Return-Path: X-Original-To: apmail-hbase-commits-archive@www.apache.org Delivered-To: apmail-hbase-commits-archive@www.apache.org Received: from mail.apache.org (hermes.apache.org [140.211.11.3]) by minotaur.apache.org (Postfix) with SMTP id 6B39010D91 for ; Fri, 11 Oct 2013 21:46:41 +0000 (UTC) Received: (qmail 77187 invoked by uid 500); 11 Oct 2013 21:46:41 -0000 Delivered-To: apmail-hbase-commits-archive@hbase.apache.org Received: (qmail 77156 invoked by uid 500); 11 Oct 2013 21:46:41 -0000 Mailing-List: contact commits-help@hbase.apache.org; run by ezmlm Precedence: bulk List-Help: List-Unsubscribe: List-Post: List-Id: Reply-To: dev@hbase.apache.org Delivered-To: mailing list commits@hbase.apache.org Received: (qmail 77149 invoked by uid 99); 11 Oct 2013 21:46:41 -0000 Received: from athena.apache.org (HELO athena.apache.org) (140.211.11.136) by apache.org (qpsmtpd/0.29) with ESMTP; Fri, 11 Oct 2013 21:46:41 +0000 X-ASF-Spam-Status: No, hits=-2000.0 required=5.0 tests=ALL_TRUSTED,T_FILL_THIS_FORM_SHORT X-Spam-Check-By: apache.org Received: from [140.211.11.4] (HELO eris.apache.org) (140.211.11.4) by apache.org (qpsmtpd/0.29) with ESMTP; Fri, 11 Oct 2013 21:46:36 +0000 Received: from eris.apache.org (localhost [127.0.0.1]) by eris.apache.org (Postfix) with ESMTP id 4FDE02388868; Fri, 11 Oct 2013 21:46:16 +0000 (UTC) Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit Subject: svn commit: r1531435 [1/2] - in /hbase/trunk: hbase-client/src/main/java/org/apache/hadoop/hbase/executor/ hbase-client/src/main/java/org/apache/hadoop/hbase/master/ hbase-it/src/test/java/org/apache/hadoop/hbase/chaos/actions/ hbase-protocol/src/main/... Date: Fri, 11 Oct 2013 21:46:15 -0000 To: commits@hbase.apache.org From: jxiang@apache.org X-Mailer: svnmailer-1.0.9 Message-Id: <20131011214616.4FDE02388868@eris.apache.org> X-Virus-Checked: Checked by ClamAV on apache.org Author: jxiang Date: Fri Oct 11 21:46:14 2013 New Revision: 1531435 URL: http://svn.apache.org/r1531435 Log: HBASE-9696 Master recovery ignores online merge znode Removed: hbase/trunk/hbase-server/src/main/java/org/apache/hadoop/hbase/master/handler/MergedRegionHandler.java hbase/trunk/hbase-server/src/main/java/org/apache/hadoop/hbase/master/handler/SplitRegionHandler.java Modified: hbase/trunk/hbase-client/src/main/java/org/apache/hadoop/hbase/executor/EventType.java hbase/trunk/hbase-client/src/main/java/org/apache/hadoop/hbase/master/RegionState.java hbase/trunk/hbase-it/src/test/java/org/apache/hadoop/hbase/chaos/actions/SplitRandomRegionOfTableAction.java hbase/trunk/hbase-protocol/src/main/java/org/apache/hadoop/hbase/protobuf/generated/ClusterStatusProtos.java hbase/trunk/hbase-protocol/src/main/protobuf/ClusterStatus.proto hbase/trunk/hbase-server/src/main/java/org/apache/hadoop/hbase/master/AssignmentManager.java hbase/trunk/hbase-server/src/main/java/org/apache/hadoop/hbase/master/RegionStates.java hbase/trunk/hbase-server/src/main/java/org/apache/hadoop/hbase/master/handler/ServerShutdownHandler.java hbase/trunk/hbase-server/src/main/java/org/apache/hadoop/hbase/regionserver/HRegion.java hbase/trunk/hbase-server/src/main/java/org/apache/hadoop/hbase/regionserver/RegionMergeTransaction.java hbase/trunk/hbase-server/src/main/java/org/apache/hadoop/hbase/regionserver/SplitTransaction.java hbase/trunk/hbase-server/src/test/java/org/apache/hadoop/hbase/master/TestMaster.java hbase/trunk/hbase-server/src/test/java/org/apache/hadoop/hbase/master/TestMasterFailover.java hbase/trunk/hbase-server/src/test/java/org/apache/hadoop/hbase/regionserver/TestSplitTransactionOnCluster.java Modified: hbase/trunk/hbase-client/src/main/java/org/apache/hadoop/hbase/executor/EventType.java URL: http://svn.apache.org/viewvc/hbase/trunk/hbase-client/src/main/java/org/apache/hadoop/hbase/executor/EventType.java?rev=1531435&r1=1531434&r2=1531435&view=diff ============================================================================== --- hbase/trunk/hbase-client/src/main/java/org/apache/hadoop/hbase/executor/EventType.java (original) +++ hbase/trunk/hbase-client/src/main/java/org/apache/hadoop/hbase/executor/EventType.java Fri Oct 11 21:46:14 2013 @@ -58,14 +58,14 @@ public enum EventType { RS_ZK_REGION_OPENED (4, ExecutorType.MASTER_OPEN_REGION), /** * RS_ZK_REGION_SPLITTING
- * - * RS has started a region split. + * + * RS has started a region split after master says it's ok to move on. */ RS_ZK_REGION_SPLITTING (5, null), /** * RS_ZK_REGION_SPLIT
- * - * RS split has completed. + * + * RS split has completed and is notifying the master. */ RS_ZK_REGION_SPLIT (6, ExecutorType.MASTER_SERVER_OPERATIONS), /** @@ -76,16 +76,30 @@ public enum EventType { RS_ZK_REGION_FAILED_OPEN (7, ExecutorType.MASTER_CLOSE_REGION), /** * RS_ZK_REGION_MERGING
- * - * RS has started merging regions. + * + * RS has started merging regions after master says it's ok to move on. */ RS_ZK_REGION_MERGING (8, null), /** * RS_ZK_REGION_MERGE
- * - * RS region merge has completed. + * + * RS region merge has completed and is notifying the master. */ RS_ZK_REGION_MERGED (9, ExecutorType.MASTER_SERVER_OPERATIONS), + /** + * RS_ZK_REQUEST_REGION_SPLIT
+ * + * RS has requested to split a region. This is to notify master + * and check with master if the region is in a state good to split. + */ + RS_ZK_REQUEST_REGION_SPLIT (10, null), + /** + * RS_ZK_REQUEST_REGION_MERGE
+ * + * RS has requested to merge two regions. This is to notify master + * and check with master if two regions is in states good to merge. + */ + RS_ZK_REQUEST_REGION_MERGE (11, null), /** * Messages originating from Master to RS.
Modified: hbase/trunk/hbase-client/src/main/java/org/apache/hadoop/hbase/master/RegionState.java URL: http://svn.apache.org/viewvc/hbase/trunk/hbase-client/src/main/java/org/apache/hadoop/hbase/master/RegionState.java?rev=1531435&r1=1531434&r2=1531435&view=diff ============================================================================== --- hbase/trunk/hbase-client/src/main/java/org/apache/hadoop/hbase/master/RegionState.java (original) +++ hbase/trunk/hbase-client/src/main/java/org/apache/hadoop/hbase/master/RegionState.java Fri Oct 11 21:46:14 2013 @@ -52,7 +52,13 @@ public class RegionState implements org. FAILED_OPEN, // failed to open, and won't retry any more FAILED_CLOSE, // failed to close, and won't retry any more MERGING, // server started merge a region - MERGED // server completed merge a region + MERGED, // server completed merge a region + SPLITTING_NEW, // new region to be created when RS splits a parent + // region but hasn't be created yet, or master doesn't + // know it's already created + MERGING_NEW // new region to be created when RS merges two + // daughter regions but hasn't be created yet, or + // master doesn't know it's already created } // Many threads can update the state at the stamp at the same time @@ -134,6 +140,10 @@ public class RegionState implements org. return state == State.SPLIT; } + public boolean isSplittingNew() { + return state == State.SPLITTING_NEW; + } + public boolean isFailedOpen() { return state == State.FAILED_OPEN; } @@ -150,10 +160,26 @@ public class RegionState implements org. return state == State.MERGED; } + public boolean isMergingNew() { + return state == State.MERGING_NEW; + } + public boolean isOpenOrMergingOnServer(final ServerName sn) { return isOnServer(sn) && (isOpened() || isMerging()); } + public boolean isOpenOrMergingNewOnServer(final ServerName sn) { + return isOnServer(sn) && (isOpened() || isMergingNew()); + } + + public boolean isOpenOrSplittingOnServer(final ServerName sn) { + return isOnServer(sn) && (isOpened() || isSplitting()); + } + + public boolean isOpenOrSplittingNewOnServer(final ServerName sn) { + return isOnServer(sn) && (isOpened() || isSplittingNew()); + } + public boolean isPendingOpenOrOpeningOnServer(final ServerName sn) { return isOnServer(sn) && isPendingOpenOrOpening(); } @@ -176,6 +202,28 @@ public class RegionState implements org. return serverName != null && serverName.equals(sn); } + // Is a region in a state ready to go offline + public boolean isReadyToOffline() { + return isMerged() || isSplit() || isOffline() + || isSplittingNew() || isMergingNew(); + } + + // Is a region in a state ready to go online + public boolean isReadyToOnline() { + return isOpened() || isSplittingNew() || isMergingNew(); + } + + // Is a region in a state not in transition but not unassignable + public boolean isNotUnassignableNotInTransition() { + return isNotUnassignableNotInTransition(state); + } + + // Check if a state is not in transition, but not unassignable + public static boolean isNotUnassignableNotInTransition(State state) { + return state == State.MERGED || state == State.SPLIT || state == State.OFFLINE + || state == State.SPLITTING_NEW || state == State.MERGING_NEW; + } + @Override public String toString() { return "{" + hri.getShortNameToLog() @@ -245,6 +293,12 @@ public class RegionState implements org. case MERGED: rs = ClusterStatusProtos.RegionState.State.MERGED; break; + case SPLITTING_NEW: + rs = ClusterStatusProtos.RegionState.State.SPLITTING_NEW; + break; + case MERGING_NEW: + rs = ClusterStatusProtos.RegionState.State.MERGING_NEW; + break; default: throw new IllegalStateException(""); } @@ -301,6 +355,12 @@ public class RegionState implements org. case MERGED: state = State.MERGED; break; + case SPLITTING_NEW: + state = State.SPLITTING_NEW; + break; + case MERGING_NEW: + state = State.MERGING_NEW; + break; default: throw new IllegalStateException(""); } Modified: hbase/trunk/hbase-it/src/test/java/org/apache/hadoop/hbase/chaos/actions/SplitRandomRegionOfTableAction.java URL: http://svn.apache.org/viewvc/hbase/trunk/hbase-it/src/test/java/org/apache/hadoop/hbase/chaos/actions/SplitRandomRegionOfTableAction.java?rev=1531435&r1=1531434&r2=1531435&view=diff ============================================================================== --- hbase/trunk/hbase-it/src/test/java/org/apache/hadoop/hbase/chaos/actions/SplitRandomRegionOfTableAction.java (original) +++ hbase/trunk/hbase-it/src/test/java/org/apache/hadoop/hbase/chaos/actions/SplitRandomRegionOfTableAction.java Fri Oct 11 21:46:14 2013 @@ -51,6 +51,11 @@ public class SplitRandomRegionOfTableAct LOG.info("Performing action: Split random region of table " + tableName); List regions = admin.getTableRegions(tableNameBytes); + if (regions == null || regions.isEmpty()) { + LOG.info("Table " + tableName + " doesn't have regions to split"); + return; + } + HRegionInfo region = PolicyBasedChaosMonkey.selectRandomItem( regions.toArray(new HRegionInfo[regions.size()])); LOG.debug("Splitting region " + region.getRegionNameAsString()); Modified: hbase/trunk/hbase-protocol/src/main/java/org/apache/hadoop/hbase/protobuf/generated/ClusterStatusProtos.java URL: http://svn.apache.org/viewvc/hbase/trunk/hbase-protocol/src/main/java/org/apache/hadoop/hbase/protobuf/generated/ClusterStatusProtos.java?rev=1531435&r1=1531434&r2=1531435&view=diff ============================================================================== --- hbase/trunk/hbase-protocol/src/main/java/org/apache/hadoop/hbase/protobuf/generated/ClusterStatusProtos.java (original) +++ hbase/trunk/hbase-protocol/src/main/java/org/apache/hadoop/hbase/protobuf/generated/ClusterStatusProtos.java Fri Oct 11 21:46:14 2013 @@ -273,6 +273,23 @@ public final class ClusterStatusProtos { * */ MERGED(12, 12), + /** + * SPLITTING_NEW = 13; + * + * 
+       * new region to be created when RS splits a parent
+       *
+ */ + SPLITTING_NEW(13, 13), + /** + * MERGING_NEW = 14; + * + * 
+       * region but hasn't be created yet, or master doesn't
+       * know it's already created
+       *
+ */ + MERGING_NEW(14, 14), ; /** @@ -379,6 +396,23 @@ public final class ClusterStatusProtos { * */ public static final int MERGED_VALUE = 12; + /** + * SPLITTING_NEW = 13; + * + * 
+       * new region to be created when RS splits a parent
+       *
+ */ + public static final int SPLITTING_NEW_VALUE = 13; + /** + * MERGING_NEW = 14; + * + * 
+       * region but hasn't be created yet, or master doesn't
+       * know it's already created
+       *
+ */ + public static final int MERGING_NEW_VALUE = 14; public final int getNumber() { return value; } @@ -398,6 +432,8 @@ public final class ClusterStatusProtos { case 10: return FAILED_CLOSE; case 11: return MERGING; case 12: return MERGED; + case 13: return SPLITTING_NEW; + case 14: return MERGING_NEW; default: return null; } } @@ -10303,48 +10339,49 @@ public final class ClusterStatusProtos { static { java.lang.String[] descriptorData = { "\n\023ClusterStatus.proto\032\013HBase.proto\032\017Clus" + - "terId.proto\032\010FS.proto\"\243\002\n\013RegionState\022 \n" + + "terId.proto\032\010FS.proto\"\307\002\n\013RegionState\022 \n" + "\013region_info\030\001 \002(\0132\013.RegionInfo\022!\n\005state" + "\030\002 \002(\0162\022.RegionState.State\022\r\n\005stamp\030\003 \001(" + - "\004\"\277\001\n\005State\022\013\n\007OFFLINE\020\000\022\020\n\014PENDING_OPEN" + + "\004\"\343\001\n\005State\022\013\n\007OFFLINE\020\000\022\020\n\014PENDING_OPEN" + "\020\001\022\013\n\007OPENING\020\002\022\010\n\004OPEN\020\003\022\021\n\rPENDING_CLO" + "SE\020\004\022\013\n\007CLOSING\020\005\022\n\n\006CLOSED\020\006\022\r\n\tSPLITTI" + "NG\020\007\022\t\n\005SPLIT\020\010\022\017\n\013FAILED_OPEN\020\t\022\020\n\014FAIL" + - "ED_CLOSE\020\n\022\013\n\007MERGING\020\013\022\n\n\006MERGED\020\014\"X\n\022R" + - "egionInTransition\022\036\n\004spec\030\001 \002(\0132\020.Region", - "Specifier\022\"\n\014region_state\030\002 \002(\0132\014.Region" + - "State\"\320\003\n\nRegionLoad\022*\n\020region_specifier" + - "\030\001 \002(\0132\020.RegionSpecifier\022\016\n\006stores\030\002 \001(\r" + - "\022\022\n\nstorefiles\030\003 \001(\r\022\"\n\032store_uncompress" + - "ed_size_MB\030\004 \001(\r\022\031\n\021storefile_size_MB\030\005 " + - "\001(\r\022\030\n\020memstore_size_MB\030\006 \001(\r\022\037\n\027storefi" + - "le_index_size_MB\030\007 \001(\r\022\033\n\023read_requests_" + - "count\030\010 \001(\004\022\034\n\024write_requests_count\030\t \001(" + - "\004\022\034\n\024total_compacting_KVs\030\n \001(\004\022\035\n\025curre" + - "nt_compacted_KVs\030\013 \001(\004\022\032\n\022root_index_siz", - "e_KB\030\014 \001(\r\022\"\n\032total_static_index_size_KB" + - "\030\r \001(\r\022\"\n\032total_static_bloom_size_KB\030\016 \001" + - "(\r\022\034\n\024complete_sequence_id\030\017 \001(\004\"\212\002\n\nSer" + - "verLoad\022\032\n\022number_of_requests\030\001 \001(\r\022 \n\030t" + - "otal_number_of_requests\030\002 \001(\r\022\024\n\014used_he" + - "ap_MB\030\003 \001(\r\022\023\n\013max_heap_MB\030\004 \001(\r\022!\n\014regi" + - "on_loads\030\005 \003(\0132\013.RegionLoad\022\"\n\014coprocess" + - "ors\030\006 \003(\0132\014.Coprocessor\022\031\n\021report_start_" + - "time\030\007 \001(\004\022\027\n\017report_end_time\030\010 \001(\004\022\030\n\020i" + - "nfo_server_port\030\t \001(\r\"O\n\016LiveServerInfo\022", - "\033\n\006server\030\001 \002(\0132\013.ServerName\022 \n\013server_l" + - "oad\030\002 \002(\0132\013.ServerLoad\"\340\002\n\rClusterStatus" + - "\022/\n\rhbase_version\030\001 \001(\0132\030.HBaseVersionFi" + - "leContent\022%\n\014live_servers\030\002 \003(\0132\017.LiveSe" + - "rverInfo\022!\n\014dead_servers\030\003 \003(\0132\013.ServerN" + - "ame\0222\n\025regions_in_transition\030\004 \003(\0132\023.Reg" + - "ionInTransition\022\036\n\ncluster_id\030\005 \001(\0132\n.Cl" + - "usterId\022)\n\023master_coprocessors\030\006 \003(\0132\014.C" + - "oprocessor\022\033\n\006master\030\007 \001(\0132\013.ServerName\022" + - "#\n\016backup_masters\030\010 \003(\0132\013.ServerName\022\023\n\013", - "balancer_on\030\t \001(\010BF\n*org.apache.hadoop.h" + - "base.protobuf.generatedB\023ClusterStatusPr" + - "otosH\001\240\001\001" + "ED_CLOSE\020\n\022\013\n\007MERGING\020\013\022\n\n\006MERGED\020\014\022\021\n\rS" + + "PLITTING_NEW\020\r\022\017\n\013MERGING_NEW\020\016\"X\n\022Regio", + "nInTransition\022\036\n\004spec\030\001 \002(\0132\020.RegionSpec" + + "ifier\022\"\n\014region_state\030\002 \002(\0132\014.RegionStat" + + "e\"\320\003\n\nRegionLoad\022*\n\020region_specifier\030\001 \002" + + "(\0132\020.RegionSpecifier\022\016\n\006stores\030\002 \001(\r\022\022\n\n" + + "storefiles\030\003 \001(\r\022\"\n\032store_uncompressed_s" + + "ize_MB\030\004 \001(\r\022\031\n\021storefile_size_MB\030\005 \001(\r\022" + + "\030\n\020memstore_size_MB\030\006 \001(\r\022\037\n\027storefile_i" + + "ndex_size_MB\030\007 \001(\r\022\033\n\023read_requests_coun" + + "t\030\010 \001(\004\022\034\n\024write_requests_count\030\t \001(\004\022\034\n" + + "\024total_compacting_KVs\030\n \001(\004\022\035\n\025current_c", + "ompacted_KVs\030\013 \001(\004\022\032\n\022root_index_size_KB" + + "\030\014 \001(\r\022\"\n\032total_static_index_size_KB\030\r \001" + + "(\r\022\"\n\032total_static_bloom_size_KB\030\016 \001(\r\022\034" + + "\n\024complete_sequence_id\030\017 \001(\004\"\212\002\n\nServerL" + + "oad\022\032\n\022number_of_requests\030\001 \001(\r\022 \n\030total" + + "_number_of_requests\030\002 \001(\r\022\024\n\014used_heap_M" + + "B\030\003 \001(\r\022\023\n\013max_heap_MB\030\004 \001(\r\022!\n\014region_l" + + "oads\030\005 \003(\0132\013.RegionLoad\022\"\n\014coprocessors\030" + + "\006 \003(\0132\014.Coprocessor\022\031\n\021report_start_time" + + "\030\007 \001(\004\022\027\n\017report_end_time\030\010 \001(\004\022\030\n\020info_", + "server_port\030\t \001(\r\"O\n\016LiveServerInfo\022\033\n\006s" + + "erver\030\001 \002(\0132\013.ServerName\022 \n\013server_load\030" + + "\002 \002(\0132\013.ServerLoad\"\340\002\n\rClusterStatus\022/\n\r" + + "hbase_version\030\001 \001(\0132\030.HBaseVersionFileCo" + + "ntent\022%\n\014live_servers\030\002 \003(\0132\017.LiveServer" + + "Info\022!\n\014dead_servers\030\003 \003(\0132\013.ServerName\022" + + "2\n\025regions_in_transition\030\004 \003(\0132\023.RegionI" + + "nTransition\022\036\n\ncluster_id\030\005 \001(\0132\n.Cluste" + + "rId\022)\n\023master_coprocessors\030\006 \003(\0132\014.Copro" + + "cessor\022\033\n\006master\030\007 \001(\0132\013.ServerName\022#\n\016b", + "ackup_masters\030\010 \003(\0132\013.ServerName\022\023\n\013bala" + + "ncer_on\030\t \001(\010BF\n*org.apache.hadoop.hbase" + + ".protobuf.generatedB\023ClusterStatusProtos" + + "H\001\240\001\001" }; com.google.protobuf.Descriptors.FileDescriptor.InternalDescriptorAssigner assigner = new com.google.protobuf.Descriptors.FileDescriptor.InternalDescriptorAssigner() { Modified: hbase/trunk/hbase-protocol/src/main/protobuf/ClusterStatus.proto URL: http://svn.apache.org/viewvc/hbase/trunk/hbase-protocol/src/main/protobuf/ClusterStatus.proto?rev=1531435&r1=1531434&r2=1531435&view=diff ============================================================================== --- hbase/trunk/hbase-protocol/src/main/protobuf/ClusterStatus.proto (original) +++ hbase/trunk/hbase-protocol/src/main/protobuf/ClusterStatus.proto Fri Oct 11 21:46:14 2013 @@ -45,6 +45,12 @@ message RegionState { FAILED_CLOSE = 10; // failed to close, and won't retry any more MERGING = 11; // server started merge a region MERGED = 12; // server completed merge of a region + SPLITTING_NEW = 13; // new region to be created when RS splits a parent + // region but hasn't be created yet, or master doesn't + // know it's already created + MERGING_NEW = 14; // new region to be created when RS merges two + // daughter regions but hasn't be created yet, or + // master doesn't know it's already created } } Modified: hbase/trunk/hbase-server/src/main/java/org/apache/hadoop/hbase/master/AssignmentManager.java URL: http://svn.apache.org/viewvc/hbase/trunk/hbase-server/src/main/java/org/apache/hadoop/hbase/master/AssignmentManager.java?rev=1531435&r1=1531434&r2=1531435&view=diff ============================================================================== --- hbase/trunk/hbase-server/src/main/java/org/apache/hadoop/hbase/master/AssignmentManager.java (original) +++ hbase/trunk/hbase-server/src/main/java/org/apache/hadoop/hbase/master/AssignmentManager.java Fri Oct 11 21:46:14 2013 @@ -68,15 +68,16 @@ import org.apache.hadoop.hbase.master.ba import org.apache.hadoop.hbase.master.handler.ClosedRegionHandler; import org.apache.hadoop.hbase.master.handler.DisableTableHandler; import org.apache.hadoop.hbase.master.handler.EnableTableHandler; -import org.apache.hadoop.hbase.master.handler.MergedRegionHandler; import org.apache.hadoop.hbase.master.handler.OpenedRegionHandler; -import org.apache.hadoop.hbase.master.handler.SplitRegionHandler; import org.apache.hadoop.hbase.regionserver.RegionAlreadyInTransitionException; +import org.apache.hadoop.hbase.regionserver.RegionMergeTransaction; import org.apache.hadoop.hbase.regionserver.RegionOpeningState; import org.apache.hadoop.hbase.regionserver.RegionServerStoppedException; +import org.apache.hadoop.hbase.regionserver.SplitTransaction; import org.apache.hadoop.hbase.util.EnvironmentEdgeManager; import org.apache.hadoop.hbase.util.KeyLocker; import org.apache.hadoop.hbase.util.Pair; +import org.apache.hadoop.hbase.util.PairOfSameType; import org.apache.hadoop.hbase.util.Threads; import org.apache.hadoop.hbase.util.Triple; import org.apache.hadoop.hbase.zookeeper.MetaRegionTracker; @@ -148,6 +149,12 @@ public class AssignmentManager extends Z private final int maximumAttempts; /** + * Map of two merging regions from the region to be created. + */ + private final Map> mergingRegions + = new HashMap>(); + + /** * The sleep time for which the assignment will wait before retrying in case of hbase:meta assignment * failure due to lack of availability of region plan */ @@ -176,9 +183,6 @@ public class AssignmentManager extends Z // For unit tests, keep track of calls to OpenedRegionHandler private Map openedRegionHandlerCalled = null; - // For unit tests, keep track of calls to SplitRegionHandler - private AtomicBoolean splitRegionHandlerCalled = null; - //Thread pool executor service for timeout monitor private java.util.concurrent.ExecutorService threadPoolExecutorService; @@ -226,6 +230,12 @@ public class AssignmentManager extends Z failedOpenTracker = new ConcurrentHashMap(); /** + * For testing only! Set to true to skip handling of split. + */ + @edu.umd.cs.findbugs.annotations.SuppressWarnings(value="MS_SHOULD_BE_FINAL") + public static boolean TEST_SKIP_SPLIT_HANDLING = false; + + /** * Constructs a new assignment manager. * * @param server @@ -552,8 +562,17 @@ public class AssignmentManager extends Z } HRegionInfo hri = regionInfo; if (hri == null) { + // Get the region from region states map/meta. However, we + // may still can't get it, for example, for online region merge, + // the znode uses the new region to be created, which may not in meta + // yet if the merging is still going on during the master recovery. hri = regionStates.getRegionInfo(rt.getRegionName()); - if (hri == null) return false; + EventType et = rt.getEventType(); + if (hri == null && et != EventType.RS_ZK_REGION_MERGING + && et != EventType.RS_ZK_REQUEST_REGION_MERGE) { + LOG.warn("Couldn't find the region in recovering " + rt); + return false; + } } processRegionsInTransition(rt, hri, stat.getVersion()); return true; @@ -575,12 +594,12 @@ public class AssignmentManager extends Z EventType et = rt.getEventType(); // Get ServerName. Could not be null. final ServerName sn = rt.getServerName(); - final String encodedRegionName = regionInfo.getEncodedName(); - final String prettyPrintedRegionName = HRegionInfo.prettyPrint(encodedRegionName); - LOG.info("Processing " + regionInfo.getRegionNameAsString() + " in state " + et); - + final byte[] regionName = rt.getRegionName(); + final String encodedName = HRegionInfo.encodeRegionName(regionName); + final String prettyPrintedRegionName = HRegionInfo.prettyPrint(encodedName); + LOG.info("Processing " + prettyPrintedRegionName + " in state " + et); - if (regionStates.isRegionInTransition(encodedRegionName)) { + if (regionStates.isRegionInTransition(encodedName)) { // Just return return; } @@ -637,7 +656,7 @@ public class AssignmentManager extends Z ReentrantLock lock = locker.acquireLock(regionInfo.getEncodedName()); try { RegionPlan plan = new RegionPlan(regionInfo, null, sn); - addPlan(encodedRegionName, plan); + addPlan(encodedName, plan); assign(rs, false, false); } finally { lock.unlock(); @@ -666,69 +685,32 @@ public class AssignmentManager extends Z new OpenedRegionHandler(server, this, regionInfo, sn, expectedVersion).process(); } break; + case RS_ZK_REQUEST_REGION_SPLIT: case RS_ZK_REGION_SPLITTING: - if (!serverManager.isServerOnline(sn)) { - // The regionserver started the split, but died before updating the status. - // It means (hopefully) that the split was not finished - // TBD - to study. In the meantime, do nothing as in the past. - LOG.warn("Processed region " + prettyPrintedRegionName + " in state : " + et + - " on a dead regionserver: " + sn + " doing nothing"); - } else { - // Splitting region should be online. We could have skipped it during - // user region rebuilding since we may consider the split is completed. - // Put it in SPLITTING state to avoid complications. - regionStates.regionOnline(regionInfo, sn); - regionStates.updateRegionState(rt, State.SPLITTING); - LOG.info("Processed " + prettyPrintedRegionName + " in state : " + et); - } - break; case RS_ZK_REGION_SPLIT: - if (!serverManager.isServerOnline(sn)) { - // The region is already in SPLIT state, do nothing - LOG.warn("Processed " + prettyPrintedRegionName - + " in state : " + et + " on a dead regionserver: " + sn - + " doing nothing"); - } else { + if (serverManager.isServerOnline(sn)) { // Splitting region should be online. We could have skipped it during // user region rebuilding since we may consider the split is completed. // Put it in SPLITTING state to avoid complications. regionStates.regionOnline(regionInfo, sn); regionStates.updateRegionState(rt, State.SPLITTING); - LOG.info("Processed " + prettyPrintedRegionName + " in state : " + et); - // Move the region to splitting state. The regionserver is supposed to update the znode - // multiple times so if it's still up we will receive an update soon. } - break; - case RS_ZK_REGION_MERGING: - if (!serverManager.isServerOnline(sn)) { - // The regionserver started the merge, but died before updating the status. - // It means (hopefully) that the merge was not finished - // This node should be gone soon since it is ephemeral. - LOG.warn("Processed " + prettyPrintedRegionName + " in state : " + et + - " on a dead regionserver: " + sn + " doing nothing"); - } else { - handleRegionMerging(rt, prettyPrintedRegionName, sn); - LOG.info("Processed region " + prettyPrintedRegionName - + " in state : " + et); + if (!handleRegionSplitting( + rt, encodedName, prettyPrintedRegionName, sn)) { + deleteSplittingNode(encodedName); } + LOG.info("Processed region " + prettyPrintedRegionName + + " in state : " + et); break; + case RS_ZK_REQUEST_REGION_MERGE: + case RS_ZK_REGION_MERGING: case RS_ZK_REGION_MERGED: - if (!serverManager.isServerOnline(sn)) { - // Do nothing, merging regions are already removed from meta, - // so they are not in region states map any more. - // The new region will be assigned by the ServerShutdownHandler - LOG.warn("Processed " + prettyPrintedRegionName - + " in state : " + et + " on a dead regionserver: " + sn - + " doing nothing"); - } else { - // Merging regions are already removed from meta. It doesn't hurt to - // do nothing here, no need to set them to merging state here. We are fine - // to put the new region to online state during user region rebuilding. - LOG.info("Processed " + prettyPrintedRegionName + " in state : " + - et + " nothing to do."); - // We don't do anything. The regionserver is supposed to update the znode - // multiple times so if it's still up we will receive an update soon. + if (!handleRegionMerging( + rt, encodedName, prettyPrintedRegionName, sn)) { + deleteMergingNode(encodedName); } + LOG.info("Processed region " + prettyPrintedRegionName + + " in state : " + et); break; default: throw new IllegalStateException("Received region in state :" + et + " is not valid."); @@ -811,7 +793,7 @@ public class AssignmentManager extends Z } RegionState regionState = - regionStates.getRegionTransitionState(encodedName); + regionStates.getRegionState(encodedName); long startTime = System.currentTimeMillis(); if (LOG.isDebugEnabled()) { boolean lateEvent = createTime < (startTime - 15000); @@ -831,7 +813,7 @@ public class AssignmentManager extends Z Lock lock = locker.acquireLock(encodedName); try { RegionState latestState = - regionStates.getRegionTransitionState(encodedName); + regionStates.getRegionState(encodedName); if ((regionState == null && latestState != null) || (regionState != null && latestState == null) || (regionState != null && latestState != null @@ -845,83 +827,24 @@ public class AssignmentManager extends Z } regionState = latestState; switch (rt.getEventType()) { + case RS_ZK_REQUEST_REGION_SPLIT: case RS_ZK_REGION_SPLITTING: - if (!isInStateForSplitting(regionState)) break; - regionStates.updateRegionState(rt, State.SPLITTING); - break; - case RS_ZK_REGION_SPLIT: - // RegionState must be null, or SPLITTING or PENDING_CLOSE. - if (!isInStateForSplitting(regionState)) break; - // If null, add SPLITTING state before going to SPLIT - if (regionState == null) { - regionState = regionStates.updateRegionState(rt, State.SPLITTING); - - String message = "Received SPLIT for region " + prettyPrintedRegionName + - " from server " + sn; - // If still null, it means we cannot find it and it was already processed - if (regionState == null) { - LOG.warn(message + " but it doesn't exist