comparision of timestamps/dates with strings
Author: Nathan Howell <nhowell@godaddy.com>
Closes#6799 from NathanHowell/spark-8093 and squashes the following commits:
76ac3e8 [Nathan Howell] [SPARK-8093] [SQL] Remove empty structs inferred from JSON documents
Author: Shilei <shilei.qian@intel.com>
Closes#6779 from qiansl127/MD5 and squashes the following commits:
11fcdb2 [Shilei] Fix the indent
04bd27b [Shilei] Add codegen
da60eb3 [Shilei] Remove checkInputDataTypes function
9509ad0 [Shilei] Format code
12c61f4 [Shilei] Accept only BinaryType for Md5
1df0b5b [Shilei] format to scala type
60ccde1 [Shilei] Add more test case
b8c73b4 [Shilei] Rewrite the type check for Md5
c166167 [Shilei] Add md5 function
I have added it for only Scala.
TODO: we should also support `in` operator in Python.
Author: Yu ISHIKAWA <yuu.ishikawa@gmail.com>
Closes#6824 from yu-iskw/SPARK-8348 and squashes the following commits:
e76d02f [Yu ISHIKAWA] Not use infix notation
6f744ac [Yu ISHIKAWA] Fit the test cases because these used the old test data set.
00077d3 [Yu ISHIKAWA] [SPARK-8348][SQL] Add in operator to DataFrame Column
Author: Sandy Ryza <sandy@cloudera.com>
Closes#6679 from sryza/sandy-spark-8135 and squashes the following commits:
c5554ff [Sandy Ryza] SPARK-8135. In SerializableWritable, don't load defaults when instantiating Configuration
This patch introduces `SparkPlanTest`, a base class for unit tests of SparkPlan physical operators. This is analogous to Spark SQL's existing `QueryTest`, which does something similar for end-to-end tests with actual queries.
These helper methods provide nicer error output when tests fail and help developers to avoid writing lots of boilerplate in order to execute manually constructed physical plans.
Author: Josh Rosen <joshrosen@databricks.com>
Author: Josh Rosen <rosenville@gmail.com>
Author: Michael Armbrust <michael@databricks.com>
Closes#6885 from JoshRosen/spark-plan-test and squashes the following commits:
f8ce275 [Josh Rosen] Fix some IntelliJ inspections and delete some dead code
84214be [Josh Rosen] Add an extra column which isn't part of the sort
ae1896b [Josh Rosen] Provide implicits automatically
a80f9b0 [Josh Rosen] Merge pull request #4 from marmbrus/pr/6885
d9ab1e4 [Michael Armbrust] Add simple resolver
c60a44d [Josh Rosen] Manually bind references
996332a [Josh Rosen] Add types so that tests compile
a46144a [Josh Rosen] WIP
JIRA: https://issues.apache.org/jira/browse/SPARK-8363
Author: Liang-Chi Hsieh <viirya@gmail.com>
Closes#6823 from viirya/move_sqrt and squashes the following commits:
8977e11 [Liang-Chi Hsieh] Remove unnecessary old tests.
d23e79e [Liang-Chi Hsieh] Explicitly indicate sqrt value sequence.
699f48b [Liang-Chi Hsieh] Use correct @since tag.
8dff6d1 [Liang-Chi Hsieh] Merge remote-tracking branch 'upstream/master' into move_sqrt
bc2ed77 [Liang-Chi Hsieh] Remove/move arithmetic expression test and expression type checking test. Remove unnecessary Sqrt type rule.
d38492f [Liang-Chi Hsieh] Now sqrt accepts boolean because type casting is handled by HiveTypeCoercion.
297cc90 [Liang-Chi Hsieh] Sqrt only accepts double input.
ef4a21a [Liang-Chi Hsieh] Move sqrt to math.
1. Add `SQLConfEntry` to store the information about a configuration. For those configurations that cannot be found in `sql-programming-guide.md`, I left the doc as `<TODO>`.
