umehrot2 commented on a change in pull request #2651:
URL: https://github.com/apache/hudi/pull/2651#discussion_r599160889
##########
File path:
hudi-client/hudi-client-common/src/main/java/org/apache/hudi/config/HoodieWriteConfig.java
##########
@@ -179,6 +179,9 @@
public static final String EXTERNAL_RECORD_AND_SCHEMA_TRANSFORMATION =
AVRO_SCHEMA + ".externalTransformation";
public static final String DEFAULT_EXTERNAL_RECORD_AND_SCHEMA_TRANSFORMATION
= "false";
+ public static final String MAX_LISTING_PARALLELISM =
"hoodie.max.list.file.parallelism";
+ public static final Integer DEFAULT_MAX_LISTING_PARALLELISM = 200;
Review comment:
- I think its fine to use the `DEFAULT_PARALLELISM` i.e `1500` as the
default. It is what we use in `FileSystemBackedTableMetadata` as well.
- We should add a method here to get this configuration just like all other
configurations.
##########
File path:
hudi-spark-datasource/hudi-spark/src/main/scala/org/apache/hudi/HoodieFileIndex.scala
##########
@@ -0,0 +1,317 @@
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one or more
+ * contributor license agreements. See the NOTICE file distributed with
+ * this work for additional information regarding copyright ownership.
+ * The ASF licenses this file to You under the Apache License, Version 2.0
+ * (the "License"); you may not use this file except in compliance with
+ * the License. You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package org.apache.hudi
+
+import java.util.Properties
+
+import scala.collection.JavaConverters._
+
+import org.apache.hadoop.fs.{FileStatus, Path}
+import org.apache.hudi.client.common.HoodieSparkEngineContext
+import org.apache.hudi.common.config.{HoodieMetadataConfig,
SerializableConfiguration}
+import org.apache.hudi.common.engine.HoodieLocalEngineContext
+import org.apache.hudi.common.fs.FSUtils
+import org.apache.hudi.common.model.HoodieBaseFile
+import org.apache.hudi.common.table.{HoodieTableMetaClient,
TableSchemaResolver}
+import org.apache.hudi.common.table.view.HoodieTableFileSystemView
+import org.apache.hudi.config.HoodieWriteConfig
+import org.apache.spark.api.java.JavaSparkContext
+import org.apache.spark.internal.Logging
+import org.apache.spark.sql.catalyst.{InternalRow, expressions}
+import org.apache.spark.sql.SparkSession
+import org.apache.spark.sql.avro.SchemaConverters
+import org.apache.spark.sql.catalyst.expressions.{AttributeReference,
BoundReference, Expression, InterpretedPredicate}
+import org.apache.spark.sql.catalyst.util.{CaseInsensitiveMap, DateTimeUtils}
+import org.apache.spark.sql.execution.datasources.{FileIndex,
PartitionDirectory, PartitionUtils}
+import org.apache.spark.sql.internal.SQLConf
+import org.apache.spark.sql.types.StructType
+import org.apache.spark.unsafe.types.UTF8String
+
+/**
+ * A File Index which support partition prune for hoodie snapshot and
read-optimized
+ * query.
+ * Main steps to get the file list for query:
+ * 1、Load all files and partition values from the table path.
+ * 2、Do the partition prune by the partition filter condition.
+ *
+ * There are 3 cases for this:
+ * 1、If the partition columns size is equal to the actually partition path
level, we
+ * read it as partitioned table.(e.g partition column is "dt", the partition
path is "2021-03-10")
+ *
+ * 2、If the partition columns size is not equal to the partition path level,
but the partition
+ * column size is "1" (e.g. partition column is "dt", but the partition path
is "2021/03/10"
+ * who'es directory level is 3).We can still read it as a partitioned table.
We will mapping the
+ * partition path (e.g. 2021/03/10) to the only partition column (e.g. "dt").
+ *
+ * 3、Else the the partition columns size is not equal to the partition
directory level and the
+ * size is great than "1" (e.g. partition column is "dt,hh", the partition
path is "2021/03/10/12")
+ * , we read it as a None Partitioned table because we cannot know how to
mapping the partition
+ * path with the partition columns in this case.
+ */
+case class HoodieFileIndex(
+ spark: SparkSession,
+ basePath: String,
+ schemaSpec: Option[StructType],
+ options: Map[String, String])
+ extends FileIndex with Logging {
+
+ @transient private val hadoopConf = spark.sessionState.newHadoopConf()
+ private lazy val metaClient = HoodieTableMetaClient
+ .builder().setConf(hadoopConf).setBasePath(basePath).build()
+
+ @transient private val queryPath = new Path(options.getOrElse("path",
"'path' option required"))
+ /**
+ * Get the schema of the table.
+ */
+ lazy val schema: StructType = schemaSpec.getOrElse({
+ val schemaUtil = new TableSchemaResolver(metaClient)
+ SchemaConverters.toSqlType(schemaUtil.getTableAvroSchema)
+ .dataType.asInstanceOf[StructType]
+ })
+
+ /**
+ * Get the partition schema from the hoodie.properties.
+ */
+ private lazy val _partitionSchemaFromProperties: StructType = {
+ val tableConfig = metaClient.getTableConfig
+ val partitionColumns = tableConfig.getPartitionColumns
+ val nameFieldMap = schema.fields.map(filed => filed.name -> filed).toMap
+
+ if (partitionColumns.isPresent) {
+ val partitionFields = partitionColumns.get().map(column =>
+ nameFieldMap.getOrElse(column, throw new
IllegalArgumentException(s"Cannot find column " +
+ s"$column in the schema[${schema.fields.mkString(",")}]")))
+ new StructType(partitionFields)
+ } else { // If the partition columns have not stored in
hoodie.properites(the table that was
+ // created earlier), we trait it as a none-partitioned table.
