[GitHub] [lucene-solr] danmuzi commented on a change in pull request #856: LUCENE-8965 SometimesConcurrentMergeScheduler

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

danmuzi commented on a change in pull request #856: LUCENE-8965 SometimesConcurrentMergeScheduler
URL: https://github.com/apache/lucene-solr/pull/856#discussion_r326267139
 
 

 ##########
 File path: lucene/core/src/java/org/apache/lucene/index/SometimesConcurrentMergeScheduler.java
 ##########
 @@ -0,0 +1,147 @@
+/*
+ * 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.lucene.index;
+
+import java.io.IOException;
+import java.util.Observable;
+import java.util.Observer;
+import java.util.concurrent.CountDownLatch;
+import java.util.concurrent.TimeUnit;
+
+import org.apache.lucene.util.ThreadInterruptedException;
+
+/**
+ * A variant of CMS: If there are cheap merges, wait for a merge to complete before continuing. This
+ * has the benefit of greatly increasing the odds that an {@link org.apache.lucene.search.IndexSearcher}
+ * will see fewer segments. Normally, CMS does all merges concurrently, and so it won't be until the
+ * next commit that the IndexSearcher benefits from fewer segments. The trade-off is less
+ * concurrency, and there will be some delay on a segment flush for a cheap merge if present.
+ *
+ * @author dsmiley
+ * @since solr.7
+ */
+public class SometimesConcurrentMergeScheduler extends ConcurrentMergeScheduler {
+
+  private long cheapMergeThresholdBytes = 2 * 1024 * 1024; // 2MB
+  private long cheapMaxMergeWaitTimeoutMs = TimeUnit.SECONDS.toMillis(2);
+
+  public long getCheapMergeThresholdBytes() {
+    return cheapMergeThresholdBytes;
+  }
+
+  /** Limit cheap merges to those less than this size in megabytes. */
+  public SometimesConcurrentMergeScheduler setCheapMergeThresholdMB(double v) {
+    v *= 1024 * 1024;
+    cheapMergeThresholdBytes = (v > Long.MAX_VALUE) ? Long.MAX_VALUE : (long) v;
+    return this;
+  }
+
+  /**
+   * If a cheap merge is taking longer than this amount time to finish, give up waiting for it. If
+   * this threshold is reached often, then the system should probably be tuned further.
+   */
+  public SometimesConcurrentMergeScheduler setCheapMaxMergeWaitTimeoutMs(long cheapMaxMergeWaitTimeoutMs) {
+    this.cheapMaxMergeWaitTimeoutMs = cheapMaxMergeWaitTimeoutMs;
+    return this;
+  }
+
+  /** Mechanism to notify when merges complete. */
+  private final Observable observableMerge = new Observable() {
+    @Override
+    public void notifyObservers(Object arg) {
+      super.setChanged(); // setChanged is protected, and "changed" must be set for notifyObservers to do anything
+      super.notifyObservers(arg);
+    }
+  };
+
+  @Override
+  public void merge(IndexWriter writer, MergeTrigger trigger, boolean newMergesFound) throws IOException {
+    // short-circuit
+    if (trigger != MergeTrigger.FULL_FLUSH || !writer.hasPendingMerges()) {
+      super.merge(writer, trigger, newMergesFound); // note: might actually do stuff; like change thread priority
+      return;
+    }
+
+    // Unfortunately Lucene doesn't give us a way to "peek" without taking the next merge.  And we can't
+    //  pass the merge we took onto CMS if we don't want to merge it eagerly.  So instead we awkwardly
+    //  hook in on the call to doMerge with an Observable, and we examine super.mergeThreads to see if
+    //  there is a cheap merge.
+    CountDownLatch latch = new CountDownLatch(1);
+    Observer observer = (o, arg) -> latch.countDown();
+    observableMerge.addObserver(observer);
+    try {
+
+      super.merge(writer, trigger, newMergesFound);
+      // at this point there should be some CMS MergeThreads
+
+      if (hasCheapMerge()) {
+        if (verbose()) message("Cheap merge; waiting synchronously");
+        try {
+          long startTimeNs = System.nanoTime();
+          if (!latch.await(cheapMaxMergeWaitTimeoutMs, TimeUnit.MILLISECONDS)) {
+            // don't guard with verbose(); this is kinda like a warning
+            message("WARN: reached timeout for cheap merge to complete");
+          } else {
+            if (verbose()) {
+              long ms = TimeUnit.NANOSECONDS.toMillis(System.nanoTime() - startTimeNs);
+              message("waited " + ms + "ms for cheap merge");
+              // note: the completed merge that triggered the latch could be an expensive merge if CMS is configured
+              //  with multiple merge threads and if it's timed right.  It doesn't matter; the point is
+              //  we waited a little for fewer segments (yay).
+            }
+          }
+        } catch (InterruptedException e) {
+          message("interrupted waiting for cheap merge");
+          // This is what IndexWriter javadocs says we should do; we don't set interrupt state.
+          throw new ThreadInterruptedException(e);
+        }
+      }
+    } finally {
+      observableMerge.deleteObserver(observer);
+    }
+  }
+
+  private boolean hasCheapMerge() throws IOException {
+    synchronized (this) {
+      for (MergeThread mergeThread : mergeThreads) {
+        MergePolicy.OneMerge oneMerge = mergeThread.merge;
+        // Unfortunately Lucene doesn't transfer the purpose of the merge so that we can further constrain
+        //   cheap merges here to a commit flush.  But I think practically speaking, if we find a "cheap"
+        //   merge, it is very likely generated by our custom MergePolicy which generates such cheap
+        //   merges exclusively at a commit.
+        // Note: maybe MP should use a custom OneMerge subclass that we see here? But it's nice to keep decoupled and not a big deal either.
+        if (!oneMerge.isExternal // unlikely but maybe we're merging external indexes
+            && !oneMerge.getMergeProgress().isPaused() // if paused, don't bother waiting for it!
+            && oneMerge.totalBytesSize() < cheapMergeThresholdBytes) { // is it cheap enough?
+          // note: infostream logs should already note size of merges happening
+          return true;
+        }
+      }
+      return false;
+    }
+  }
+
+  // called by a CMS MergeThread to do the actual merging
+  @Override
+  protected void doMerge(IndexWriter writer, MergePolicy.OneMerge merge) throws IOException {
+    super.doMerge(writer, merge);
+    observableMerge.notifyObservers(merge);
+  }
+
 
 Review comment:
   The new line here seems unnecessary.

----------------------------------------------------------------
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:
users@infra.apache.org


With regards,
Apache Git Services

---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscribe@lucene.apache.org
For additional commands, e-mail: issues-help@lucene.apache.org