hbase-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "jiraposter@reviews.apache.org (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (HBASE-4015) Refactor the TimeoutMonitor to make it less racy
Date Wed, 07 Sep 2011 05:39:11 GMT

    [ https://issues.apache.org/jira/browse/HBASE-4015?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13098605#comment-13098605

jiraposter@reviews.apache.org commented on HBASE-4015:

This is an automatically generated e-mail. To reply, visit:

I have some feedback below.  On how to write a test, I'm not sure how -- so many systems are
touched.  We'd need to break out AM so it standalone but that is too much to ask of this patch.
 Good work Ram.  Thanks for persevering with this one.


    I see.  You have full path here to distingush from o.a.h.h.executor.ExecutorService.


    I wonder about passing this in rather than create it in AM on construction?  If you need
to do a close on the ExecutorService, add a close to AM and have the host of the AM class
call the close on the way out?  Otherwise, you are exposing the implementation.  If you need
configuration setting up the ExecutorService, you have the conf object by doing master.getConfiguration?
 There may be other reasons for passing in the ExecutorService that I'm not seeing... Above
is just a thought.


    How do we know this will not result in a double assignment?  What if the .META. is OPENING
and the master crashes.   A new master comes up but meantime the .META. has moved to OPENED
and here we will assign the new .META. out without checking if .META. successfully OPENED?

    Should we at least check the znode again before going ahead with assign?  Should we actually
try grabbing the znode completely before we reassign so if .META. is opening, it will fail
to complete?


    Thanks for making this little method disableRegionIfInRIT


    If you make a new patch, can you fix formatting of this comment?  Its strange the way
its broke over a number of lines.
    So, the flag here is 'reassign'.  Does that mean 'called from timeout monitor'?  If so,
its not clear (Change name of flag or at least fixup javadoc).  Else this comment is hard
to understand.  But even so, do we care where its called from?  All thats important is if
we are to reassign or not (What if we figure we need to set reassign to true elsewhere than
in timeout monitor?  Then this comment would be extra confusing?)


    This is great.


    Should this method be called isDisabledOrDisabling?  Then a reader will expect the boolean


    This javadoc is better in that it says the flag 'reassign' is set if we were invoked by
timeout monitor.   Is it important that we're called by the timeout monitor?   Do we need
to know about the caller when this method is run?  What if we decide to call this method from
elsewhere than from timeout monitor? 


    Again is it important that we know we are being called from timeout monitor?  What does
it mean when we are called from timeout monitor?  Is it that we are trying to interrupt normal
processing because it seems to have failed?  If so, 'reassign' may be the wrong name for this
flag?  Would 'hijack' be a better flag name?  Or 'appropriate' or some such?
    Sorry for harping on the flag name Ram.  If you have good names that convey intent/purpose,
it makes it easier on the reader.


    Can you fix formatting of this comment?


    Should we change znode state here?  Read it, get the znode id and try to change it.  If
we fail, then RS still has control?


    Making a new method is good here.


    This is nice.


    Try not to make these changes going forward (you are moving the + from end of line to
start of next line and thats all -- it bulks up your patch unnecessarily).




    Does doing this clear the watch on the znode such that we will not notice if it moves


    Is there something wrong here?  The comment says do not make znode OFFLINE, wait on timeout
monitor, but we are queuing up an ASSIGN it looks like?


    Yeah, I wonder if this even belongs out here.  Just have it inside in AM?
    Do we even call close on it?  Should we?


    No biggie but you could have written this as:
    return (versionOfflineNode == -1)? openRegion(region): openRegion(region, versionOffilneNode);


    I'd suggest you move the enum here from AM rather than have this strange import of an
inner type.
    Maybe you'd have two classes in here.  AssignCallable and UnassignCallable then no need
of the enum?
    For this class name, do you think we need TimeOutManager in its name?  In fact, is there
a TimeoutManager?   It seems to be the ExcecutorService?  Or is it the method invokeTimeoutManager?
    Seems like method should be something like invoke and you queue AssignCallable or UnassignCallable?




    I think I asked this already and I should check the response above, but is there any duplication
here with current openRegion?  Should we factor out common code?


    For sure this method should not know about timeout monitor.  This is a class in another
package from master AM code.  Should this flag be named 'hijack' or 'appropriate' or 'preempt'


    You do not javadoc your return.