2. Verify the value when setting a configuration if this is in SQLConf.
3. Use `SET -v` to display all public configurations.
Author: zsxwing <zsxwing@gmail.com>
Closes#6747 from zsxwing/sqlconf and squashes the following commits:
7d09bad [zsxwing] Use SQLConfEntry in HiveContext
49f6213 [zsxwing] Add getConf, setConf to SQLContext and HiveContext
e014f53 [zsxwing] Merge branch 'master' into sqlconf
93dad8e [zsxwing] Fix the unit tests
cf950c1 [zsxwing] Fix the code style and tests
3c5f03e [zsxwing] Add unsetConf(SQLConfEntry) and fix the code style
a2f4add [zsxwing] getConf will return the default value if a config is not set
037b1db [zsxwing] Add schema to SetCommand
0520c3c [zsxwing] Merge branch 'master' into sqlconf
7afb0ec [zsxwing] Fix the configurations about HiveThriftServer
7e728e3 [zsxwing] Add doc for SQLConfEntry and fix 'toString'
5e95b10 [zsxwing] Add enumConf
c6ba76d [zsxwing] setRawString => setConfString, getRawString => getConfString
4abd807 [zsxwing] Fix the test for 'set -v'
6e47e56 [zsxwing] Fix the compilation error
8973ced [zsxwing] Remove floatConf
1fc3a8b [zsxwing] Remove the 'conf' command and use 'set -v' instead
99c9c16 [zsxwing] Fix tests that use SQLConfEntry as a string
88a03cc [zsxwing] Add new lines between confs and return types
ce7c6c8 [zsxwing] Remove seqConf
f3c1b33 [zsxwing] Refactor SQLConf to display better error message
This PR is a improvement for https://github.com/apache/spark/pull/5189.
The resolution rule for ORDER BY is: first resolve based on what comes from the select clause and then fall back on its child only when this fails.
There are 2 steps. First, try to resolve `Sort` in `ResolveReferences` based on select clause, and ignore exceptions. Second, try to resolve `Sort` in `ResolveSortReferences` and add missing projection.
However, the way we resolve `SortOrder` is wrong. We just resolve `UnresolvedAttribute` and use the result to indicate if we can resolve `SortOrder`. But `UnresolvedAttribute` is only part of `GetField` chain(broken by `GetItem`), so we need to go through the whole chain to indicate if we can resolve `SortOrder`.
With this change, we can also avoid re-throw GetField exception in `CheckAnalysis` which is little ugly.
Author: Wenchen Fan <cloud0fan@outlook.com>
Closes#5659 from cloud-fan/order-by and squashes the following commits:
cfa79f8 [Wenchen Fan] update test
3245d28 [Wenchen Fan] minor improve
465ee07 [Wenchen Fan] address comment
1fc41a2 [Wenchen Fan] fix SPARK-7067
1. Given a query
`select coalesce(null, 1, '1') from dual` will cause exception:
java.lang.RuntimeException: Could not determine return type of Coalesce for IntegerType,StringType
2. Given a query:
`select case when true then 1 else '1' end from dual` will cause exception:
java.lang.RuntimeException: Types in CASE WHEN must be the same or coercible to a common type: StringType != IntegerType
I checked the code, the main cause is the HiveTypeCoercion doesn't do implicit convert when there is a IntegerType and StringType.
Numeric types can be promoted to string type
Hive will always do this implicit conversion.
Author: OopsOutOfMemory <victorshengli@126.com>
Closes#6551 from OopsOutOfMemory/pnts and squashes the following commits:
7a209d7 [OopsOutOfMemory] rebase master
6018613 [OopsOutOfMemory] convert function to method
4cd5618 [OopsOutOfMemory] limit the data type to primitive type
df365d2 [OopsOutOfMemory] refine
95cbd58 [OopsOutOfMemory] fix style
403809c [OopsOutOfMemory] promote non-string to string when can not found tighestCommonTypeOfTwo
chenghao-intel adrian-wang
Author: dragonli <lisurprise@gmail.com>
Author: zhichao.li <zhichao.li@intel.com>
Closes#6838 from zhichao-li/positive and squashes the following commits:
e1032a0 [dragonli] remove useless import and refactor code
624d438 [zhichao.li] add positive identify function
[SQL][DOC] I found it a bit confusing when I came across it for the first time in the docs
Author: Radek Ostrowski <dest.hawaii@gmail.com>
Author: radek <radek@radeks-MacBook-Pro-2.local>
Closes#6332 from radek1st/master and squashes the following commits:
dae3347 [Radek Ostrowski] fixed typo
c76bb3a [radek] improved a comment
In order to have better performance out of box, this PR turn on codegen by default, then codegen can be tested by sql/test and hive/test.
This PR also fix some corner cases for codegen.
