[
https://issues.apache.org/jira/browse/DRILL-5080?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15850665#comment-15850665
]
ASF GitHub Bot commented on DRILL-5080:
---------------------------------------
Github user paul-rogers commented on a diff in the pull request:
https://github.com/apache/drill/pull/717#discussion_r99182115
--- 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>
+ * <li>Each spill set is merged to create a new, sorted collection of
+ * batches, and each is spilled to disk.</li>
+ * <li>To allow the use of multiple disk storage, each spill group is
written
+ * round-robin to a set of spill directories.</li>
+ * </ul>
+ * <p>
+ * During the sort/merge phase:
+ * <p>
+ * <ul>
+ * <li>When the input operator is complete, this operator merges the
accumulated
+ * batches (which may be all in memory or partially on disk), and returns
+ * them to the output (downstream) operator in chunks of no more than
+ * 32K records.</li>
+ * <li>The final merge must combine a collection of in-memory and spilled
+ * batches. Several limits apply to the maximum "width" of this merge. For
+ * example, we each open spill run consumes a file handle, and we may wish
+ * to limit the number of file handles. A consolidation phase combines
+ * in-memory and spilled batches prior to the final merge to control final
+ * merge width.</li>
+ * <li>A special case occurs if no batches were spilled. In this case, the
input
+ * batches are sorted in memory without merging.</li>
+ * </ul>
+ * <p>
+ * Many complex details are involved in doing the above; the details are
explained
+ * in the methods of this class.
+ * <p>
+ * <h4>Configuration Options</h4>
+ * <dl>
+ * <dt>drill.exec.sort.external.spill.fs</dt>
+ * <dd>The file system (file://, hdfs://, etc.) of the spill
directory.</dd>
+ * <dt>drill.exec.sort.external.spill.directories</dt>
+ * <dd>The (comma? space?) separated list of directories, on the above file
+ * system, to which to spill files in round-robin fashion. The query will
+ * fail if any one of the directories becomes full.</dt>
+ * <dt>drill.exec.sort.external.spill.file_size</dt>
+ * <dd>Target size for first-generation spill files Set this to large
+ * enough to get nice long writes, but not so large that spill directories
+ * are overwhelmed.</dd>
+ * <dt>drill.exec.sort.external.mem_limit</dt>
+ * <dd>Maximum memory to use for the in-memory buffer. (Primarily for
testing.)</dd>
+ * <dt>drill.exec.sort.external.batch_limit</dt>
+ * <dd>Maximum number of batches to hold in memory. (Primarily for
testing.)</dd>
+ * <dt>drill.exec.sort.external.spill.max_count</dt>
+ * <dd>Maximum number of batches to add to “first generation” files.
+ * Defaults to 0 (no limit). (Primarily for testing.)</dd>
+ * <dt>drill.exec.sort.external.spill.min_count</dt>
+ * <dd>Minimum number of batches to add to “first generation” files.
+ * Defaults to 0 (no limit). (Primarily for testing.)</dd>
+ * <dt>drill.exec.sort.external.merge_limit</dt>
+ * <dd>Sets the maximum number of runs to be merged in a single pass
(limits
+ * the number of open files.)</dd>
+ * </dl>
+ * <p>
+ * The memory limit observed by this operator is the lesser of:
+ * <ul>
+ * <li>The maximum allocation allowed the the allocator assigned to this
batch, or</li>
--- End diff --
Fixed.
> Create a memory-managed version of the External Sort operator
> -------------------------------------------------------------
>
> Key: DRILL-5080
> URL: https://issues.apache.org/jira/browse/DRILL-5080
> Project: Apache Drill
> Issue Type: Improvement
> Affects Versions: 1.8.0
> Reporter: Paul Rogers
> Assignee: Paul Rogers
> Fix For: 1.10.0
>
> Attachments: ManagedExternalSortDesign.pdf
>
>
> We propose to create a "managed" version of the external sort operator that
> works to a clearly-defined memory limit. Attached is a design specification
> for the work.
> The project will include fixing a number of bugs related to the external
> sort, include as sub-tasks of this umbrella task.
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