    Can you clean up the formatting on this comment?  Do we need to know about timeout monitor
in here?


    Should this check be done up in timeout monitor or in a callable so its not inline?  If
it still is in OPENING state up in timeout monitor, then come in here?  Just a suggestion
(May not be a good one).


    Is it correct adding 1?


    This is a strange location for a comment...  I'd say put it into the else clause.

- Michael

On 2011-08-29 12:11:23, ramkrishna vasudevan wrote:
bq.  -----------------------------------------------------------
bq.  This is an automatically generated e-mail. To reply, visit:
bq.  https://reviews.apache.org/r/1668/
bq.  -----------------------------------------------------------
bq.  (Updated 2011-08-29 12:11:23)
bq.  Review request for Ted Yu, Michael Stack, Jean-Daniel Cryans, and Jonathan Gray.
bq.  Summary
bq.  -------
bq.  HBASE-4015 - updated patch.
bq.          invokeTimeOutManager(regionState.getRegion(),
bq.              TimeOutOperationType.ASSIGN);
bq.  Instead of storing the state when timeout was deducted we process it
bq.  using future task.
bq.  ================================================================
bq.        case RS_ZK_REGION_OPENING:
bq.          // TODO: Could check if it was on deadServers.  If it was, then we could
bq.          // do what happens in TimeoutMonitor when it sees this condition.
bq.          // Just insert region into RIT
bq.          // If this never updates the timeout will trigger new assignment
bq.          if (regionInfo.isMetaRegion() || regionInfo.isRootRegion()) {
bq.            regionsInTransition.put(encodedRegionName, new RegionState(
bq.                regionInfo, RegionState.State.OPENING, data.getStamp(), data
bq.                    .getOrigin()));
bq.            processOpeningState(regionInfo);
bq.            break;
bq.          }
bq.          regionsInTransition.put(encodedRegionName, new RegionState(regionInfo,
bq.              RegionState.State.OPENING, data.getStamp(), data.getOrigin()));
bq.          break;
bq.  This change is for HBASE-4203. META and ROOT table need not wait till timeout
bq.  ======================================================================
bq.  In forceRegionStateToOffline()
bq.      } else {
bq.        // If invoked from timeout monitor donot force it to OFFLINE. Based on the
bq.        // state we will decide if to change in-memory state to OFFLINE or not.  It will
bq.        // be done before setting the znode to OFFLINE state.
bq.        if (!timeOutMonitorReAllocate) {
bq.          LOG.debug("Forcing OFFLINE; was=" + state);
bq.          state.update(RegionState.State.OFFLINE);
bq.        }
bq.  If the timeout monitor tries to reallcoate the node then dont make
bq.  the inmemory state to OFFLINE.
bq.  But the noraml assign flow doesnot expect the inmemory state to OFFLINE.
bq.  Hence the above change.  This is continued with the check in 
bq.  assign()
bq.  int setOfflineInZooKeeper(final RegionState state,
bq.        boolean timeOutMonitorReAllocate) {
bq.      // If invoked from timeoutmonitor the current state in memory need not be
bq.      // OFFLINE.  
bq.      if (!timeOutMonitorReAllocate && !state.isClosed() && !state.isOffline())
bq.            this.master.abort("Unexpected state trying to OFFLINE; " + state,
bq.            new IllegalStateException());
bq.        return -1;
bq.      }
bq.  ======================================================================
bq.      boolean allowCreation = false;
bq.      // If the isReAllocate is true and the current state is PENDING_OPEN
bq.      // or OPENING then update the inmemory state to PENDING_OPEN. This is
bq.      // important because
bq.      // if timeoutmonitor deducts that a region was in OPENING state for a long
bq.      // time but by the
bq.      // time timeout monitor tranits the node to OFFLINE the RS would have opened
bq.      // the node and the
bq.      // state in znode will be RS_ZK_REGION_OPENED. Inorder to invoke the
bq.      // OpenedRegionHandler
bq.      // we expect the inmemeory state to be PENDING_OPEN or OPENING.
bq.      // For all other cases we can change the inmemory state to OFFLINE.
bq.      if (timeOutMonitorReAllocate
bq.          && (state.getState().equals(RegionState.State.PENDING_OPEN) || state
bq.              .getState().equals(RegionState.State.OPENING))) {
bq.        state.update(RegionState.State.PENDING_OPEN);
bq.        allowCreation = false;
bq.      } else {
bq.        state.update(RegionState.State.OFFLINE);
bq.        allowCreation = true;
bq.      }
bq.  This change is quite tricky.  
bq.  In normal assign flow the unassigned node for the region will not be present
bq.  Hence we need to allow the creation of the node newly.
bq.  But in timeout monitor case we will have the node present in some state hence 
bq.  we decide whether to create node newly or not inside ZKAssign.createOrForceNodeOffline
bq.  The above code also updates the inmemory state of OFFLINE or PENDING_OPEN
bq.  which was not update in the previous forceRegionStateToOffline() call.
bq.  ==============================================================================
bq.  In ZKAssign.java()
bq.      if (version == -1) {
bq.        // If timeoutmonitor deducts a node to be in OPENING state but before it
bq.        // could
bq.        // transit to OFFLINE state if RS had opened the region then the Master
bq.        // deletes the
bq.        // assigned region znode. In that case the znode will not exist. So we
bq.        // should not
bq.        // create the znode again which will lead to double assignment.
bq.        if (timeOutMonitorReAllocate && !allowCreation) {
bq.          return -1;
bq.        }
bq.  this part prevents double assignment
