hbase-commits mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From nspiegelb...@apache.org
Subject svn commit: r1181383 - /hbase/branches/0.89/src/main/java/org/apache/hadoop/hbase/regionserver/HRegion.java
Date Tue, 11 Oct 2011 02:03:59 GMT
Author: nspiegelberg
Date: Tue Oct 11 02:03:58 2011
New Revision: 1181383

URL: http://svn.apache.org/viewvc?rev=1181383&view=rev
Log:
HBASE-2933 : Skip EOF Errors during Log Recovery

Summary:
If we get an EOF error during log recovery, the file was transferred fine but
contains bad application-level information, either due to a write/parse error or
master death while splitting.  Either way, we should continue instead of failing
and having the master reassign this region to another server and continue the
cycle.

Test Plan:
mvn test

DiffCamp Revision: 161212
Reviewed By: kannan
CC: nspiegelberg, kannan
Revert Plan:
OK

Modified:
    hbase/branches/0.89/src/main/java/org/apache/hadoop/hbase/regionserver/HRegion.java

Modified: hbase/branches/0.89/src/main/java/org/apache/hadoop/hbase/regionserver/HRegion.java
URL: http://svn.apache.org/viewvc/hbase/branches/0.89/src/main/java/org/apache/hadoop/hbase/regionserver/HRegion.java?rev=1181383&r1=1181382&r2=1181383&view=diff
==============================================================================
--- hbase/branches/0.89/src/main/java/org/apache/hadoop/hbase/regionserver/HRegion.java (original)
+++ hbase/branches/0.89/src/main/java/org/apache/hadoop/hbase/regionserver/HRegion.java Tue
Oct 11 02:03:58 2011
@@ -19,6 +19,7 @@
  */
 package org.apache.hadoop.hbase.regionserver;
 