+ new StructType()
+ }
+ }
+
+ @transient @volatile private var fileSystemView: HoodieTableFileSystemView =
_
+ @transient @volatile private var cachedAllInputFiles: Array[HoodieBaseFile]
= _
+ @transient @volatile private var cachedFileSize: Long = 0L
+ @transient @volatile private var cachedAllPartitionPaths: Seq[PartitionPath]
= _
+
+ @volatile private var queryAsNonePartitionedTable: Boolean = _
+
+ refresh()
+
+ override def rootPaths: Seq[Path] = queryPath :: Nil
+
+ override def listFiles(partitionFilters: Seq[Expression],
+ dataFilters: Seq[Expression]):
Seq[PartitionDirectory] = {
+ if (queryAsNonePartitionedTable) { // Read as None Partitioned table.
+ Seq(PartitionDirectory(InternalRow.empty, allFiles))
+ } else {
+ // Prune the partition path by the partition filters
+ val prunedPartitions = prunePartition(cachedAllPartitionPaths,
partitionFilters)
+ prunedPartitions.map { partition =>
+ val fileStatues =
fileSystemView.getLatestBaseFiles(partition.partitionPath).iterator()
+ .asScala.toSeq
+ .map(_.getFileStatus)
+ PartitionDirectory(partition.values, fileStatues)
+ }
+ }
+ }
+
+ override def inputFiles: Array[String] = {
+ cachedAllInputFiles.map(_.getFileStatus.getPath.toString)
+ }
+
+ override def refresh(): Unit = {
+ val startTime = System.currentTimeMillis()
+ val partitionFiles = loadPartitionPathFiles()
+ val allFiles = partitionFiles.values.reduceOption(_ ++ _)
+ .getOrElse(Array.empty[FileStatus])
+
+ metaClient.reloadActiveTimeline()
+ val activeInstants =
metaClient.getActiveTimeline.getCommitsTimeline.filterCompletedInstants
+ fileSystemView = new HoodieTableFileSystemView(metaClient, activeInstants,
allFiles)
+ cachedAllInputFiles =
fileSystemView.getLatestBaseFiles.iterator().asScala.toArray
+ cachedAllPartitionPaths = partitionFiles.keys.toSeq
+ cachedFileSize = cachedAllInputFiles.map(_.getFileLen).sum
+
+ // If the partition value contains InternalRow.empty, we query it as a
none partitioned table.
+ queryAsNonePartitionedTable = cachedAllPartitionPaths
+ .exists(p => p.values == InternalRow.empty)
+ val flushSpend = System.currentTimeMillis() - startTime
+ logInfo(s"Refresh for table ${metaClient.getTableConfig.getTableName}," +
+ s" spend: $flushSpend ms")
+ }
+
+ override def sizeInBytes: Long = {
+ cachedFileSize
+ }
+
+ override def partitionSchema: StructType = {
+ if (queryAsNonePartitionedTable) {
+ // If we read it as None Partitioned table, we should not
+ // return the partition schema.
+ new StructType()
+ } else {
+ _partitionSchemaFromProperties
+ }
+ }
+
+ /**
+ * Get the data schema of the table.
+ * @return
+ */
+ def dataSchema: StructType = {
+ val partitionColumns = partitionSchema.fields.map(_.name).toSet
+ StructType(schema.fields.filterNot(f => partitionColumns.contains(f.name)))
+ }
+
+ def allFiles: Seq[FileStatus] = cachedAllInputFiles.map(_.getFileStatus)
+
+ private def prunePartition(partitionPaths: Seq[PartitionPath],
+ predicates: Seq[Expression]): Seq[PartitionPath]
= {
+
+ val partitionColumnNames = partitionSchema.fields.map(_.name).toSet
+ val partitionPruningPredicates = predicates.filter {
+ _.references.map(_.name).toSet.subsetOf(partitionColumnNames)
+ }
+ if (partitionPruningPredicates.nonEmpty) {
+ val predicate = partitionPruningPredicates.reduce(expressions.And)
+
+ val boundPredicate = InterpretedPredicate.create(predicate.transform {
+ case a: AttributeReference =>
+ val index = partitionSchema.indexWhere(a.name == _.name)
+ BoundReference(index, partitionSchema(index).dataType, nullable =
true)
+ })
+
+ val partitionPruned = partitionPaths.filter {
+ case PartitionPath(values, _) => boundPredicate.eval(values)
+ }
+ logInfo(s"Total partition size is: ${partitionPaths.size}," +
+ s" after partition prune size is: ${partitionPruned.size}")
+ partitionPruned
+ } else {
+ partitionPaths
+ }
+ }
+
+ /**
+ * Load all partition paths and it's files under the query table path.
+ */
+ private def loadPartitionPathFiles(): Map[PartitionPath, Array[FileStatus]]
= {
+ val sparkEngine = new HoodieSparkEngineContext(new
JavaSparkContext(spark.sparkContext))
+ val properties = new Properties()
+ properties.putAll(options.asJava)
+ val metadataConfig =
HoodieMetadataConfig.newBuilder.fromProperties(properties).build()
+
+ val queryPartitionPath = FSUtils.getRelativePartitionPath(new
Path(basePath), queryPath)
+ // Load all the partition path from the basePath, and filter by the query
partition path.
+ val partitionPaths = FSUtils.getAllPartitionPaths(sparkEngine,
metadataConfig, basePath).asScala
+ .filter(_.startsWith(queryPartitionPath))
+
+ val maxListParallelism =
options.get(HoodieWriteConfig.MAX_LISTING_PARALLELISM)
+
.map(_.toInt).getOrElse(HoodieWriteConfig.DEFAULT_MAX_LISTING_PARALLELISM.intValue())
+ val serializableConf = new
SerializableConfiguration(spark.sessionState.newHadoopConf())
+ val partitionSchema = _partitionSchemaFromProperties
+ val timeZoneId = CaseInsensitiveMap(options)
+ .get(DateTimeUtils.TIMEZONE_OPTION)
+ .getOrElse(SQLConf.get.sessionLocalTimeZone)
+
+ // List files in all of the partition path.