anymore," + - " probably already processed its split"); - break; - } - LOG.info(message + - " but region was not first in SPLITTING state; continuing"); - } - // Check it has daughters. - byte [] payload = rt.getPayload(); - List daughters; - try { - daughters = HRegionInfo.parseDelimitedFrom(payload, 0, payload.length); - } catch (IOException e) { - LOG.error("Dropped split! Failed reading split payload for " + - prettyPrintedRegionName); - break; + if (!handleRegionSplitting( + rt, encodedName, prettyPrintedRegionName, sn)) { + deleteSplittingNode(encodedName); } - assert daughters.size() == 2; - // Assert that we can get a serverinfo for this server. - if (!this.serverManager.isServerOnline(sn)) { - LOG.error("Dropped split! ServerName=" + sn + " unknown."); - break; - } - // Run handler to do the rest of the SPLIT handling. - new SplitRegionHandler(server, this, regionState.getRegion(), sn, daughters).process(); - updateSplitHandlerTracker(); break; + case RS_ZK_REQUEST_REGION_MERGE: case RS_ZK_REGION_MERGING: + case RS_ZK_REGION_MERGED: // Merged region is a new region, we can't find it in the region states now. // However, the two merging regions are not new. They should be in state for merging. - handleRegionMerging(rt, prettyPrintedRegionName, sn); - break; - - case RS_ZK_REGION_MERGED: - // Assert that we can get a serverinfo for this server. - if (!this.serverManager.isServerOnline(sn)) { - LOG.error("Dropped merge! ServerName=" + sn + " unknown."); - break; - } - // Get merged and merging regions. - byte[] payloadOfMerge = rt.getPayload(); - List mergeRegions; - try { - mergeRegions = HRegionInfo.parseDelimitedFrom(payloadOfMerge, 0, - payloadOfMerge.length); - } catch (IOException e) { - LOG.error("Dropped merge! Failed reading merge payload for " + - prettyPrintedRegionName); - break; - } - assert mergeRegions.size() == 3; - HRegionInfo merge_a = mergeRegions.get(1); - HRegionInfo merge_b = mergeRegions.get(2); - if (!isInStateForMerging(sn, merge_a, merge_b)) { - // Move on. Merge already happened (passed PONR), no point to stop now - LOG.warn("Got merge event, but not in state good for MERGED; rs_a=" - + merge_a + ", rs_b=" + merge_b); + if (!handleRegionMerging( + rt, encodedName, prettyPrintedRegionName, sn)) { + deleteMergingNode(encodedName); } - // Run handler to do the rest of the MERGED handling. - new MergedRegionHandler(server, this, sn, mergeRegions).process(); break; case M_ZK_REGION_CLOSING: @@ -1056,18 +979,9 @@ public class AssignmentManager extends Z } //For unit tests only - boolean wasSplitHandlerCalled() { - //compareAndSet to be sure that unit tests don't see stale values. Means, - //we will return true exactly once unless the handler code resets to true - //this value. - return splitRegionHandlerCalled.compareAndSet(true, false); - } - - //For unit tests only void initializeHandlerTrackers() { closedRegionHandlerCalled = new HashMap(); openedRegionHandlerCalled = new HashMap(); - splitRegionHandlerCalled = new AtomicBoolean(false); } void updateClosedRegionHandlerTracker(HRegionInfo hri) { @@ -1082,36 +996,6 @@ public class AssignmentManager extends Z } } - void updateSplitHandlerTracker() { - if (splitRegionHandlerCalled != null) { //only for unit tests this is true - splitRegionHandlerCalled.set(true); - } - } - - /** - * @return Returns true if this RegionState is splittable; i.e. the - * RegionState is currently in splitting state or pending_close or - * null (Anything else will return false). (Anything else will return false). - */ - private boolean isInStateForSplitting(final RegionState rs) { - if (rs == null) return true; - if (rs.isSplitting()) return true; - if (convertPendingCloseToSplitting(rs)) return true; - LOG.warn("Dropped region split! Not in state good for SPLITTING; rs=" + rs); - return false; - } - - /** - * @return Returns true if both regions are merging/open on specified server - */ - private boolean isInStateForMerging(final ServerName sn, - final HRegionInfo a, final HRegionInfo b) { - RegionState rs_a = regionStates.getRegionState(a); - RegionState rs_b = regionStates.getRegionState(b); - return ((rs_a == null || rs_a.isOpenOrMergingOnServer(sn)) - && (rs_b == null || rs_b.isOpenOrMergingOnServer(sn))); - } - // TODO: processFavoredNodes might throw an exception, for e.g., if the // meta could not be contacted/updated. We need to see how seriously to treat // this problem as. Should we fail the current assignment. We should be able @@ -1131,25 +1015,6 @@ public class AssignmentManager extends Z } /** - * If the passed regionState is in PENDING_CLOSE, clean up PENDING_CLOSE - * state and convert it to SPLITTING instead. - * This can happen in case where master wants to close a region at same time - * a regionserver starts a split. The split won. Clean out old PENDING_CLOSE - * state. - * @param rs - * @return True if we converted from PENDING_CLOSE to SPLITTING - */ - private boolean convertPendingCloseToSplitting(final RegionState rs) { - if (!rs.isPendingClose()) return false; - LOG.debug("Converting PENDING_CLOSE to SPLITTING; rs=" + rs); - regionStates.updateRegionState(rs.getRegion(), State.SPLITTING); - // Clean up existing state. Clear from region plans seems all we - // have to do here by way of clean up of PENDING_CLOSE. - clearRegionPlan(rs.getRegion()); - return true; - } - - /** * Handle a ZK unassigned node transition triggered by HBCK repair tool. *

* This is handled in a separate code path because it breaks the normal rules. @@ -1305,38 +1170,56 @@ public class AssignmentManager extends Z Lock lock = locker.acquireLock(regionName); try { RegionState rs = regionStates.getRegionTransitionState(regionName); - if (rs == null) return; + if (rs == null) { + rs = regionStates.getRegionState(regionName); + if (rs == null || !rs.isMergingNew()) { + // MergingNew is an offline state + return; + } + } HRegionInfo regionInfo = rs.getRegion(); String regionNameStr = regionInfo.getRegionNameAsString(); LOG.debug("Znode " + regionNameStr + " deleted, state: " + rs); - if (rs.isOpened()) { - ServerName serverName = rs.getServerName(); - regionOnline(regionInfo, serverName); - boolean disabled = getZKTable().isDisablingOrDisabledTable(regionInfo.getTable()); - if (!serverManager.isServerOnline(serverName) && !disabled) { - LOG.info("Opened " + regionNameStr - + "but the region server is offline, reassign the region"); - assign(regionInfo, true); - } else if (disabled) { - // if server is offline, no hurt to unassign again - LOG.info("Opened " + regionNameStr - + "but this table is disabled, triggering close of region"); - unassign(regionInfo); + boolean disabled = getZKTable().isDisablingOrDisabledTable(regionInfo.getTable()); + ServerName serverName = rs.getServerName(); + if (serverManager.isServerOnline(serverName)) { + if (rs.isOnServer(serverName) + && (rs.isOpened() || rs.isSplitting())) { + regionOnline(regionInfo, serverName); + if (disabled) { + // if server is offline, no hurt to unassign again + LOG.info("Opened " + regionNameStr + + "but this table is disabled, triggering close of region"); + unassign(regionInfo); + } + } else if (rs.isMergingNew()) { + synchronized (regionStates) { + String p = regionInfo.getEncodedName(); + PairOfSameType regions = mergingRegions.get(p); + if (regions != null) { + onlineMergingRegion(disabled, regions.getFirst(), serverName); + onlineMergingRegion(disabled, regions.getSecond(), serverName); + } + } } - } else if (rs.isSplitting()) { - LOG.debug("Ephemeral node deleted. Found in SPLITTING state. " + "Removing from RIT " - + rs.getRegion()); - // it can be either SPLIT fail, or RS dead. - regionStates.regionOnline(rs.getRegion(), rs.getServerName()); } - // RS does not delete the znode in case SPLIT, it only means RS died which - // will be handled by SSH - // in region merge we do not put merging regions to MERGING state } finally { lock.unlock(); } } + + private void onlineMergingRegion(boolean disabled, + final HRegionInfo hri, final ServerName serverName) { + RegionState regionState = regionStates.getRegionState(hri); + if (regionState != null && regionState.isMerging() + && regionState.isOnServer(serverName)) { + regionOnline(regionState.getRegion(), serverName); + if (disabled) { + unassign(hri); + } + } + } }); } } @@ -1371,23 +1254,7 @@ public class AssignmentManager extends Z // on it, so no need to watch it again. So, as I know for now, // this is needed to watch splitting nodes only. if (!regionStates.isRegionInTransition(child)) { - stat.setVersion(0); - byte[] data = ZKAssign.getDataAndWatch(watcher, - ZKUtil.joinZNode(watcher.assignmentZNode, child), stat); - if (data != null && stat.getVersion() > 0) { - try { - RegionTransition rt = RegionTransition.parseFrom(data); - - //See HBASE-7551, handle splitting too, in case we miss the node change event - EventType type = rt.getEventType(); - if (type == EventType.RS_ZK_REGION_SPLITTING - || type == EventType.RS_ZK_REGION_MERGING) { - handleRegion(rt, stat.getVersion()); - } - } catch (DeserializationException de) { - LOG.error("error getting data for " + child, de); - } - } + ZKAssign.getDataAndWatch(watcher, child, stat); } } } @@ -2374,7 +2241,8 @@ public class AssignmentManager extends Z */ public void unassign(HRegionInfo region, boolean force, ServerName dest) { // TODO: Method needs refactoring. Ugly buried returns throughout. Beware! - LOG.debug("Starting unassign of " + region.getRegionNameAsString() + " (offlining)"); + LOG.debug("Starting unassign of " + region.getRegionNameAsString() + + " (offlining), current state: " + regionStates.getRegionState(region)); String encodedName = region.getEncodedName(); // Grab the state of this region and synchronize on it @@ -2389,8 +2257,7 @@ public class AssignmentManager extends Z // Region is not in transition. // We can unassign it only if it's not SPLIT/MERGED. state = regionStates.getRegionState(encodedName); - if (state != null && (state.isMerged() - || state.isSplit() || state.isOffline())) { + if (state != null && state.isNotUnassignableNotInTransition()) { LOG.info("Attempting to unassign " + state + ", ignored"); // Offline region will be reassigned below return; @@ -2484,27 +2351,9 @@ public class AssignmentManager extends Z * @param region regioninfo of znode to be deleted. */ public void deleteClosingOrClosedNode(HRegionInfo region) { - String encodedName = region.getEncodedName(); - try { - if (!ZKAssign.deleteNode(watcher, encodedName, - EventType.M_ZK_REGION_CLOSING)) { - boolean deleteNode = ZKAssign.deleteNode(watcher, - encodedName, EventType.RS_ZK_REGION_CLOSED); - // TODO : We don't abort if the delete node returns false. Is there any - // such corner case? - if (!deleteNode) { - LOG.error("The deletion of the CLOSED node for " - + encodedName + " returned " + deleteNode); - } - } - } catch (NoNodeException e) { - LOG.debug("CLOSING/CLOSED node for " + encodedName - + " already deleted"); - } catch (KeeperException ke) { - server.abort( - "Unexpected ZK exception deleting node CLOSING/CLOSED for the region " - + encodedName, ke); - } + String regionName = region.getEncodedName(); + deleteNodeInStates(regionName, "closing", EventType.M_ZK_REGION_CLOSING, + EventType.RS_ZK_REGION_CLOSED); } /** @@ -2519,16 +2368,22 @@ public class AssignmentManager extends Z // This may fail if the SPLIT or SPLITTING or MERGED or MERGING znode gets // cleaned up before we can get data from it. byte [] data = ZKAssign.getData(watcher, path); - if (data == null) return false; + if (data == null) { + LOG.info("Node " + path + " is gone"); + return false; + } RegionTransition rt = RegionTransition.parseFrom(data); switch (rt.getEventType()) { + case RS_ZK_REQUEST_REGION_SPLIT: case RS_ZK_REGION_SPLIT: case RS_ZK_REGION_SPLITTING: + case RS_ZK_REQUEST_REGION_MERGE: case RS_ZK_REGION_MERGED: case RS_ZK_REGION_MERGING: result = true; break; default: + LOG.info("Node " + path + " is in " + rt.getEventType()); break; } return result; @@ -2818,6 +2673,7 @@ public class AssignmentManager extends Z // Region is being served and on an active server // add only if region not in disabled or enabling table if (!disabledOrEnablingTables.contains(tableName)) { + regionStates.updateRegionState(regionInfo, State.OPEN, regionLocation); regionStates.regionOnline(regionInfo, regionLocation); } // need to enable the table if not disabled or disabling or enabling @@ -3275,8 +3131,9 @@ public class AssignmentManager extends Z server.abort("Unexpected ZK exception deleting node " + hri, ke); } if (zkTable.isDisablingOrDisabledTable(hri.getTable())) { - it.remove(); + regionStates.updateRegionState(hri, State.OFFLINE); regionStates.regionOffline(hri); + it.remove(); continue; } // Mark the region offline and assign it again by SSH @@ -3290,38 +3147,6 @@ public class AssignmentManager extends Z } /** - * Update inmemory structures. - * @param sn Server that reported the split - * @param parent Parent region that was split - * @param a Daughter region A - * @param b Daughter region B - */ - public void handleSplitReport(final ServerName sn, final HRegionInfo parent, - final HRegionInfo a, final HRegionInfo b) { - synchronized (regionStates) { - regionOffline(parent, State.SPLIT); - onlineNewRegion(a, sn); - onlineNewRegion(b, sn); - } - } - - /** - * Update inmemory structures. - * @param sn Server that reported the merge - * @param merged regioninfo of merged - * @param a region a - * @param b region b - */ - public void handleRegionsMergeReport(final ServerName sn, - final HRegionInfo merged, final HRegionInfo a, final HRegionInfo b) { - synchronized (regionStates) { - regionOffline(a, State.MERGED); - regionOffline(b, State.MERGED); - onlineNewRegion(merged, sn); - } - } - - /** * @param plan Plan to execute. */ public void balance(final RegionPlan plan) { @@ -3397,33 +3222,283 @@ public class AssignmentManager extends Z return true; } + private boolean deleteNodeInStates( + String regionName, String desc, EventType... types) { + try { + for (EventType et: types) { + if (ZKAssign.deleteNode(watcher, regionName, et)) { + return true; + } + } + LOG.info("Failed to delete the " + desc + " node for " + + regionName + ". The node type may not match"); + } catch (NoNodeException e) { + LOG.debug("The " + desc + " node for " + regionName + " already deleted"); + } catch (KeeperException ke) { + server.abort("Unexpected ZK exception deleting " + desc + + " node for the region " + regionName, ke); + } + return false; + } + + private void deleteMergingNode(String encodedName) { + deleteNodeInStates(encodedName, "merging", EventType.RS_ZK_REGION_MERGING, + EventType.RS_ZK_REQUEST_REGION_MERGE, EventType.RS_ZK_REGION_MERGED); + } + + private void deleteSplittingNode(String encodedName) { + deleteNodeInStates(encodedName, "splitting", EventType.RS_ZK_REGION_SPLITTING, + EventType.RS_ZK_REQUEST_REGION_SPLIT, EventType.RS_ZK_REGION_SPLIT); + } + /** * A helper to handle region merging transition event. * It transitions merging regions to MERGING state. */ - private boolean handleRegionMerging(final RegionTransition rt, + private boolean handleRegionMerging(final RegionTransition rt, final String encodedName, final String prettyPrintedRegionName, final ServerName sn) { + if (!serverManager.isServerOnline(sn)) { + LOG.warn("Dropped merging! ServerName=" + sn + " unknown."); + return false; + } byte [] payloadOfMerging = rt.getPayload(); List mergingRegions; try { mergingRegions = HRegionInfo.parseDelimitedFrom( payloadOfMerging, 0, payloadOfMerging.length); } catch (IOException e) { - LOG.error("Dropped merging! Failed reading merging payload for " - + prettyPrintedRegionName); + LOG.error("Dropped merging! Failed reading " + rt.getEventType() + + " payload for " + prettyPrintedRegionName); + return false; + } + assert mergingRegions.size() == 3; + HRegionInfo p = mergingRegions.get(0); + HRegionInfo hri_a = mergingRegions.get(1); + HRegionInfo hri_b = mergingRegions.get(2); + + RegionState rs_p = regionStates.getRegionState(p); + RegionState rs_a = regionStates.getRegionState(hri_a); + RegionState rs_b = regionStates.getRegionState(hri_b); + + if (!((rs_a == null || rs_a.isOpenOrMergingOnServer(sn)) + && (rs_b == null || rs_b.isOpenOrMergingOnServer(sn)) + && (rs_p == null || rs_p.isOpenOrMergingNewOnServer(sn)))) { + LOG.warn("Dropped merging! Not in state good for MERGING; rs_p=" + + rs_p + ", rs_a=" + rs_a + ", rs_b=" + rs_b); + return false; + } + + EventType et = rt.getEventType(); + if (et == EventType.RS_ZK_REQUEST_REGION_MERGE) { + try { + if (RegionMergeTransaction.transitionMergingNode(watcher, p, + hri_a, hri_b, sn, -1, EventType.RS_ZK_REQUEST_REGION_MERGE, + EventType.RS_ZK_REGION_MERGING) == -1) { + byte[] data = ZKAssign.getData(watcher, encodedName); + EventType currentType = null; + if (data != null) { + RegionTransition newRt = RegionTransition.parseFrom(data); + currentType = newRt.getEventType(); + } + if (currentType == null || (currentType != EventType.RS_ZK_REGION_MERGED + && currentType != EventType.RS_ZK_REGION_MERGING)) { + LOG.warn("Failed to transition pending_merge node " + + encodedName + " to merging, it's now " + currentType); + return false; + } + } + } catch (Exception e) { + LOG.warn("Failed to transition pending_merge node " + + encodedName + " to merging", e); + return false; + } + } + + synchronized (regionStates) { + if (regionStates.getRegionState(p) == null) { + regionStates.createRegionState(p); + } + regionStates.updateRegionState(hri_a, State.MERGING); + regionStates.updateRegionState(hri_b, State.MERGING); + + if (et != EventType.RS_ZK_REGION_MERGED) { + regionStates.updateRegionState(p, State.MERGING_NEW, sn);; + regionStates.regionOffline(p, State.MERGING_NEW); + this.mergingRegions.put(encodedName, + new PairOfSameType(hri_a, hri_b)); + } else { + this.mergingRegions.remove(encodedName); + regionStates.updateRegionState(hri_a, State.MERGED); + regionStates.updateRegionState(hri_b, State.MERGED); + regionOffline(hri_a, State.MERGED); + regionOffline(hri_b, State.MERGED); + regionOnline(p, sn); + } + } + + if (et == EventType.RS_ZK_REGION_MERGED) { + LOG.debug("Handling MERGED event for " + encodedName + "; deleting node"); + // Remove region from ZK + try { + boolean successful = false; + while (!successful) { + // It's possible that the RS tickles in between the reading of the + // znode and the deleting, so it's safe to retry. + successful = ZKAssign.deleteNode( + watcher, encodedName, EventType.RS_ZK_REGION_MERGED); + } + } catch (KeeperException e) { + if (e instanceof NoNodeException) { + String znodePath = ZKUtil.joinZNode(watcher.splitLogZNode, encodedName); + LOG.debug("The znode " + znodePath + " does not exist. May be deleted already."); + } else { + server.abort("Error deleting MERGED node " + encodedName, e); + } + } + LOG.info("Handled MERGED event; merged=" + p.getRegionNameAsString() + + ", region_a=" + hri_a.getRegionNameAsString() + ", region_b=" + + hri_b.getRegionNameAsString() + ", on " + sn); + + // User could disable the table before master knows the new region. + if (zkTable.isDisablingOrDisabledTable(p.getTable())) { + unassign(p); + } + } + return true; + } + + /** + * A helper to handle region splitting transition event. + */ + private boolean handleRegionSplitting(final RegionTransition rt, final String encodedName, + final String prettyPrintedRegionName, final ServerName sn) { + if (!serverManager.isServerOnline(sn)) { + LOG.warn("Dropped splitting! ServerName=" + sn + " unknown."); + return false; + } + byte [] payloadOfSplitting = rt.getPayload(); + List splittingRegions; + try { + splittingRegions = HRegionInfo.parseDelimitedFrom( + payloadOfSplitting, 0, payloadOfSplitting.length); + } catch (IOException e) { + LOG.error("Dropped splitting! Failed reading " + rt.getEventType() + + " payload for " + prettyPrintedRegionName); return false; } - assert mergingRegions.size() == 2; - HRegionInfo merging_a = mergingRegions.get(0); - HRegionInfo merging_b = mergingRegions.get(1); - - if (!isInStateForMerging(sn, merging_a, merging_b)) { - LOG.warn("Dropped merging! Not in state good for MERGING; rs_a=" - + merging_a + ", rs_b=" + merging_b); + assert splittingRegions.size() == 2; + HRegionInfo hri_a = splittingRegions.get(0); + HRegionInfo hri_b = splittingRegions.get(1); + + RegionState rs_p = regionStates.getRegionState(encodedName); + RegionState rs_a = regionStates.getRegionState(hri_a); + RegionState rs_b = regionStates.getRegionState(hri_b); + + if (!