Before 1.5 release, we should re-visit this, turn it off if it's not stable or causing regressions.
cc rxin JoshRosen
Author: Davies Liu <davies@databricks.com>
Closes#6726 from davies/enable_codegen and squashes the following commits:
f3b25a5 [Davies Liu] fix warning
73750ea [Davies Liu] fix long overflow when compare
3017a47 [Davies Liu] Merge branch 'master' of github.com:apache/spark into enable_codegen
a7d75da [Davies Liu] Merge branch 'master' of github.com:apache/spark into enable_codegen
ff5b75a [Davies Liu] Merge branch 'master' of github.com:apache/spark into enable_codegen
f4cf2c2 [Davies Liu] fix style
99fc139 [Davies Liu] Merge branch 'enable_codegen' of github.com:davies/spark into enable_codegen
91fc7a2 [Davies Liu] disable codegen for ScalaUDF
207e339 [Davies Liu] Update CodeGenerator.scala
44573a3 [Davies Liu] check thread safety of expression
f3886fa [Davies Liu] don't inline primitiveTerm for null literal
c8e7cd2 [Davies Liu] address comment
a8618c9 [Davies Liu] enable codegen by default
Add aggregates in ORDER BY clauses to the `Aggregate` operator beneath. Project these results away after the Sort.
Based on work by watermen. Also Closes#5290.
Author: Yadong Qi <qiyadong2010@gmail.com>
Author: Michael Armbrust <michael@databricks.com>
Closes#6816 from marmbrus/pr/5290 and squashes the following commits:
3226a97 [Michael Armbrust] consistent ordering
eb8938d [Michael Armbrust] no vars
c8b25c1 [Yadong Qi] move the test data.
7f9b736 [Yadong Qi] delete Substring case
a1e87c1 [Yadong Qi] fix conflict
f119849 [Yadong Qi] order by aggregated function
Author: Michael Armbrust <michael@databricks.com>
Closes#6811 from marmbrus/aliasExplodeStar and squashes the following commits:
fbd2065 [Michael Armbrust] more style
806a373 [Michael Armbrust] fix style
7cbb530 [Michael Armbrust] [SPARK-8358][SQL] Wait for child resolution when resolving generatorsa
Also addressed code review feedback from #6754
Author: Reynold Xin <rxin@databricks.com>
Closes#6803 from rxin/abs and squashes the following commits:
d07beba [Reynold Xin] [SPARK-8347] Add unit tests for abs.
This patch updates two pieces of logic that are related to handling of keyOrderings in ShuffleDependencies:
- The Tungsten ShuffleManager falls back to regular SortShuffleManager whenever the shuffle dependency specifies a key ordering, but technically we only need to fall back when an aggregator is also specified. This patch updates the fallback logic to reflect this so that the Tungsten optimizations can apply to more workloads.
- The SQL Exchange operator performs defensive copying of shuffle inputs when a key ordering is specified, but this is unnecessary. The copying was added to guard against cases where ExternalSorter would buffer non-serialized records in memory. When ExternalSorter is configured without an aggregator, it uses the following logic to determine whether to buffer records in a serialized or deserialized format:
```scala
private val useSerializedPairBuffer =
ordering.isEmpty &&
conf.getBoolean("spark.shuffle.sort.serializeMapOutputs", true) &&
ser.supportsRelocationOfSerializedObjects
```
The `newOrdering.isDefined` branch in `ExternalSorter.needToCopyObjectsBeforeShuffle`, removed by this patch, is not necessary:
- It was checked even if we weren't using sort-based shuffle, but this was unnecessary because only SortShuffleManager performs map-side sorting.
- Map-side sorting during shuffle writing is only performed for shuffles that perform map-side aggregation as part of the shuffle (to see this, look at how SortShuffleWriter constructs ExternalSorter). Since SQL never pushes aggregation into Spark's shuffle, we can guarantee that both the aggregator and ordering will be empty and Spark SQL always uses serializers that support relocation, so sort-shuffle will use the serialized pair buffer unless the user has explicitly disabled it via the SparkConf feature-flag. Therefore, I think my optimization in Exchange should be safe.
Author: Josh Rosen <joshrosen@databricks.com>
Closes#6773 from JoshRosen/SPARK-8319 and squashes the following commits:
7a14129 [Josh Rosen] Revise comments; add handler to guard against future ShuffleManager implementations
07bb2c9 [Josh Rosen] Update comment to clarify circumstances under which shuffle operates on serialized records
269089a [Josh Rosen] Avoid unnecessary copy in SQL Exchange
34e526e [Josh Rosen] Enable Tungsten shuffle for non-agg shuffles w/ key orderings
cc rxin marmbrus
Author: Davies Liu <davies@databricks.com>
Closes#6802 from davies/cleanup_internalrow and squashes the following commits:
769d2aa [Davies Liu] remove not needed cast
4acbbe4 [Davies Liu] catalyst.Internal -> InternalRow
The original fix uses DecimalType.Unlimited, which is harder to
handle afterwards. There is no scale and most data should fit into
a long, thus DecimalType(20,0) should be better.
