From 1415130b41be49186b0aa4a8ae4e5855a5c35721 Mon Sep 17 00:00:00 2001
From: Jean-Daniel Cryans <jdcryans@apache.org>
Date: Fri, 23 Jul 2010 22:44:21 +0000
Subject: [PATCH] HBASE-2866  Region permanently offlined

git-svn-id: https://svn.apache.org/repos/asf/hbase/trunk@967290 13f79535-47bb-0310-9956-ffa450edef68
---
 CHANGES.txt                                   |  1 +
 .../hbase/master/ZKUnassignedWatcher.java     |  1 +
 .../hbase/zookeeper/ZooKeeperWrapper.java     | 44 ++++++++++++-------
 3 files changed, 31 insertions(+), 15 deletions(-)

diff --git a/CHANGES.txt b/CHANGES.txt
index a5ce08295a7..d3e7d5f9553 100644
--- a/CHANGES.txt
+++ b/CHANGES.txt
@@ -449,6 +449,7 @@ Release 0.21.0 - Unreleased
    HBASE-2859  Cleanup deprecated stuff in TestHLog (Alex Newman via Stack)
    HBASE-2858  TestReplication.queueFailover fails half the time
    HBASE-2863  HBASE-2553 removed an important edge case
+   HBASE-2866  Region permanently offlined
 
   IMPROVEMENTS
    HBASE-1760  Cleanup TODOs in HTable
diff --git a/src/main/java/org/apache/hadoop/hbase/master/ZKUnassignedWatcher.java b/src/main/java/org/apache/hadoop/hbase/master/ZKUnassignedWatcher.java
index 75eb21e421d..acd23d11458 100644
--- a/src/main/java/org/apache/hadoop/hbase/master/ZKUnassignedWatcher.java
+++ b/src/main/java/org/apache/hadoop/hbase/master/ZKUnassignedWatcher.java
@@ -169,6 +169,7 @@ public class ZKUnassignedWatcher implements Watcher {
     String region = zNodePath.substring(
         zNodePath.indexOf(rgnInTransitNode) + rgnInTransitNode.length() + 1);
     HBaseEventType rsEvent = HBaseEventType.fromByte(data[0]);
+    LOG.debug("Got event type [ " + rsEvent + " ] for region " + region);
 
     // if the node was CLOSED then handle it
     if(rsEvent == HBaseEventType.RS2ZK_REGION_CLOSED) {
diff --git a/src/main/java/org/apache/hadoop/hbase/zookeeper/ZooKeeperWrapper.java b/src/main/java/org/apache/hadoop/hbase/zookeeper/ZooKeeperWrapper.java
index 82251780e5b..d19c881fae8 100644
--- a/src/main/java/org/apache/hadoop/hbase/zookeeper/ZooKeeperWrapper.java
+++ b/src/main/java/org/apache/hadoop/hbase/zookeeper/ZooKeeperWrapper.java
@@ -1133,22 +1133,36 @@ public class ZooKeeperWrapper implements Watcher {
       return;
     }
     
-    if(LOG.isDebugEnabled()) {
-      // Check existing state for logging purposes.
-      Stat stat = new Stat();
-      byte[] oldData = null;
-      try {
-        oldData = readZNode(znode, stat);
-      } catch (IOException e) {
-        LOG.error("Error reading data for " + znode);
-      }
-      if(oldData == null) {
-        LOG.debug("While updating UNASSIGNED region " + regionName + " - node exists with no data" );          
-      }
-      else {
-        LOG.debug("While updating UNASSIGNED region " + regionName + " exists, state = " + (HBaseEventType.fromByte(oldData[0])));
-      }
+    Stat stat = new Stat();
+    byte[] oldData = null;
+    try {
+      oldData = readZNode(znode, stat);
+    } catch (IOException e) {
+      LOG.error("Error reading data for " + znode);
     }
+    // If there is no data in the ZNode, then update it
+    if(oldData == null) {
+      LOG.debug("While updating UNASSIGNED region " + regionName + " - node exists with no data" );
+    }
+    // If there is data in the ZNode, do not update if it is already correct
+    else {
+      HBaseEventType curState = HBaseEventType.fromByte(oldData[0]);
+      HBaseEventType newState = HBaseEventType.fromByte(data[0]);
+      // If the znode has the right state already, do not update it. Updating
+      // the znode again and again will bump up the zk version. This may cause
+      // the region server to fail. The RS expects that the znode is never
+      // updated by anyone else while it is opening/closing a region.
+      if(curState == newState) {
+        LOG.debug("No need to update UNASSIGNED region " + regionName +
+                  " as it already exists in state = " + curState);
+        return;
+      }
+
+      // If the ZNode is in another state, then update it
+      LOG.debug("UNASSIGNED region " + regionName + " is currently in state = " +
+                curState + ", updating it to " + newState);
+    }
+    // Update the ZNode
     synchronized(unassignedZNodesWatched) {
       unassignedZNodesWatched.add(znode);
       try {