((rs_p == null || rs_p.isOpenOrSplittingOnServer(sn)) + && (rs_a == null || rs_a.isOpenOrSplittingNewOnServer(sn)) + && (rs_b == null || rs_b.isOpenOrSplittingNewOnServer(sn)))) { + LOG.warn("Dropped splitting! Not in state good for SPLITTING; rs_p=" + + rs_p + ", rs_a=" + rs_a + ", rs_b=" + rs_b); return false; } - regionStates.updateRegionState(merging_a, State.MERGING); - regionStates.updateRegionState(merging_b, State.MERGING); + + if (rs_p == null) { + // Splitting region should be online + rs_p = regionStates.updateRegionState(rt, State.OPEN); + if (rs_p == null) { + LOG.warn("Received splitting for region " + prettyPrintedRegionName + + " from server " + sn + " but it doesn't exist anymore," + + " probably already processed its split"); + return false; + } + regionStates.regionOnline(rs_p.getRegion(), sn); + } + + HRegionInfo p = rs_p.getRegion(); + EventType et = rt.getEventType(); + if (et == EventType.RS_ZK_REQUEST_REGION_SPLIT) { + try { + if (SplitTransaction.transitionSplittingNode(watcher, p, + hri_a, hri_b, sn, -1, EventType.RS_ZK_REQUEST_REGION_SPLIT, + EventType.RS_ZK_REGION_SPLITTING) == -1) { + byte[] data = ZKAssign.getData(watcher, encodedName); + EventType currentType = null; + if (data != null) { + RegionTransition newRt = RegionTransition.parseFrom(data); + currentType = newRt.getEventType(); + } + if (currentType == null || (currentType != EventType.RS_ZK_REGION_SPLIT + && currentType != EventType.RS_ZK_REGION_SPLITTING)) { + LOG.warn("Failed to transition pending_split node " + + encodedName + " to splitting, it's now " + currentType); + return false; + } + } + } catch (Exception e) { + LOG.warn("Failed to transition pending_split node " + + encodedName + " to splitting", e); + return false; + } + } + + synchronized (regionStates) { + if (regionStates.getRegionState(hri_a) == null) { + regionStates.createRegionState(hri_a); + } + if (regionStates.getRegionState(hri_b) == null) { + regionStates.createRegionState(hri_b); + } + + regionStates.updateRegionState(hri_a, State.SPLITTING_NEW, sn); + regionStates.updateRegionState(hri_b, State.SPLITTING_NEW, sn); + regionStates.regionOffline(hri_a, State.SPLITTING_NEW); + regionStates.regionOffline(hri_b, State.SPLITTING_NEW); + regionStates.updateRegionState(rt, State.SPLITTING); + + // The below is for testing ONLY! We can't do fault injection easily, so + // resort to this kinda uglyness -- St.Ack 02/25/2011. + if (TEST_SKIP_SPLIT_HANDLING) { + LOG.warn("Skipping split message, TEST_SKIP_SPLIT_HANDLING is set"); + return true; // return true so that the splitting node stays + } + + if (et == EventType.RS_ZK_REGION_SPLIT) { + regionStates.updateRegionState(p, State.SPLIT); + regionOffline(p, State.SPLIT); + regionOnline(hri_a, sn); + regionOnline(hri_b, sn); + } + } + + if (et == EventType.RS_ZK_REGION_SPLIT) { + LOG.debug("Handling SPLIT event for " + encodedName + "; deleting node"); + // Remove region from ZK + try { + boolean successful = false; + while (!successful) { + // It's possible that the RS tickles in between the reading of the + // znode and the deleting, so it's safe to retry. + successful = ZKAssign.deleteNode( + watcher, encodedName, EventType.RS_ZK_REGION_SPLIT); + } + } catch (KeeperException e) { + if (e instanceof NoNodeException) { + String znodePath = ZKUtil.joinZNode(watcher.splitLogZNode, encodedName); + LOG.debug("The znode " + znodePath + " does not exist. May be deleted already."); + } else { + server.abort("Error deleting SPLIT node " + encodedName, e); + } + } + LOG.info("Handled SPLIT event; parent=" + p.getRegionNameAsString() + + ", daughter a=" + hri_a.getRegionNameAsString() + ", daughter b=" + + hri_b.getRegionNameAsString() + ", on " + sn); + + // User could disable the table before master knows the new region. + if (zkTable.isDisablingOrDisabledTable(p.getTable())) { + unassign(hri_a); + unassign(hri_b); + } + } return true; } @@ -3438,21 +3513,4 @@ public class AssignmentManager extends Z // remove the region plan as well just in case. clearRegionPlan(regionInfo); } - - /** - * Online a newly created region, which is usually from split/merge. - */ - private void onlineNewRegion(final HRegionInfo region, final ServerName sn) { - synchronized (regionStates) { - // Someone could find the region from meta and reassign it. - if (regionStates.getRegionState(region) == null) { - regionStates.createRegionState(region); - regionOnline(region, sn); - } - } - // User could disable the table before master knows the new region. - if (zkTable.isDisablingOrDisabledTable(region.getTable())) { - unassign(region); - } - } } Modified: hbase/trunk/hbase-server/src/main/java/org/apache/hadoop/hbase/master/RegionStates.java URL: http://svn.apache.org/viewvc/hbase/trunk/hbase-server/src/main/java/org/apache/hadoop/hbase/master/RegionStates.java?rev=1531435&r1=1531434&r2=1531435&view=diff ============================================================================== --- hbase/trunk/hbase-server/src/main/java/org/apache/hadoop/hbase/master/RegionStates.java (original) +++ hbase/trunk/hbase-server/src/main/java/org/apache/hadoop/hbase/master/RegionStates.java Fri Oct 11 21:46:14 2013 @@ -190,7 +190,15 @@ public class RegionStates { */ public synchronized boolean isRegionInState( final HRegionInfo hri, final State... states) { - RegionState regionState = getRegionState(hri); + return isRegionInState(hri.getEncodedName(), states); + } + + /** + * @return True if specified region is in one of the specified states. + */ + public synchronized boolean isRegionInState( + final String regionName, final State... states) { + RegionState regionState = getRegionState(regionName); State s = regionState != null ? regionState.getState() : null; for (State state: states) { if (s == state) return true; @@ -358,11 +366,11 @@ public class RegionStates { if (oldState == null) { LOG.warn("Online region not in RegionStates: " + hri.getShortNameToLog()); } else { - State state = oldState.getState(); ServerName sn = oldState.getServerName(); - if (state != State.OPEN || sn == null || !sn.equals(serverName)) { - LOG.debug("Online " + hri.getShortNameToLog() + " with current state=" + state + - ", expected state=OPEN" + ", assigned to server: " + sn + " expected " + serverName); + if (!oldState.isReadyToOnline() || sn == null || !sn.equals(serverName)) { + LOG.debug("Online " + hri.getShortNameToLog() + " with current state=" + + oldState.getState() + ", expected state=OPEN/MERGING_NEW/SPLITTING_NEW" + + ", assigned to server: " + sn + " expected " + serverName); } } updateRegionState(hri, State.OPEN, serverName); @@ -434,29 +442,28 @@ public class RegionStates { } /** - * A region is offline, won't be in transition any more. - * Its state should be the specified expected state, which - * can be Split/Merged/Offline/null(=Offline) only. + * A region is offline, won't be in transition any more. Its state + * should be the specified expected state, which can only be + * Split/Merged/Offline/null(=Offline)/SplittingNew/MergingNew. */ public synchronized void regionOffline( final HRegionInfo hri, final State expectedState) { Preconditions.checkArgument(expectedState == null - || expectedState == State.OFFLINE || expectedState == State.SPLIT - || expectedState == State.MERGED, "Offlined region should be in state" - + " OFFLINE/SPLIT/MERGED instead of " + expectedState); + || RegionState.isNotUnassignableNotInTransition(expectedState), + "Offlined region should be in state OFFLINE/SPLIT/MERGED/" + + "SPLITTING_NEW/MERGING_NEW instead of " + expectedState); String regionName = hri.getEncodedName(); RegionState oldState = regionStates.get(regionName); if (oldState == null) { LOG.warn("Offline region not in RegionStates: " + hri.getShortNameToLog()); } else if (LOG.isDebugEnabled()) { - State state = oldState.getState(); ServerName sn = oldState.getServerName(); - if (state != State.OFFLINE - && state != State.SPLITTING && state != State.MERGING) { + if (!oldState.isReadyToOffline()) { LOG.debug("Offline " + hri.getShortNameToLog() + " with current state=" - + state + ", expected state=OFFLINE/SPLITTING/MERGING"); + + oldState.getState() + ", expected state=OFFLINE/SPLIT/" + + "MERGED/SPLITTING_NEW/MERGING_NEW"); } - if (sn != null && state == State.OFFLINE) { + if (sn != null && oldState.isOffline()) { LOG.debug("Offline " + hri.getShortNameToLog() + " with current state=OFFLINE, assigned to server: " + sn + ", expected null"); @@ -497,9 +504,8 @@ public class RegionStates { if (isRegionOnline(region)) { regionsToOffline.add(region); } else { - RegionState state = getRegionState(region); - if (state.isSplitting() || state.isMerging()) { - LOG.debug("Offline splitting/merging region " + state); + if (isRegionInState(region, State.SPLITTING, State.MERGING)) { + LOG.debug("Offline splitting/merging region " + getRegionState(region)); try { // Delete the ZNode if exists ZKAssign.deleteNodeFailSilent(watcher, region); @@ -512,6 +518,7 @@ public class RegionStates { } for (HRegionInfo hri : regionsToOffline) { + updateRegionState(hri, State.OFFLINE); regionOffline(hri); } Modified: hbase/trunk/hbase-server/src/main/java/org/apache/hadoop/hbase/master/handler/ServerShutdownHandler.java URL: