Vectorized Read Issue

[asheeshgarg]

Apache Iceberg version

0.14.0

Query engine

None

Please describe the bug 🐞

@nastra

I am using Vectorized read from Java API and able to load the data from Vectorized root from iceberg batches to arrow.

Some of the records I am not able to build in Arrow cache as the schema return are different

VectorizedTableScanIterable iter = new VectorizedTableScanIterable(filteredScan, batchSize, false);) {
for (ColumnarBatch batch : iter) {
VectorSchemaRoot batchRoot = batch.createVectorSchemaRootFromVectors();

batchRoot schema are returned different for a given table.

Do we type cast the results to adhere to actual schema of Iceberg when we read the base files returned by the read?

[nastra]

@asheeshgarg thanks for the ticket. For anyone looking at this ticket it would be quite helpful if you could add additional details about the issue you're seeing (what is expected, what differs in the schema exactly, how does the schema look like, ...)

[asheeshgarg]

@nastra filled in the missing bits
So this schema that is define in Iceberg entity_status is UTF8
Schema<entity_id: Utf8, entity_name: Utf8, entity_status: Utf8,
This is what is been generated by batchRoot Schema
Schema<entity_id: Utf8, entity_name: Utf8, entity_status: Int(32, true),

There are other fields also coming as Int which are UTF8 it throws a error like below because of type mismatch. SPark Able to read the data fine

java.lang.IndexOutOfBoundsException: index: 0, length: 8388608 (expected: range(0, 15888))
at org.apache.arrow.memory.ArrowBuf.checkIndex(ArrowBuf.java:701)
at org.apache.arrow.memory.ArrowBuf.setBytes(ArrowBuf.java:955)
at org.apache.arrow.vector.BaseFixedWidthVector.reAlloc(BaseFixedWidthVector.java:451)
at org.apache.arrow.vector.BaseFixedWidthVector.setValueCount(BaseFixedWidthVector.java:732)
at org.apache.arrow.vector.VectorSchemaRoot.setRowCount(VectorSchemaRoot.java:240)
at org.apache.arrow.vector.VectorLoader.load(VectorLoader.java:86)

Expected behavior batches should return the data of schema type defined in iceberg/hive.

[nastra]

See

iceberg/arrow/src/test/java/org/apache/iceberg/arrow/vectorized/ArrowReaderTest.java

Lines 149 to 210 in c07f2aa

	/**
	* This test writes each partition with constant value rows. The Arrow vectors returned are mostly
	* of type int32 which is unexpected. This is happening because of dictionary encoding at the
	* storage level.
	*
	* <p>Following are the expected and actual Arrow schema:
	*
	* <pre>
	* Expected Arrow Schema:
	* timestamp: Timestamp(MICROSECOND, null) not null,
	* timestamp_nullable: Timestamp(MICROSECOND, null),
	* boolean: Bool not null,
	* boolean_nullable: Bool,
	* int: Int(32, true) not null,
	* int_nullable: Int(32, true),
	* long: Int(64, true) not null,
	* long_nullable: Int(64, true),
	* float: FloatingPoint(SINGLE) not null,
	* float_nullable: FloatingPoint(SINGLE),
	* double: FloatingPoint(DOUBLE) not null,
	* double_nullable: FloatingPoint(DOUBLE),
	* timestamp_tz: Timestamp(MICROSECOND, UTC) not null,
	* timestamp_tz_nullable: Timestamp(MICROSECOND, UTC),
	* string: Utf8 not null,
	* string_nullable: Utf8,
	* bytes: Binary not null,
	* bytes_nullable: Binary,
	* date: Date(DAY) not null,
	* date_nullable: Date(DAY),
	* int_promotion: Int(32, true) not null
	*
	* Actual Arrow Schema:
	* timestamp: Int(32, true) not null,
	* timestamp_nullable: Int(32, true),
	* boolean: Bool not null,
	* boolean_nullable: Bool,
	* int: Int(32, true) not null,
	* int_nullable: Int(32, true),
	* long: Int(32, true) not null,
	* long_nullable: Int(32, true),
	* float: Int(32, true) not null,
	* float_nullable: Int(32, true),
	* double: Int(32, true) not null,
	* double_nullable: Int(32, true),
	* timestamp_tz: Int(32, true) not null,
	* timestamp_tz_nullable: Int(32, true),
	* string: Int(32, true) not null,
	* string_nullable: Int(32, true),
	* bytes: Int(32, true) not null,
	* bytes_nullable: Int(32, true),
	* date: Date(DAY) not null,
	* date_nullable: Date(DAY),
	* int_promotion: Int(32, true) not null
	* </pre>
	*
	* <p>TODO: fix the returned Arrow vectors to have vector types consistent with Iceberg types.
	*
	* <p>Read all rows and columns from the table without any filter. The test asserts that the Arrow
	* {@link VectorSchemaRoot} contains the expected schema and expected vector types. Then the test
	* asserts that the vectors contains expected values. The test also asserts the total number of
	* rows match the expected value.
	*/

for some background why the issue currently exists.
I have opened #3024 a while ago to address this issue.

