Would you like to help test Confluence and Jira updates? Please fill out the form here: https://forms.gle/5TbSqvYYCnTyyxSN7
This message was received:
Event type: L1Proc.unLockRun Event timestamp: 2017-10-30 21:57:42.474089 Spacecraft ID: -1 Target: None Trigger name: L1-processing Cluster size: 1 Message text: Can't open lockfile cd /r0531060188.lock.
Status in data monitoring pages:
The process was listed as terminated and has this java error:
org.srs.jobcontrol.JobControlException: Remote Exception performing operation at org.srs.jobcontrol.JobControlClient.checkException(JobControlClient.java:219) at org.srs.jobcontrol.JobControlClient.submit(JobControlClient.java:77) at org.srs.pipeline.server.process.batch.BatchProcessExecutor.run(BatchProcessExecutor.java:202) at java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1145) at java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:615) at java.lang.Thread.run(Thread.java:745) Caused by: java.net.SocketTimeoutException: Read timed out at java.net.SocketInputStream.socketRead0(Native Method) at java.net.SocketInputStream.read(SocketInputStream.java:152) at java.net.SocketInputStream.read(SocketInputStream.java:122) at java.io.BufferedInputStream.fill(BufferedInputStream.java:235) at java.io.BufferedInputStream.read(BufferedInputStream.java:254) at java.io.DataInputStream.readByte(DataInputStream.java:265) at sun.rmi.transport.StreamRemoteCall.executeCall(StreamRemoteCall.java:214) at sun.rmi.server.UnicastRef.invoke(UnicastRef.java:161) at java.rmi.server.RemoteObjectInvocationHandler.invokeRemoteMethod(RemoteObjectInvocationHandler.java:194) at java.rmi.server.RemoteObjectInvocationHandler.invoke(RemoteObjectInvocationHandler.java:148) at com.sun.proxy.$Proxy7.submit(Unknown Source) at org.srs.jobcontrol.JobControlClient.submit(JobControlClient.java:74) ... 4 more
Seems to be an NFS error. I can't write to that directory as myself or the glast account. Need to check on proprer account/permissions.
A stuck job usually shows up in a bjobs list as having consumed no CPU or memory, and in an UNKWN state and has been sitting around for a long time (depends on job, need more data).
The procedure is to log on as glastraw an bkill them. As Warren explained:
You can't roll back a running job. You have to bkill them and wit for the reaper to notice that they're dead. The pipeline will automatically retry them once for L1 (0 is the default, you can ask for up to 3). If the jobs you killed were retries, then you'll have to roll them back. I've never seen a job in unknown state recover by itself.
Use bhosts to check the host status (although you might want to grep with the host name as the list is long).
You can just immediately run a command on any batch host with lsrun. Obviously this should be used sparingly. uptime is a good cheap thing that will show you if the host is responsive and what the load is. Or check to see if a particular filesystem is causing trouble, or if some command works.
When a script gets terminated and viewing messages includes the error message "failure when executing java...(Need to get exact text next time it happens)". This is typically an indication of a bad variable in the environment/database that the process is running it. This bad value is typically set by the upstream process having run in some strange manner and mangling the variable. In every case so far that I've seen, the upstream process ran twice simultaneously on two different hosts and so the two processes were overwriting each other. This is fairly easy to identify as the output log contains all the execution output (except possibly the environment output) twice and there are two different LSF summary blocks at the end of the log.
The solution is to simply roll back the upstream process so it gets a clean execution.
Check run failed and I thought at first it was a case of bad data as the process ran simultaneously on two different hosts. However a simple rollback did not solve the problem as there was a lock left from running the mergeCalChunks upstream process. This lock was left because of an infrastructure glitch preventing proper access to the xroot data disks. Rerunning that process removed the lock file but checkRun still didn't work, complaining about the existence of a dontCleanUp file. This file was actually left from the bad mergeCalChunks run. Before deleting it we need to check that the files listed there actually exist. This command:
for ff in $(egrep '^root' <path to>/dontCleanUp) ; do xrd.pl --where stat $ff ; done
will poll the xroot server for information on the files listed in the dontCleanUp file.
