Apache
/
hbase
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

HBASE-20173 [AMv2] DisableTableProcedure concurrent to ServerCrashProcedure can deadlock 

Allow that DisableTableProcedue can grab a region lock before
ServerCrashProcedure can. Cater to this cricumstance where SCP
was not unable to make progress by running the search for RIT
against the crashed server a second time, post creation of all
crashed-server assignemnts. The second run will uncover such as
the above DisableTableProcedure unassign and will interrupt its
suspend allowing both procedures to make progress.

M hbase-protocol-shaded/src/main/protobuf/MasterProcedure.proto
 Add new procedure step post-assigns that reruns the RIT finder method.

M hbase-server/src/main/java/org/apache/hadoop/hbase/master/assignment/AssignmentManager.java
 Make this important log more specific as to what is going on.

M hbase-server/src/main/java/org/apache/hadoop/hbase/master/assignment/UnassignProcedure.java
 Better explanation as to what is going on.

M hbase-server/src/main/java/org/apache/hadoop/hbase/master/procedure/ServerCrashProcedure.java
 Add extra step and run handleRIT a second time after we've queued up
 all SCP assigns. Also fix a but. SCP was adding an assign of a RIT
 that was actually trying to unassign (made the deadlock more likely).
This commit is contained in:
Michael Stack 2018-03-11 13:15:16 -07:00
parent e80ce3d91b
commit 260ee0da60
6 changed files with 104 additions and 80 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
hbase-it/src/test/java/org/apache/hadoop/hbase/IntegrationTestDDLMasterFailover.java
View File

@ -53,12 +53,12 @@ import org.slf4j.LoggerFactory;
/**
*
* Integration test that verifies Procedure V2. <br/><br/>
* Integration test that verifies Procedure V2.
*
* DDL operations should go through (rollforward or rollback) when primary master is killed by
* ChaosMonkey (default MASTER_KILLING)<br/><br/>
* ChaosMonkey (default MASTER_KILLING).
*
* Multiple Worker threads are started to randomly do the following Actions in loops:<br/>
* <p></p>Multiple Worker threads are started to randomly do the following Actions in loops:
* Actions generating and populating tables:
* <ul>
* <li>CreateTableAction</li>

1
hbase-protocol-shaded/src/main/protobuf/MasterProcedure.proto
View File

@ -302,6 +302,7 @@ enum ServerCrashState {
// Removed SERVER_CRASH_CALC_REGIONS_TO_ASSIGN = 7;
SERVER_CRASH_ASSIGN = 8;
SERVER_CRASH_WAIT_ON_ASSIGN = 9;
SERVER_CRASH_HANDLE_RIT2 = 20;
SERVER_CRASH_FINISH = 100;
}

2
hbase-server/src/main/java/org/apache/hadoop/hbase/master/assignment/AssignmentManager.java
View File

@ -1196,7 +1196,7 @@ public class AssignmentManager implements ServerListener {
private void handleRegionOverStuckWarningThreshold(final RegionInfo regionInfo) {
final RegionStateNode regionNode = regionStates.getRegionStateNode(regionInfo);
//if (regionNode.isStuck()) {
LOG.warn("TODO Handle stuck in transition: " + regionNode);
LOG.warn("STUCK Region-In-Transition {}", regionNode);
}
// ============================================================================================

13
hbase-server/src/main/java/org/apache/hadoop/hbase/master/assignment/UnassignProcedure.java
View File

@ -249,10 +249,10 @@ public class UnassignProcedure extends RegionTransitionProcedure {
final IOException exception) {
// TODO: Is there on-going rpc to cleanup?
if (exception instanceof ServerCrashException) {
// This exception comes from ServerCrashProcedure after log splitting.
// This exception comes from ServerCrashProcedure AFTER log splitting.
// SCP found this region as a RIT. Its call into here says it is ok to let this procedure go
// on to a complete close now. This will release lock on this region so subsequent action on
// region can succeed; e.g. the assign that follows this unassign when a move (w/o wait on SCP
// complete. This complete will release lock on this region so subsequent action on region
// can succeed; e.g. the assign that follows this unassign when a move (w/o wait on SCP
// the assign could run w/o logs being split so data loss).
try {
reportTransition(env, regionNode, TransitionCode.CLOSED, HConstants.NO_SEQNUM);
@ -263,7 +263,6 @@ public class UnassignProcedure extends RegionTransitionProcedure {
} else if (exception instanceof RegionServerAbortedException ||
exception instanceof RegionServerStoppedException ||
exception instanceof ServerNotRunningYetException) {
// TODO
// RS is aborting, we cannot offline the region since the region may need to do WAL
// recovery. Until we see the RS expiration, we should retry.
// TODO: This should be suspend like the below where we call expire on server?
@ -276,8 +275,10 @@ public class UnassignProcedure extends RegionTransitionProcedure {
LOG.warn("Expiring server " + this + "; " + regionNode.toShortString() +
", exception=" + exception);
env.getMasterServices().getServerManager().expireServer(regionNode.getRegionLocation());
// Return false so this procedure stays in suspended state. It will be woken up by a
// ServerCrashProcedure when it notices this RIT.
// Return false so this procedure stays in suspended state. It will be woken up by the
// ServerCrashProcedure that was scheduled when we called #expireServer above. SCP calls
// #handleRIT which will call this method only the exception will be a ServerCrashException
// this time around (See above).
// TODO: Add a SCP as a new subprocedure that we now come to depend on.
return false;
}

