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).
We assumed that we can run for loop from 0 to lastStep sequentially. MergeTableRegionProcedure skips step 2. So, when i is 0 the procedure is already at step 3.
Added a method StateMachineProcedure#getCurrentStateId that can be used from test code only.
On failed RPC we expire the server and suspend expecting the
resultant ServerCrashProcedure to wake us back up again. In tests,
TestRSGroup hung because it failed to schedule a server expiration
because the server was already expired undergoing processing (the
test was shutting down). Deal with this case by having expire
servers return false if unable to expire. Callers will then know
where a ServerCrashProcedure has been scheduled or not.
M hbase-server/src/main/java/org/apache/hadoop/hbase/master/ServerManager.java
Have expireServer return true if successful.
M hbase-server/src/main/java/org/apache/hadoop/hbase/master/assignment/RegionTransitionProcedure.java
The log that included an exception whose message was the current
procedure as a String totally baffled me. Make it more obvious what
exception is.
M hbase-server/src/main/java/org/apache/hadoop/hbase/master/assignment/UnassignProcedure.java
If failed expire of a server, wake our procedure -- do not suspend --
and presume ok to move region to CLOSED state (because going down or
concurrent crashed server processing ongoing).
This de-couples the snapshot size calculation from the
SpaceQuotaObserverChore into another API which both the periodically
invoked Master chore and the Master service endpoint can invoke. This
allows for multiple sources of snapshot size to reported (from the
multiple sources we have in HBase).
When a file is archived, snapshot sizes can be more quickly realized and
the Master can still perform periodical computations of the total
snapshot size to account for any delayed/missing/lost file archival RPCs.
Signed-off-by: Ted Yu <yuzhihong@gmail.com>
* rely on git plumbing commands when checking if we've built the site for a particular commit already
* switch to forcing '-e' for bash
* add command line switches for: path to hbase, working directory, and publishing
* only export JAVA/MAVEN HOME if they aren't already set.
* add some docs about assumptions
* Update javadoc plugin to consistently be version 3.0.0
* avoid duplicative site invocations on reactor modules
* update use of cp command so it works both on linux and mac
* manually skip enforcer plugin during build
* still doing install of all jars due to MJAVADOC-490, but then skip rebuilding during aggregate reports.
* avoid the pager on git-diff by teeing to a log file, which also helps later reviewing in the case of big changesets.
Signed-off-by: Michael Stack <stack@apache.org>
Signed-off-by: Misty Stanley-Jones <misty@apache.org>
Master never left waitForMasterActive because it never checked state of
the clusterUp flag. The test here was aborting regionserver and then
just exiting. The minihbasecluster shutdown sets the cluster down flag
but we were never looking at it so Master thread was staying up.
M hbase-server/src/main/java/org/apache/hadoop/hbase/master/HMaster.java
The tableOnMaster check in waitForMasterActive looks wrong. It was
making it so a 'normal' Master was getting stuck in here. This is not
the place to worry about tablesOnMaster. That is for the balancer to be
concerned with. There is a problem with Master hosting
system-tables-only. After further study, Master can carry regions like a
regionserver but making it so it carries system tables only is tricky
given meta assign happens ahead of all others which means that the
Master needs to have checked-in as a regionserver super early... It
needs work. Punted for now.
M hbase-server/src/main/java/org/apache/hadoop/hbase/master/assignment/AssignmentManager.java
Mostly renaming so lists and maps of region infos have same name as they
have elsewhere in code base and cleaning up confusion that may arise
when we talk of servers-for-system-tables....It is talking about
something else in the code changes here that is other than the normal
understanding. It is about filtering regionservers by their version
numbers so we favor regions with higher version numbers. Needs to go
back up into the balancer.
M hbase-server/src/main/java/org/apache/hadoop/hbase/master/balancer/BaseLoadBalancer.java
It was possible for the Master to be given regions if no regionservers
available (as per the failing unit test in this case).
M hbase-server/src/main/java/org/apache/hadoop/hbase/regionserver/HRegionServer.java
Minor reordering moving the waitForMasterActive later in the initialize
and wrapping each test in a check if we are to keep looping (which
checks cluster status flag).
M hbase-server/src/test/java/org/apache/hadoop/hbase/master/TestAssignmentManagerMetrics.java
This was an old test from the days when Master carried system tables.
Updated test and fixed metrics. Metrics count the hbase:meta along with
the userspace region so upped expected numbers (previously the
hbase:meta was hosted on the master so metrics were not incremented).
M hbase-server/src/test/java/org/apache/hadoop/hbase/master/balancer/TestRegionsOnMasterOptions.java
I took a look at this test again but nope, needs a load of work still to
make it pass.
M hbase-zookeeper/src/main/java/org/apache/hadoop/hbase/zookeeper/ZKUtil.java
Stop being so whiney.
