The kudu master failure was at 16:36:28 UTC, but at 13:57:55 and 15:58:41 we have ntptime printing "status 0x2001 (PLL,NANO),".
On Thu, Nov 17, 2016 at 10:14 AM, Todd Lipcon <[email protected]> wrote: > On Thu, Nov 17, 2016 at 10:00 AM, Jim Apple <[email protected]> wrote: > >> I now have a Kudu master failing with "F1117 16:36:26.940562 113049 >> hybrid_clock.cc:227] Couldn't get the current time: Clock >> unsynchronized. Status: Service unavailable: Error reading clock. >> Clock considered unsynchronized" >> >> Todd, when you say "we need to wait for sync before running tests or >> allow an unsafe flag to disable the check," do you mean "we" Impala, >> "we" Kudu, or both? If Kudu or both, is there a Kudu bug I should >> follow? I don't see a likely candidate at >> >> https://issues.apache.org/jira/browse/KUDU-1202?jql= >> project%20%3D%20KUDU%20AND%20text%20~%20%22%5C%22Clock% >> 20considered%20unsynchronized%5C%22%22 > > > Sorry, I was on my phone so wasn't as precise as I should have been. I > think that either: > > (a) Impala could add a script which runs prior to tests which waits for > synchronization. Apparently there is a builtin "ntp-wait" command in some > distros. We use the following which also includes restarting NTP a few > times which seems to speed things up: > https://gist.github.com/toddlipcon/97f7c8a4cf1d9c2551bd4289b97dfe48 > > (b) Kudu could add a flag like --ignore_bad_ntp_sync_status which you could > use in minicluster tests. In a minicluster where all the daemons run on one > node, it's sort of guaranteed that the clocks are in sync, so maybe it's > not too bad of an idea. The only risk I can see is that the clock might be > adjusted backwards if NTP is not in sync status yet, and we may have some > assertions which would be unhappy if that happened. We could give it a try, > though, if the "wait-for-ntp" approach doesn't help. > > -Todd > > > > > > >> >> >> On Wed, Nov 16, 2016 at 2:19 PM, Todd Lipcon <[email protected]> wrote: >> > This particular error message isn't related to the clock error itself. >> It's >> > the sync status flag which indicates whether NTP thinks its phase lock >> loop >> > is in sync or not. So, changing the allowed error shouldn't have any >> > effect. Instead we need to wait for sync before running tests or allow an >> > unsafe flag to disable the check. >> > >> > On Nov 16, 2016 1:55 PM, "Matthew Jacobs" <[email protected]> wrote: >> > >> >> Yeah, I don't doubt that the ntp output when you checked it was wrong, >> >> but it may be different from the time when Kudu failed to start, since >> >> that was probably at least several minutes later (assuming you had to >> >> find out about the failed job, ssh into the machine, discover the ntp >> >> warning and then check it at that time). >> >> >> >> I agree adding some logging at the beginning of jobs is the best thing >> >> to do now. >> >> >> >> Thanks! >> >> >> >> On Wed, Nov 16, 2016 at 11:38 AM, Jim Apple <[email protected]> >> wrote: >> >> > I found a place where I suppose that could go >> >> > (testdata/cluster/node_templates/cdh5/etc/kudu/), but I'm not sure >> >> > what I want to set it at. If "maximum error 197431 us" is to be >> >> > believed, it needs to be at least 0.2 seconds, and if >> >> > https://kudu.apache.org/docs/troubleshooting.html is to be believed, >> >> > it is already 10 seconds. >> >> > >> >> > I added some more logging to try and see what is going on, and if I >> >> > hit this again I'll re-open the thread. For now, I am not planning to >> >> > submit a patch to "fix" it because I'm not sure increasing the number >> >> > is the real solution. >> >> > >> >> > On Wed, Nov 16, 2016 at 11:00 AM, Matthew Jacobs <[email protected]> >> >> wrote: >> >> >> According to the error message, it looks like we can specify the >> >> >> '--max_clock_sync_error_usec' flag when starting the Kudu processes. >> >> >> We may want to start by printing ntptime output at the beginning of >> >> >> jobs so we can see how far off it is. If it's off by days then maybe >> >> >> changing the error isn't a good idea, and we'll need to figure out >> >> >> something else. >> >> >> >> >> >> On Wed, Nov 16, 2016 at 10:55 AM, Jim Apple <[email protected]> >> >> wrote: >> >> >>> How do we bump up the allowable error? >> >> >>> >> >> >>> On Wed, Nov 16, 2016 at 10:20 AM, Matthew Jacobs <[email protected]> >> >> wrote: >> >> >>>> I asked on the Kudu slack channel, they have seen issues where >> freshly >> >> >>>> provisioned ec2 nodes take some time for ntp to quiesce, but they >> >> >>>> didn't have a sense of