[nastra]

@rdblue thoughts on getting the above issue fixed?

[nazq]

Happy to create a PR @rdblue , just want to make sure we're on the right track here

[nastra]

@nazq just FYI, there's already #3024 that addresses this issue.

[nazq]

Excellent. Thanks for the pointer @nastra

[asheeshgarg]

@nastra thanks for the references seems to be the case. @rdblue this seems to be really the case with lot of datasets. Do we have any time line when #3024 will be merged.

[rdblue]

I agree with the analysis that the problem is that we are returning dictionary-encoded Arrow vectors. Maybe we're not doing that the right way. I'll take a look at #3024.

[asheeshgarg]

@nastra @nazq so in the mean time we have merged the pull request and bundle a local jar with the #3024
It work fine for most of the columns but we are getting
java.lang.IndexOutOfBoundsException: index: 32749, length: 32 (expected: range(0, 32768))
at org.apache.arrow.memory.ArrowBuf.checkIndex(ArrowBuf.java:701)
at org.apache.arrow.memory.ArrowBuf.setBytes(ArrowBuf.java:765)
at org.apache.arrow.vector.BaseVariableWidthVector.setBytes(BaseVariableWidthVector.java:1244)
at org.apache.arrow.vector.BaseVariableWidthVector.set(BaseVariableWidthVector.java:1025)
at org.apache.iceberg.arrow.DictEncodedArrowConverter.lambda$toVarCharVector$5(DictEncodedArrowConverter.java:153)
at org.apache.iceberg.arrow.DictEncodedArrowConverter.initVector(DictEncodedArrowConverter.java:201)
at org.apache.iceberg.arrow.DictEncodedArrowConverter.toVarCharVector(DictEncodedArrowConverter.java:150)
at org.apache.iceberg.arrow.DictEncodedArrowConverter.toArrowVector(DictEncodedArrowConverter.java:47)
at org.apache.iceberg.arrow.vectorized.ColumnVector.getArrowVector(ColumnVector.java:66)
at java.base/java.util.stream.ReferencePipeline$3$1.accept(ReferencePipeline.java:197)
at java.base/java.util.Spliterators$ArraySpliterator.forEachRemaining(Spliterators.java:992)
at java.base/java.util.stream.AbstractPipeline.copyInto(AbstractPipeline.java:509)
at java.base/java.util.stream.AbstractPipeline.wrapAndCopyInto(AbstractPipeline.java:499)
at java.base/java.util.stream.AbstractPipeline.evaluate(AbstractPipeline.java:575)
at java.base/java.util.stream.AbstractPipeline.evaluateToArrayNode(AbstractPipeline.java:260)
at java.base/java.util.stream.ReferencePipeline.toArray(ReferencePipeline.java:616)
at org.apache.iceberg.arrow.vectorized.ColumnarBatch.createVectorSchemaRootFromVectors(ColumnarBatch.java:58)
at com.ReadIcebergTableTestV3.main(ReadIcebergTableTestV3.java:54)
when reading columns where distinct count is large for the dictionary. Will try to create a test case to replicate it.

@rdblue Arrow also added dataset where we can read the tabular data Arrow Vectors I was able to read the parquet file directly not sure if we like to add read support using the Arrow Dataset.

[marun224]

@nastra, @mayursrivastava, @amogh-jahagirdar, @rdblue

It seems above error is occurring because of issue at line no 173 in the class DictEncodedArrowConverter. I have commented on the PR #3024.

Can somebody take a look it and advise?

[github-actions]

This issue has been automatically marked as stale because it has been open for 180 days with no activity. It will be closed in next 14 days if no further activity occurs. To permanently prevent this issue from being considered stale, add the label 'not-stale', but commenting on the issue is preferred when possible.

[github-actions]

This issue has been closed because it has not received any activity in the last 14 days since being marked as 'stale'