[SPARK-12854][SQL] Implement complex types support in ColumnarBatch #10820

nongli · 2016-01-18T23:42:32Z

This patch adds support for complex types for ColumnarBatch. ColumnarBatch supports structs
and arrays. There is a simple mapping between the richer catalyst types to these two. Strings
are treated as an array of bytes.

ColumnarBatch will contain a column for each node of the schema. Non-complex schemas consists
of just leaf nodes. Structs represent an internal node with one child for each field. Arrays
are internal nodes with one child. Structs just contain nullability. Arrays contain offsets
and lengths into the child array. This structure is able to handle arbitrary nesting. It has
the key property that we maintain columnar throughout and that primitive types are only stored
in the leaf nodes and contiguous across rows. For example, if the schema is

array<array<int>>

There are three columns in the schema. The internal nodes each have one children. The leaf node contains all the int data stored consecutively.

As part of this, this patch adds append APIs in addition to the Put APIs (e.g. putLong(rowid, v)
vs appendLong(v)). These APIs are necessary when the batch contains variable length elements.
The vectors are not fixed length and will grow as necessary. This should make the usage a lot
simpler for the writer.

…atch. WIP: this patch adds some random row generation. The test code needs to be cleaned up as it duplicates functionality from else where. The non-test code should be good to review. This patch adds support for complex types for ColumnarBatch. ColumnarBatch supports structs and arrays. There is a simple mapping between the richer catalyst types to these two. Strings are treated as an array of bytes. ColumnarBatch will contain a column for each node of the schema. Non-complex schemas consists of just leaf nodes. Structs represent an internal node with one child for each field. Arrays are internal nodes with one child. Structs just contain nullability. Arrays contain offsets and lengths into the child array. This structure is able to handle arbitrary nesting. It has the key property that we maintain columnar throughout and that primitive types are only stored in the leaf nodes and contiguous across rows. For example, if the schema is array<array<int>>, all of the int data is stored consecutively. As part of this, this patch adds append APIs in addition to the Put APIs (e.g. putLong(rowid, v) vs appendLong(v)). These APIs are necessary when the batch contains variable length elements. The vectors are not fixed length and will grow as necessary. This should make the usage a lot simpler for the writer.

SparkQA · 2016-01-19T02:07:54Z

Test build #49636 has finished for PR 10820 at commit 518d9bc.

This patch fails Spark unit tests.
This patch merges cleanly.
This patch adds the following public classes (experimental):
- public static final class Array
- public static final class Struct
- public class ColumnVectorUtils
- public static final class Row extends InternalRow

rxin · 2016-01-19T07:42:42Z

if the schema is array>,? array of int?

rxin · 2016-01-19T07:45:25Z

sql/core/src/main/java/org/apache/spark/sql/execution/vectorized/ColumnVector.java

can you explain how lengths and offsets are stored? also is there a single "parent" column that encodes nullability, length, and offset?

I'll update the comment when I rev the PR but the answer is:

I'm not sure what single "parent" means. The array is a column that stores nullability, lengths and offsets. The child column stores the values, including their nullability. Either one can be independently nullable or not.

Lengths and offsets are encoded like 'ints'.

kiszk · 2016-01-20T09:17:08Z

Will ColumnarBatch become a column version of UnsafeRow? Or, it will become more general columnar store?

nongli · 2016-01-25T22:23:20Z

@kiszk We're looking to make the execution engine represent data in a columnar way. I'm not sure what you mean by more general columnar store. I usually think of that more for storage.

SparkQA · 2016-01-26T00:40:09Z

Test build #50029 has finished for PR 10820 at commit f579889.

This patch fails Spark unit tests.
This patch merges cleanly.
This patch adds no public classes.

davies · 2016-01-26T00:55:40Z

unsafe/src/main/java/org/apache/spark/unsafe/Platform.java

Should this also work for on-heap?

I have this in the same group of functions as freeMemory, allocateMemory. What would the on heap version look like?

davies · 2016-01-26T01:05:28Z

sql/core/src/main/java/org/apache/spark/sql/execution/vectorized/OnHeapColumnVector.java

How does it compare to Unsafe.copyMemory()?

SparkQA · 2016-01-26T02:13:41Z

Test build #50044 has finished for PR 10820 at commit c58eedf.

This patch fails Spark unit tests.
This patch merges cleanly.
This patch adds no public classes.

SparkQA · 2016-01-26T02:20:45Z

Test build #50046 has finished for PR 10820 at commit f579889.