how long that might take. If you checked >> >> >>>> ntptime after the job failed, it may be that ntp had enough time. >> We >> >> >>>> can probably consider bumping up the allowable error. >> >> >>>> >> >> >>>> On Wed, Nov 16, 2016 at 9:24 AM, Jim Apple <[email protected]> >> >> wrote: >> >> >>>>> This is the second time I have seen it, but it doesn't happen >> every >> >> >>>>> time. It could very well be a difference on ec2; already I've seen >> >> >>>>> some bugs due to my ec2 instances being Etc/UTC timezone while >> most >> >> >>>>> Impala developers work in America/Los_Angeles. >> >> >>>>> >> >> >>>>> On Wed, Nov 16, 2016 at 9:10 AM, Matthew Jacobs <[email protected]> >> >> wrote: >> >> >>>>>> No problem. If this happens again we should ask the Kudu >> >> developers. I >> >> >>>>>> haven't seen this before - I wonder if it could be some >> weirdness on >> >> >>>>>> ec2... >> >> >>>>>> >> >> >>>>>> Thanks >> >> >>>>>> >> >> >>>>>> On Wed, Nov 16, 2016 at 9:01 AM, Jim Apple <[email protected] >> > >> >> wrote: >> >> >>>>>>> Thank you for your help! >> >> >>>>>>> >> >> >>>>>>> This was on an AWS machine that has expired, but I can see from >> the >> >> >>>>>>> logs that "IMPALA_KUDU_VERSION=88b023" and >> >> >>>>>>> "KUDU_JAVA_VERSION=1.0.0-SNAPSHOT" and "Downloading >> >> >>>>>>> kudu-python-0.3.0.tar.gz" and "URL >> >> >>>>>>> https://native-toolchain.s3.amazonaws.com/build/264- >> >> e9d44349ba/kudu/88b023-gcc-4.9.2/kudu-88b023-gcc-4.9.2-ec2- >> >> package-ubuntu-14-04.tar.gz". >> >> >>>>>>> I'll add "ps aux | grep kudu" to the logging this machine does >> on >> >> >>>>>>> error, so we'll have it next time, but I did "ps -Afly" on exit >> and >> >> >>>>>>> there were no kudu processes running, it looks like. >> >> >>>>>>> >> >> >>>>>>> On Wed, Nov 16, 2016 at 8:52 AM, Matthew Jacobs < >> [email protected]> >> >> wrote: >> >> >>>>>>>> Can you check which version of the client you're building >> against >> >> >>>>>>>> (KUDU_VERSION env var) vs what Kudu version is running (ps aux >> | >> >> grep >> >> >>>>>>>> kudu >> >> >>>>>>>> >> >> >>>>>>>> On Wed, Nov 16, 2016 at 8:48 AM, Jim Apple < >> [email protected]> >> >> wrote: >> >> >>>>>>>>> Yes. >> >> >>>>>>>>> >> >> >>>>>>>>> On Wed, Nov 16, 2016 at 7:45 AM, Matthew Jacobs < >> [email protected]> >> >> wrote: >> >> >>>>>>>>>> Do you have NTP installed? >> >> >>>>>>>>>> >> >> >>>>>>>>>> On Tue, Nov 15, 2016 at 9:22 PM, Jim Apple < >> >> [email protected]> wrote: >> >> >>>>>>>>>>> I have a machine where Kudu failed to start: >> >> >>>>>>>>>>> >> >> >>>>>>>>>>> F1116 05:02:00.173629 71098 tablet_server_main.cc:64] Check >> >> failed: >> >> >>>>>>>>>>> _s.ok() Bad status: Service unavailable: Cannot initialize >> >> clock: >> >> >>>>>>>>>>> Error reading clock. Clock considered unsynchronized >> >> >>>>>>>>>>> >> >> >>>>>>>>>>> https://kudu.apache.org/docs/troubleshooting.html says: >> >> >>>>>>>>>>> >> >> >>>>>>>>>>> "For the master and tablet server daemons, the server’s >> clock >> >> must be >> >> >>>>>>>>>>> synchronized using NTP. In addition, the maximum clock error >> >> (not to >> >> >>>>>>>>>>> be mistaken with the estimated error) be below a >> configurable >> >> >>>>>>>>>>> threshold. The default value is 10 seconds, but it can be >> set >> >> with the >> >> >>>>>>>>>>> flag --max_clock_sync_error_usec." >> >> >>>>>>>>>>> >> >> >>>>>>>>>>> and >> >> >>>>>>>>>>> >> >> >>>>>>>>>>> "If NTP is installed the user can monitor the >> synchronization >> >> status >> >> >>>>>>>>>>> by running ntptime. The relevant value is what is reported >> for >> >> maximum >> >> >>>>>>>>>>> error." >> >> >>>>>>>>>>> >> >> >>>>>>>>>>> ntptime reports: >> >> >>>>>>>>>>> >> >> >>>>>>>>>>> ntp_gettime() returns code 0 (OK) >> >> >>>>>>>>>>> time dbd66a6a.59bca948 Wed, Nov 16 2016 5:17:30.350, >> >> (.350535824), >> >> >>>>>>>>>>> maximum error 197431 us, estimated error 71015 us, TAI >> >> offset 0 >> >> >>>>>>>>>>> ntp_adjtime() returns code 0 (OK) >> >> >>>>>>>>>>> modes 0x0 (), >> >> >>>>>>>>>>> offset 74989.459 us, frequency 19.950 ppm, interval 1 s, >> >> >>>>>>>>>>> maximum error 197431 us, estimated error 71015 us, >> >> >>>>>>>>>>> status 0x2001 (PLL,NANO), >> >> >>>>>>>>>>> time constant 6, precision 0.001 us, tolerance 500 ppm, >> >> >>>>>>>>>>> >> >> >>>>>>>>>>> So it looks like this error is anticipated, but the expected >> >> >>>>>>>>>>> conditions for it to occur are absent. Any ideas what could >> be >> >> going >> >> >>>>>>>>>>> on here? This is with a recent checkout of Impala master. >> >> >> > > > > -- > Todd Lipcon > Software Engineer, Cloudera