+ val partition2Files =
+ spark.sparkContext.parallelize(partitionPaths,
Math.min(partitionPaths.size, maxListParallelism))
+ .map { partitionPath =>
+ val partitionRow = if (partitionSchema.fields.length == 0) {
+ // This is a none partitioned table
+ InternalRow.empty
+ } else {
+ val partitionFragments = partitionPath.split("/")
+
+ if (partitionFragments.length != partitionSchema.fields.length &&
+ partitionSchema.fields.length == 1) {
+ // If the partition column size is not equal to the partition
fragment size
+ // and the partition column size is 1, we map the whole
partition path
+ // to the partition column which can benefit from the partition
prune.
+ InternalRow.fromSeq(Seq(UTF8String.fromString(partitionPath)))
+ } else if (partitionFragments.length !=
partitionSchema.fields.length &&
+ partitionSchema.fields.length > 1) {
+ // If the partition column size is not equal to the partition
fragments size
+ // and the partition column size > 1, we do not know how to map
the partition
+ // fragments to the partition columns. So we trait it as a None
Partitioned Table
+ // for the query which do not benefit from the partition prune.
+ InternalRow.empty
Review comment:
Shall we log a warning that `Partition pruning will not work`
##########
File path:
hudi-spark-datasource/hudi-spark/src/main/scala/org/apache/hudi/HoodieSparkSqlWriter.scala
##########
@@ -112,12 +112,15 @@ private[hudi] object HoodieSparkSqlWriter {
val archiveLogFolder = parameters.getOrElse(
HoodieTableConfig.HOODIE_ARCHIVELOG_FOLDER_PROP_NAME, "archived")
+ val partitionColumns =
parameters.getOrElse(DataSourceWriteOptions.PARTITIONPATH_FIELD_OPT_KEY, null)
+
val tableMetaClient = HoodieTableMetaClient.withPropertyBuilder()
.setTableType(tableType)
.setTableName(tblName)
.setArchiveLogFolder(archiveLogFolder)
.setPayloadClassName(parameters(PAYLOAD_CLASS_OPT_KEY))
.setPreCombineField(parameters.getOrDefault(PRECOMBINE_FIELD_OPT_KEY, null))
+ .setPartitionColumns(partitionColumns)
Review comment:
We should make similar change at other places as well:
- There is `bootstrap` code in the same file where we create the meta client
- Making the same change in `DeltaSync` is important so tables created via
`HoodieDeltaStreamer` can also make use of this feature
##########
File path:
hudi-spark-datasource/hudi-spark/src/main/scala/org/apache/hudi/HoodieFileIndex.scala
##########
@@ -0,0 +1,317 @@
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one or more
+ * contributor license agreements. See the NOTICE file distributed with
+ * this work for additional information regarding copyright ownership.
+ * The ASF licenses this file to You under the Apache License, Version 2.0
+ * (the "License"); you may not use this file except in compliance with
+ * the License. You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package org.apache.hudi
+
+import java.util.Properties
+
+import scala.collection.JavaConverters._
+
+import org.apache.hadoop.fs.{FileStatus, Path}
+import org.apache.hudi.client.common.HoodieSparkEngineContext
+import org.apache.hudi.common.config.{HoodieMetadataConfig,
SerializableConfiguration}
+import org.apache.hudi.common.engine.HoodieLocalEngineContext
+import org.apache.hudi.common.fs.FSUtils
+import org.apache.hudi.common.model.HoodieBaseFile
+import org.apache.hudi.common.table.{HoodieTableMetaClient,
TableSchemaResolver}
+import org.apache.hudi.common.table.view.HoodieTableFileSystemView
+import org.apache.hudi.config.HoodieWriteConfig
+import org.apache.spark.api.java.JavaSparkContext
+import org.apache.spark.internal.Logging
+import org.apache.spark.sql.catalyst.{InternalRow, expressions}
+import org.apache.spark.sql.SparkSession
+import org.apache.spark.sql.avro.SchemaConverters
+import org.apache.spark.sql.catalyst.expressions.{AttributeReference,
BoundReference, Expression, InterpretedPredicate}
+import org.apache.spark.sql.catalyst.util.{CaseInsensitiveMap, DateTimeUtils}
+import org.apache.spark.sql.execution.datasources.{FileIndex,
PartitionDirectory, PartitionUtils}
+import org.apache.spark.sql.internal.SQLConf
+import org.apache.spark.sql.types.StructType
+import org.apache.spark.unsafe.types.UTF8String
+
+/**
+ * A File Index which support partition prune for hoodie snapshot and
read-optimized
+ * query.
+ * Main steps to get the file list for query:
+ * 1、Load all files and partition values from the table path.
+ * 2、Do the partition prune by the partition filter condition.
+ *
+ * There are 3 cases for this:
+ * 1、If the partition columns size is equal to the actually partition path
level, we
+ * read it as partitioned table.(e.g partition column is "dt", the partition
path is "2021-03-10")
+ *
+ * 2、If the partition columns size is not equal to the partition path level,
but the partition
+ * column size is "1" (e.g. partition column is "dt", but the partition path
is "2021/03/10"
+ * who'es directory level is 3).We can still read it as a partitioned table.
We will mapping the
+ * partition path (e.g. 2021/03/10) to the only partition column (e.g. "dt").
