[gpfsug-discuss] A cautionary tale of upgrades

[Simon Thompson]

I’ll start by saying this is our experience, maybe we did something stupid along the way, but just in case others see similar issues …

We have a cluster which contains protocol nodes, these were all happily running GPFS 5.0.1-2 code. But the cluster was a only 4 nodes + 1 quorum node – manager and quorum functions were handled by the 4 protocol nodes.

Then one day we needed to reboot a protocol node. We did so and its disk controller appeared to have failed. Oh well, we thought we’ll fix that another day, we still have three other quorum nodes.

As they are all getting a little long in the tooth and were starting to struggle, we thought, well we have DME, lets add some new nodes for quorum and token functions. Being shiny and new they were all installed with GPFS 5.0.2-1 code.

All was well.

The some-time later, we needed to restart another of the CES nodes, when we started GPFS on the node, it was causing havock in our cluster – CES IPs were constantly being assigned, then removed from the remaining nodes in the cluster. Crap we thought and disabled the node in the cluster. This made things stabilise and as we’d been having other GPFS issues, we didn’t want service to be interrupted whilst we dug into this. Besides, it was nearly Christmas and we had conferences and other work to content with.

More time passes and we’re about to cut over all our backend storage to some shiny new DSS-G kit, so we plan a whole system maintenance window. We finish all our data sync’s and then try to start our protocol nodes to test them. No dice … we can’t get any of the nodes to bring up IPs, the logs look like they start the assignment process, but then gave up.

A lot of digging in the mm korn shell scripts, and some studious use of DEBUG=1 when testing, we find that mmcesnetmvaddress is calling “tsctl shownodes up”. On our protocol nodes, we find output of the form:
bear-er-dtn01.bb2.cluster.cluster,rds-aw-ctdb01-data.bb2.cluster.cluster,rds-er-ctdb01-data.bb2.cluster.cluster,bber-irods-ires01-data.bb2.cluster.cluster,bber-irods-icat01-data.bb2.cluster.cluster,bbaw-irods-icat01-data.bb2.cluster.cluster,proto-pg-mgr01.bear.cluster.cluster,proto-pg-pf01.bear.cluster.cluster,proto-pg-dtn01.bear.cluster.cluster,proto-er-mgr01.bear.cluster.cluster,proto-er-pf01.bear.cluster.cluster,proto-aw-mgr01.bear.cluster.cluster,proto-aw-pf01.bear.cluster.cluster

Now our DNS name for these nodes is bb2.cluster … something is repeating the DNS name.

So we dig around, resolv.conf, /etc/hosts etc all look good and name resolution seems fine.

We look around on the manager/quorum nodes and they don’t do this cluster.cluster thing. We can’t find anything else Linux config wise that looks bad. In fact the only difference is that our CES nodes are running 5.0.1-2 and the manager nodes 5.0.2-1. Given we’re changing the whole storage hardware, we didn’t want to change the GPFS/NFS/SMB code on the CES nodes, (we’ve been bitten before with SMB packages not working properly in our environment), but we go ahead and do GPFS and NFS packages.

Suddenly, magically all is working again. CES starts fine and IPs get assigned OK. And tsctl gives the correct output.

So, my supposition is that there is some incompatibility between 5.0.1-2 and 5.0.2-1 when running CES and the cluster manager is running on 5.0.2-1. As I said before, I don’t have hard evidence we did something stupid, but it certainly is fishy. We’re guessing this same “feature” was the cause of the CES issues we saw when we rebooted a CES node and the IPs kept deassigning… It looks like all was well as we added the manager nodes after CES was started, but when a CES node restarted, things broke.

We got everything working again in house so didn’t raise a PMR, but if you find yourself in this upgrade path, beware!

Simon


-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://gpfsug.org/pipermail/gpfsug-discuss_gpfsug.org/attachments/20190111/c9d4a7fb/attachment-0001.htm>