1. Jobs on k8s should have dependency relation.
   A job will run if and only the state of it's dependency job is Complete.
   A job can be configurated without CRON if it has an dependency job.

There was an idea for that, but was rejected eventually from core. You'll have to use some 3-rd party solution to provide such functionality or write your own controller on top of TPR. You can sync with @sdminonne, he's interested in this as well

Support CRON syntax, CronJob can do this.

Already available.

Support configuration a maximum number of failures backoff policy.

For now, RestartPolicy=Never will start a new pod, #30243 also discusses this.

Yup, #30243 is the way to solve this.

We can add the ETL jobs with kubectl create -f xxx.yaml, at the same time, apiserver will check whether a loop will be formed.

Not sure what you mean with this, so I'm leaving this open until that gets cleared out.

@Yancey1989 if you have this kind of job (please confirm) and you want to state the dependency in your manifest files you may need a custom controller with built on top of Third Party Resources (TPR). My company has some interested in this kind of tool and we're trying to build one even if very simple. Don't hesitate to contact me (on slack) if you need more info.

@soltysh @sdminonne , thanks for your reply.

We can add the ETL jobs with kubectl create -f xxx.yaml, at the same time, apiserver will check whether a loop will be formed.

This mean, we assume each job is a vertex, each dependency relation is an edge. The graph consisting of these vertices and edges can not have a ring, otherwise these jobs will fall into deadlock.

@sdminonne The repo is very helpful for us, I'll have a thorough understanding of this, and maybe contact you on slack, thank you! :)

Since workflow is not part of k8s and other problems have already appropriate issues, I'm closing this one.