+ *
+ * 3、Else the the partition columns size is not equal to the partition
directory level and the
+ * size is great than "1" (e.g. partition column is "dt,hh", the partition
path is "2021/03/10/12")
+ * , we read it as a None Partitioned table because we cannot know how to
mapping the partition
+ * path with the partition columns in this case.
+ */
+case class HoodieFileIndex(
+ spark: SparkSession,
+ basePath: String,
+ schemaSpec: Option[StructType],
+ options: Map[String, String])
+ extends FileIndex with Logging {
+
+ @transient private val hadoopConf = spark.sessionState.newHadoopConf()
+ private lazy val metaClient = HoodieTableMetaClient
+ .builder().setConf(hadoopConf).setBasePath(basePath).build()
Review comment:
Instead of re-creating it here, can we pass the metaclient created in
`DefaultSource` ?
##########
File path:
hudi-spark-datasource/hudi-spark/src/main/scala/org/apache/hudi/DefaultSource.scala
##########
@@ -79,39 +82,58 @@ class DefaultSource extends RelationProvider
val allPaths = path.map(p => Seq(p)).getOrElse(Seq()) ++ readPaths
val fs = FSUtils.getFs(allPaths.head,
sqlContext.sparkContext.hadoopConfiguration)
- val globPaths = HoodieSparkUtils.checkAndGlobPathIfNecessary(allPaths, fs)
-
- val tablePath = DataSourceUtils.getTablePath(fs, globPaths.toArray)
+ // Use the HoodieFileIndex only if the 'path' has specified with no "*"
contains.
+ // And READ_PATHS_OPT_KEY has not specified.
+ // Or else we use the original way to read hoodie table.
+ val useHoodieFileIndex = path.isDefined && !path.get.contains("*") &&
+ !parameters.contains(DataSourceReadOptions.READ_PATHS_OPT_KEY)
+ val globPaths = if (useHoodieFileIndex) {
+ None
+ } else {
+ Some(HoodieSparkUtils.checkAndGlobPathIfNecessary(allPaths, fs))
+ }
+ // Get the table base path
+ val tablePath = if (globPaths.isDefined) {
+ DataSourceUtils.getTablePath(fs, globPaths.get.toArray)
+ } else {
+ val tablePathOpt = TablePathUtils.getTablePath(fs, new Path(path.get))
+ if (tablePathOpt.isPresent) {
+ tablePathOpt.get().toString
+ } else {
+ throw new HoodieException(s"Failed to get the table path, " +
+ s"Maybe ${allPaths.head} is not a hudi table path")
+ }
Review comment:
Please use `DataSourceUtils.getTablePath` here as well. It already
handles when table path is not found and throws the appropriate
`TableNotFoundException`.
##########
File path:
hudi-spark-datasource/hudi-spark/src/main/scala/org/apache/hudi/DefaultSource.scala
##########
@@ -162,18 +184,26 @@ class DefaultSource extends RelationProvider
override def shortName(): String = "hudi"
- private def getBaseFileOnlyView(sqlContext: SQLContext,
+ private def getBaseFileOnlyView(useHoodieFileIndex: Boolean,
+ sqlContext: SQLContext,
optParams: Map[String, String],
schema: StructType,
+ tablePath: String,
extraReadPaths: Seq[String],
- isBootstrappedTable: Boolean,
- globPaths: Seq[Path],
metaClient: HoodieTableMetaClient):
BaseRelation = {
log.warn("Loading Base File Only View.")
+ if (useHoodieFileIndex) {
+ val fileIndex = HoodieFileIndex(sqlContext.sparkSession, tablePath,
+ if (schema == null) Option.empty[StructType] else Some(schema),
optParams)
- if (isBootstrappedTable) {
- // For bootstrapped tables, use our custom Spark relation for querying
- new HoodieBootstrapRelation(sqlContext, schema, globPaths, metaClient,
optParams)
+ log.info("Constructing hoodie (as parquet) data source with options :" +
optParams)
Review comment:
This log seems common in both if and else and can be moved out.
##########
File path:
hudi-spark-datasource/hudi-spark/src/main/scala/org/apache/hudi/HoodieFileIndex.scala
##########
@@ -0,0 +1,317 @@
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one or more
+ * contributor license agreements. See the NOTICE file distributed with
+ * this work for additional information regarding copyright ownership.
+ * The ASF licenses this file to You under the Apache License, Version 2.0
+ * (the "License"); you may not use this file except in compliance with
+ * the License. You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package org.apache.hudi
+
+import java.util.Properties
+
+import scala.collection.JavaConverters._
+
+import org.apache.hadoop.fs.{FileStatus, Path}
+import org.apache.hudi.client.common.HoodieSparkEngineContext
+import org.apache.hudi.common.config.{HoodieMetadataConfig,
SerializableConfiguration}
+import org.apache.hudi.common.engine.HoodieLocalEngineContext
+import org.apache.hudi.common.fs.FSUtils
+import org.apache.hudi.common.model.HoodieBaseFile
+import org.apache.hudi.common.table.{HoodieTableMetaClient,
TableSchemaResolver}
+import org.apache.hudi.common.table.view.HoodieTableFileSystemView
+import org.apache.hudi.config.HoodieWriteConfig
+import org.apache.spark.api.java.JavaSparkContext
+import org.apache.spark.internal.Logging
+import org.apache.spark.sql.catalyst.{InternalRow, expressions}
+import org.apache.spark.sql.SparkSession
+import org.apache.spark.sql.avro.SchemaConverters
+import org.apache.spark.sql.catalyst.expressions.{AttributeReference,
BoundReference, Expression, InterpretedPredicate}
+import org.apache.spark.sql.catalyst.util.{CaseInsensitiveMap, DateTimeUtils}
+import org.apache.spark.sql.execution.datasources.{FileIndex,
PartitionDirectory, PartitionUtils}
+import org.apache.spark.sql.internal.SQLConf
+import org.apache.spark.sql.types.StructType
+import org.apache.spark.unsafe.types.UTF8String
+
+/**
+ * A File Index which support partition prune for hoodie snapshot and
read-optimized
+ * query.
