[
https://issues.apache.org/jira/browse/HBASE-4064?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13069098#comment-13069098
]
Jean-Daniel Cryans commented on HBASE-4064:
-------------------------------------------
This work should really be done for 0.92, not 0.90, because in its current
state the master is _bound_ to have race conditions.
As I have commented in other jiras, the TimeoutMonitor acts out of band with
regards to the other threads. While most of the code in the master assumes that
every operation is atomic on a single region, the TM doesn't. It can figure out
that a RIT expired and then only process it minutes later because of how that
code was written. Here's how it works:
# iterate on all RIT
## if the timed out
### Put the region in the assign/unassign list (depends on last state)
### also depending on state, do some ZK operation (which can take a few tens
of ms)
# Process all the regions to unassign one by one
# Process all the regions to assign one by one
Here the issue is that at any time a region can change state while the TM runs.
Let's say the TM saw region x timing out on PENDING_CLOSE, but then the
millisecond later the master receives the notification that the region was
finally closed. What happens is that the region will be forced unassigned in
the loop at the end (while the region might already be reopened elsewhere). All
of this is made even worse by the fact that ZK calls, in aggregate, really take
a lot of time. Plus during the assign and unassign loops you need to talk to
region servers.
The window for races is so large basically anything can happen.
One reason those operations are slow is that the master is suffering for a lot
of lock contention, for reference see HBASE-3789. This is only in 0.92, and
that's one more reason to focus our work there.
Something we did late in 0.90 to make it less prone to races is to bump the
timeout itself, as I was mentioning in HBASE-3669.
If anything else should be done in 0.90, it would be to disable the
TimeoutMonitor as it does more harm than good.
In 0.92 it should be completely redone to work atomically on regions, instead
of doing its own thing.
> Two concurrent unassigning of the same region caused the endless loop of
> "Region has been PENDING_CLOSE for too long..."
> ------------------------------------------------------------------------------------------------------------------------
>
> Key: HBASE-4064
> URL: https://issues.apache.org/jira/browse/HBASE-4064
> Project: HBase
> Issue Type: Bug
> Components: master
> Affects Versions: 0.90.3
> Reporter: Jieshan Bean
> Fix For: 0.90.5
>
> Attachments: HBASE-4064-v1.patch, HBASE-4064_branch90V2.patch
>
>
> 1. If there is a "rubbish" RegionState object with "PENDING_CLOSE" in
> regionsInTransition(The RegionState was remained by some exception which
> should be removed, that's why I called it as "rubbish" object), but the
> region is not currently assigned anywhere, TimeoutMonitor will fall into an
> endless loop:
> 2011-06-27 10:32:21,326 INFO
> org.apache.hadoop.hbase.master.AssignmentManager: Regions in transition timed
> out: test2,070712,1308971310309.9a6e26d40293663a79523c58315b930f.
> state=PENDING_CLOSE, ts=1309141555301
> 2011-06-27 10:32:21,326 INFO
> org.apache.hadoop.hbase.master.AssignmentManager: Region has been
> PENDING_CLOSE for too long, running forced unassign again on
> region=test2,070712,1308971310309.9a6e26d40293663a79523c58315b930f.
> 2011-06-27 10:32:21,438 DEBUG
> org.apache.hadoop.hbase.master.AssignmentManager: Starting unassignment of
> region test2,070712,1308971310309.9a6e26d40293663a79523c58315b930f.
> (offlining)
> 2011-06-27 10:32:21,441 DEBUG
> org.apache.hadoop.hbase.master.AssignmentManager: Attempted to unassign
> region test2,070712,1308971310309.9a6e26d40293663a79523c58315b930f. but it is
> not currently assigned anywhere
> 2011-06-27 10:32:31,207 INFO
> org.apache.hadoop.hbase.master.AssignmentManager: Regions in transition timed
> out: test2,070712,1308971310309.9a6e26d40293663a79523c58315b930f.
> state=PENDING_CLOSE, ts=1309141555301
> 2011-06-27 10:32:31,207 INFO
> org.apache.hadoop.hbase.master.AssignmentManager: Region has been
> PENDING_CLOSE for too long, running forced unassign again on
> region=test2,070712,1308971310309.9a6e26d40293663a79523c58315b930f.
> 2011-06-27 10:32:31,215 DEBUG
> org.apache.hadoop.hbase.master.AssignmentManager: Starting unassignment of
> region test2,070712,1308971310309.9a6e26d40293663a79523c58315b930f.
