Do you mean that after an OOME, the tserver process didn't die and got
into this bad state with an permanently offline tablet?
Christopher wrote:
You might be seeing https://issues.apache.org/jira/browse/ACCUMULO-4160
On Mon, Apr 11, 2016 at 5:52 PM Jayesh Patel <[email protected]
<mailto:[email protected]>> wrote:
There really aren't a lot of log messages that can explain why
tablets for other tables went offline except the following:
2016-04-11 13:32:18,258
[tserver.TabletServerResourceManager$AssignmentWatcher] WARN :
tserver:instance-accumulo-3 Assignment for 2<< has been running for
at least 973455566ms
java.lang.Exception: Assignment of 2<<
at sun.misc.Unsafe.park(Native Method)
at java.util.concurrent.locks.LockSupport.park(Unknown Source)
at
java.util.concurrent.locks.AbstractQueuedSynchronizer.parkAndCheckInterrupt(Unknown
Source)
at
java.util.concurrent.locks.AbstractQueuedSynchronizer.acquireQueued(Unknown
Source)
at
java.util.concurrent.locks.AbstractQueuedSynchronizer.acquire(Unknown
Source)
at
java.util.concurrent.locks.ReentrantLock$FairSync.lock(Unknown Source)
at java.util.concurrent.locks.ReentrantLock.lock(Unknown Source)
at
org.apache.accumulo.tserver.TabletServer.acquireRecoveryMemory(TabletServer.java:2230)
at
org.apache.accumulo.tserver.TabletServer.access$2600(TabletServer.java:252)
at
org.apache.accumulo.tserver.TabletServer$AssignmentHandler.run(TabletServer.java:2150)
at
org.apache.accumulo.fate.util.LoggingRunnable.run(LoggingRunnable.java:35)
at
org.apache.accumulo.tserver.ActiveAssignmentRunnable.run(ActiveAssignmentRunnable.java:61)
at
org.apache.htrace.wrappers.TraceRunnable.run(TraceRunnable.java:57)
at java.util.concurrent.ThreadPoolExecutor.runWorker(Unknown
Source)
at java.util.concurrent.ThreadPoolExecutor$Worker.run(Unknown
Source)
at
org.apache.accumulo.fate.util.LoggingRunnable.run(LoggingRunnable.java:35)
at java.lang.Thread.run(Unknown Source)
Table 2<< here doesn't have the issue with minc failing and so
shouldn’t be offline. These messages happened on a restart of a
tserver if that offers any clues. All the nodes were rebooted at
that time due to a power failure. I'm assuming that it's tablet
went offline soon after this message first appeared in the logs.
Other tidbit of note is that the Accumulo operates for hours/days
without taking the tablets offline even though minc is failing and
it's the crash of a tserver due to OutOfMemory situation in one case
that seems to have taken the tablet offline. Is it safe to assume
that other tservers are not able to pick up the tablets that are
failing minc from a crashed tserver?
-----Original Message-----
From: Josh Elser [mailto:[email protected]
<mailto:[email protected]>]
Sent: Friday, April 08, 2016 10:52 AM
To: [email protected] <mailto:[email protected]>
Subject: Re: Fwd: why compaction failure on one table brings other
tables offline, how to recover
Billie Rinaldi wrote:
> *From:* Jayesh Patel
> *Sent:* Thursday, April 07, 2016 4:36 PM
> *To:* '[email protected] <mailto:[email protected]>
<mailto:[email protected] <mailto:[email protected]>>'
> <[email protected] <mailto:[email protected]>
<mailto:[email protected] <mailto:[email protected]>>>
> *Subject:* RE: why compaction failure on one table brings other
tables
> offline, how to recover____
>
> __ __
>
> I have a 3 node Accumulo 1.7 cluster with a few small tables (few MB
> in size at most).____
>
> __ __
>
> I had one of those table fail minc because I had configured a
> SummingCombiner with FIXEDLEN but had smaller values:____
>
> MinC failed (trying to convert to long, but byte array isn't long
> enough, wanted 8 found 1) to create
>
hdfs://instance-accumulo:8020/accumulo/tables/1/default_tablet/F0002bc
> s.rf_tmp
> retrying ...____
>
> __ __
>
> I have learned since to set the ‘lossy’ parameter to true to
avoid this.
> *Why is the default value for it false* if it can cause catastrophic
> failure that you’ll read about ahead.____
I'm pretty sure I told you this on StackOverflow, but if you're not
writing 8-byte long values, don't used FIXEDLEN. Use VARLEN instead.
> However, this brought other the tablets for other tables offline
> without any apparent errors or warnings. *Can someone please explain
> why?*____
Can you provide logs? We are not wizards :)
> In order to recover from this, I did a ‘droptable’ from the shell on
> the affected tables, but they all got stuck in the ‘DELETING’ state.
> I was able to finally delete them using zkcli ‘rmr’ command. *Is
there
> a better way?____*
Again, not sure why they would have gotten stuck in the deleting
phase without more logs/context (nor how far along in the deletion
process they got). It's possible that there were still entries in
the accumulo.metadata table.
> I’m assuming there is a more proper way because when I created the
> tables again (with the same name), they went back to having a single
> offline tablet right away. *Is this because there are “traces” of the
> old table left behind that affect the new table even though the new
> table has a different table id?* I ended up wiping out hdfs and
> recreating the accumulo instance. ____
Accumulo uses monotonically increasing IDs to identify tables. The
human-readable names are only there for your benefit. Creating a
table with the same name would not cause a problem. It sounds like
you got the metadata table in a bad state or have tabletservers in a
bad state (if you haven't restarted them).
> It seems that a small bug, writing 1 byte value instead of 8 bytes,
> caused us to dump the whole accumulo instance. Luckily the data
> wasn’t that important, but this whole episode makes us wonder why
> doing things the right way (assuming there is a right way) wasn’t
> obvious or if Accumulo is just very fragile.____
>
Causing Accumulo to be unable to flush data from memory to disk in a
minor compaction is a very bad idea. One that we cannot
automatically recover from because of the combiner configuration you
set.
If you can provide logs and stack traces from the Accumulo services,
we can try to help you further. This is not normal. If you don't
believe me, take a look at the distributed tests we run each release
where we write hundreds of gigabytes of data across many servers
while randomly killing Accumulo processes.
>
> Please ask away any questions/clarification you might have. We’ll
> appreciate any input you might have so we make educated decisions
> about using Accumulo going forward.____
>
> __ __
>
> Thank you,____
>
> Jayesh____
>
>