+ * Main steps to get the file list for query:
+ * 1、Load all files and partition values from the table path.
+ * 2、Do the partition prune by the partition filter condition.
+ *
+ * There are 3 cases for this:
+ * 1、If the partition columns size is equal to the actually partition path
level, we
+ * read it as partitioned table.(e.g partition column is "dt", the partition
path is "2021-03-10")
+ *
+ * 2、If the partition columns size is not equal to the partition path level,
but the partition
+ * column size is "1" (e.g. partition column is "dt", but the partition path
is "2021/03/10"
+ * who'es directory level is 3).We can still read it as a partitioned table.
We will mapping the
+ * partition path (e.g. 2021/03/10) to the only partition column (e.g. "dt").
+ *
+ * 3、Else the the partition columns size is not equal to the partition
directory level and the
+ * size is great than "1" (e.g. partition column is "dt,hh", the partition
path is "2021/03/10/12")
+ * , we read it as a None Partitioned table because we cannot know how to
mapping the partition
+ * path with the partition columns in this case.
+ */
+case class HoodieFileIndex(
+ spark: SparkSession,
+ basePath: String,
+ schemaSpec: Option[StructType],
+ options: Map[String, String])
+ extends FileIndex with Logging {
+
+ @transient private val hadoopConf = spark.sessionState.newHadoopConf()
+ private lazy val metaClient = HoodieTableMetaClient
+ .builder().setConf(hadoopConf).setBasePath(basePath).build()
+
+ @transient private val queryPath = new Path(options.getOrElse("path",
"'path' option required"))
+ /**
+ * Get the schema of the table.
+ */
+ lazy val schema: StructType = schemaSpec.getOrElse({
+ val schemaUtil = new TableSchemaResolver(metaClient)
+ SchemaConverters.toSqlType(schemaUtil.getTableAvroSchema)
+ .dataType.asInstanceOf[StructType]
+ })
+
+ /**
+ * Get the partition schema from the hoodie.properties.
+ */
+ private lazy val _partitionSchemaFromProperties: StructType = {
+ val tableConfig = metaClient.getTableConfig
+ val partitionColumns = tableConfig.getPartitionColumns
+ val nameFieldMap = schema.fields.map(filed => filed.name -> filed).toMap
+
+ if (partitionColumns.isPresent) {
+ val partitionFields = partitionColumns.get().map(column =>
+ nameFieldMap.getOrElse(column, throw new
IllegalArgumentException(s"Cannot find column " +
+ s"$column in the schema[${schema.fields.mkString(",")}]")))
+ new StructType(partitionFields)
+ } else { // If the partition columns have not stored in
hoodie.properites(the table that was
+ // created earlier), we trait it as a none-partitioned table.
+ new StructType()
+ }
+ }
+
+ @transient @volatile private var fileSystemView: HoodieTableFileSystemView =
_
+ @transient @volatile private var cachedAllInputFiles: Array[HoodieBaseFile]
= _
+ @transient @volatile private var cachedFileSize: Long = 0L
+ @transient @volatile private var cachedAllPartitionPaths: Seq[PartitionPath]
= _
+
+ @volatile private var queryAsNonePartitionedTable: Boolean = _
+
+ refresh()
+
+ override def rootPaths: Seq[Path] = queryPath :: Nil
+
+ override def listFiles(partitionFilters: Seq[Expression],
+ dataFilters: Seq[Expression]):
Seq[PartitionDirectory] = {
+ if (queryAsNonePartitionedTable) { // Read as None Partitioned table.
+ Seq(PartitionDirectory(InternalRow.empty, allFiles))
+ } else {
+ // Prune the partition path by the partition filters
+ val prunedPartitions = prunePartition(cachedAllPartitionPaths,
partitionFilters)
+ prunedPartitions.map { partition =>
+ val fileStatues =
fileSystemView.getLatestBaseFiles(partition.partitionPath).iterator()
+ .asScala.toSeq
+ .map(_.getFileStatus)
+ PartitionDirectory(partition.values, fileStatues)
+ }
+ }
+ }
+
+ override def inputFiles: Array[String] = {
+ cachedAllInputFiles.map(_.getFileStatus.getPath.toString)
+ }
+
+ override def refresh(): Unit = {
+ val startTime = System.currentTimeMillis()
+ val partitionFiles = loadPartitionPathFiles()
+ val allFiles = partitionFiles.values.reduceOption(_ ++ _)
+ .getOrElse(Array.empty[FileStatus])
+
+ metaClient.reloadActiveTimeline()
+ val activeInstants =
metaClient.getActiveTimeline.getCommitsTimeline.filterCompletedInstants
+ fileSystemView = new HoodieTableFileSystemView(metaClient, activeInstants,
allFiles)
+ cachedAllInputFiles =
fileSystemView.getLatestBaseFiles.iterator().asScala.toArray
+ cachedAllPartitionPaths = partitionFiles.keys.toSeq
+ cachedFileSize = cachedAllInputFiles.map(_.getFileLen).sum
+
+ // If the partition value contains InternalRow.empty, we query it as a
none partitioned table.
+ queryAsNonePartitionedTable = cachedAllPartitionPaths
+ .exists(p => p.values == InternalRow.empty)
+ val flushSpend = System.currentTimeMillis() - startTime
+ logInfo(s"Refresh for table ${metaClient.getTableConfig.getTableName}," +
+ s" spend: $flushSpend ms")
+ }
+
+ override def sizeInBytes: Long = {
+ cachedFileSize
+ }
+
+ override def partitionSchema: StructType = {
+ if (queryAsNonePartitionedTable) {
+ // If we read it as None Partitioned table, we should not
+ // return the partition schema.