bq.  If timeoutmonitor tries to force to OFFLINE state an existing region which was in RIT
bq.  but before it could do that if the node was opened then openedregionhandler will delete
bq.  the node hence we should not create the node again.
bq.  =======================================================================================
bq.  In ZkAssign.java()
bq.      } else {
bq.        RegionTransitionData curDataInZNode = ZKAssign.getDataNoWatch(zkw, region
bq.            .getEncodedName(), stat);
bq.        // Do not move the node to OFFLINE if znode is in any of the following
bq.        // state.
bq.        // Because these are already executed states.
bq.        if (timeOutMonitorReAllocate && null != curDataInZNode) {
bq.          EventType eventType = curDataInZNode.getEventType();
bq.          if (eventType.equals(EventType.RS_ZK_REGION_CLOSING)
bq.              || eventType.equals(EventType.RS_ZK_REGION_CLOSED)
bq.              || eventType.equals(EventType.RS_ZK_REGION_OPENED)) {
bq.            return -1;
bq.          }
bq.        }
bq.  This check prevents from moving the node to OFFLINE state if just before
bq.  the node is tried to force to OFFLINE the RS would have changed the state 
bq.  to either CLOSING or CLOSED or OPENED.  now again moving to OFFLINE
bq.  will lead to douoble assignment and will an additional operaition.
bq.  ====================================================================================
bq.  In ZKassign.java()
bq.        boolean setData = false;
bq.        try {
bq.          setData = ZKUtil.setData(zkw, node, data.getBytes(), version);
bq.          // Setdata throws KeeperException which aborts the Master. So we are
bq.          // catching it here.
bq.          // If just before setting the znode to OFFLINE if the RS has made any
bq.          // change to the
bq.          // znode state then we need to return -1.
bq.        } catch (KeeperException kpe) {
bq.          LOG.info("Version mismatch while setting the node to OFFLINE state.");
bq.          return -1;
bq.        }
bq.  This change is actually to avoid the master from abortng. If the forceful OFFLINE is
bq.  progrss but just before setting the RS has changed the state.
bq.  Then the setData will fail leading to master abort.
bq.  Hence we are catching the exception.
bq.  ====================================================================================
bq.  In assignmentManager.java
bq.  +      if (setOfflineInZK && versionOfOfflineNode == -1)
bq.  +        return;
bq.  This is nothing but the refactoring done in the existing code.
bq.  -     if (setOfflineInZK && !setOfflineInZooKeeper(state)) return;
bq.  So if setting the version is unsuccessful return.
bq.  =====================================================================================
bq.  In ZKassign.java()
bq.  // the below check ensure that double assignment doesnot happen.
bq.      // When the node is created for the first time then the expected version
bq.      // that is
bq.      // passed will be -1 and the version in znode will be 0.
bq.      // In all other cases the version in znode will be > 0.
bq.      else if (beginState.equals(EventType.M_ZK_REGION_OFFLINE)
bq.          && endState.equals(EventType.RS_ZK_REGION_OPENING)
bq.          && expectedVersion == -1 && stat.getVersion() != 0) {
bq.        LOG.warn(zkw.prefix("Attempt to transition the " + "unassigned node for "
bq.            + encoded + " from " + beginState + " to " + endState + " failed, "
bq.            + "the node existed but was version " + stat.getVersion()
bq.            + " not the expected version " + expectedVersion));
bq.        return -1;
bq.      }
bq.  As the comment explains when the node is created for first time the expectedversion will
bq.  be -1 but the actual version will be 0.  Here the scenario is
bq.  If RS1 has not tranitioned the node from OFFLINE to OPENING if
bq.  RS2 gets the call from Master after forcefully chaning to OFFLINE the
bq.  Rs2 will take the control of the node.
bq.  At that time if RS1 starts transmitting the node then we should not allow
bq.  it.
bq.  ===============================================================================
bq.  This addresses bug HBASE-4015.
bq.      https://issues.apache.org/jira/browse/HBASE-4015
bq.  Diffs
bq.  -----
bq.    http://svn.apache.org/repos/asf/hbase/trunk/src/main/java/org/apache/hadoop/hbase/master/AssignmentManager.java
bq.    http://svn.apache.org/repos/asf/hbase/trunk/src/main/java/org/apache/hadoop/hbase/ipc/HRegionInterface.java
bq.    http://svn.apache.org/repos/asf/hbase/trunk/src/main/java/org/apache/hadoop/hbase/master/HMaster.java
bq.    http://svn.apache.org/repos/asf/hbase/trunk/src/main/java/org/apache/hadoop/hbase/master/ServerManager.java
bq.    http://svn.apache.org/repos/asf/hbase/trunk/src/main/java/org/apache/hadoop/hbase/master/TimeOutManagerCallable.java
bq.    http://svn.apache.org/repos/asf/hbase/trunk/src/main/java/org/apache/hadoop/hbase/regionserver/HRegionServer.java
bq.    http://svn.apache.org/repos/asf/hbase/trunk/src/main/java/org/apache/hadoop/hbase/regionserver/handler/OpenMetaHandler.java
bq.    http://svn.apache.org/repos/asf/hbase/trunk/src/main/java/org/apache/hadoop/hbase/regionserver/handler/OpenRegionHandler.java
bq.    http://svn.apache.org/repos/asf/hbase/trunk/src/main/java/org/apache/hadoop/hbase/regionserver/handler/OpenRootHandler.java
bq.    http://svn.apache.org/repos/asf/hbase/trunk/src/main/java/org/apache/hadoop/hbase/zookeeper/ZKAssign.java
bq.  Diff: https://reviews.apache.org/r/1668/diff
bq.  Testing
bq.  -------
bq.  Yes.  But could not add new test case.
bq.  TestMasterFailOver is passing with the current changes also.
bq.  Thanks,
bq.  ramkrishna