2
hbase-server/src/main/java/org/apache/hadoop/hbase/master/procedure/RecoverMetaProcedure.java
View File

@ -159,6 +159,8 @@ public class RecoverMetaProcedure
* just as we do over in ServerCrashProcedure#handleRIT except less to do here; less context
* to carry.
*/
// NOTE: Make sure any fix or improvement done here is also done in SCP#handleRIT; the methods
// have overlap.
private void handleRIT(MasterProcedureEnv env, RegionInfo ri, ServerName crashedServerName) {
AssignmentManager am = env.getAssignmentManager();
RegionTransitionProcedure rtp = am.getRegionStates().getRegionTransitionProcedure(ri);

46
hbase-server/src/main/java/org/apache/hadoop/hbase/master/procedure/ServerCrashProcedure.java
View File

@ -19,6 +19,7 @@ package org.apache.hadoop.hbase.master.procedure;
import java.io.IOException;
import java.util.ArrayList;
import java.util.Collections;
import java.util.Iterator;
import java.util.List;
@ -28,7 +29,6 @@ import org.apache.hadoop.hbase.client.RegionInfo;
import org.apache.hadoop.hbase.client.RegionReplicaUtil;
import org.apache.hadoop.hbase.master.MasterServices;
import org.apache.hadoop.hbase.master.MasterWalManager;
import org.apache.hadoop.hbase.master.assignment.AssignProcedure;
import org.apache.hadoop.hbase.master.assignment.AssignmentManager;
import org.apache.hadoop.hbase.master.assignment.RegionTransitionProcedure;
import org.apache.hadoop.hbase.procedure2.ProcedureMetrics;
@ -161,11 +161,22 @@ implements ServerProcedureInterface {
RegionInfo.getShortNameToLog(regionsOnCrashedServer) + ", " + this +
"; cycles=" + getCycles());
}
handleRIT(env, regionsOnCrashedServer);
// Handle RIT against crashed server. Will cancel any ongoing assigns/unassigns.
// Returns list of regions we need to reassign.
List<RegionInfo> toAssign = handleRIT(env, regionsOnCrashedServer);
AssignmentManager am = env.getAssignmentManager();
// createAssignProcedure will try to use the old location for the region deploy.
addChildProcedure(am.createAssignProcedures(regionsOnCrashedServer));
// CreateAssignProcedure will try to use the old location for the region deploy.
addChildProcedure(am.createAssignProcedures(toAssign));
setNextState(ServerCrashState.SERVER_CRASH_HANDLE_RIT2);
} else {
setNextState(ServerCrashState.SERVER_CRASH_FINISH);
}
break;
case SERVER_CRASH_HANDLE_RIT2:
// Run the handleRIT again for case where another procedure managed to grab the lock on
// a region ahead of this crash handling procedure. Can happen in rare case. See
handleRIT(env, regionsOnCrashedServer);
setNextState(ServerCrashState.SERVER_CRASH_FINISH);
break;
@ -360,18 +371,24 @@ implements ServerProcedureInterface {
* otherwise we have two assigns going on and they will fight over who has lock.
* Notify Unassigns. If unable to unassign because server went away, unassigns block waiting
* on the below callback from a ServerCrashProcedure before proceeding.
* @param env
* @param regions Regions that were on crashed server
* @param regions Regions on the Crashed Server.
* @return List of regions we should assign to new homes (not same as regions on crashed server).
*/
private void handleRIT(final MasterProcedureEnv env, final List<RegionInfo> regions) {
if (regions == null) return;
private List<RegionInfo> handleRIT(final MasterProcedureEnv env, List<RegionInfo> regions) {
if (regions == null || regions.isEmpty()) {
return Collections.emptyList();
}
AssignmentManager am = env.getMasterServices().getAssignmentManager();
final Iterator<RegionInfo> it = regions.iterator();
List<RegionInfo> toAssign = new ArrayList<RegionInfo>(regions);
// Get an iterator so can remove items.
final Iterator<RegionInfo> it = toAssign.iterator();
ServerCrashException sce = null;
while (it.hasNext()) {
final RegionInfo hri = it.next();
RegionTransitionProcedure rtp = am.getRegionStates().getRegionTransitionProcedure(hri);
if (rtp == null) continue;
if (rtp == null) {
continue;
}
// Make sure the RIT is against this crashed server. In the case where there are many
// processings of a crashed server -- backed up for whatever reason (slow WAL split) --
// then a previous SCP may have already failed an assign, etc., and it may have a new
@ -389,11 +406,14 @@ implements ServerProcedureInterface {
sce = new ServerCrashException(getProcId(), getServerName());
}
rtp.remoteCallFailed(env, this.serverName, sce);
if (rtp instanceof AssignProcedure) {
// If an assign, include it in our return and remove from passed-in list of regions.
// If an assign, remove from passed-in list of regions so we subsequently do not create
// a new assign; the exisitng assign after the call to remoteCallFailed will recalibrate
// and assign to a server other than the crashed one; no need to create new assign.
// If an unassign, do not return this region; the above cancel will wake up the unassign and
// it will complete. Done.
it.remove();
}
}
return toAssign;
}
@Override