Hello fellow community members,
In our organization, we have developed, deployed and utilized an API-based MLOps pipeline using Azure DevOps.
The CI/CD pipeline has been developed and refined for about 18 months or so, and I have to say that it is pretty stable and we don't have any issues. Everything is seamless and works like charm!
We have MLOps Engineers within the team that want us to redo the CI/CD pipeline using either Databricks SDK or new Databricks standalone CLI. We discussed this within the team and still want to hear more arguments around if this is a good decision or not? Here are some argument bullet points:
Databricks API:
- We already have a working version using APIs. Why should we even switch? takes a lot of effort for no apparent gain
- Potentially most stable and up to date
- Better community support
- Language agnostic
- Gives us the most utility/flexibility
- Most integrability
- Con: Can potentially take longer to setup if updates are out. For example jobs API2.0. This argument is not really concerns me though. Took us only two days to develop, test and integrate with CI/CD pipeline. But an argument from the team.
Databricks SDK for Python:
- Simplicity!?
- Con: Limitation, we can always end up going back and utilizing APIs if we need something that is not supported.
- Con: community support. Not sure how good it is
- Con: Now team has to understand and maintain another fork of the CI/CD pipeline. Adds to tech debt.
- What else? pros or cons?
Databricks CLI (new):
- Not sure about this to be honest?
- Any arguments here?
Any suggestions and/or insights are helpful here.
Thank you