+ new StructType()
+ } else {
+ _partitionSchemaFromProperties
+ }
+ }
+
+ /**
+ * Get the data schema of the table.
+ * @return
+ */
+ def dataSchema: StructType = {
+ val partitionColumns = partitionSchema.fields.map(_.name).toSet
+ StructType(schema.fields.filterNot(f => partitionColumns.contains(f.name)))
+ }
+
+ def allFiles: Seq[FileStatus] = cachedAllInputFiles.map(_.getFileStatus)
+
+ private def prunePartition(partitionPaths: Seq[PartitionPath],
+ predicates: Seq[Expression]): Seq[PartitionPath]
= {
+
+ val partitionColumnNames = partitionSchema.fields.map(_.name).toSet
+ val partitionPruningPredicates = predicates.filter {
+ _.references.map(_.name).toSet.subsetOf(partitionColumnNames)
+ }
+ if (partitionPruningPredicates.nonEmpty) {
+ val predicate = partitionPruningPredicates.reduce(expressions.And)
+
+ val boundPredicate = InterpretedPredicate.create(predicate.transform {
+ case a: AttributeReference =>
+ val index = partitionSchema.indexWhere(a.name == _.name)
+ BoundReference(index, partitionSchema(index).dataType, nullable =
true)
+ })
+
+ val partitionPruned = partitionPaths.filter {
+ case PartitionPath(values, _) => boundPredicate.eval(values)
+ }
+ logInfo(s"Total partition size is: ${partitionPaths.size}," +
+ s" after partition prune size is: ${partitionPruned.size}")
+ partitionPruned
+ } else {
+ partitionPaths
+ }
+ }
+
+ /**
+ * Load all partition paths and it's files under the query table path.
+ */
+ private def loadPartitionPathFiles(): Map[PartitionPath, Array[FileStatus]]
= {
+ val sparkEngine = new HoodieSparkEngineContext(new
JavaSparkContext(spark.sparkContext))
+ val properties = new Properties()
+ properties.putAll(options.asJava)
+ val metadataConfig =
HoodieMetadataConfig.newBuilder.fromProperties(properties).build()
+
+ val queryPartitionPath = FSUtils.getRelativePartitionPath(new
Path(basePath), queryPath)
+ // Load all the partition path from the basePath, and filter by the query
partition path.
+ val partitionPaths = FSUtils.getAllPartitionPaths(sparkEngine,
metadataConfig, basePath).asScala
+ .filter(_.startsWith(queryPartitionPath))
+
+ val maxListParallelism =
options.get(HoodieWriteConfig.MAX_LISTING_PARALLELISM)
+
.map(_.toInt).getOrElse(HoodieWriteConfig.DEFAULT_MAX_LISTING_PARALLELISM.intValue())
+ val serializableConf = new
SerializableConfiguration(spark.sessionState.newHadoopConf())
+ val partitionSchema = _partitionSchemaFromProperties
+ val timeZoneId = CaseInsensitiveMap(options)
+ .get(DateTimeUtils.TIMEZONE_OPTION)
+ .getOrElse(SQLConf.get.sessionLocalTimeZone)
+
+ // List files in all of the partition path.
+ val partition2Files =
+ spark.sparkContext.parallelize(partitionPaths,
Math.min(partitionPaths.size, maxListParallelism))
+ .map { partitionPath =>
+ val partitionRow = if (partitionSchema.fields.length == 0) {
+ // This is a none partitioned table
+ InternalRow.empty
+ } else {
+ val partitionFragments = partitionPath.split("/")
+
+ if (partitionFragments.length != partitionSchema.fields.length &&
+ partitionSchema.fields.length == 1) {
+ // If the partition column size is not equal to the partition
fragment size
+ // and the partition column size is 1, we map the whole
partition path
+ // to the partition column which can benefit from the partition
prune.
+ InternalRow.fromSeq(Seq(UTF8String.fromString(partitionPath)))
+ } else if (partitionFragments.length !=
partitionSchema.fields.length &&
+ partitionSchema.fields.length > 1) {
+ // If the partition column size is not equal to the partition
fragments size
+ // and the partition column size > 1, we do not know how to map
the partition
+ // fragments to the partition columns. So we trait it as a None
Partitioned Table
+ // for the query which do not benefit from the partition prune.
+ InternalRow.empty
+ } else { // If partitionSeqs.length ==
partitionSchema.fields.length
+
+ // Append partition name to the partition value if the
+ // HIVE_STYLE_PARTITIONING_OPT_KEY is disable.
+ // e.g. convert "/xx/xx/2021/02" to "/xx/xx/year=2021/month=02"
+ val partitionWithName =
+ partitionFragments.zip(partitionSchema).map {
+ case (partition, field) =>
+ if (partition.indexOf("=") == -1) {
+ s"${field.name}=$partition"
+ } else {
+ partition
+ }
+ }.mkString("/")
+ val pathWithPartitionName = new Path(basePath, partitionWithName)
+ val partitionDataTypes = partitionSchema.fields.map(f => f.name
-> f.dataType).toMap
+ val partitionValues =
PartitionUtils.parsePartition(pathWithPartitionName,
+ typeInference = false, Set(new Path(basePath)),
partitionDataTypes,
+ DateTimeUtils.getTimeZone(timeZoneId))
+
+ // Convert partitionValues to InternalRow
+ partitionValues.map(_.literals.map(_.value))
+ .map(InternalRow.fromSeq)
+ .getOrElse(InternalRow.empty)
+ }
+ }
+
+ // Get files from the partition path.
+ val fullPartitionPath = if (partitionPath.isEmpty) {
+ new Path(basePath) // This is a none partition path
+ } else {
+ new Path(basePath, partitionPath)
+ }
+ // Here we use a LocalEngineContext to get the files in the
partition.
