hbase-commits mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From nspiegelb...@apache.org
Subject svn commit: r1181592 - /hbase/branches/0.89/src/main/java/org/apache/hadoop/hbase/io/hfile/HFileReaderV2.java
Date Tue, 11 Oct 2011 02:23:30 GMT
Author: nspiegelberg
Date: Tue Oct 11 02:23:27 2011
New Revision: 1181592

URL: http://svn.apache.org/viewvc?rev=1181592&view=rev
Log:
Don't rewind on a reseek operation in HFile v2 reader

Summary: Not doing a rewind to the first key in block on a reseek operation in
HFile v2 reader, because reseek implies going forward in the file. This fixes a
performance regression introduced in HFile v2 code. This evaded my performance
tests because it only applies to a case of sequential reads with skipping keys,
but I only did random read and contiguous sequential read tests.
Test Plan: Run unit tests. Run HBaseTest.
Reviewed By: kannan
Reviewers: kannan, nspiegelberg
CC: hbase@lists, , kannan
Revert Plan: OK
Differential Revision: 275483

Modified:
    hbase/branches/0.89/src/main/java/org/apache/hadoop/hbase/io/hfile/HFileReaderV2.java

Modified: hbase/branches/0.89/src/main/java/org/apache/hadoop/hbase/io/hfile/HFileReaderV2.java
URL: http://svn.apache.org/viewvc/hbase/branches/0.89/src/main/java/org/apache/hadoop/hbase/io/hfile/HFileReaderV2.java?rev=1181592&r1=1181591&r2=1181592&view=diff
==============================================================================
--- hbase/branches/0.89/src/main/java/org/apache/hadoop/hbase/io/hfile/HFileReaderV2.java
(original)
+++ hbase/branches/0.89/src/main/java/org/apache/hadoop/hbase/io/hfile/HFileReaderV2.java
Tue Oct 11 02:23:27 2011
@@ -501,20 +501,41 @@ public class HFileReaderV2 extends Abstr
       return seekTo(key, 0, key.length);
     }
 
-    @Override
-    public int seekTo(byte[] key, int offset, int length) throws IOException {
+    /**
+     * An internal API function. Seek to the given key, optionally rewinding to
+     * the first key of the block before doing the seek.
+     *
+     * @param key key byte array
+     * @param offset key offset in the key byte array
+     * @param length key length
+     * @param rewind whether to rewind to the first key of the block before
+     *        doing the seek. If this is false, we are assuming we never go
+     *        back, otherwise the result is undefined.
+     * @return -1 if the key is earlier than the first key of the file,
+     *         0 if we are at the given key, and 1 if we are past the given key
+     * @throws IOException
+     */
+    private int seekTo(byte[] key, int offset, int length, boolean rewind)
+        throws IOException {
       HFileBlock seekToBlock =
-          ((HFileReaderV2) reader).getDataBlockIndexReader().seekToDataBlock(
-              key, offset, length, block);
+        ((HFileReaderV2) reader).getDataBlockIndexReader().seekToDataBlock(
+            key, offset, length, block);
       if (seekToBlock == null) {
         // This happens if the key e.g. falls before the beginning of the file.
         return -1;
       }
-      return loadBlockAndSeekToKey(seekToBlock, true, key, offset, length,
+      return loadBlockAndSeekToKey(seekToBlock, rewind, key, offset, length,
           false);
     }
 
     @Override
+    public int seekTo(byte[] key, int offset, int length) throws IOException {
+      // Always rewind to the first key of the block, because the given key
+      // might be before or after the current key.
+      return seekTo(key, offset, length, true);
+    }
+
+    @Override
     public int reseekTo(byte[] key) throws IOException {
       return reseekTo(key, 0, key.length);
     }
@@ -531,7 +552,10 @@ public class HFileReaderV2 extends Abstr
           return compared;
         }
       }
-      return seekTo(key, offset, length);
+
+      // Don't rewind on a reseek operation, because reseek implies that we are
+      // always going forward in the file.
+      return seekTo(key, offset, length, false);
     }
 
     private int loadBlockAndSeekToKey(HFileBlock seekToBlock, boolean rewind,



Mime
View raw message