> (offlining)
> 2011-06-27 10:32:31,215 DEBUG
> org.apache.hadoop.hbase.master.AssignmentManager: Attempted to unassign
> region test2,070712,1308971310309.9a6e26d40293663a79523c58315b930f. but it is
> not currently assigned anywhere
> 2011-06-27 10:32:41,164 INFO
> org.apache.hadoop.hbase.master.AssignmentManager: Regions in transition timed
> out: test2,070712,1308971310309.9a6e26d40293663a79523c58315b930f.
> state=PENDING_CLOSE, ts=1309141555301
> 2011-06-27 10:32:41,164 INFO
> org.apache.hadoop.hbase.master.AssignmentManager: Region has been
> PENDING_CLOSE for too long, running forced unassign again on
> region=test2,070712,1308971310309.9a6e26d40293663a79523c58315b930f.
> 2011-06-27 10:32:41,172 DEBUG
> org.apache.hadoop.hbase.master.AssignmentManager: Starting unassignment of
> region test2,070712,1308971310309.9a6e26d40293663a79523c58315b930f.
> (offlining)
> 2011-06-27 10:32:41,172 DEBUG
> org.apache.hadoop.hbase.master.AssignmentManager: Attempted to unassign
> region test2,070712,1308971310309.9a6e26d40293663a79523c58315b930f. but it is
> not currently assigned anywhere
> .....
> 2 In the following scenario, two concurrent unassigning call of the same
> region may lead to the above problem:
> the first unassign call send rpc call success, the master watched the event
> of "RS_ZK_REGION_CLOSED", process this event, will create a
> ClosedRegionHandler to remove the state of the region in master.eg.
> while ClosedRegionHandler is running in
> "hbase.master.executor.closeregion.threads" thread (A), another unassign call
> of same region run in another thread(B).
> while thread B run "if (!regions.containsKey(region))", this.regions have
> the region info, now cpu switch to thread A.
> The thread A will remove the region from the sets of "this.regions" and
> "regionsInTransition", then switch to thread B. the thread B run continue,
> will throw an exception with the msg of "Server null returned
> java.lang.NullPointerException: Passed server is null for
> 9a6e26d40293663a79523c58315b930f", but without removing the new-adding
> RegionState from "regionsInTransition",and it can not be removed for ever.
> public void unassign(HRegionInfo region, boolean force) {
> LOG.debug("Starting unassignment of region " +
> region.getRegionNameAsString() + " (offlining)");
> synchronized (this.regions) {
> // Check if this region is currently assigned
> if (!regions.containsKey(region)) {
> LOG.debug("Attempted to unassign region " +
> region.getRegionNameAsString() + " but it is not " +
> "currently assigned anywhere");
> return;
> }
> }
> String encodedName = region.getEncodedName();
> // Grab the state of this region and synchronize on it
> RegionState state;
> long stamp = -1;
> synchronized (regionsInTransition) {
> state = regionsInTransition.get(encodedName);
> if (state == null) {
> state = new RegionState(region, RegionState.State.PENDING_CLOSE);
> stamp =state.getStamp();
> regionsInTransition.put(encodedName, state);
> } else if (force && state.isPendingClose()) {
> LOG.debug("Attempting to unassign region " +
> region.getRegionNameAsString() + " which is already pending close
> "
> + "but forcing an additional close");
> state.update(RegionState.State.PENDING_CLOSE);
> } else {
> LOG.debug("Attempting to unassign region " +
> region.getRegionNameAsString() + " but it is " +
> "already in transition (" + state.getState() + ")");
> return;
> }
> }
>
> // Send CLOSE RPC
> HServerInfo server = null;
> synchronized (this.regions) {
> server = regions.get(region);
> }
> }
> 2011-06-27 10:20:59,583 DEBUG
> org.apache.hadoop.hbase.master.AssignmentManager: Starting unassignment of
> region test2,070712,1308971310309.9a6e26d40293663a79523c58315b930f.
> (offlining)
> 2011-06-27 10:20:59,585 DEBUG
> org.apache.hadoop.hbase.master.AssignmentManager: Sent CLOSE to
> serverName=158-1-91-101,20020,1308983865292, load=(requests=0, regions=0,
> usedHeap=0, maxHeap=0) for region
> test2,070712,1308971310309.9a6e26d40293663a79523c58315b930f.
