Hey Luca, Thanks for getting back to us. That sounds very promising.
One thing I think would be great to have is a doc version of the steps for upgrade and rollback. The benefits: 1. Anything unexpected happened during automation, you do have folks can quickly understand what's going on and get into the investigation. 2. Share the doc with us to help the others OSS users for doing the migration. For the env specific things I think that's fine. We can left comment on it. At least all the other users can get a high level view of a proven solution. And then they can go and find out the rest of the pieces by themselves. For automations, I suggest to split up the automation into several stages, and apply some validation steps(manually is ok) before kicking of the next stage. Best, Evans Luca Toscano <toscano.l...@gmail.com> 於 2020年7月15日 週三 下午9:07寫道： > Hi everybody, > > I didn't get the time to work on this until recently, but I finally > managed to have a reliable procedure to upgrade from CDH to Bigtop 1.4 > and rollback if needed. The assumptions are: > > 1) It is ok to have (limited) cluster downtime. > 2) Rolling upgrade is not needed. > 3) QJM is used. > > The procedure is listed in these two scripts: > > > https://github.com/wikimedia/operations-cookbooks/blob/master/cookbooks/sre/hadoop/stop-cluster.py > > https://github.com/wikimedia/operations-cookbooks/blob/master/cookbooks/sre/hadoop/change-distro-from-cdh.py > > The code is highly dependent on my working environment, but it should > be clear to follow when writing a tutorial about how to migrate from > CDH to Bigtop. All the suggestions given by this mailing list were > really useful to reach a solution! > > My next steps will be: > > 1) Keep testing Bigtop 1.4 (finalize HDFS upgrade, run more hadoop > jobs, test Hive 2, etc..). > 2) Upgrade the production Hadoop cluster to Bigtop 1.4 on Debian 9 > (HDFS 2.6.0-cdh -> 2.8.5). > 3) Upgrade to Bigtop 1.5 on Debian 9 (HDFS 2.8.5 -> 2.10). > 4) Upgrade to Debian 10. > > With automation it shouldn't be very difficult, I'll report progress once > made. > > Thanks a lot! > > Luca > > On Mon, Apr 13, 2020 at 9:25 AM Luca Toscano <toscano.l...@gmail.com> > wrote: > > > > Hi Evans, > > > > thanks a lot for the feedback, it was exactly what I needed. The > > simpler the better is definitely a good advice in this use case, I'll > > try this week another rollout/rollback and report back :) > > > > Luca > > > > On Thu, Apr 9, 2020 at 8:09 PM Evans Ye <evan...@apache.org> wrote: > > > > > > Hi Luca, > > > > > > Thanks for reporting back and let us know how it goes. > > > I don't have the exactly HDFS with QJM HA upgrade experience. The > experience I had was 0.20 non-HA upgrade to 2.0 non-HA and then enable QJM > HA, which was back in 2014. > > > > > > Regarding to rollback, I think you're right: > > > > > > it is possible to rollback to HDFS’ state before the upgrade in case > of unexpected problems. > > > > > > My previous experience is the same that the rollback is merely a > snapshot before the upgrade. If you've gone far, then rollback cost more > data lost... Our runbook is if our sanity check failed during upgrade > downtime, we perform the rollback immediately. > > > > > > Regarding to that FSImage hole issue, I've experienced it as well. > > > I managed to fix it by manually edit the FSImage with offline image > viewer[1] and delete that missing editLog in FSImage. That actually brought > my cluster back with a little number of missing blocks. > > > > > > Our experience says that the more the steps, the more the chance you > failed the upgrade. We did good on dozen times of testing, DEV cluster, > STAGING cluster, but still got missing blocks when upgrading Production... > > > > > > The suggestion is to get your production in good shape first(the less > decommissioned, offline DNs, disk failures, the better). > > > Also, maybe you can switch to non-HA mode and do the upgrade to > simplify the things? > > > > > > Not many helps but please let us know if any progress. > > > Last one, have you reached out to Hadoop community? the authors should > know the most :) > > > > > > - Evans > > > > > > [1] > https://hadoop.apache.org/docs/r2.8.5/hadoop-project-dist/hadoop-hdfs/HdfsImageViewer.html > > > > > > Luca Toscano <toscano.l...@gmail.com> 於 2020年4月8日 週三 21:03 寫道： > > >> > > >> Hi everybody, > > >> > > >> most of the bugs/issues/etc.. that I found while upgrading from CDH 5 > > >> to BigTop 1.4 are fixed, I am now testing (as suggested also in here) > > >> upgrade/rollback procedures for HDFS (all written in > > >> https://phabricator.wikimedia.org/T244499, will add documentation > > >> about this at the end I promise). > > >> > > >> I initially followed [1][2] in my Test cluster, choosing the Rolling > > >> upgrade, but when I tried to rollback (after days since the initial > > >> upgrade) I ended up in an inconsistent state and I wasn't able to > > >> recover the previous HDFS state. I didn't save the exact error > > >> messages but the situation was more or less the following: > > >> > > >> FS-Image-rollback (created at the time of the upgrade) - up to > transaction X > > >> FS-Image-current - up to transaction Y, with Y = X + 10000 (number > > >> totally made up for the example) > > >> QJM cluster: first available transaction Z = X + 10000 + 1 > > >> > > >> When I tried to rolling rollback, the Namenode complained about a hole > > >> in the transaction log, namely at X + 1, so it refused to start. I > > >> tried to force a regular rollback, but the Namenode refused again > > >> saying that there was no available FS Image to roll back to. I checked > > >> in the Hadoop code and indeed the Namenode saves the fs image with > > >> different naming/path in case of a rolling upgrade or a regular > > >> upgrade. Both cases make sense, especially the first one since there > > >> was indeed a hole between the last transaction of the > > >> FS-Image-rollback and the first available transaction to reply on the > > >> QJM cluster. I chose the rolling upgrade initially since it was > > >> appealing: it promises to bring back the Namenodes to their previous > > >> versions, but keeping the data modified between upgrade and rollback. > > >> > > >> I then found [3], in which it is said that with QJM everything is more > > >> complicated, and a regular rollback is the only option available. What > > >> I think this mean is that due to the Edit log spread among multiple > > >> nodes, a rollback that keeps data between upgrade and rollback is not > > >> available, so worst case scenario the data modified during that > > >> timeframe is lost. Not a big deal in my case, but I want to triple > > >> check with you if this is the correct interpretation or if there is > > >> another tutorial/guide/etc.. that I haven't read with a different > > >> procedure :) > > >> > > >> Is my interpretation correct? If not, is there anybody with experience > > >> in HDFS upgrades that could shed some light on the subject? > > >> > > >> Thanks in advance! > > >> > > >> Luca > > >> > > >> > > >> > > >> [1] > https://hadoop.apache.org/docs/r2.8.5/hadoop-project-dist/hadoop-hdfs/HdfsUserGuide.html#Upgrade_and_Rollback > > >> [2] > https://hadoop.apache.org/docs/r2.8.5/hadoop-project-dist/hadoop-hdfs/HdfsRollingUpgrade.html > > >> [3] > https://hadoop.apache.org/docs/r2.8.5/hadoop-project-dist/hadoop-hdfs/HDFSHighAvailabilityWithQJM.html#HDFS_UpgradeFinalizationRollback_with_HA_Enabled >