+ // We can do this because the TableMetadata.getAllFilesInPartition
only rely on the
+ // hadoopConf of the EngineContext.
+ val engineContext = new
HoodieLocalEngineContext(serializableConf.get())
+ val filesInPartition = FSUtils.getFilesInPartition(engineContext,
metadataConfig,
+ basePath, fullPartitionPath)
Review comment:
As I had mentioned earlier, I wanted to re-use the same `FileSystemView`
that you created earlier because we don't want to create `HoodieTableMetadata`
again and again because it will then open the reader
(https://github.com/apache/hudi/blob/master/hudi-common/src/main/java/org/apache/hudi/metadata/HoodieBackedTableMetadata.java#L79)
N times for N partitions. This will result in N file system calls for N
partitions and looses the benefit of metadata based listing as we would want to
read the table just once.
However, because the reader is `transient` even if we re-use the file system
view at the executors it will be `null` and have to be re-created. So, with the
current implementation for `TableMetadata` it seems the only way to read the
table just once is to loop through all the partitions at the driver itself (in
case metadata listing is enabled).
@vinothchandar @prashantwason Would like your inputs here. How can we make
the metadata table be read once while doing it in a distributed way ?
- Do we want to utilize something like timeline server for the query side
too ?
- Or we need to introduce an API that accepts all partitions and is able to
read all of them at once ?
##########
File path:
hudi-spark-datasource/hudi-spark/src/main/scala/org/apache/hudi/HoodieFileIndex.scala
##########
@@ -0,0 +1,317 @@
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one or more
+ * contributor license agreements. See the NOTICE file distributed with
+ * this work for additional information regarding copyright ownership.
+ * The ASF licenses this file to You under the Apache License, Version 2.0
+ * (the "License"); you may not use this file except in compliance with
+ * the License. You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package org.apache.hudi
+
+import java.util.Properties
+
+import scala.collection.JavaConverters._
+
+import org.apache.hadoop.fs.{FileStatus, Path}
+import org.apache.hudi.client.common.HoodieSparkEngineContext
+import org.apache.hudi.common.config.{HoodieMetadataConfig,
SerializableConfiguration}
+import org.apache.hudi.common.engine.HoodieLocalEngineContext
+import org.apache.hudi.common.fs.FSUtils
+import org.apache.hudi.common.model.HoodieBaseFile
+import org.apache.hudi.common.table.{HoodieTableMetaClient,
TableSchemaResolver}
+import org.apache.hudi.common.table.view.HoodieTableFileSystemView
+import org.apache.hudi.config.HoodieWriteConfig
+import org.apache.spark.api.java.JavaSparkContext
+import org.apache.spark.internal.Logging
+import org.apache.spark.sql.catalyst.{InternalRow, expressions}
+import org.apache.spark.sql.SparkSession
+import org.apache.spark.sql.avro.SchemaConverters
+import org.apache.spark.sql.catalyst.expressions.{AttributeReference,
BoundReference, Expression, InterpretedPredicate}
+import org.apache.spark.sql.catalyst.util.{CaseInsensitiveMap, DateTimeUtils}
+import org.apache.spark.sql.execution.datasources.{FileIndex,
PartitionDirectory, PartitionUtils}
+import org.apache.spark.sql.internal.SQLConf
+import org.apache.spark.sql.types.StructType
+import org.apache.spark.unsafe.types.UTF8String
+
+/**
+ * A File Index which support partition prune for hoodie snapshot and
read-optimized
+ * query.
+ * Main steps to get the file list for query:
+ * 1、Load all files and partition values from the table path.
+ * 2、Do the partition prune by the partition filter condition.
+ *
+ * There are 3 cases for this:
+ * 1、If the partition columns size is equal to the actually partition path
level, we
+ * read it as partitioned table.(e.g partition column is "dt", the partition
path is "2021-03-10")
+ *
+ * 2、If the partition columns size is not equal to the partition path level,
but the partition
+ * column size is "1" (e.g. partition column is "dt", but the partition path
is "2021/03/10"
+ * who'es directory level is 3).We can still read it as a partitioned table.
We will mapping the
+ * partition path (e.g. 2021/03/10) to the only partition column (e.g. "dt").
+ *
+ * 3、Else the the partition columns size is not equal to the partition
directory level and the
+ * size is great than "1" (e.g. partition column is "dt,hh", the partition
path is "2021/03/10/12")
+ * , we read it as a None Partitioned table because we cannot know how to
mapping the partition
+ * path with the partition columns in this case.
+ */
+case class HoodieFileIndex(
+ spark: SparkSession,
+ basePath: String,
+ schemaSpec: Option[StructType],
+ options: Map[String, String])
+ extends FileIndex with Logging {
+
+ @transient private val hadoopConf = spark.sessionState.newHadoopConf()
+ private lazy val metaClient = HoodieTableMetaClient
+ .builder().setConf(hadoopConf).setBasePath(basePath).build()
+
+ @transient private val queryPath = new Path(options.getOrElse("path",
"'path' option required"))
+ /**
+ * Get the schema of the table.
+ */
+ lazy val schema: StructType = schemaSpec.getOrElse({
+ val schemaUtil = new TableSchemaResolver(metaClient)
+ SchemaConverters.toSqlType(schemaUtil.getTableAvroSchema)
+ .dataType.asInstanceOf[StructType]
+ })
+
+ /**
+ * Get the partition schema from the hoodie.properties.