If they are all there, we are good to go and can remove the dontCleanUp file and rollback the checkRun process. If not, we need to look further upstream in the data processing chain.
Note: the xrd.pl in the above command is found at /afs/slac.stanford.edu/g/glast/applications/xrootd/PROD/bin/xrd.pl
The particular event that triggered this scenario was the two deliveries of run 543796948. The first delivery (180315009) had several recon jobs that ended up on a dead host and had to be killed. While they were hung, the rest of the events in the run arrived in delivery 180315010 and the findChunks task was in a pending state waiting for its pre-exec command (/afs/slac/g/glast/isoc/flightOps/rhel6_gcc44/ISOC_PROD/bin/isoc run /afs/slac.stanford.edu/g/glast/ground/PipelineConfig/Level1/5.6/lockFile.py) to give it the go ahead. After the hung recon tasks were killed, the reaper properly restarted them and the run completed and properly cleaned up after itself. However, the findChunks task never started up. LSF normally retries the pre-exec command regularly (every 2.5 minutes in this case - how to we check?) but had not run it in over 2 hours (again how to we check? is there a max retries?).
When this happens:
This is only a partial description of handling this. This comes up when there are a large number of jobs in an UNKWN state all on the same host. Often, looking at the status of the host (bhosts <hostname>) will tell you it is unavailable, in which case you're probably safe. If it is still open and accepting jobs, however, you want to prevent it from getting new batch jobs as they will just hang as well. From Warren your options are:
Close the host.
sudo badmin -C "Witty comment" <hostname>
Which requires sudo permission which you probably don't have yet and they'll want to limit to some machine we own and we should find a box that's not my workstation to use for that.
Fill it up with dummy jobs.
for ii in $(seq 8) ; do bsub -q glastdataq -sp 90 -m <hostname> sleep 86400
# in bash, as glastraw
So probably really only one realistic option at the moment.
Also you should probably notify unix-admin.
You may need to do something with files on XrootD. See the Intro to using xrootd.
This one came up the incident reference by this thread in the opsprob mailing list: https://www-glast.stanford.edu/protected/mail/opsprob/11338.html
Basically several jobs were hung, when they were killed, they still didn't work and rolling things back made it worse. In the end, we needed to copy the files back from xroot to NFS and restart the process. Some questions, comments, and answers from the thread (some of these still need answers):
Here's an example the output for the issue in https://www-glast.stanford.edu/protected/mail/opsprob/15272.html
The log file is of the findChunks instance that copied the evt files but didn't finish for whatever reason.
It's prudent to check that the script looks reasonable. I execute it as me, it never was neccessary to become glastraw. After that remove the run lock and roll back findChunks.
If you ever need to limit the amount of work being done on the pipeline (like we wanted to with the LAT restart in April 2018), you can manually create throttle locks to limit the number of simultaneous runs that can be worked on at a time. Right now the pipeline is set to allow up to 6 runs to be worked on at once. If you want to limit that, simply create lock files in the /nfs/farm/g/glast/u41/L1/throttle directory of the form 0.lock, 1.lock, ... up to 5.lock. The contents can be anything you want. It is just the presence of the file that stops things from running. Each lock file created will reduce the number of simultaneous runs by one. Creating all six will stop the pipeline from processing anything.
There are some variables found on the JMX page of the pipeline that control the operation of the reaper (the process that looks for dead process, cleans them up, and restarts them).
The first is the ReaperFrequencySeconds in the Control section at the top. This parameter controls how often the reaper actually runs. Currently set to 60 (once per minute).