http://svn.apache.org/viewvc/hbase/trunk/hbase-server/src/main/java/org/apache/hadoop/hbase/master/handler/ServerShutdownHandler.java?rev=1531435&r1=1531434&r2=1531435&view=diff ============================================================================== --- hbase/trunk/hbase-server/src/main/java/org/apache/hadoop/hbase/master/handler/ServerShutdownHandler.java (original) +++ hbase/trunk/hbase-server/src/main/java/org/apache/hadoop/hbase/master/handler/ServerShutdownHandler.java Fri Oct 11 21:46:14 2013 @@ -44,6 +44,7 @@ import org.apache.hadoop.hbase.master.As import org.apache.hadoop.hbase.master.DeadServer; import org.apache.hadoop.hbase.master.MasterServices; import org.apache.hadoop.hbase.master.RegionState; +import org.apache.hadoop.hbase.master.RegionState.State; import org.apache.hadoop.hbase.master.RegionStates; import org.apache.hadoop.hbase.master.ServerManager; import org.apache.hadoop.hbase.zookeeper.ZKAssign; @@ -246,10 +247,14 @@ public class ServerShutdownHandler exten //clean zk node LOG.info("Reassigning region with rs = " + rit + " and deleting zk node if exists"); ZKAssign.deleteNodeFailSilent(services.getZooKeeper(), hri); + regionStates.updateRegionState(hri, State.OFFLINE); } catch (KeeperException ke) { this.server.abort("Unexpected ZK exception deleting unassigned node " + hri, ke); return; } + } else if (regionStates.isRegionInState( + hri, State.SPLITTING_NEW, State.MERGING_NEW)) { + regionStates.regionOffline(hri); } toAssignRegions.add(hri); } else if (rit != null) { @@ -260,8 +265,9 @@ public class ServerShutdownHandler exten // The rit that we use may be stale in case the table was in DISABLING state // but though we did assign we will not be clearing the znode in CLOSING state. // Doing this will have no harm. See HBASE-5927 + regionStates.updateRegionState(hri, State.OFFLINE); am.deleteClosingOrClosedNode(hri); - am.regionOffline(hri); + am.offlineDisabledRegion(hri); } else { LOG.warn("THIS SHOULD NOT HAPPEN: unexpected region in transition " + rit + " not to be assigned by SSH of server " + serverName); Modified: hbase/trunk/hbase-server/src/main/java/org/apache/hadoop/hbase/regionserver/HRegion.java URL: http://svn.apache.org/viewvc/hbase/trunk/hbase-server/src/main/java/org/apache/hadoop/hbase/regionserver/HRegion.java?rev=1531435&r1=1531434&r2=1531435&view=diff ============================================================================== --- hbase/trunk/hbase-server/src/main/java/org/apache/hadoop/hbase/regionserver/HRegion.java (original) +++ hbase/trunk/hbase-server/src/main/java/org/apache/hadoop/hbase/regionserver/HRegion.java Fri Oct 11 21:46:14 2013 @@ -66,7 +66,6 @@ import org.apache.hadoop.fs.FileStatus; import org.apache.hadoop.fs.FileSystem; import org.apache.hadoop.fs.Path; import org.apache.hadoop.hbase.Cell; -import org.apache.hadoop.hbase.CellScanner; import org.apache.hadoop.hbase.CellUtil; import org.apache.hadoop.hbase.CompoundConfiguration; import org.apache.hadoop.hbase.DroppedSnapshotException; @@ -114,12 +113,9 @@ import org.apache.hadoop.hbase.master.As import org.apache.hadoop.hbase.monitoring.MonitoredTask; import org.apache.hadoop.hbase.monitoring.TaskMonitor; import org.apache.hadoop.hbase.protobuf.generated.AdminProtos.GetRegionInfoResponse.CompactionState; -import org.apache.hadoop.hbase.protobuf.generated.AdminProtos.WALEntry; import org.apache.hadoop.hbase.protobuf.generated.ClientProtos.CoprocessorServiceCall; -import org.apache.hadoop.hbase.protobuf.generated.ClientProtos.MutationProto.MutationType; import org.apache.hadoop.hbase.protobuf.generated.HBaseProtos.SnapshotDescription; import org.apache.hadoop.hbase.protobuf.generated.WALProtos.CompactionDescriptor; -import org.apache.hadoop.hbase.protobuf.generated.WALProtos.WALKey; import org.apache.hadoop.hbase.regionserver.MultiVersionConsistencyControl.WriteEntry; import org.apache.hadoop.hbase.regionserver.compactions.CompactionContext; import org.apache.hadoop.hbase.regionserver.wal.HLog; @@ -2781,6 +2777,11 @@ public class HRegion implements HeapSize FileSystem fs = this.fs.getFileSystem(); NavigableSet files = HLogUtil.getSplitEditFilesSorted(fs, regiondir); + if (LOG.isDebugEnabled()) { + LOG.debug("Found " + (files == null ? 0 : files.size()) + + " recovered edits file(s) under " + regiondir); + } + if (files == null || files.isEmpty()) return seqid; for (Path edits: files) { @@ -2794,10 +2795,12 @@ public class HRegion implements HeapSize String fileName = edits.getName(); maxSeqId = Math.abs(Long.parseLong(fileName)); if (maxSeqId <= minSeqIdForTheRegion) { - String msg = "Maximum sequenceid for this log is " + maxSeqId + if (LOG.isDebugEnabled()) { + String msg = "Maximum sequenceid for this log is " + maxSeqId + " and minimum sequenceid for the region is " + minSeqIdForTheRegion + ", skipped the whole file, path=" + edits; - LOG.debug(msg); + LOG.debug(msg); + } continue; } Modified: hbase/trunk/hbase-server/src/main/java/org/apache/hadoop/hbase/regionserver/RegionMergeTransaction.java URL: http://svn.apache.org/viewvc/hbase/trunk/hbase-server/src/main/java/org/apache/hadoop/hbase/regionserver/RegionMergeTransaction.java?rev=1531435&r1=1531434&r2=1531435&view=diff ============================================================================== --- hbase/trunk/hbase-server/src/main/java/org/apache/hadoop/hbase/regionserver/RegionMergeTransaction.java (original) +++ hbase/trunk/hbase-server/src/main/java/org/apache/hadoop/hbase/regionserver/RegionMergeTransaction.java Fri Oct 11 21:46:14 2013 @@ -18,6 +18,10 @@ */ package org.apache.hadoop.hbase.regionserver; +import static org.apache.hadoop.hbase.executor.EventType.RS_ZK_REGION_MERGED; +import static org.apache.hadoop.hbase.executor.EventType.RS_ZK_REGION_MERGING; +import static org.apache.hadoop.hbase.executor.EventType.RS_ZK_REQUEST_REGION_MERGE; + import java.io.IOException; import java.util.ArrayList; import java.util.List; @@ -45,6 +49,7 @@ import org.apache.hadoop.hbase.zookeeper import org.apache.hadoop.hbase.zookeeper.ZooKeeperWatcher; import org.apache.zookeeper.KeeperException; import org.apache.zookeeper.KeeperException.NodeExistsException; +import org.apache.zookeeper.data.Stat; /** * Executes region merge as a "transaction". It is similar with @@ -261,25 +266,16 @@ public class RegionMergeTransaction { createNodeMerging(server.getZooKeeper(), this.mergedRegionInfo, server.getServerName(), region_a.getRegionInfo(), region_b.getRegionInfo()); } catch (KeeperException e) { - throw new IOException("Failed creating MERGING znode on " + throw new IOException("Failed creating PENDING_MERGE znode on " + this.mergedRegionInfo.getRegionNameAsString(), e); } } this.journal.add(JournalEntry.SET_MERGING_IN_ZK); if (server != null && server.getZooKeeper() != null) { - try { - // Transition node from MERGING to MERGING after creating the merge - // node. Master will get the callback for node change only if the - // transition is successful. - // Note that if the transition fails then the rollback will delete the - // created znode as the journal entry SET_MERGING_IN_ZK is added. - this.znodeVersion = transitionNodeMerging(server.getZooKeeper(), - this.mergedRegionInfo, server.getServerName(), -1, - region_a.getRegionInfo(), region_b.getRegionInfo()); - } catch (KeeperException e) { - throw new IOException("Failed setting MERGING znode on " - + this.mergedRegionInfo.getRegionNameAsString(), e); - } + // After creating the merge node, wait for master to transition it + // from PENDING_MERGE to MERGING so that we can move on. We want master + // knows about it and won't transition any region which is merging. + znodeVersion = getZKNode(server, services); } this.region_a.getRegionFileSystem().createMergesDir(); @@ -303,9 +299,10 @@ public class RegionMergeTransaction { try { // Do one more check on the merging znode (before it is too late) in case // any merging region is moved somehow. If so, the znode transition will fail. - this.znodeVersion = transitionNodeMerging(server.getZooKeeper(), - this.mergedRegionInfo, server.getServerName(), this.znodeVersion, - region_a.getRegionInfo(), region_b.getRegionInfo()); + this.znodeVersion = transitionMergingNode(server.getZooKeeper(), + this.mergedRegionInfo, region_a.getRegionInfo(), region_b.getRegionInfo(), + server.getServerName(), this.znodeVersion, + RS_ZK_REGION_MERGING, RS_ZK_REGION_MERGING); } catch (KeeperException e) { throw new IOException("Failed setting MERGING znode on " + this.mergedRegionInfo.getRegionNameAsString(), e); @@ -489,9 +486,10 @@ public class RegionMergeTransaction { // Tell master about merge by updating zk. If we fail, abort. try { - this.znodeVersion = transitionNodeMerge(server.getZooKeeper(), - this.mergedRegionInfo, region_a.getRegionInfo(), - region_b.getRegionInfo(), server.getServerName(), this.znodeVersion); + this.znodeVersion = transitionMergingNode(server.getZooKeeper(), + this.mergedRegionInfo, region_a.getRegionInfo(), + region_b.getRegionInfo(), server.getServerName(), this.znodeVersion, + RS_ZK_REGION_MERGING, RS_ZK_REGION_MERGED); long startTime = EnvironmentEdgeManager.currentTimeMillis(); int spins = 0; @@ -506,9 +504,10 @@ public class RegionMergeTransaction { } Thread.sleep(100); // When this returns -1 it means the znode doesn't exist - this.znodeVersion = tickleNodeMerge(server.getZooKeeper(), - this.mergedRegionInfo, region_a.getRegionInfo(), - region_b.getRegionInfo(), server.getServerName(), this.znodeVersion); + this.znodeVersion = transitionMergingNode(server.getZooKeeper(), + this.mergedRegionInfo, region_a.getRegionInfo(), + region_b.getRegionInfo(), server.getServerName(), this.znodeVersion, + RS_ZK_REGION_MERGED, RS_ZK_REGION_MERGED); spins++; } while (this.znodeVersion != -1 && !server.isStopped() && !services.isStopping()); @@ -520,13 +519,84 @@ public class RegionMergeTransaction { + mergedRegionInfo.getEncodedName(), e); } - // Leaving here, the mergedir with its dross will be in place but since the // merge was successful, just leave it; it'll be cleaned when region_a is // cleaned up by CatalogJanitor on master } /** + * Wait for the merging node to be transitioned from pending_merge + * to merging by master. That's how we are sure master has processed + * the event and is good with us to move on. If we don't get any update, + * we periodically transition the node so that master gets the callback. + * If the node is removed or is not in pending_merge state any more, + * we abort the merge. + */ + private int getZKNode(final Server server, + final RegionServerServices services) throws IOException { + // Wait for the master to process the pending_merge. + try { + int spins = 0; + Stat stat = new Stat(); + ZooKeeperWatcher zkw = server.getZooKeeper(); + ServerName expectedServer = server.getServerName(); + String node = mergedRegionInfo.getEncodedName(); + while (!