+import java.io.EOFException;
 import java.io.IOException;
 import java.io.UnsupportedEncodingException;
 import java.lang.reflect.Constructor;
@@ -1954,83 +1955,85 @@ public class HRegion implements HeapSize
     LOG.info("Replaying edits from " + edits + "; minSeqId=" + minSeqId);
     HLog.Reader reader = HLog.getReader(this.fs, edits, conf);
     try {
-      return replayRecoveredEdits(reader, minSeqId, reporter);
-    } finally {
-      reader.close();
-    }
-  }
-
- /* @param reader Reader against file of recovered edits.
-  * @param minSeqId Any edit found in split editlogs needs to be in excess of
-  * this minSeqId to be applied, else its skipped.
-  * @param reporter
-  * @return the sequence id of the last edit added to this region out of the
-  * recovered edits log or <code>minSeqId</code> if nothing added from editlogs.
-  * @throws IOException
-  */
-  private long replayRecoveredEdits(final HLog.Reader reader,
-    final long minSeqId, final Progressable reporter)
-  throws IOException {
     long currentEditSeqId = minSeqId;
     long firstSeqIdInLog = -1;
     long skippedEdits = 0;
     long editsCount = 0;
     HLog.Entry entry;
     Store store = null;
-    // How many edits to apply before we send a progress report.
-    int interval = this.conf.getInt("hbase.hstore.report.interval.edits", 2000);
-    while ((entry = reader.next()) != null) {
-      HLogKey key = entry.getKey();
-      WALEdit val = entry.getEdit();
-      if (firstSeqIdInLog == -1) {
-        firstSeqIdInLog = key.getLogSeqNum();
-      }
-      // Now, figure if we should skip this edit.
-      if (key.getLogSeqNum() <= currentEditSeqId) {
-        skippedEdits++;
-        continue;
-      }
-      currentEditSeqId = key.getLogSeqNum();
-      boolean flush = false;
-      for (KeyValue kv: val.getKeyValues()) {
-        // Check this edit is for me. Also, guard against writing the special
-        // METACOLUMN info such as HBASE::CACHEFLUSH entries
-        if (kv.matchingFamily(HLog.METAFAMILY) ||
-            !Bytes.equals(key.getRegionName(), this.regionInfo.getRegionName())) {
-          skippedEdits++;
-          continue;
+
+    try {
+      // How many edits to apply before we send a progress report.
+      int interval = this.conf.getInt("hbase.hstore.report.interval.edits", 2000);
+      while ((entry = reader.next()) != null) {
+        HLogKey key = entry.getKey();
+        WALEdit val = entry.getEdit();
+        if (firstSeqIdInLog == -1) {
+          firstSeqIdInLog = key.getLogSeqNum();
         }
-        // Figure which store the edit is meant for.
-        if (store == null || !kv.matchingFamily(store.getFamily().getName())) {
-          store = this.stores.get(kv.getFamily());
-        }
-        if (store == null) {
-          // This should never happen.  Perhaps schema was changed between
-          // crash and redeploy?
-          LOG.warn("No family for " + kv);
+        // Now, figure if we should skip this edit.
+        if (key.getLogSeqNum() <= currentEditSeqId) {
           skippedEdits++;
           continue;
         }
-        // Once we are over the limit, restoreEdit will keep returning true to
-        // flush -- but don't flush until we've played all the kvs that make up
-        // the WALEdit.
-        flush = restoreEdit(store, kv);
-        editsCount++;
-     }
-     if (flush) internalFlushcache(null, currentEditSeqId);
-
-      // Every 'interval' edits, tell the reporter we're making progress.
-      // Have seen 60k edits taking 3minutes to complete.
-      if (reporter != null && (editsCount % interval) == 0) {
-        reporter.progress();
+        currentEditSeqId = key.getLogSeqNum();
+        boolean flush = false;
+        for (KeyValue kv: val.getKeyValues()) {
+          // Check this edit is for me. Also, guard against writing the special
+          // METACOLUMN info such as HBASE::CACHEFLUSH entries
+          if (kv.matchingFamily(HLog.METAFAMILY) ||
+              !Bytes.equals(key.getRegionName(), this.regionInfo.getRegionName())) {
+            skippedEdits++;
+            continue;
+          }
+          // Figure which store the edit is meant for.
+          if (store == null || !kv.matchingFamily(store.getFamily().getName())) {
+            store = this.stores.get(kv.getFamily());
+          }
+          if (store == null) {
+            // This should never happen.  Perhaps schema was changed between
+            // crash and redeploy?
+            LOG.warn("No family for " + kv);
+            skippedEdits++;
+            continue;
+          }
+          // Once we are over the limit, restoreEdit will keep returning true to
+          // flush -- but don't flush until we've played all the kvs that make up
+          // the WALEdit.
+          flush = restoreEdit(store, kv);
+          editsCount++;
+        }
+        if (flush) internalFlushcache(null, currentEditSeqId);
+
+        // Every 'interval' edits, tell the reporter we're making progress.
+        // Have seen 60k edits taking 3minutes to complete.
+        if (reporter != null && (editsCount % interval) == 0) {
+          reporter.progress();
+        }
+      }
+    } catch (EOFException eof) {
+      Path p = HLog.moveAsideBadEditsFile(fs, edits);
+      LOG.warn("Encountered EOF. Most likely due to Master failure during " +
+          "log spliting, so we have this data in another edit.  " +
+          "Continuing, but renaming " + edits + " as " + p, eof);
+    } catch (IOException ioe) {
+      if (ioe.getMessage().startsWith("File is corrupt")) {
+        Path p = HLog.moveAsideBadEditsFile(fs, edits);
+        LOG.warn("File corruption encountered!  " +
+            "Continuing, but renaming " + edits + " as " + p, ioe);
+      } else {
+        throw ioe;
       }
     }
     if (LOG.isDebugEnabled()) {
       LOG.debug("Applied " + editsCount + ", skipped " + skippedEdits +
-        ", firstSeqIdInLog=" + firstSeqIdInLog +
-        ", maxSeqIdInLog=" + currentEditSeqId);
+          ", firstSeqIdInLog=" + firstSeqIdInLog +
+          ", maxSeqIdInLog=" + currentEditSeqId);
     }
     return currentEditSeqId;
+    } finally {
+      reader.close();
+    }
   }
 
   /**



Mime
View raw message