kangkaisen commented on a change in pull request #3296: spark data preparation process, issue #3295 URL: https://github.com/apache/incubator-doris/pull/3296#discussion_r407070343
########## File path: fe/src/main/java/org/apache/doris/load/loadv2/dpp/SparkDpp.java ########## @@ -0,0 +1,848 @@ +// 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.doris.load.loadv2.dpp; + +import com.google.common.base.Strings; +import com.google.gson.Gson; +import org.apache.doris.common.UserException; +import org.apache.doris.load.loadv2.etl.EtlJobConfig; +import org.apache.hadoop.fs.FileStatus; +import org.apache.hadoop.fs.Path; +import org.apache.hadoop.conf.Configuration; +import org.apache.parquet.column.ParquetProperties; +import org.apache.parquet.example.data.Group; +import org.apache.parquet.example.data.simple.SimpleGroupFactory; +import org.apache.parquet.hadoop.ParquetWriter; +import org.apache.parquet.hadoop.example.GroupWriteSupport; +import org.apache.parquet.hadoop.metadata.CompressionCodecName; +import org.apache.parquet.schema.MessageType; +import org.apache.parquet.schema.OriginalType; +import org.apache.parquet.schema.PrimitiveType; +import org.apache.parquet.schema.Types; +import org.apache.spark.Partitioner; +import org.apache.spark.api.java.function.ForeachPartitionFunction; +import org.apache.spark.api.java.function.PairFlatMapFunction; +import org.apache.spark.sql.functions; +import org.apache.spark.sql.RowFactory; +import org.apache.spark.sql.Row; +import org.apache.spark.sql.Dataset; +import org.apache.spark.sql.Column; +import org.apache.spark.sql.SparkSession; +import org.apache.spark.sql.functions; +import org.apache.spark.sql.types.DataType; +import org.apache.spark.sql.types.DataTypes; +import org.apache.spark.sql.types.StructField; +import org.apache.spark.sql.types.StructType; +import org.apache.spark.sql.expressions.UserDefinedAggregateFunction; +import org.apache.spark.util.LongAccumulator; +import scala.Tuple2; +import scala.collection.Seq; + +import org.apache.spark.api.java.JavaPairRDD; +import org.apache.spark.api.java.JavaRDD; + +import java.math.BigInteger; +import java.net.URI; +import java.util.Map; +import java.util.Set; +import java.util.List; +import java.util.ArrayList; +import java.util.Date; +import java.util.HashMap; +import java.util.Iterator; +import java.util.Queue; +import java.util.LinkedList; +import java.util.HashSet; +import java.util.Arrays; + +import org.apache.hadoop.fs.FileSystem; +import org.apache.hadoop.fs.FSDataOutputStream; +import scala.collection.JavaConverters; + +// This class is a Spark-based data preprocessing program, +// which will make use of the distributed compute framework of spark to +// do ETL job/sort/preaggregate/convert data format job in spark job +// to boost the process of large amount of data load. +public final class SparkDpp implements java.io.Serializable { + private static final String NULL_FLAG = "\\N"; + private static final String DPP_RESULT_FILE = "dpp_result.json"; + private static final String BITMAP_TYPE = "bitmap"; + private EtlJobConfig etlJobConfig = null; + private LongAccumulator abnormalRowAcc = null; + private LongAccumulator unselectedRowAcc = null; + private LongAccumulator scannedRowsAcc = null; + private LongAccumulator fileNumberAcc = null; + private LongAccumulator fileSizeAcc = null; + private DppResult dppResult; + private SparkSession spark = null; + private UserDefinedAggregateFunction bitmap_union = null; + + public SparkDpp(SparkSession spark, EtlJobConfig etlJobConfig) { + this.spark = spark; + this.etlJobConfig = etlJobConfig; + } + + public