+ */
+ private lazy val _partitionSchemaFromProperties: StructType = {
+ val tableConfig = metaClient.getTableConfig
+ val partitionColumns = tableConfig.getPartitionColumns
+ val nameFieldMap = schema.fields.map(filed => filed.name -> filed).toMap
+
+ if (partitionColumns.isPresent) {
+ val partitionFields = partitionColumns.get().map(column =>
+ nameFieldMap.getOrElse(column, throw new
IllegalArgumentException(s"Cannot find column " +
+ s"$column in the schema[${schema.fields.mkString(",")}]")))
+ new StructType(partitionFields)
+ } else { // If the partition columns have not stored in
hoodie.properites(the table that was
+ // created earlier), we trait it as a none-partitioned table.
+ new StructType()
+ }
+ }
+
+ @transient @volatile private var fileSystemView: HoodieTableFileSystemView =
_
+ @transient @volatile private var cachedAllInputFiles: Array[HoodieBaseFile]
= _
+ @transient @volatile private var cachedFileSize: Long = 0L
+ @transient @volatile private var cachedAllPartitionPaths: Seq[PartitionPath]
= _
+
+ @volatile private var queryAsNonePartitionedTable: Boolean = _
+
+ refresh()
+
+ override def rootPaths: Seq[Path] = queryPath :: Nil
+
+ override def listFiles(partitionFilters: Seq[Expression],
+ dataFilters: Seq[Expression]):
Seq[PartitionDirectory] = {
+ if (queryAsNonePartitionedTable) { // Read as None Partitioned table.
+ Seq(PartitionDirectory(InternalRow.empty, allFiles))
+ } else {
+ // Prune the partition path by the partition filters
+ val prunedPartitions = prunePartition(cachedAllPartitionPaths,
partitionFilters)
+ prunedPartitions.map { partition =>
+ val fileStatues =
fileSystemView.getLatestBaseFiles(partition.partitionPath).iterator()
+ .asScala.toSeq
+ .map(_.getFileStatus)
+ PartitionDirectory(partition.values, fileStatues)
+ }
+ }
+ }
+
+ override def inputFiles: Array[String] = {
+ cachedAllInputFiles.map(_.getFileStatus.getPath.toString)
+ }
+
+ override def refresh(): Unit = {
Review comment:
- Do you know when this `refresh()` is called apart from the internal
call you make from within this index class at the time of index creation ?
- The `InMemoryFileIndex` mantains `refresh()` and `refresh0()`. The
`refresh0()` is called from within the index class at the time of creation.
While `refresh()` is called from other places too. Do you know why we don't
need to follow similar model ?
I will do more reading of the spark code to understand this part as well,
but if you already know the answers you can respond to the questions I have.
##########
File path:
hudi-spark-datasource/hudi-spark/src/main/scala/org/apache/hudi/HoodieFileIndex.scala
##########
@@ -0,0 +1,317 @@
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one or more
+ * contributor license agreements. See the NOTICE file distributed with
+ * this work for additional information regarding copyright ownership.
+ * The ASF licenses this file to You under the Apache License, Version 2.0
+ * (the "License"); you may not use this file except in compliance with
+ * the License. You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package org.apache.hudi
+
+import java.util.Properties
+
+import scala.collection.JavaConverters._
+
+import org.apache.hadoop.fs.{FileStatus, Path}
+import org.apache.hudi.client.common.HoodieSparkEngineContext
+import org.apache.hudi.common.config.{HoodieMetadataConfig,
SerializableConfiguration}
+import org.apache.hudi.common.engine.HoodieLocalEngineContext
+import org.apache.hudi.common.fs.FSUtils
+import org.apache.hudi.common.model.HoodieBaseFile
+import org.apache.hudi.common.table.{HoodieTableMetaClient,
TableSchemaResolver}
+import org.apache.hudi.common.table.view.HoodieTableFileSystemView
+import org.apache.hudi.config.HoodieWriteConfig
+import org.apache.spark.api.java.JavaSparkContext
+import org.apache.spark.internal.Logging
+import org.apache.spark.sql.catalyst.{InternalRow, expressions}
+import org.apache.spark.sql.SparkSession
+import org.apache.spark.sql.avro.SchemaConverters
+import org.apache.spark.sql.catalyst.expressions.{AttributeReference,
BoundReference, Expression, InterpretedPredicate}
+import org.apache.spark.sql.catalyst.util.{CaseInsensitiveMap, DateTimeUtils}
+import org.apache.spark.sql.execution.datasources.{FileIndex,
PartitionDirectory, PartitionUtils}
+import org.apache.spark.sql.internal.SQLConf
+import org.apache.spark.sql.types.StructType
+import org.apache.spark.unsafe.types.UTF8String
+
+/**
+ * A File Index which support partition prune for hoodie snapshot and
read-optimized
+ * query.
+ * Main steps to get the file list for query:
+ * 1、Load all files and partition values from the table path.
+ * 2、Do the partition prune by the partition filter condition.
+ *
+ * There are 3 cases for this:
+ * 1、If the partition columns size is equal to the actually partition path
level, we
+ * read it as partitioned table.(e.g partition column is "dt", the partition
path is "2021-03-10")
+ *
+ * 2、If the partition columns size is not equal to the partition path level,
but the partition
+ * column size is "1" (e.g. partition column is "dt", but the partition path
is "2021/03/10"
+ * who'es directory level is 3).We can still read it as a partitioned table.
We will mapping the
+ * partition path (e.g. 2021/03/10) to the only partition column (e.g. "dt").
+ *
+ * 3、Else the the partition columns size is not equal to the partition
directory level and the
+ * size is great than "1" (e.g. partition column is "dt,hh", the partition
path is "2021/03/10/12")
+ * , we read it as a None Partitioned table because we cannot know how to
mapping the partition
+ * path with the partition columns in this case.
+ */
+case class HoodieFileIndex(
Review comment:
I see that `InMemoryFileIndex` uses a `FileStatusCache` to cache
FileStatus across queries. Don't we want to include something like that ?
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
us...@infra.apache.org
