Issue 86 : Add support for Datasource V2 : ORC by maheshk114 · Pull Request #85 · qubole/spark-acid

maheshk114 · 2020-07-25T16:55:59Z

No description provided.

amoghmargoor · 2020-08-18T23:46:26Z

src/main/scala/com/qubole/spark/hiveacid/reader/hive/HiveAcidReader.scala

+      colNames, colTypes) _
+
+    //TODO :Its a ugly hack, but avoids lots of duplicate code.
+    val rdd = new HiveAcidRDD(


Are we doing this only to get Splits ?

amoghmargoor

I have given one pass and see there is load code lift from Spark code base. We would like to understand more about why was that required and why we couldn't have just extended those classes or composed it or use them as it is.

amoghmargoor · 2020-08-18T23:47:53Z

src/main/scala/com/qubole/spark/hiveacid/datasource/HiveAcidDataSourceV2.scala

+  }
+
+  def keyPrefix() : String = {
+    "HiveAcidV2"


Would table created using 'USING HiveAcid' would never be able to use DSv2 untill not recreated ?

Yes. Unless we make V2 the default reader.

amoghmargoor · 2020-08-18T23:53:09Z

src/main/scala/com/qubole/spark/hiveacid/hive/HiveAcidMetadata.scala

+ * @param identifier - table identifier
+ * @param hiveConf hiveConf - hive conf
 */
-class HiveAcidMetadata(sparkSession: SparkSession,


This is an open API and it can break clients using it.

looks like the changes are not required, will revert it.

amoghmargoor · 2020-08-19T00:41:11Z

src/main/scala/com/qubole/spark/hiveacid/reader/TableReader.scala

+      partitions)
+  }
+
+  def getReader(requiredColumns: Array[String],


I think we are returning list of Partitions here. We should not be calling it as getReader.

makes sense, shall change it to getPartitionV2

amoghmargoor · 2020-08-19T00:46:17Z

src/main/scala/com/qubole/spark/hiveacid/reader/hive/HiveAcidReader.scala

   *
   * @return - Returns RDD on top of partitioned hive acid table
   */
-  def makeRDDForPartitionedTable(hiveAcidMetadata: HiveAcidMetadata,


Any reason for removing partitions from method and making it member variable of class. I think one HiveAcidReader can be used multiple times to read 1, 2 or more partitions. HiveAcidReader should map to just tables and not to <Table, partitions>

I think the logic is same. Earlier getTableReader used to create the HiveAcidReader and then call getRDD. Now i have just split it into two methods. The readerOptions is already created using the filters passed by user and from that the partitions list is generated. So i am not sure if we should change the partitions list after HiveAcidReader object is created.

amoghmargoor · 2020-08-19T00:50:42Z

src/main/scala/com/qubole/spark/hiveacid/reader/v2/HiveAcidColumnVector.java

+ * Spark's vectorized.ColumnVector, this column vector is used to adapt Hive ColumnVector with
+ * Spark ColumnarVector. This class is a copy of spark ColumnVector which is declared private.
+ */
+public class HiveAcidColumnVector extends org.apache.spark.sql.vectorized.ColumnVector {


This looks similar to org.apache.spark.sql.execution.datasources.orc.OrcColumnVector. Should we not be extending it or using it directly here. We should also be Format agnostic but we can use something like Factory etc to get ColumnVector based on format.

OrcColumnVector in spark is in private scope so had to copy it.

@maheshk114 : The class org.apache.spark.sql.vectorized.ColumnVector is public in spark 2.4.x. Check this https://github.com/apache/spark/blob/branch-2.4/sql/core/src/main/java/org/apache/spark/sql/execution/datasources/orc/OrcColumnVector.java#L36

this file is removed

amoghmargoor · 2020-08-19T00:56:25Z

src/main/scala/com/qubole/spark/hiveacid/reader/v2/OrcColumnarBatchReader.java

+ * After creating, `initialize` and `initBatch` should be called sequentially. This internally uses
+ * the Hive ACID Vectorized ORC reader to support reading of deleted and updated data.
+ */
+public class OrcColumnarBatchReader extends RecordReader<Void, ColumnarBatch> {


This looks similar to org.apache.spark.sql.execution.datasources.orc.OrcColumnarBatchReader ? Should we use that instead . what are the extra changes here ?

No it has some code from OrcColumnarBatchReader , but it has logic to support sargs and ACID table reader.

@maheshk114 : Can't we set searchArguments outside somewhere outside i.e before initializing this reader? There doesn't seem to be anything specific to the reader which we are setting in search arguments.