Certain operations in HBase are known to directly affect
the utilization of tables on HDFS. When these actions
occur, we can circumvent the normal path and notify the
Master directly. This results in a much faster response to
changes in HDFS usage.
This requires FS scanning by the RS to be decoupled from
the reporting of sizes to the Master. An API inside each
RS is made so that any operation can hook into this call
in the face of other operations (e.g. compaction, flush,
bulk load).
Signed-off-by: Ted Yu <yuzhihong@gmail.com>
Allow OPEN as a possible state when update region transition state.
Usually state is OPENING but if crash before finish step is completed,
on replay, master may have read that the state is OPEN from meta table
and so will think it open... When we replay the procedure finish, allow
that the region is already OPEN.
Fix MoveRandomRegionOfTableAction. It depended on old AM behavior.
Make it do explicit move as is required in AMv3; w/o it, it was just
closing region causing test to fail.
Fix pom so hadoop3 profile specifies a different netty3 version.
Bunch of logging format change that came of trying trying to read
the spew from this test.
This reverts commit f1a81618fd.
Conflicts:
hbase-server/src/main/java/org/apache/hadoop/hbase/regionserver/HStore.java
hbase-server/src/main/java/org/apache/hadoop/hbase/regionserver/Segment.java
patch reverted changes that happened in parallel without explanation. see jira.
Negative Remaining KVs and progress percent greater than 100 is because CompactionProgress#totalCompactingKVs is sometimes less than CompactionProgress#currentCompactedKVs.
Changes add a getter to CompactionProgress#totalCompactingKVs and from inside getter warning is logged. currentCompactedKVs are return when totalCompactingKVs are less than current.
Signed-off-by: Michael Stack <stack@apache.org>
Revert "Revert "HBASE-2004 TestClientClusterStatus is flakey""
This is a revert of a revert, i.e. a reapplication, just so I can fix
the JIRA number.
This reverts commit 6796b8e21f.
Negative Remaining KVs and progress percent greater than 100 is because CompactionProgress#totalCompactingKVs is sometimes less than CompactionProgress#currentCompactedKVs.
Changes add a getter to CompactionProgress#totalCompactingKVs and from inside getter warning is logged. currentCompactedKVs are return when totalCompactingKVs are less than current.
LocalHBaseCluster forces random port assignment for sake of concurrent unit test
execution friendliness, but we still need a positive test for RPC and info port
assignment.
The 1.x functionality of Master DDL operations is that "post" observer hooks
are only invoked when the DDL action was successful. With the async-ness of
ProcV2, we find ourselves in a case where the post-hook may be invoked before
the Procedure runs and fails. We need to introduce some blocking to wait and
see if the Procedure is going to fail on a precondition before invoking the hook.
Signed-off-by: Michael Stack <stack@apache.org>
- Added ADMIN permission check for following rpc calls:
normalize, setNormalizerRunning, runCatalogScan, enableCatalogJanitor, runCleanerChore,
setCleanerChoreRunning, execMasterService, execProcedure, execProcedureWithRet
- Moved authorizationEnabled check to start of AccessChecker's functions. Currently, and IDK why,
we call authManager.authorize() first and then discard its result if authorizationEnabled is false. Weird.
Remove assert in splittableregionprocedure. It was in the prepare.
Was causing fail in legit case where a region split follows a
table split BEFORE the parent has been GC'd. The region split
finds the parent in SPLIT state which is right. The assert was
having us fail. No need.
Also disabled TestHTrace since not supported in 2.0.0 and flakey.
Only call server.checkIfShouldMoveSystemRegionAsync if a node has been
added. Do not call it if only one regionserver in cluster. Make it
so ServerCrashProcedure runs before it. Add logging if
server.checkIfShouldMoveSystemRegionAsync was responsible for
MOVE (Previous was a mystery when it cut in).
Previous we'd call it when there was a nodeChildrenChanged. These
happen before nodeDeleted. If a server crashed,
checkIfShouldMoveSystemRegionAsync could run first, find the
server that had not yet registered as crashed, find system
tables on it and then try to move them. It would fail because
server would not respond to RPC. The region move would then
be waiting on the servercrashprocedure to wake it up when
done processing but this move had locked the region so
SCP couldn't run....
Serializing, if appropriate, write the hbase-1.x version of the
Comparator to the hfile trailer so hbase-1.x files can read hbase-2.x
hfiles (they are the same format).
Sahil Aggarwal
Michael Stack
BELUGA BEHR
zhangduo
Umesh Agashe
Guangxu Cheng
Ashish Singhi
Josh Elser
haxiaolin
Chia-Ping Tsai
Sean Busbey
huzheng
tedyu
Sergey Soldatov
Vladimir Rodionov
gsheffi
Vasudevan
Toshihiro Suzuki
Apekshit Sharma
Esteban Gutierrez
Andrew Purtell
Mike Drob
anastas
Reid Chan
eshcar
Thiruvel Thirumoolan