The other parameter is set on a per submission source basis. This is the ReaperDelayMinutes parameter and is found in each of the later sections on that page. This controls how long a process has to be dead for before the reaper kills it. It is typically set to 60 or 120 minutes. NOTE (from Warren): I'm not convinced that ReaperDelayMinutes actually does anything. Restarting the pipeline set both of them back to default.
This came up with delivery 180409011. In this case the launchL1 task was started simultaneously on two different hosts in the batch queue. The first on ran successfully but since the second one failed with an error saying the L1 stream already existed, the process was marked as failed. To clean up the data display do the following:
This has come up twice now under slightly different circumstances. See the following threads for the OpsProbList chatter:
The end result seems to be that we need to repipe the runs in question. Details (from Warren) on that process (based on the second event):
The evt files weren't on xroot because cleanupChunks Stream 180502013.546962321.546962321 deleted them. I've started a RePipe:
$> run=546962321 $> dl=180502012 $> ~glast/pipeline-II/prod/pipeline createStream --stream $run --define downlinkID=$dl,runID=$run RePipe
RePipe finished. then:
$> cd /nfs/farm/g/glast/u28
$> mv RePipe/$dl/r0$run/*.evt stage/$dl/r0$run
$> cd /nfs/farm/g/glast/u41/L1/runs/546/r0546962321/
$> cd /nfs/farm/g/glast/u41/L1/runs/546/r0546962321/
$> /afs/slac.stanford.edu/g/glast/ground/PipelineConfig/Level1/5.6/newChunkList.py /nfs/farm/g/glast/u28/stage/$dl/r0$run
$> mv r0546962321_180502012_chunkList.txt.new r0546962321_180502012_chunkList.txt
You'll have to be glastraw for the last 2 steps.
But then I messed up. I rolled back fondChunks 180502012.546962321 and it failed complaining about overlapping data. The right thing is to roll back the whole doRun stream while defining deliveriesToIgnore=180502013 (How?). Which I've now done.
And now there are more errors, which I'll have to investigate later, but probably involve the magic7 file.
This is usually indicated by one or more instances of the LSF summary report appearing near the end (although not always) of the file. The summary looks like this:
------------------------------------------------------------ Sender: LSF System <lsf@hequ0119> Subject: Job 217648: <findChunks> in cluster <slac> Done Job <findChunks> was submitted from host <fermilnx-v08> by user <glastraw> in cluster <slac> at Thu Aug 2 11:37:06 2018. Job was executed on host(s) <hequ0119>, in queue <glastdataq>, as user <glastraw> in cluster <slac> at Thu Aug 2 11:37:08 2018. </u/gl/glastraw> was used as the home directory. </nfs/farm/g/glast/u41/L1/logs/PROD/L1Proc/5.6/doRun/findChunks/180xxxxxx/802xxx/015/554xxxxxx/915xxx/869> was used as the working directory. Started at Thu Aug 2 11:37:08 2018. Terminated at Thu Aug 2 11:39:29 2018. Results reported at Thu Aug 2 11:39:29 2018. Your job looked like: ------------------------------------------------------------ # LSBATCH: User input bash pipeline_wrapper ------------------------------------------------------------ Successfully completed. Resource usage summary: CPU time : 8.33 sec. Max Memory : 66 MB Average Memory : 24.33 MB Total Requested Memory : - Delta Memory : - Max Swap : - Max Processes : 8 Max Threads : 13 Run time : 141 sec. Turnaround time : 143 sec. The output (if any) is above this job summary.
In 95% of the cases, all that needs to be done is a simple rollback of that affected process. In the other 5%, there is some other underlying problem that is also affecting the completion of the job and you'll need to search through the log files for the error.
This occurs when there are no files for the half-pipe to process in the delivery. There are three causes for this:
They symptom of this is that you get an error of the form:
LSEReader::read: unknown LSE_Keys typeid 551075433 from /nfs/farm/g/glast/u42/ISOC-flight/Downlinks/180619008/0000f5a2-20d8be69-03bd-00a21.evt
In this particular case the LSE_Key was set to the run number instead of it's proper value (-1 through 3 from an enum in the code).