(server.isStopped() || services.isStopping())) { + if (spins % 5 == 0) { + LOG.debug("Still waiting for master to process " + + "the pending_merge for " + node); + transitionMergingNode(zkw, mergedRegionInfo, region_a.getRegionInfo(), + region_b.getRegionInfo(), expectedServer, -1, RS_ZK_REQUEST_REGION_MERGE, + RS_ZK_REQUEST_REGION_MERGE); + } + Thread.sleep(100); + spins++; + byte [] data = ZKAssign.getDataNoWatch(zkw, node, stat); + if (data == null) { + throw new IOException("Data is null, merging node " + + node + " no longer exists"); + } + RegionTransition rt = RegionTransition.parseFrom(data); + EventType et = rt.getEventType(); + if (et == RS_ZK_REGION_MERGING) { + ServerName serverName = rt.getServerName(); + if (!serverName.equals(expectedServer)) { + throw new IOException("Merging node " + node + " is for " + + serverName + ", not us " + expectedServer); + } + byte [] payloadOfMerging = rt.getPayload(); + List mergingRegions = HRegionInfo.parseDelimitedFrom( + payloadOfMerging, 0, payloadOfMerging.length); + assert mergingRegions.size() == 3; + HRegionInfo a = mergingRegions.get(1); + HRegionInfo b = mergingRegions.get(2); + HRegionInfo hri_a = region_a.getRegionInfo(); + HRegionInfo hri_b = region_b.getRegionInfo(); + if (!(hri_a.equals(a) && hri_b.equals(b))) { + throw new IOException("Merging node " + node + " is for " + a + ", " + + b + ", not expected regions: " + hri_a + ", " + hri_b); + } + // Master has processed it. + return stat.getVersion(); + } + if (et != RS_ZK_REQUEST_REGION_MERGE) { + throw new IOException("Merging node " + node + + " moved out of merging to " + et); + } + } + // Server is stopping/stopped + throw new IOException("Server is " + + (services.isStopping() ? "stopping" : "stopped")); + } catch (Exception e) { + if (e instanceof InterruptedException) { + Thread.currentThread().interrupt(); + } + throw new IOException("Failed getting MERGING znode on " + + mergedRegionInfo.getRegionNameAsString(), e); + } + } + + /** * Create reference file(s) of merging regions under the region_a merges dir * @param hstoreFilesOfRegionA * @param hstoreFilesOfRegionB @@ -566,6 +636,7 @@ public class RegionMergeTransaction { * of no return and so now need to abort the server to minimize * damage. */ + @SuppressWarnings("deprecation") public boolean rollback(final Server server, final RegionServerServices services) throws IOException { assert this.mergedRegionInfo != null; @@ -653,20 +724,22 @@ public class RegionMergeTransaction { private static void cleanZK(final Server server, final HRegionInfo hri) { try { // Only delete if its in expected state; could have been hijacked. - ZKAssign.deleteNode(server.getZooKeeper(), hri.getEncodedName(), - EventType.RS_ZK_REGION_MERGING); + if (!ZKAssign.deleteNode(server.getZooKeeper(), hri.getEncodedName(), + RS_ZK_REQUEST_REGION_MERGE)) { + ZKAssign.deleteNode(server.getZooKeeper(), hri.getEncodedName(), + RS_ZK_REGION_MERGING); + } } catch (KeeperException.NoNodeException e) { LOG.warn("Failed cleanup zk node of " + hri.getRegionNameAsString(), e); } catch (KeeperException e) { server.abort("Failed cleanup zk node of " + hri.getRegionNameAsString(),e); } - } /** - * Creates a new ephemeral node in the MERGING state for the merged region. + * Creates a new ephemeral node in the PENDING_MERGE state for the merged region. * Create it ephemeral in case regionserver dies mid-merge. - * + * *

* Does not transition nodes from other states. If a node already exists for * this region, a {@link NodeExistsException} will be thrown. @@ -674,32 +747,27 @@ public class RegionMergeTransaction { * @param zkw zk reference * @param region region to be created as offline * @param serverName server event originates from - * @return Version of znode created. * @throws KeeperException * @throws IOException */ - int createNodeMerging(final ZooKeeperWatcher zkw, final HRegionInfo region, + public static void createNodeMerging(final ZooKeeperWatcher zkw, final HRegionInfo region, final ServerName serverName, final HRegionInfo a, final HRegionInfo b) throws KeeperException, IOException { LOG.debug(zkw.prefix("Creating ephemeral node for " - + region.getEncodedName() + " in MERGING state")); - byte [] payload = HRegionInfo.toDelimitedByteArray(a, b); + + region.getEncodedName() + " in PENDING_MERGE state")); + byte [] payload = HRegionInfo.toDelimitedByteArray(region, a, b); RegionTransition rt = RegionTransition.createRegionTransition( - EventType.RS_ZK_REGION_MERGING, region.getRegionName(), serverName, payload); + RS_ZK_REQUEST_REGION_MERGE, region.getRegionName(), serverName, payload); String node = ZKAssign.getNodeName(zkw, region.getEncodedName()); if (!ZKUtil.createEphemeralNodeAndWatch(zkw, node, rt.toByteArray())) { throw new IOException("Failed create of ephemeral " + node); } - // Transition node from MERGING to MERGING and pick up version so we - // can be sure this znode is ours; version is needed deleting. - return transitionNodeMerging(zkw, region, serverName, -1, a, b); } /** - * Transitions an existing node for the specified region which is currently in - * the MERGING state to be in the MERGE state. Converts the ephemeral MERGING - * znode to an ephemeral MERGE node. Master cleans up MERGE znode when it - * reads it (or if we crash, zk will clean it up). + * Transitions an existing ephemeral node for the specified region which is + * currently in the begin state to be in the end state. Master cleans up the + * final MERGE znode when it reads it (or if we crash, zk will clean it up). * *

* Does not transition nodes from other states. If for some reason the node @@ -710,19 +778,18 @@ public class RegionMergeTransaction { * This method can fail and return false for three different reasons: *

    *
  • Node for this region does not exist
    • - *
  • Node for this region is not in MERGING state
    • - *
  • After verifying MERGING state, update fails because of wrong version + *
  • Node for this region is not in the begin state
    • + *
  • After verifying the begin state, update fails because of wrong version * (this should never actually happen since an RS only does this transition - * following a transition to MERGING. if two RS are conflicting, one would - * fail the original transition to MERGING and not this transition)
    • + * following a transition to the begin state. If two RS are conflicting, one would + * fail the original transition to the begin state and not this transition) *
* *

* Does not set any watches. * *

- * This method should only be used by a RegionServer when completing the open - * of merged region. + * This method should only be used by a RegionServer when merging two regions. * * @param zkw zk reference * @param merged region to be transitioned to opened @@ -730,45 +797,19 @@ public class RegionMergeTransaction { * @param b merging region B * @param serverName server event originates from * @param znodeVersion expected version of data before modification + * @param beginState the expected current state the znode should be + * @param endState the state to be transition to * @return version of node after transition, -1 if unsuccessful transition * @throws KeeperException if unexpected zookeeper exception * @throws IOException */ - private static int transitionNodeMerge(ZooKeeperWatcher zkw, + public static int transitionMergingNode(ZooKeeperWatcher zkw, HRegionInfo merged, HRegionInfo a, HRegionInfo b, ServerName serverName, - final int znodeVersion) throws KeeperException, IOException { + final int znodeVersion, final EventType beginState, + final EventType endState) throws KeeperException, IOException { byte[] payload = HRegionInfo.toDelimitedByteArray(merged, a, b); return ZKAssign.transitionNode(zkw, merged, serverName, - EventType.RS_ZK_REGION_MERGING, EventType.RS_ZK_REGION_MERGED, - znodeVersion, payload); - } - - /** - * - * @param zkw zk reference - * @param parent region to be transitioned to merging - * @param serverName server event originates from - * @param version znode version - * @return version of node after transition, -1 if unsuccessful transition - * @throws KeeperException - * @throws IOException - */ - int transitionNodeMerging(final ZooKeeperWatcher zkw, - final HRegionInfo parent, final ServerName serverName, final int version, - final HRegionInfo a, final HRegionInfo b) throws KeeperException, IOException { - byte[] payload = HRegionInfo.toDelimitedByteArray(a, b); - return ZKAssign.transitionNode(zkw, parent, serverName, - EventType.RS_ZK_REGION_MERGING, EventType.RS_ZK_REGION_MERGING, - version, payload); - } - - private static int tickleNodeMerge(ZooKeeperWatcher zkw, HRegionInfo merged, - HRegionInfo a, HRegionInfo b, ServerName serverName, - final int znodeVersion) throws KeeperException, IOException { - byte[] payload = HRegionInfo.toDelimitedByteArray(a, b); - return ZKAssign.transitionNode(zkw, merged, serverName, - EventType.RS_ZK_REGION_MERGED, EventType.RS_ZK_REGION_MERGED, - znodeVersion, payload); + beginState, endState, znodeVersion, payload); } /**