> 2011-06-27 10:22:15,299 DEBUG
> org.apache.hadoop.hbase.master.AssignmentManager: Handling new unassigned
> node: /hbase/unassigned/9a6e26d40293663a79523c58315b930f
> (region=test2,070712,1308971310309.9a6e26d40293663a79523c58315b930f.,
> server=158-1-91-101,20020,1308983865292, state=RS_ZK_REGION_CLOSED)
> 2011-06-27 10:22:15,299 DEBUG
> org.apache.hadoop.hbase.master.AssignmentManager: Handling
> transition=RS_ZK_REGION_CLOSED, server=158-1-91-101,20020,1308983865292,
> region=9a6e26d40293663a79523c58315b930f
> 2011-06-27 10:25:22,636 INFO
> org.apache.hadoop.hbase.master.AssignmentManager: Regions in transition timed
> out: test2,070712,1308971310309.9a6e26d40293663a79523c58315b930f.
> state=CLOSED, ts=1309141335247
> 2011-06-27 10:25:22,636 INFO
> org.apache.hadoop.hbase.master.AssignmentManager: Region
> 9a6e26d40293663a79523c58315b930f has been CLOSED for too long, waiting on
> queued ClosedRegionHandler to run or server shutdown
> 2011-06-27 10:25:55,022 DEBUG
> org.apache.hadoop.hbase.master.handler.ClosedRegionHandler: Handling CLOSED
> event for 9a6e26d40293663a79523c58315b930f
> 2011-06-27 10:25:55,022 DEBUG
> org.apache.hadoop.hbase.master.AssignmentManager: Table being disabled so
> deleting ZK node and removing from regions in transition, skipping assignment
> of region test2,070712,1308971310309.9a6e26d40293663a79523c58315b930f.
> 2011-06-27 10:25:55,022 DEBUG org.apache.hadoop.hbase.zookeeper.ZKAssign:
> master:20000-0x230ba8b85230036 Deleting existing unassigned node for
> 9a6e26d40293663a79523c58315b930f that is in expected state
> RS_ZK_REGION_CLOSED
> 2011-06-27 10:25:55,101 DEBUG org.apache.hadoop.hbase.zookeeper.ZKAssign:
> master:20000-0x230ba8b85230036 Successfully deleted unassigned node for
> region 9a6e26d40293663a79523c58315b930f in expected state RS_ZK_REGION_CLOSED
> 2011-06-27 10:25:55,301 DEBUG
> org.apache.hadoop.hbase.master.AssignmentManager: Starting unassignment of
> region test2,070712,1308971310309.9a6e26d40293663a79523c58315b930f.
> (offlining)
> 2011-06-27 10:25:55,302 INFO
> org.apache.hadoop.hbase.master.AssignmentManager: Server null returned
> java.lang.NullPointerException: Passed server is null for
> 9a6e26d40293663a79523c58315b930f
> 2011-06-27 10:32:21,326 INFO
> org.apache.hadoop.hbase.master.AssignmentManager: Regions in transition timed
> out: test2,070712,1308971310309.9a6e26d40293663a79523c58315b930f.
> state=PENDING_CLOSE, ts=1309141555301
> 2011-06-27 10:32:21,326 INFO
> org.apache.hadoop.hbase.master.AssignmentManager: Region has been
> PENDING_CLOSE for too long, running forced unassign again on
> region=test2,070712,1308971310309.9a6e26d40293663a79523c58315b930f.
> 2011-06-27 10:32:21,438 DEBUG
> org.apache.hadoop.hbase.master.AssignmentManager: Starting unassignment of
> region test2,070712,1308971310309.9a6e26d40293663a79523c58315b930f.
> (offlining)
> 2011-06-27 10:32:21,441 DEBUG
> org.apache.hadoop.hbase.master.AssignmentManager: Attempted to unassign
> region test2,070712,1308971310309.9a6e26d40293663a79523c58315b930f. but it is
> not currently assigned anywhere
> 3 The following scenario shows how the above problem 2 happened:
> (1)A table have a lot of regions, more than 70000 in my test.
> (2)Disable the table, if 'BulkDisabler.waitUntilDone' timeout,
> 'DisableTableHandler.process' will create another BulkDisabler object and
> start its thread pool. The region which was still online will call
> AssignmentManager.unassign again. so the same region
> "AssignmentManager.unassign" could be called concurrentlly more than 1.
--
This message is automatically generated by JIRA.
For more information on JIRA, see: http://www.atlassian.com/software/jira
