open-feature/open-feature-operator

[feature] Support in-process evaluator in OFO

toddbaert opened this issue · 9 comments

toddbaert commented

We now have JS, Java, and Go in-process flagd providers. These providers connect directly to a sync.proto-compliant gRPC server (such as flagd-proxy). OFO should be enhanced to support the use of these providers. This would require:

  • ability to specify this in APIs
  • OFO sets the corresponding env-var which configures the provider to use in-process evaluation (see flagd-injector), FLAGD_SOURCE_RESOLVER - the SA injection/patch should NOT take place.
    • host/port are configured similarly, through ENV_VARS
  • flagd is not deployed as a sidecar to the associated workload, since it's not needed
  • associated tests, including KUTTL tests
  • associated documentation

cc @odubajDT @thisthat @beeme1mr @Kavindu-Dodan

odubajDT commented

After some discussions, it would be beneficial to do a refactoring of the existing FeatureFlagSource CRD structure to be able to distinguish which parameters are set for RPC or in-process evaluation. We agreed to draft a PoC with the potential solution (with bumping the API version to v1beta2) also containing the conversion webhooks, so users running v1beta1 versions won't need to adapt their resources in case they want to use only RPC evaluation.

odubajDT commented

Implemented a PoC supporting the in-process evaluation and additionally refactoring the actual API.

During the implementation, I got the feeling that it would be possible to move the in-process evaluation parameters into a new separate CRD to decrease the complexity of FeatureFlagSource CRD and therefore reducing the need to introduce conversion webhooks and new API version.

Also it can potentially bring the advantage of easier testability of the code as well as decreasing the complexity of the mutating webhook logic (especially the part handling the decision if the flagd sidecar injection should happen).

On the other side, it requires introducing a new standardized annotation, which will be used by deployments using OFO (like openfeature.dev/featureflagsource) for in-process evaluation

Open for any opinions!

@toddbaert @thisthat @bacherfl @Kavindu-Dodan @beeme1mr

  • 👍1
bacherfl commented

In general I also have the impression that the current FeatureFlagSource is handling too many different responsibilities, so if we can extract something into a separate CRD i think that would be beneficial

  • 👍1
Kavindu-Dodan commented

@odubajDT what you are proposing is to keep FeatureFlagSource untouched and introduce a new CRD dedicated for in-process mode isn't it ?

I like the idea. We should also consider the ability to utilize flagd's grpc streaming capability 1 which was added recently (@toddbaert we could refine the issue on this).

Footnotes

  1. ↩
  • 👍1
odubajDT commented

@odubajDT what you are proposing is to keep FeatureFlagSource untouched and introduce a new CRD dedicated for in-process mode isn't it ?

Yes exactly!

I like the idea. We should also consider the ability to utilize flagd's grpc streaming capability 1 which was added recently (@toddbaert we could refine the issue on this).

Footnotes

1. * https://flagd.dev/reference/grpc-sync-service/
   
   [↩](#user-content-fnref-1-59ee0d967e0e1eb93c01c9f33d6dfa32)
odubajDT commented

Drafted a PoC to showcase how the possible solution with a separate CRD can look like. It shows that using a separate CRD makes the code way cleaner and reduces the complexity of the whole solution

odubajDT commented

If we decide to go with the new CRD, we should consider picking a new name. My suggestion would be: FeatureFlagInProcessConfiguration. Any other ideas?

@toddbaert @Kavindu-Dodan @beeme1mr @thisthat @bacherfl

Kavindu-Dodan commented

@odubajDT FeatureFlagInProcessConfiguration sounds good (but please wait for feedback from others)

Besides, I assume the workflow here is you add FeatureFlagInProcessConfiguration CRD to your workload and the operator sets environment variables to the deployment. I think we need to check lang specific flagd in-proces provider capability to swithc processing mode (remote vs in-process) based on env configuration 🤔 (cc @toddbaert to be followed up if we don't do this already)

  • 👍2
thisthat commented

Regarding the naming, we would have the following then:

  • FeatureFlag
  • FeatureFlagSource
  • FeatureFlagInProcessConfiguration

I think it is consistent and explains the purpose clearly enough.

  • 👍2