To date, I've only ever seen this in the mergeEvt task. This is symptomatic of a problem upstream in the doChunk streams (running the makeEvt task). In every instance, I've found one or more of those makeEvt tasks had a multiple submission. Rolling all the doChunk streams back with the multiple submissions fixes the problem.
Sometimes you need to roll back a stream but there are jobs still running that prevent the rollback from occurring immediately. Warren provided this little tidbit to set up the rollback and not have to keep checking back:
When a run is not ready for rollback, I usually do this:
$> del=<delivery #>
$> until ~glast/pipeline-II/prod/pipeline rollbackStream --minimum --force L1Proc[$del] ; do sleep 300 ; done
(in bash) so I don't have to keep checking.
I haven't actually used this yet so I'll update it (or remove this line) when I can verify that it works.
For each chunk, the individual checkChunk should remove the respective chunk token. CheckChunk of chunk 1757481 reports it can't find the token (nfs hiccup) but proceeds by design.
After all the merges etc. checkRun checks for remaining tokens and if present chokes on it. To resolve it I usually just remove the chunk token in question (checking that the chunk was processed properly, of course), and roll back the youngest (if there are more than one) checkRun. If there isn't any other issue checkRun sets the status to Complete, which triggers cleanupCompleteRun.
It's overkill to roll back from the checkChunk. Maybe even harmful, but I guess the chunk data are still on xrootd. But this should be verified.
This has happened a few times. For instance,
A typical rollback probably won't work. It will likely need to be rolled back from an earlier process like scanDigi.
This mysql server hosts the GlastCalibDB database where the L1 pipeline looks up the calibration file to use. If it's down or needs to be taken down (e.g., to install an update), Brian left instructions to halt job submission for L1Proc so that other tasks may run as normal.
Stopping a single task:
This should suspend job submission and leave processes in the Waiting state.
To clear it (and all other limits that were set for the jobsite):
When this happens, it will leave a lock file behind (e.g., /nfs/farm/g/glast/u41/L1/runs/697/r0697981861/r0697981861.lock) that you will need to delete before rolling the job back. Otherwise, the rollback will stay queued but not run. If you do the rollback before removing the lock file you should bkill the job, wait for the reaper to terminate it, then remove the lock file, and do the rollback again.
This will cause an error in the logs like:
MootSvc FATAL Hw key 8517 in data does not match hw key belonging to Moot Config #3063
This has happened in 2023, 2017, and 2013 due to a problem with an update to the hot strip mask on the satellite, e.g., the change was reverted but the pipeline was expecting the MOOT key for the new configuration. Following what was first done in 2013, the fix is to update the databases to change the MOOT key for the failed runs back to the old value and rollback the failures. Other stuff mentioned in the original thread doesn't seem necessary.
1) Change the moot_key in glastops_acqsummary in the Oracle database.
2) Change the config_fk (mootkey) in MOOD's Acq_summary in the MOOD/FMX MySQL database.
To access the database, you need to login to fermilnx02 (in 2023) and issue the command
mysql -h glastMOOD -P8208 -u <username> -p. You will need to get a password from the secret/fpd/fermi/isoc vault. Richard may need to put in a ticket for you to get permission. In 2023, we used Steve Tether's account (tether) since the "magician" account is overpowered for what we need (it's basically the super-user account).
3) In some receipts (e.g. in older emails) is written that a repipe is neccessary. Michael doesn't recall repipes for the 2017 key mismatch. At least no helpful repipes. In 2023 for sure no repipes were performed. Don't repipe unless really neccessary!
4) Rollback the affected deliveries
If they are in a Failed state, you can just do a standard rollback.
If findChunks is stuck, then bkill the process and wait for the reaper to get it. Most of them should then auto rollback without intervention.