+1 Not updating the generation id when the effective deployment spec does not change is an oversight, which we missed here, but it can be fixed.
On Wed, Feb 1, 2023 at 2:07 AM Thomas Weise <t...@apache.org> wrote: > > Hi, > > Sorry for the late reply to this thread, but in the meantime we learned that > the assumption based on which the above mentioned change to upgradeMode was > approved does not hold true. The assumption was that the generation id in > spec metadata and reconciled spec can be used to determine if changes are > reconciled or not, rather than comparing the full specs. > > https://issues.apache.org/jira/browse/FLINK-30858 > > One way to get around the issue is to repeat the operator's spec diff logic > on the client side, but that introduces a tighter coupling with the operator > implementation than desirable. It would be good to fix the observed > generation update in the operator so that generation ids can be compared > safely, including the scenario where the diff is empty but the generation id > has increased. > > Thanks, > Thomas > > > On Mon, Jan 9, 2023 at 11:27 AM Maximilian Michels <m...@apache.org> wrote: >> >> +1 (binding) >> >> Thanks for clarifying. I wanted to make sure this is not an unintended >> regression. >> >> On Mon, Jan 9, 2023 at 4:26 PM Gyula Fóra <gyula.f...@gmail.com> wrote: >> >> > @Maximilian Michels <m...@apache.org> this is a completely intentional >> > improvement and it is required to ensure consistency for some operations >> > within the operator logic. >> > >> > On Mon, Jan 9, 2023 at 4:06 PM Maximilian Michels <m...@apache.org> wrote: >> > >> > > +0 >> > > >> > > 1. Downloaded the source archive release staged at >> > > >> > > >> > https://dist.apache.org/repos/dist/dev/flink/flink-kubernetes-operator-1.3.1-rc1/ >> > > 2. Verified the signature >> > > 3. Inspected the extracted source code for binaries >> > > 4. Compiled the source code >> > > 5. Verified license files / headers >> > > 6. Deployed to test environment >> > > >> > > I see an issue with (6), I noticed that if "upgradeMode" gets set to >> > > "last-state" for a fresh deployment, the `lastReconciledSpec` field >> > yields >> > > `stateless`. This is an issue when users compare the current spec to the >> > > lastReconciledSpec to assess whether the spec was reconciled. I suppose >> > > there are other means to ensure reconciliation, e.g. by looking at the >> > > generation id or similar. Just wanted to double check that this is what >> > we >> > > want. >> > > >> > > -Max >> > > >> > > On Wed, Jan 4, 2023 at 10:07 PM Hao t Chang <htch...@us.ibm.com> wrote: >> > > >> > > > I did the following: >> > > > Ran OLM bundle CI test suite for Kubernetes. >> > > > Generated and Deployed OLM bundle. >> > > > Created standalone/session jobs. >> > > > All Look good. Thanks for managing the release! >> > > > >> > > > -- >> > > > Best, >> > > > Ted Chang | Software Engineer | htch...@us.ibm.com >> > > > >> > > > >> > > > >> > > >> >