> Refactor the TimeoutMonitor to make it less racy
> ------------------------------------------------
>                 Key: HBASE-4015
>                 URL: https://issues.apache.org/jira/browse/HBASE-4015
>             Project: HBase
>          Issue Type: Sub-task
>    Affects Versions: 0.90.3
>            Reporter: Jean-Daniel Cryans
>            Assignee: ramkrishna.s.vasudevan
>            Priority: Blocker
>             Fix For: 0.92.0
>         Attachments: HBASE-4015_1_trunk.patch, HBASE-4015_2_trunk.patch, HBASE-4015_reprepared_trunk_2.patch,
Timeoutmonitor with state diagrams.pdf
> The current implementation of the TimeoutMonitor acts like a race condition generator,
mostly making things worse rather than better. It does it's own thing for a while without
caring for what's happening in the rest of the master.
> The first thing that needs to happen is that the regions should not be processed in one
big batch, because that sometimes can take minutes to process (meanwhile a region that timed
out opening might have opened, then what happens is it will be reassigned by the TimeoutMonitor
generating the never ending PENDING_OPEN situation).
> Those operations should also be done more atomically, although I'm not sure how to do
it in a scalable way in this case.

This message is automatically generated by JIRA.
For more information on JIRA, see: http://www.atlassian.com/software/jira


View raw message