Hi Hao, I think he's exactly talking about the usecase where the JM/TM restart and they come back up from the latest savepoint which might be stale by that time.
On Tue, 24 Sep 2019, 19:24 Hao Sun, <ha...@zendesk.com> wrote: > We always make a savepoint before we shutdown the job-cluster. So the > savepoint is always the latest. When we fix a bug or change the job graph, > it can resume well. > We only use checkpoints for unplanned downtime, e.g. K8S killed JM/TM, > uncaught exception, etc. > > Maybe I do not understand your use case well, I do not see a need to start > from checkpoint after a bug fix. > From what I know, currently you can use checkpoint as a savepoint as well > > Hao Sun > > > On Tue, Sep 24, 2019 at 7:48 AM Yuval Itzchakov <yuva...@gmail.com> wrote: > >> AFAIK there's currently nothing implemented to solve this problem, but >> working on a possible fix can be implemented on top of >> https://github.com/lyft/flinkk8soperator which already has a pretty >> fancy state machine for rolling upgrades. I'd love to be involved as this >> is an issue I've been thinking about as well. >> >> Yuval >> >> On Tue, Sep 24, 2019 at 5:02 PM Sean Hester <sean.hes...@bettercloud.com> >> wrote: >> >>> hi all--we've run into a gap (knowledge? design? tbd?) for our use cases >>> when deploying Flink jobs to start from savepoints using the job-cluster >>> mode in Kubernetes. >>> >>> we're running a ~15 different jobs, all in job-cluster mode, using a mix >>> of Flink 1.8.1 and 1.9.0, under GKE (Google Kubernetes Engine). these are >>> all long-running streaming jobs, all essentially acting as microservices. >>> we're using Helm charts to configure all of our deployments. >>> >>> we have a number of use cases where we want to restart jobs from a >>> savepoint to replay recent events, i.e. when we've enhanced the job logic >>> or fixed a bug. but after the deployment we want to have the job resume >>> it's "long-running" behavior, where any unplanned restarts resume from the >>> latest checkpoint. >>> >>> the issue we run into is that any obvious/standard/idiomatic Kubernetes >>> deployment includes the savepoint argument in the configuration. if the Job >>> Manager container(s) have an unplanned restart, when they come back up they >>> will start from the savepoint instead of resuming from the latest >>> checkpoint. everything is working as configured, but that's not exactly >>> what we want. we want the savepoint argument to be transient somehow (only >>> used during the initial deployment), but Kubernetes doesn't really support >>> the concept of transient configuration. >>> >>> i can see a couple of potential solutions that either involve custom >>> code in the jobs or custom logic in the container (i.e. a custom entrypoint >>> script that records that the configured savepoint has already been used in >>> a file on a persistent volume or GCS, and potentially when/why/by which >>> deployment). but these seem like unexpected and hacky solutions. before we >>> head down that road i wanted to ask: >>> >>> - is this is already a solved problem that i've missed? >>> - is this issue already on the community's radar? >>> >>> thanks in advance! >>> >>> -- >>> *Sean Hester* | Senior Staff Software Engineer | m. 404-828-0865 >>> 3525 Piedmont Rd. NE, Building 6, Suite 500, Atlanta, GA 30305 >>> <http://www.bettercloud.com> <http://www.bettercloud.com> >>> *Altitude 2019 in San Francisco | Sept. 23 - 25* >>> It’s not just an IT conference, it’s “a complete learning and networking >>> experience” >>> <https://altitude.bettercloud.com/?utm_source=gmail&utm_medium=signature&utm_campaign=2019-altitude> >>> >>> >> >> -- >> Best Regards, >> Yuval Itzchakov. >> >