The search arguments are set in reader option, so its better to set with the reader code.

amoghmargoor · 2020-08-19T00:57:35Z

src/main/scala/com/qubole/spark/hiveacid/reader/v2/HiveAcidInputPartitionV2.scala

+import org.apache.spark.sql.catalyst.InternalRow
+import org.apache.spark.sql.types._
+
+private[hiveacid] class HiveAcidInputPartitionV2(split: HiveAcidPartition,


Instead of calling it HiveAcidInputPartitionV2, we should it ColumnarHiveAcidPartition.

I have kept the name inline with "org.apache.spark.sql.sources.v2" .

sourabh912 · 2020-08-09T01:34:32Z

src/main/scala/com/qubole/spark/hiveacid/HiveAcidAutoConvert.scala


+  private def convertV2(relation: HiveTableRelation): LogicalPlan = {
+    val serde = relation.tableMeta.storage.serde.getOrElse("").toLowerCase(Locale.ROOT)
+    if (!serde.equals("org.apache.hadoop.hive.ql.io.orc.orcserde")) {


nit: Use HiveSerde.sourceToSerDe() api to do the comparison.

sourabh912 · 2020-09-14T18:58:21Z

src/main/scala/com/qubole/spark/hiveacid/HiveAcidAutoConvert.scala

+  private def convertV2(relation: HiveTableRelation): LogicalPlan = {
+    val serde = relation.tableMeta.storage.serde.getOrElse("").toLowerCase(Locale.ROOT)
+    if (!serde.equals("org.apache.hadoop.hive.ql.io.orc.orcserde")) {
+      // Only ORC formatted is supported as of now. If its not ORC, then fallback to


Add a log line stating the reason to fallback to v1?

sourabh912 · 2020-09-14T19:07:47Z

src/main/scala/com/qubole/spark/hiveacid/HiveAcidAutoConvert.scala

      case relation: HiveTableRelation
        if DDLUtils.isHiveTable(relation.tableMeta) && isConvertible(relation) =>
-        convert(relation)
+          if (spark.conf.get("spark.acid.use.datasource.v2", "false").toBoolean) {


nit: change config to spark.hive-acid.read.datasource.version. The config will accept "v1" and "v2" as valid values.

make sense. This way the config can be extended if later we want to support "v3" reader. Done

sourabh912 · 2020-09-15T01:30:57Z

src/main/scala/com/qubole/spark/hiveacid/datasource/HiveAcidDataSourceV2.scala

+
+class HiveAcidDataSourceV2 extends DataSourceV2 with ReadSupport with Logging {
+  override def  createReader (options: DataSourceOptions) : DataSourceReader = {
+    logInfo("Using Datasource V2 for table" + options.tableName.get)


nit: Change logging to "Creating datasource v2 reader for table ....."

sourabh912 · 2020-09-15T23:00:42Z

src/main/scala/com/qubole/spark/hiveacid/datasource/HiveAcidDataSourceV2.scala

+                                  options.databaseName.get, options.tableName.get)
+  }
+
+  def keyPrefix() : String = {


@maheshk114 : Where is this method used? Since it is not an overridden method, I am not sure if it is getting used anywhere in spark/acid.

i have kept it for future use. But i think we should use the same "HiveAcid" for both the readers and switch is based on config.

sourabh912 · 2020-09-16T23:46:43Z

src/main/scala/com/qubole/spark/hiveacid/reader/v2/HiveAcidInputPartitionReaderV2.scala

+    //TODO: Need to generalize it for supporting other kind of file format.
+    orcColumnarBatchReader.initialize(fileSplit, taskAttemptContext)
+    orcColumnarBatchReader.initBatch(readerLocal.getSchema, requestedColIds,
+      requiredFields, partitionSchema, partitionValues, isFullAcidTable && !fileSplit.isOriginal)


Shouldn't we be creating orcColumnarBatchReader only if these conditions are met:
isFullAcidTable && !fileSplit.isOriginal

Right now orcColumnarBatchReader would work for both acid and non acid reads because we are creating a baseRecordReader but creating AcidRecordReader only if fullAcidScan is true.

Here we should only create vectorized reader for acid reads. For normal reads, we should simply throw an exception.

fullAcidScan is for full ACID table. The base reader will be used for insert-only table. In that case there will not be any filtering from Hive ACID reader as it does not support update/delete. I think spark native reader does not support read from insert-only Hive ACID tables. If the user has a ACID table with old file format(original), then HiveAcid reader should be able to read it. I don't think we should throw exception.

sourabh912 · 2020-09-16T23:58:26Z

src/main/scala/com/qubole/spark/hiveacid/reader/v2/OrcColumnarBatchReader.java

+import com.qubole.shaded.hadoop.hive.ql.io.sarg.*;
+
+/**
+ * After creating, `initialize` and `initBatch` should be called sequentially. This internally uses


To avoid code duplication, we can extend spark's OrcColumnarBatchReader and override the methods required. For example -

initialize() - where we can initialize our vectorizedAcidRecordReader.