Author: Rene Treffer <treffer@measite.de>
Closes#6789 from rtreffer/spark-7897-unsigned-bigint-as-decimal and squashes the following commits:
2006613 [Rene Treffer] Fix type for "unsigned bigint" jdbc loading.
Author: Michael Armbrust <michael@databricks.com>
Closes#6786 from marmbrus/optionsParser and squashes the following commits:
e7d18ef [Michael Armbrust] add dots
99a3452 [Michael Armbrust] [SPARK-8329][SQL] Allow _ in DataSource options
Currently, we use o.a.s.sql.Row both internally and externally. The external interface is wider than what the internal needs because it is designed to facilitate end-user programming. This design has proven to be very error prone and cumbersome for internal Row implementations.
As a first step, we create an InternalRow interface in the catalyst module, which is identical to the current Row interface. And we switch all internal operators/expressions to use this InternalRow instead. When we need to expose Row, we convert the InternalRow implementation into Row for users.
For all public API, we use Row (for example, data source APIs), which will be converted into/from InternalRow by CatalystTypeConverters.
For all internal data sources (Json, Parquet, JDBC, Hive), we use InternalRow for better performance, casted into Row in buildScan() (without change the public API). When create a PhysicalRDD, we cast them back to InternalRow.
cc rxin marmbrus JoshRosen
Author: Davies Liu <davies@databricks.com>
Closes#6792 from davies/internal_row and squashes the following commits:
f2abd13 [Davies Liu] fix scalastyle
a7e025c [Davies Liu] move InternalRow into catalyst
30db8ba [Davies Liu] Merge branch 'master' of github.com:apache/spark into internal_row
7cbced8 [Davies Liu] separate Row and InternalRow
It's a follow up of https://github.com/apache/spark/pull/6173, for expressions like `Coalesce` that have a `Seq[Expression]`, when we do semantic equal check for it, we need to do semantic equal check for all of its children.
Also we can just use `Seq[(Expression, NamedExpression)]` instead of `Map[Expression, NamedExpression]` as we only search it with `find`.
chenghao-intel, I agree that we probably never knows `semanticEquals` in a general way, but I think we have done that in `TreeNode`, so we can use similar logic. Then we can handle something like `Coalesce(children: Seq[Expression])` correctly.
Author: Wenchen Fan <cloud0fan@outlook.com>
Closes#6261 from cloud-fan/tmp and squashes the following commits:
4daef88 [Wenchen Fan] address comments
dd8fbd9 [Wenchen Fan] correct semanticEquals
This brings in major improvement in that footers are not read on the driver. This also cleans up the code in parquetTableOperations, where we had to override getSplits to eliminate multiple listStatus calls.
cc liancheng
are there any other changes we need for this ?
Author: Yash Datta <Yash.Datta@guavus.com>
Closes#5889 from saucam/parquet_1.6 and squashes the following commits:
d1bf41e [Yash Datta] SPARK-7340: Fix scalastyle and incorporate review comments
c9aa042 [Yash Datta] SPARK-7340: Use the new user defined filter predicate for pushing down inset into parquet
56bc750 [Yash Datta] SPARK-7340: Change parquet version to latest release
In some cases, Spark SQL pushes sorting operations into the shuffle layer by specifying a key ordering as part of the shuffle dependency. I think that we should not do this:
- Since we do not delegate aggregation to Spark's shuffle, specifying the keyOrdering as part of the shuffle has no effect on the shuffle map side.
- By performing the shuffle ourselves (by inserting a sort operator after the shuffle instead), we can use the Exchange planner to choose specialized sorting implementations based on the types of rows being sorted.
- We can remove some complexity from SqlSerializer2 by not requiring it to know about sort orderings, since SQL's own sort operators will already perform the necessary defensive copying.
This patch removes Exchange's `canSortWithShuffle` path and the associated code in `SqlSerializer2`. Shuffles that used to go through the `canSortWithShuffle` path would always wind up using Spark's `ExternalSorter` (inside of `HashShuffleReader`); to avoid a performance regression as a result of handling these shuffles ourselves, I've changed the SQLConf defaults so that external sorting is enabled by default.
Author: Josh Rosen <joshrosen@databricks.com>
Closes#6772 from JoshRosen/SPARK-8317 and squashes the following commits:
ebf9c0f [Josh Rosen] Do not push sort into shuffle in Exchange operator
bf3b4c8 [Josh Rosen] Enable external sort by default
When df.cache() method called, the `withCachedData` of `QueryExecution` has been created, which mean it will not look up the cached tables when action method called afterward.
