moritzkiefer-da
commented
2 months ago What if services are split into two, "init" and "serve", where "init" would be exposed as soon as container live and "serve" would be exposed only when serving port is ready? This both ports would be accessible when they are needed and the app would only become ready when it's initialized and ready to serve traffic?
They already are. The "init" API is on port 5009, the "serve" API is on port 5008. But a readiness probe is per pod not per service so this doesn't help. You could overwrite the k8s behavior that it doesn't route to non-ready ports but imho that is more sketchy than useful. Or we could try to split it into two separate pods but that is a very large undertaking and isn't going to happen anytime soon.
stas-sbi
While having a comprehensive monitoring is a good thing I believe that obvious problems should manifest themselves in obvious ways. More sophisticated tooling comes especially useful when the problem is already detected and needs further investigation. This approach does not only help during initial setup but also during upgrades, experiments, when an extensive monitoring is not setup yet or when it (extensive monitoring) just fails for whatever reason.
Some of SV components (such as sequencer) become ready and get exposed on a service endpoints before the port which is supposed to serve the traffic opens. In case of sequencer this is done to make sure that another port becomes accessible before sequencer is initialized which is used to initialize sequencer.
What if services are split into two, "init" and "serve", where "init" would be exposed as soon as container live and "serve" would be exposed only when serving port is ready? This both ports would be accessible when they are needed and the app would only become ready when it's initialized and ready to serve traffic?