This patch fails Spark unit tests.
This patch merges cleanly.
This patch adds no public classes.

SparkQA · 2016-01-26T03:04:05Z

Test build #50051 has finished for PR 10820 at commit 2918077.

This patch fails Spark unit tests.
This patch merges cleanly.
This patch adds no public classes.

SparkQA · 2016-01-26T09:08:50Z

Test build #50080 has finished for PR 10820 at commit f378335.

This patch fails Spark unit tests.
This patch merges cleanly.
This patch adds the following public classes (experimental):
- public static final class Array extends ArrayData
- public static final class Struct extends InternalRow

SparkQA · 2016-01-26T19:52:36Z

Test build #50119 has finished for PR 10820 at commit 3d11c37.

This patch fails RAT tests.
This patch merges cleanly.
This patch adds no public classes.

rxin · 2016-01-26T20:07:51Z

sql/core/src/test/scala/org/apache/spark/sql/execution/vectorized/ColumnarBatchBenchmark.scala

utf8string supports reading offheap too so you can probably improve this substantially.

SparkQA · 2016-01-26T22:16:47Z

Test build #2460 has finished for PR 10820 at commit 3d11c37.

This patch fails PySpark unit tests.
This patch merges cleanly.
This patch adds no public classes.

SparkQA · 2016-01-27T00:49:40Z

Test build #2464 has finished for PR 10820 at commit 3d11c37.

This patch passes all tests.
This patch merges cleanly.
This patch adds no public classes.

rxin · 2016-01-27T01:33:50Z

Going to merge this in master. Thanks.

cloud-fan · 2017-11-16T13:57:38Z

sql/core/src/main/java/org/apache/spark/sql/execution/vectorized/ColumnVector.java

+        childCapacity *= DEFAULT_ARRAY_LENGTH;
+      }
+      this.childColumns = new ColumnVector[1];
+      this.childColumns[0] = ColumnVector.allocate(childCapacity, childType, memMode);


Why only grow the capacity for non-array type?

rxin reviewed Jan 19, 2016
View reviewed changes

Updates

f579889

nongli changed the title ~~[WIP] [SPARK-12854][SQL] Implement complex types support in ColumnarBatch~~ [SPARK-12854][SQL] Implement complex types support in ColumnarBatch Jan 25, 2016

nongli force-pushed the spark-12854 branch from c58eedf to f579889 Compare January 26, 2016 00:33

davies reviewed Jan 26, 2016
View reviewed changes

Fix asserts.

2918077

davies reviewed Jan 26, 2016
View reviewed changes

CR: made struct/array extend the catalyst types

f378335

Temporarily disable string types in random testing.

3d11c37

rxin reviewed Jan 26, 2016
View reviewed changes

asfgit closed this in 5551273 Jan 27, 2016

nongli deleted the spark-12854 branch January 27, 2016 04:11

cloud-fan reviewed Nov 16, 2017

View reviewed changes

[SPARK-12854][SQL] Implement complex types support in ColumnarBatch #10820

[SPARK-12854][SQL] Implement complex types support in ColumnarBatch #10820

Uh oh!

Conversation

nongli commented Jan 18, 2016

Uh oh!

SparkQA commented Jan 19, 2016

Uh oh!

rxin commented Jan 19, 2016

Uh oh!

rxin Jan 19, 2016

Choose a reason for hiding this comment

Uh oh!

nongli Jan 19, 2016

Choose a reason for hiding this comment

Uh oh!

kiszk commented Jan 20, 2016

Uh oh!

nongli commented Jan 25, 2016

Uh oh!

SparkQA commented Jan 26, 2016

Uh oh!

davies Jan 26, 2016

Choose a reason for hiding this comment

Uh oh!

nongli Jan 26, 2016

Choose a reason for hiding this comment

Uh oh!

davies Jan 26, 2016

Choose a reason for hiding this comment

Uh oh!

SparkQA commented Jan 26, 2016

Uh oh!

SparkQA commented Jan 26, 2016

Uh oh!

SparkQA commented Jan 26, 2016

Uh oh!

SparkQA commented Jan 26, 2016

Uh oh!

SparkQA commented Jan 26, 2016

Uh oh!

rxin Jan 26, 2016

Choose a reason for hiding this comment

Uh oh!

SparkQA commented Jan 26, 2016

Uh oh!

SparkQA commented Jan 27, 2016

Uh oh!

rxin commented Jan 27, 2016

Uh oh!

cloud-fan Nov 16, 2017

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

6 participants