Github user paul-rogers commented on a diff in the pull request: https://github.com/apache/drill/pull/717#discussion_r99067046 --- Diff: exec/java-exec/src/main/java/org/apache/drill/exec/physical/impl/xsort/managed/ExternalSortBatch.java --- @@ -0,0 +1,1321 @@ +/* + * 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.drill.exec.physical.impl.xsort.managed; + +import java.io.IOException; +import java.util.Collection; +import java.util.LinkedList; +import java.util.List; + +import org.apache.drill.common.AutoCloseables; +import org.apache.drill.common.config.DrillConfig; +import org.apache.drill.common.exceptions.UserException; +import org.apache.drill.exec.ExecConstants; +import org.apache.drill.exec.exception.OutOfMemoryException; +import org.apache.drill.exec.exception.SchemaChangeException; +import org.apache.drill.exec.memory.BufferAllocator; +import org.apache.drill.exec.ops.FragmentContext; +import org.apache.drill.exec.ops.MetricDef; +import org.apache.drill.exec.physical.config.ExternalSort; +import org.apache.drill.exec.physical.impl.sort.RecordBatchData; +import org.apache.drill.exec.physical.impl.xsort.MSortTemplate; +import org.apache.drill.exec.physical.impl.xsort.SingleBatchSorter; +import org.apache.drill.exec.physical.impl.xsort.managed.BatchGroup.InputBatch; +import org.apache.drill.exec.record.AbstractRecordBatch; +import org.apache.drill.exec.record.BatchSchema; +import org.apache.drill.exec.record.BatchSchema.SelectionVectorMode; +import org.apache.drill.exec.record.RecordBatch; +import org.apache.drill.exec.record.SchemaUtil; +import org.apache.drill.exec.record.VectorContainer; +import org.apache.drill.exec.record.VectorWrapper; +import org.apache.drill.exec.record.WritableBatch; +import org.apache.drill.exec.record.selection.SelectionVector2; +import org.apache.drill.exec.record.selection.SelectionVector4; +import org.apache.drill.exec.testing.ControlsInjector; +import org.apache.drill.exec.testing.ControlsInjectorFactory; +import org.apache.drill.exec.vector.ValueVector; +import org.apache.drill.exec.vector.complex.AbstractContainerVector; + +import com.google.common.collect.Lists; + +/** + * External sort batch: a sort batch which can spill to disk in + * order to operate within a defined memory footprint. + * <p> + * <h4>Basic Operation</h4> + * The operator has three key phases: + * <p> + * <ul> + * <li>The load phase in which batches are read from upstream.</li> + * <li>The merge phase in which spilled batches are combined to + * reduce the number of files below the configured limit. (Best + * practice is to configure the system to avoid this phase.) + * <li>The delivery phase in which batches are combined to produce + * the final output.</li> + * </ul> + * During the load phase: + * <p> + * <ul> + * <li>The incoming (upstream) operator provides a series of batches.</li> + * <li>This operator sorts each batch, and accumulates them in an in-memory + * buffer.</li> + * <li>If the in-memory buffer becomes too large, this operator selects + * a subset of the buffered batches to spill.</li> --- End diff -- We write spill files of size 300 MB. Any gains from long disk writes have already been gotten at that spill size. (That's 3 sec. of I/O for a spinning disk at 100 MB/s.) Drill handles big data. A sort might be given 10 GB of memory to buffer batches. Spilling this entire amount, just because we need a small amount of room, does not make much sense. If we have 10 GB and read 11 GB of data, we'd want to write just four spill files of 300 MB each, then merge the in-memory and on-disk runs to avoid writing the other 10 GB. This assumes, of course, that merging many in-batch runs is faster than spilling, then merging from disk a smaller set of spilled runs. This seems to be the case. So, we do achieve writing long spill runs (300 MB), but we don't dump the entire buffer unless necessary.
--- If your project is set up for it, you can reply to this email and have your reply appear on GitHub as well. If your project does not have this feature enabled and wishes so, or if the feature is enabled but not working, please contact infrastructure at infrastruct...@apache.org or file a JIRA ticket with INFRA. ---