Thanks for the response, Is there a limit on how long I can run in mixed mode? Let's say if datacenter 1 is upgraded and upgradesstables was run on day 1 and datacenter 3 is upgraded and upgradesstables runs on day 10. Is that going to be a big concern?
> 2 might be strictly safer if you trust internode mixed mode AND have a way to fail out of a dc and rebuild it without violating consistency. I tested the mixed mode and it works, but are there any catches that won't work? I am okay to disable repair during this time. On Wed, May 19, 2021 at 6:02 PM Jeff Jirsa <jji...@gmail.com> wrote: > > > > On May 19, 2021, at 5:15 PM, Jai Bheemsen Rao Dhanwada < > jaibheem...@gmail.com> wrote: > > > > > > Hello, > > > > I am trying to get some suggestions on running upgradesstables(for a > major version upgrade where there is a SSTable format change). > > > > The Cassandra cluster is set up in 3 datacenters with 18 nodes in each > region and here are the 3 options that I am evaluating. > > > > 1. Run "nodetool upgradesstables" after every Cassandra node is upgraded > to the new version. > > 2. Upgrade cassandra version on one of the data centers, run "nodetool > upgradesstables" and proceed to the next datacenter and so on. > > 3. Upgrade cassandra version on all the datacenters and run "nodetool > upgradesstables" at the end on all the data centers. > > > > I am more interested in understanding option-2 where I upgrade the > cassandra version and run upgradesstables for one data center and then > proceed to the next datacenter and so on. Are there any disadvantages with > this option? > > > > The difference between 2 and 3 is how long you stay in mixed mode > > In 2, you could theoretically roll back the whole DC if you detect a > failure and the app can fail out of that dc. It’d be hard but possible. > > In 3, you minimize mixed mode (internode), and the sstables largely > upgrade themselves in high churn clusters through compaction. > > In theory 3 is probably faster, if you disable repair in mixed mode and > until sstables are upgraded, 3 gets you done “faster” so repair can run > again. 2 might be strictly safer if you trust internode mixed mode AND have > a way to fail out of a dc and rebuild it without violating consistency. > > > > >