nextKeyValue() -> to read next batch using vectorizedAcidRecordReader.

getProgress() -> return progress of acid reader.

i think most of the code is required ..like putRepeatingValues,putNonNullValues is required as spark reader does not take care of the "selected" rows. In case of Hive ACID row batch, some of the rows needs to be ignored. I think that logic is not present in spark currently. Anyways i have extended it to avoid few duplicate method like putDecimalWritables

sourabh912 · 2020-09-16T23:59:19Z

src/main/scala/com/qubole/spark/hiveacid/reader/v2/HiveAcidInputPartitionReaderV2.scala

+        extends InputPartitionReader[ColumnarBatch] {
+  //TODO : Need to get a unique id to cache the jobConf.
+  private val jobConf = new JobConf(broadcastedConf.value.value)
+  private val orcColumnarBatchReader = new OrcColumnarBatchReader(1024)


Shouldn't we make capacity configurable? Currently it is hard coded to 1024.

i am not sure if anybody can use this configuration.But for perf test and all i have am making it configurable using "spark.hive.acid.default.row.batch.size"

sourabh912 · 2020-09-17T01:23:32Z

src/main/scala/com/qubole/spark/hiveacid/reader/TableReader.scala

                                    hiveAcidMetadata: HiveAcidMetadata) extends Logging {

-  def getRdd(requiredColumns: Array[String],
+  def getTableReader(requiredColumns: Array[String],


make this private?

sourabh912 · 2020-09-17T01:25:36Z

src/main/scala/com/qubole/spark/hiveacid/reader/TableReader.scala

+      partitions)
+  }

+  def getPartitionsV2(requiredColumns: Array[String],


nit: We can rename it to getInputPartitionsInBatch.

I think it is better to keep the naming with "V2" to make it inline with spark naming for v2 reader.

sourabh912 · 2020-09-30T18:52:13Z

@maheshk114 : I was testing the changes locally. I created a simple unpartitioned table and tried reading data in a range. I got the wrong results. Here is the query I tried:

sql("select * from t1 where id > 15 and id < 25").show

the output was:

+----------+---+
|      name| id|
+----------+---+
|new_name11| 11|
|new_name12| 12|
|new_name13| 13|
|new_name14| 14|
|new_name15| 15|
|new_name16| 16|
|new_name17| 17|
|new_name18| 18|
|new_name19| 19|
|new_name20| 20|
|new_name21| 21|
|new_name22| 22|
|new_name23| 23|
|new_name24| 24|
|new_name25| 25|
|new_name26| 26|
|new_name27| 27|
|new_name28| 28|
|new_name29| 29|
|new_name30| 30|
+----------+---+

Table schema is:

org.apache.spark.sql.types.StructType = StructType(StructField(name,StringType,true), StructField(id,IntegerType,true))

I get the expected result if I disable dsv2 reader. Let me know if you need more details regarding the above query.

Have you guys tested this PR in your environment?

maheshk114 · 2020-10-16T03:58:55Z

@sourabh912 thanks for pointing it out. The issue is that ORC does not support row level filtering and thus the filtering has to be done again in spark. We have done some testing internally. But the hive/spark/ORC version is different. For this specific issue, this was already fixed internally, i forgot to merge it to this PR.

maheshk114 changed the title ~~Dv2 2~~ Issue 86 : Add support for Datasource V2 Jul 25, 2020

maheshk114 changed the title ~~Issue 86 : Add support for Datasource V2~~ Issue 86 : Add support for Datasource V2 : ORC Jul 25, 2020

maheshk114 force-pushed the dv2_2 branch from 3fd63a3 to 0fe9c79 Compare July 25, 2020 18:26

amoghmargoor self-requested a review July 26, 2020 02:45

amoghmargoor self-assigned this Jul 26, 2020

maheshk114 force-pushed the dv2_2 branch from 55ad669 to 8fc2efe Compare July 26, 2020 05:42

amoghmargoor reviewed Aug 18, 2020

View reviewed changes

amoghmargoor reviewed Aug 19, 2020

View reviewed changes

maheshk114 force-pushed the dv2_2 branch 3 times, most recently from d16e9e1 to e7abcb5 Compare August 24, 2020 07:13

sourabh912 reviewed Sep 17, 2020

View reviewed changes

maheshk114 and others added 5 commits October 16, 2020 09:24

Issue 86 : Add support for Datasource V2 : ORC

24e869b

Some code refactoring

4b7d107

review comment fix

4f9d40a

review comment fix 2

ef00a5d

Fixed issue with range filter

217d8fd

maheshk114 force-pushed the dv2_2 branch from 5a84118 to 217d8fd Compare October 16, 2020 03:55

Comments

Conversation

maheshk114 commented Jul 25, 2020

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

amoghmargoor left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

sourabh912 commented Sep 30, 2020 •

edited

Loading