aokolnychyi commented on code in PR #8123: URL: https://github.com/apache/iceberg/pull/8123#discussion_r1270143466
########## spark/v3.4/spark/src/main/java/org/apache/iceberg/SparkDistributedDataBatchScan.java: ########## @@ -0,0 +1,332 @@ +/* + * 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.iceberg; + +import java.io.Serializable; +import java.nio.ByteBuffer; +import java.util.Iterator; +import java.util.List; +import java.util.Map; +import java.util.stream.Collectors; +import org.apache.iceberg.expressions.Expression; +import org.apache.iceberg.io.ClosingIterator; +import org.apache.iceberg.io.FileIO; +import org.apache.iceberg.spark.SparkReadConf; +import org.apache.iceberg.spark.source.SerializableTableWithSize; +import org.apache.iceberg.util.ByteBuffers; +import org.apache.spark.api.java.JavaSparkContext; +import org.apache.spark.api.java.function.FlatMapFunction; +import org.apache.spark.broadcast.Broadcast; +import org.apache.spark.sql.SparkSession; + +public class SparkDistributedDataBatchScan extends DistributedDataBatchScan { + + private final SparkSession spark; + private final JavaSparkContext sparkContext; + private final SparkReadConf readConf; + + private Broadcast<Table> tableBroadcast = null; + + public SparkDistributedDataBatchScan(SparkSession spark, Table table, SparkReadConf readConf) { + this(spark, table, readConf, table.schema(), TableScanContext.empty()); + } + + private SparkDistributedDataBatchScan( + SparkSession spark, + Table table, + SparkReadConf readConf, + Schema schema, + TableScanContext context) { + super(table, schema, context); + this.spark = spark; + this.sparkContext = JavaSparkContext.fromSparkContext(spark.sparkContext()); + this.readConf = readConf; + } + + @Override + protected BatchScan newRefinedScan( + Table newTable, Schema newSchema, TableScanContext newContext) { + return new SparkDistributedDataBatchScan(spark, newTable, readConf, newSchema, newContext); + } + + @Override + protected int remoteParallelism() { + return sparkContext.defaultParallelism(); + } + + @Override + protected PlanningMode dataPlanningMode() { + return readConf.dataPlanningMode(); + } + + @Override + protected List<DataFile> planDataRemotely(List<ManifestFile> manifests) { Review Comment: My initial though was to query the `data_files` metadata table. However, that faced a number of issues. - Need to translate a predicate on the main table into a predicate on the metadata table. - Need to implement Iceberg to Spark filter conversion. - Hard to populate scan metrics. It was also expensive as `DataFile` is wrapped as `GenericInternalRow`, then converted to `UnsafeRow`, then bytes are collected to the driver, deserialized on the driver, then converted to public `Row`, then wrapped into `DataFile` again. Not to mention the required complexity. Right now, I am using `RDD` so there is only one round of serialization (either Java or Kryo) and no conversion. The current approach performs exceptionally well for selective queries. The cost to serialize the entire content of a 10 MB manifest is around 0.2-0.3 s and is not an issue. The full table scan performance depends on how quickly the driver can fetch the result from other nodes. The most critical part is the size of serialized data. **Option 1**: Java serialization of `DataFile` and `DeleteFile`. **Option 2**: Kryo serialization of `DataFile` and `DeleteFile`. **Option 3**: Converting `DataFile` and `DeleteFile` to `UnsafeRow` and using Java/Kryo serialization on top. Option 2 (serializing files with Kryo) produced around 15% smaller chunks compared to Option 1 (serializing files with Java) and did not require any extra logic to convert to and from `UnsafeRow`. Option 3 does not require Kryo to be efficient but requires the conversion logic and gave only 3-5% size reduction in size compared to Option 2. I was surprised how well Kryo worked on top of files but I still debate whether conversion to `UnsafeRow` makes sense to not depend on Kryo. That said, using `UnsafeRow` will yield only marginally smaller chunks. The cost of the serialization is not an issue, like I said earlier. The most efficient approach was to implement a custom Kryo serializer for `DataFile[]` and apply dictionary encoding for partitions. That gave 15% size reduction when a big number of files belonged to the same partition but is extremely complicated and required the user to register a custom serializer. All in all, if someone expects to query 20 PB in one job, I doubt 30 seconds to plan the job will be their biggest issue. That's why I opted for the simplest solution that works really well for the first two use cases mentioned in the PR description (still faster in the third use case but not drastically). If we want to support such use cases, we need to change how Spark plans jobs and not collect data files back to the driver. -- 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. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
