[GitHub] Ben-Zvi commented on a change in pull request #1606: Drill 6845: Semi-Hash-Join to skip incoming build duplicates, automatically stop skipping if too few

[GitBox <gi...@apache.org>]

Ben-Zvi commented on a change in pull request #1606: Drill 6845: Semi-Hash-Join to skip incoming build duplicates, automatically stop skipping if too few
URL: https://github.com/apache/drill/pull/1606#discussion_r247004221
 
 

 ##########
 File path: exec/java-exec/src/main/java/org/apache/drill/exec/physical/impl/join/HashJoinSpillControlImpl.java
 ##########
 @@ -0,0 +1,52 @@
+/*
+ * 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.join;
+
+  import org.apache.drill.exec.memory.BufferAllocator;
+  import org.apache.drill.exec.record.RecordBatchSizer;
+  import org.apache.drill.exec.record.VectorContainer;
+
+/**
+ * This class is currently used only for Semi-Hash-Join that avoids duplicates by the use of a hash table
+ * The method {@link HashJoinMemoryCalculator.HashJoinSpillControl#shouldSpill(VectorContainer)} returns true if the memory available now to the allocator if not enough
+ * to hold (a multiple of, for safety) a new allocated batch
+ */
+public class HashJoinSpillControlImpl implements HashJoinMemoryCalculator.HashJoinSpillControl {
+  private BufferAllocator allocator;
+  private int recordsPerBatch;
+  private int minBatchesInAvailableMemory;
+
+  HashJoinSpillControlImpl(BufferAllocator allocator, int recordsPerBatch, int minBatchesInAvailableMemory) {
+    this.allocator = allocator;
+    this.recordsPerBatch = recordsPerBatch;
+    this.minBatchesInAvailableMemory = minBatchesInAvailableMemory;
+  }
+
+  @Override
+  public boolean shouldSpill(VectorContainer currentVectorContainer) {
+    assert currentVectorContainer.hasRecordCount();
 
 Review comment:
   About the three missing issues:
   * Reserving enough space for the output batch: Indeed was missing. I modified `HashJoinSpillControlImpl` to also take the batchMemoryManager, and then reserve the output batch size during the calculations.
   Note that just like a regular Hash-Join, this size is used while reading all the build side, before `prefetchFirstProbeBatch()` is called; so this size may not be relying on real probe side data. 
   *  Reserving space for the incoming probe batch: Not needed - The incoming size is not charged to the Hash-Join's available memory, as the incoming is just read and deallocated. (Except that optimization with num_partitions equal 1, but then there is no spilling).
   * Reserving room for the partial batches of each partition: This in a way is covered by this code, which is called prior to allocating these batches. One point that is not covered is variable length data types, but in this special case we are dealing with join keys only, which tend to be of a reasonable uniform sizes.
   
   So all this new memory spill control code is only needed for special cases. In most reasonable uses of Semi-Join, the build side keys are mostly unique, and the Semi would stop skipping duplicates and fall back to the regular memory calculator. In extreme cases of duplicates (e.g., some benchmarks), then those would be skipped and consume zero extra memory. So the cases left are in the middle, with enough duplicates to skip, but still very many diverged keys.
   
   While performing an aggregation like Hash-Aggr, the case of hash-tables here is much simpler than in Hash-Aggr, as there are no "values" to be matched. So memory need not be computed for the hash-Tables; any operation (put, or doubling) that would OOM, would be caught, then that partition would spill and reset, and then continue.  The phase of "creating hash tables" in the regular Hash-Join build does not exist here (as the tables already exist).
   
   Last: I changed that hard-coded %20 into an option: `semi_percent_duplicates_to_skip` (default 20). So now there are three options (also the spill triggering "num of batches in available mem", and disabling the whole feature).  So if any user ever gets an OOM from the skip-duplicates semi, the user can easily tune some option to avoid that OOM.
   

----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on GitHub and use the
URL above to go to the specific comment.
 
For queries about this service, please contact Infrastructure at:
users@infra.apache.org


With regards,
Apache Git Services