Author: Cheng Hao <hao.cheng@intel.com>
Closes#5714 from chenghao-intel/SPARK-7158 and squashes the following commits:
58ea8aa [Cheng Hao] style issue
2bf740f [Cheng Hao] create new QueryExecution instance for CacheManager
a5647d9 [Cheng Hao] hide the queryExecution of DataFrame
fbfd3c5 [Cheng Hao] make the DataFrame.queryExecution mutable for cache/persist/unpersist
Unit test is still in Scala.
Author: Reynold Xin <rxin@databricks.com>
Closes#6738 from rxin/utf8string-java and squashes the following commits:
562dc6e [Reynold Xin] Flag...
98e600b [Reynold Xin] Another try with encoding setting ..
cfa6bdf [Reynold Xin] Merge branch 'master' into utf8string-java
a3b124d [Reynold Xin] Try different UTF-8 encoded characters.
1ff7c82 [Reynold Xin] Enable UTF-8 encoding.
82d58cc [Reynold Xin] Reset run-tests.
2cb3c69 [Reynold Xin] Use utf-8 encoding in set bytes.
53f8ef4 [Reynold Xin] Hack Jenkins to run one test.
9a48e8d [Reynold Xin] Fixed runtime compilation error.
911c450 [Reynold Xin] Moved unit test also to Java.
4eff7bd [Reynold Xin] Improved unit test coverage.
8e89a3c [Reynold Xin] Fixed tests.
77c64bd [Reynold Xin] Fixed string type codegen.
ffedb62 [Reynold Xin] Code review feedback.
0967ce6 [Reynold Xin] Fixed import ordering.
45a123d [Reynold Xin] [SPARK-8286] Rewrite UTF8String in Java and move it into unsafe package.
Spark SQL does not support timezone, and Pyrolite does not support timezone well. This patch will convert datetime into POSIX timestamp (without confusing of timezone), which is used by SQL. If the datetime object does not have timezone, it's treated as local time.
The timezone in RDD will be lost after one round trip, all the datetime from SQL will be local time.
Because of Pyrolite, datetime from SQL only has precision as 1 millisecond.
This PR also drop the timezone in date, convert it to number of days since epoch (used in SQL).
Author: Davies Liu <davies@databricks.com>
Closes#6250 from davies/tzone and squashes the following commits:
44d8497 [Davies Liu] add timezone support for DateType
99d9d9c [Davies Liu] use int for timestamp
10aa7ca [Davies Liu] Merge branch 'master' of github.com:apache/spark into tzone
6a29aa4 [Davies Liu] support datetime with timezone
Author: Daoyuan Wang <daoyuan.wang@intel.com>
This patch had conflicts when merged, resolved by
Committer: Reynold Xin <rxin@databricks.com>
Closes#6718 from adrian-wang/udflog2 and squashes the following commits:
3909f48 [Daoyuan Wang] math function: log2
Author: Cheng Hao <hao.cheng@intel.com>
Closes#6724 from chenghao-intel/length and squashes the following commits:
aaa3c31 [Cheng Hao] revert the additional change
97148a9 [Cheng Hao] remove the codegen testing temporally
ae08003 [Cheng Hao] update the comments
1eb1fd1 [Cheng Hao] simplify the code as commented
3e92d32 [Cheng Hao] use the selectExpr in unit test intead of SQLQuery
3c729aa [Cheng Hao] fix bug for constant null value in codegen
3641f06 [Cheng Hao] keep the length() method for registered function
8e30171 [Cheng Hao] update the code as comment
db604ae [Cheng Hao] Add code gen support
548d2ef [Cheng Hao] register the length()
09a0738 [Cheng Hao] add length support
Currently we only support `Seq[Expression]`, we should handle cases like `Seq[Seq[Expression]]` so that we can remove the unnecessary `GroupExpression`.
Author: Wenchen Fan <cloud0fan@outlook.com>
Closes#6706 from cloud-fan/clean and squashes the following commits:
60a1193 [Wenchen Fan] support nested expression sequence and remove GroupExpression
case cs CombineSum(expr) =>
val calcType = expr.dataType
expr.dataType match {
case DecimalType.Fixed(_, _) =>
DecimalType.Unlimited
case _ =>
expr.dataType
}
calcType is always expr.dataType. credits are all belong to IntelliJ
Author: navis.ryu <navis@apache.org>
Closes#6736 from navis/SPARK-8285 and squashes the following commits:
20382c1 [navis.ryu] [SPARK-8285] [SQL] CombineSum should be calculated as unlimited decimal first