void init() { + bitmap_union = spark.udf().register("bitmap_union", new BitmapUnion()); + abnormalRowAcc = spark.sparkContext().longAccumulator(); + unselectedRowAcc = spark.sparkContext().longAccumulator(); + scannedRowsAcc = spark.sparkContext().longAccumulator(); + fileNumberAcc = spark.sparkContext().longAccumulator(); + fileSizeAcc = spark.sparkContext().longAccumulator(); + } + + private Dataset<Row> processDataframeAgg(Dataset<Row> dataframe, EtlJobConfig.EtlIndex indexMeta) { + dataframe.createOrReplaceTempView("base_table"); + StringBuilder sb = new StringBuilder(); + sb.append("select "); + sb.append(DppUtils.BUCKET_ID + ","); + + // assume that keys are all before values + StringBuilder groupBySb = new StringBuilder(); + groupBySb.append(DppUtils.BUCKET_ID + ","); + Map<String, DataType> valueColumnsOriginalType = new HashMap<>(); + for (EtlJobConfig.EtlColumn column : indexMeta.columns) { + if (column.isKey) { + sb.append(column.columnName + ","); + groupBySb.append(column.columnName + ","); + } else { + // get the value columns's original type + DataType originalType = dataframe.schema().apply(column.columnName).dataType(); + valueColumnsOriginalType.put(column.columnName, originalType); + if (column.aggregationType.equalsIgnoreCase("MAX")) { + sb.append("max(" + column.columnName + ") as " + column.columnName); + sb.append(","); + } else if (column.aggregationType.equalsIgnoreCase("MIN")) { + sb.append("min(" + column.columnName + ") as " + column.columnName); + sb.append(","); + } else if (column.aggregationType.equalsIgnoreCase("SUM")) { + sb.append("sum(" + column.columnName + ") as " + column.columnName); + sb.append(","); + } else if (column.aggregationType.equalsIgnoreCase("BITMAP_UNION")) { + sb.append("bitmap_union(" + column.columnName + ") as " + column.columnName); + sb.append(","); + } + } + } + sb.deleteCharAt(sb.length() - 1); + groupBySb.deleteCharAt(groupBySb.length() - 1); + sb.append(" from base_table "); + sb.append(" group by "); + sb.append(groupBySb.toString()); + String aggSql = sb.toString(); + Dataset<Row> aggDataFrame = spark.sql(aggSql); + // after agg, the type of sum column maybe be changed, so should add type cast for value column + for (Map.Entry<String, DataType> entry : valueColumnsOriginalType.entrySet()) { + DataType currentType = aggDataFrame.schema().apply(entry.getKey()).dataType(); + if (!currentType.equals(entry.getValue())) { + aggDataFrame = aggDataFrame.withColumn(entry.getKey(), aggDataFrame.col(entry.getKey()).cast(entry.getValue())); + } + } + return aggDataFrame; + } + + private void writePartitionedAndSortedDataframeToParquet(Dataset<Row> dataframe, String pathPattern, long tableId, + EtlJobConfig.EtlIndex indexMeta) { + dataframe.foreachPartition(new ForeachPartitionFunction<Row>() { + @Override + public void call(Iterator<Row> t) throws Exception { + // write the data to dst file + Configuration conf = new Configuration(); + FileSystem fs = FileSystem.get(URI.create(etlJobConfig.outputPath), conf); + String lastBucketKey = null; + ParquetWriter<Group> writer = null; + Types.MessageTypeBuilder builder = Types.buildMessage(); + for (EtlJobConfig.EtlColumn column : indexMeta.columns) { + if (column.isAllowNull) { + if (column.columnType.equals("SMALLINT") ||column.columnType.equals("INT")) { + builder.optional(PrimitiveType.PrimitiveTypeName.INT32).named(column.columnName); + } else if (column.columnType.equals("BIGINT")) { + builder.optional(PrimitiveType.PrimitiveTypeName.INT64).named(column.columnName); + } else if (column.columnType.equals("BOOL")) { + builder.optional(PrimitiveType.PrimitiveTypeName.BOOLEAN).named(column.columnName); + } else if (column.columnType.equals("VARCHAR")) { + // should use as(OriginalType.UTF8), or result will be binary + builder.optional(PrimitiveType.PrimitiveTypeName.BINARY).as(OriginalType.UTF8).named(column.columnName); + } else if (column.columnType.equals("FLOAT")) { + builder.optional(PrimitiveType.PrimitiveTypeName.FLOAT).named(column.columnName); + } else if (column.columnType.equals("DOUBLE")) { + builder.optional(PrimitiveType.PrimitiveTypeName.DOUBLE).named(column.columnName); + } else { + // TODO: exception handle + System.err.println("invalid column type:" + column); + } + } else { + if (column.columnType.equals("SMALLINT") ||column.columnType.equals("INT")) { + builder.required(PrimitiveType.PrimitiveTypeName.INT32).named(column.columnName); + } else if (column.columnType.equals("BIGINT")) { + builder.required(PrimitiveType.PrimitiveTypeName.INT64).named(column.columnName); + } else if (column.columnType.equals("BOOL")) { + builder.required(PrimitiveType.PrimitiveTypeName.BOOLEAN).named(column.columnName); + } else if (column.columnType.equals("VARCHAR")) { + // should use as(OriginalType.UTF8), or result will be binary + builder.required(PrimitiveType.PrimitiveTypeName.BINARY).as(OriginalType.UTF8).named(column.columnName); + } else if (column.columnType.equals("FLOAT")) { + builder.required(PrimitiveType.PrimitiveTypeName.FLOAT).named(column.columnName); + } else if (column.columnType.equals("DOUBLE")) { + builder.required(PrimitiveType.PrimitiveTypeName.DOUBLE).named(column.columnName); + } else { + // TODO: exception handle + System.err.println("invalid column type:" + column); + } + } + } + MessageType index_schema = builder.named("index_" + indexMeta.indexId); + while (t.hasNext()) { + Row row = t.next(); + if (row.length() <= 1) { + System.err.println("invalid row:" + row); + continue; + } + String curBucketKey = row.getString(0); + if (lastBucketKey == null || !curBucketKey.equals(lastBucketKey)) { + if (writer != null) { + System.out.println("close writer"); + writer.close(); + } + // flush current writer and create a new writer + String[] bucketKey = curBucketKey.split("_"); + if (bucketKey.length != 2) { + System.err.println("invalid bucket key:" + curBucketKey); + continue; + } + int partitionId = Integer.parseInt(bucketKey[0]); + int bucketId = Integer.parseInt(bucketKey[1]); + String path = String.format(pathPattern, tableId, partitionId, indexMeta.indexId, + bucketId, indexMeta.schemaHash); + GroupWriteSupport.setSchema(index_schema, conf); + writer = new ParquetWriter<Group>(new Path(path), new GroupWriteSupport(), + CompressionCodecName.SNAPPY, 1024, 1024, 512, + true, false, + ParquetProperties.WriterVersion.PARQUET_1_0, conf); + if(writer != null){ + System.out.println("[HdfsOperate]>> initialize writer succeed! path:" + path); + } + lastBucketKey = curBucketKey; + } + SimpleGroupFactory groupFactory = new SimpleGroupFactory(index_schema); + Group group = groupFactory.newGroup(); + for (int i = 1; i < row.length(); i++) { + Object columnObject = row.get(i); + if (columnObject instanceof Short) { + group.add(indexMeta.columns.get(i - 1).columnName, row.getShort(i)); + } else if (columnObject instanceof Integer) { + group.add(indexMeta.columns.get(i - 1).columnName, row.getInt(i)); + } else if (columnObject instanceof String) { + group.add(indexMeta.columns.get(i - 1).columnName, row.getString(i)); + } else if (columnObject instanceof Long) { + group.add(indexMeta.columns.get(i - 1).columnName, row.getLong(i)); + } + } + try { + writer.write(group); + } catch (Exception e) { + System.err.println("exception caught:" + e); + e.printStackTrace(); + } + } + if (writer != null) { + writer.close(); + } + } + }); + } + + private void processRollupTree(RollupTreeNode rootNode, + Dataset<Row> rootDataframe, + long tableId, EtlJobConfig.EtlTable tableMeta, EtlJobConfig.EtlIndex baseIndex) { + Queue<RollupTreeNode> nodeQueue = new LinkedList<>(); + nodeQueue.offer(rootNode); + int currentLevel = 0; + // level travel the tree + Map<Long, Dataset<Row>> parentDataframeMap = new HashMap<>(); + parentDataframeMap.put(baseIndex.indexId, rootDataframe); + Map<Long, Dataset<Row>> childrenDataframeMap = new HashMap<>(); + String pathPattern = etlJobConfig.outputPath + "/" + etlJobConfig.outputFilePattern; + while (!nodeQueue.isEmpty()) { + RollupTreeNode curNode = nodeQueue.poll(); + System.err.println("start to process index:" + curNode.indexId); + if (curNode.children != null) { + for (RollupTreeNode child : curNode.children) { + nodeQueue.offer(child); + } + } + Dataset<Row> curDataFrame = null; + // column select for rollup + if (curNode.level != currentLevel) { + // need to unpersist parent dataframe + currentLevel = curNode.level; + parentDataframeMap.clear(); + parentDataframeMap = childrenDataframeMap; + childrenDataframeMap = new HashMap<>(); + } + + long parentIndexId = baseIndex.indexId; + if (curNode.parent != null) { + parentIndexId = curNode.parent.indexId; + } + + Dataset<Row> parentDataframe = parentDataframeMap.get(parentIndexId); + List<Column> columns = new ArrayList<>(); + List<Column> keyColumns = new ArrayList<>(); + Column bucketIdColumn = new Column(DppUtils.BUCKET_ID); + keyColumns.add(bucketIdColumn); + columns.add(bucketIdColumn); + for (String keyName : curNode.keyColumnNames) { + columns.add(new Column(keyName)); + keyColumns.add(new Column(keyName)); + } + for (String valueName : curNode.valueColumnNames) { + columns.add(new Column(valueName)); + } + Seq<Column> columnSeq = JavaConverters.asScalaIteratorConverter(columns.iterator()).asScala().toSeq(); + curDataFrame = parentDataframe.select(columnSeq); + if (curNode.indexMeta.indexType.equals("AGGREGATE")) { + // do aggregation + curDataFrame = processDataframeAgg(curDataFrame, curNode.indexMeta); + } + Seq<Column> keyColumnSeq = JavaConverters.asScalaIteratorConverter(keyColumns.iterator()).asScala().toSeq(); + // should use sortWithinPartitions, not sort + // because sort will modify the partition number which will lead to bug + curDataFrame = curDataFrame.sortWithinPartitions(keyColumnSeq); + childrenDataframeMap.put(curNode.indexId, curDataFrame); + + if (curNode.children != null && curNode.children.size() > 1) { + // if the children number larger than 1, persist the dataframe for performance + curDataFrame.persist(); + } + writePartitionedAndSortedDataframeToParquet(curDataFrame, pathPattern, tableId, curNode.indexMeta); + } + } + + private Dataset<Row> repartitionDataframeByBucketId(SparkSession spark, Dataset<Row> dataframe, + EtlJobConfig.EtlPartitionInfo partitionInfo, + List<Integer> partitionKeyIndex, + List<Class> partitionKeySchema, + List<DorisRangePartitioner.PartitionRangeKey> partitionRangeKeys, + List<String> keyColumnNames, + List<String> valueColumnNames, + StructType dstTableSchema, + EtlJobConfig.EtlIndex baseIndex, + List<Long> validPartitionId) throws UserException { + List<String> distributeColumns = partitionInfo.distributionColumnRefs; + Partitioner partitioner = new DorisRangePartitioner(partitionInfo, partitionKeyIndex, partitionRangeKeys); + Set<Integer> validPartitionIndex = new HashSet<>(); + if (validPartitionId == null) { + for (int i = 0; i < partitionInfo.partitions.size(); ++i) { + validPartitionIndex.add(i); + } + } else { + for (int i = 0; i < partitionInfo.partitions.size(); ++i) { + if (validPartitionId.contains(partitionInfo.partitions.get(i).partitionId)) { + validPartitionIndex.add(i); + } + } + } + JavaPairRDD<String, DppColumns> pairRDD = dataframe.javaRDD().flatMapToPair( + new PairFlatMapFunction<Row, String, DppColumns>() { + @Override + public Iterator<Tuple2<String, DppColumns>> call(Row row) { + List<Object> columns = new ArrayList<>(); + List<Class> classes = new ArrayList<>(); + List<Object> keyColumns = new ArrayList<>(); + List<Class> keyClasses = new ArrayList<>(); + for (String columnName : keyColumnNames) { + Object columnObject = row.get(row.fieldIndex(columnName)); + columns.add(columnObject); + classes.add(DppUtils.dataTypeToClass(dstTableSchema.apply(dstTableSchema.fieldIndex(columnName)).dataType())); + keyColumns.add(columnObject); + keyClasses.add(DppUtils.dataTypeToClass(dstTableSchema.apply(dstTableSchema.fieldIndex(columnName)).dataType())); + } + + for (String columnName : valueColumnNames) { + columns.add(row.get(row.fieldIndex(columnName))); + classes.add(DppUtils.dataTypeToClass(dstTableSchema.apply(dstTableSchema.fieldIndex(columnName)).dataType())); + } + DppColumns dppColumns = new DppColumns(columns, classes); + DppColumns key = new DppColumns(keyColumns, keyClasses); + List<Tuple2<String, DppColumns>> result = new ArrayList<>(); + int pid = partitioner.getPartition(key); + if (!validPartitionIndex.contains(pid)) { + System.err.println("invalid partition for row:" + row + ", pid:" + pid); + abnormalRowAcc.add(1); + } else { + long hashValue = DppUtils.getHashValue(row, distributeColumns, dstTableSchema); + int bucketId = (int) ((hashValue & 0xffffffff) % partitionInfo.partitions.get(pid).bucketNum); + long partitionId = partitionInfo.partitions.get(pid).partitionId; + // bucketKey is partitionId_bucketId + String bucketKey = Long.toString(partitionId) + "_" + Integer.toString(bucketId); + Tuple2<String, DppColumns> newTuple = new Tuple2<String, DppColumns>(bucketKey, dppColumns); + result.add(newTuple); + } + return result.iterator(); + } + }); + + JavaRDD<Row> resultRdd = pairRDD.map( record -> { + String bucketKey = record._1; + List<Object> row = new ArrayList<>(); + // bucketKey as the first key + row.add(bucketKey); + row.addAll(record._2.columns); + return RowFactory.create(row.toArray()); + } + ); + + StructType tableSchemaWithBucketId = DppUtils.createDstTableSchema(baseIndex.columns, true); + dataframe = spark.createDataFrame(resultRdd, tableSchemaWithBucketId); + // use bucket number as the parallel number + int reduceNum = 0; + for (EtlJobConfig.EtlPartition partition : partitionInfo.partitions) { + reduceNum += partition.bucketNum; + } + dataframe = dataframe.repartition(reduceNum, new Column(DppUtils.BUCKET_ID)); + return dataframe; + } + + private Dataset<Row> convertSrcDataframeToDstDataframe(EtlJobConfig.EtlIndex baseIndex, + Dataset<Row> srcDataframe, + StructType dstTableSchema, + EtlJobConfig.EtlFileGroup fileGroup) throws UserException { + System.out.println("convertSrcDataframeToDstDataframe for index:" + baseIndex); Review comment: Use Log ---------------------------------------------------------------- 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 With regards, Apache Git Services --------------------------------------------------------------------- To unsubscribe, e-mail: commits-unsubscr...@doris.apache.org For additional commands, e-mail: commits-h...@doris.apache.org
