Closed beeme1mr closed 4 months ago
This OFEP potentially addresses @tcarrio pluggable flagd OFEP in an alternative way.
I'm really not sold on defining this protocol as part of OpenFeature. I think it distracts from the main focus of OF - the spec and SDK - and I'm skeptical that it will get vendor support.
I'm not sure why we wouldn't just use the existing flagd protocol.
@moredip
I understand this perspective and shared it up until last week, but I think @oleg-nenashev in this thread helped change my view. As @thomaspoignant says here, there are limitations with the SDK+provider model. It works fantastic for first party apps that developers are interested in adding feature flag support to. It works less-well with self-hosted "off-the-shelf" applications (think Jenkins, Grafana, Keycloak Server) that could reasonably define sets of flags to be administered by whoever hosts them. Requiring these apps to maintain tens of plugins (providers) seems like a non-starter, compared to the relatively easy solution of allowing them to conform to this evaluation protocol, which could be made standard.
Practically, there's no actually that much work for us here. The flagd schema largely mirrors the SDK primitives. It could be made more neutral and generic, along the flagd providers, and we could ship both as "OpenFeature RPC" components, which anyone could build compliant backends for (also made easier by gRPC server generation).
I think this is a good initiative and reduced complexity when OF needs to be adopted by other systems. And I tried to visualize how this would fit into our SDKs and given below is my understanding visualization in a diagram,
flowchart TD
A[SDK] --> CTB(SDK Contrib)
CTB --> VA(Vendor A) --> Backend(Vendor Server)
CTB --> VB(Vendor B) --> ....
A--> W(Wire Protocol)
W --> IMPLA(Server A)
W --> PR(Proxy A)
Left side (in-process)
Right side (wire protocol)
Besides, I think both approaches will live side-by-side through OF SDK as they have their own benefits.
Just to re-iterate, I can (sort of) see the value of this for that open source/platform use case where the product that's having its feature flags managed might want to support multiple providers. However, I'm not a fan of adding this protocol as a core part of OpenFeature.
Firstly, I believe the use case is really a bit of a corner case. Is it really that important that a platform like Jenkins allows its internal feature flags to be managed by a variety of feature flagging frameworks? If you're just a consumer of the platform then the feature flag ruleset is really just going to be another type of configuration - you're not going to need dynamic evaluation or anything like that. It seems much more likely that Jenkins would just pick a specific fixed provider and ask its users to use that framework if they want to manage the flags.
Even if Jenkins could support multiple providers via this new capability, it would still add a bunch of complexity. Jenkins maintainers would need to provide an ongoing list of feature flag rulesets ("as of this release, flag foo_bar
should be on by default, and there's also a new flag baz_chirp
that should be off for now") AND they'd now be requiring their consumers to run a new standalone process for flagging (very likely flagd).
I don't know why we wouldn't just encourage a platform like Jenkins to just use flagd, and that's the end of it.
Practically, there's no actually that much work for us here. The flagd schema largely mirrors the SDK primitives. It could be made more neutral and generic, along the flagd providers, and we could ship both as "OpenFeature RPC" components, which anyone could build compliant backends for (also made easier by gRPC server generation).
agree it's not much work - really just ratifying some of the flagd work - but there's also an ongoing carrying cost, and more importantly it broadens the entire scope of the OpenFeature to include this wire protocol, which is more confusing for new people coming to OpenFeature.
If you're just a consumer of the platform then the feature flag ruleset is really just going to be another type of configuration - you're not going to need dynamic evaluation or anything like that
I see this differently. Configurations are a good start and this is the minimal setup of feature flags. But if we think about platform integrations, there is a need for dynamic configurability. This is the missing piece I see wire protocol can fulfill
It seems much more likely that Jenkins would just pick a specific fixed provider and ask its users to use that framework if they want to manage the flags
Wire protocol will avoid vendor locking, which is favorable for commercial/open-source platforms. Plus, this gives flexibility to adoptions so that they can use any feature flag vendor that supports the standard (more or less similar to OF SDKs)
AND they'd now be requiring their consumers to run a new standalone process for flagging (very likely flagd)
I don't think this is mandatory for everyone. Default values will be the standard operating behavior. Flagging service will be only needed for advanced use cases such as platform offerings (ex:- Jenkins as a service, K8s as a service). Since wire protocol is standardized, this can be a custom or vendor-backed implementation.
... broadens the entire scope of the OpenFeature to include this wire protocol, which is more confusing for new people coming to OpenFeature
Agree on the complexity - we will have to maintain the wire protocol and provide a reference implementation(flagd). But I don't think this will confuse newcomers to feature flagging. SDK will continue to support conventional in-process evaluations [1]. Rather than confusing, I think this will enable more use cases.
[1] - https://github.com/open-feature/ofep/pull/67#issuecomment-1615204626
I don't believe we need 100% consensus in all things, however @moredip has been key in the project since it's inception, and this does represent a sort of new direction for OpenFeature.
@oleg-nenashev Do you have any thoughts in response to @moredip 's comments here? And @Kavindu-Dodan 's responses here? I think these are both full of good points.
Some points worth cross-referencing from the community meeting on 6th July 2023 [1] [2]
flagd at the current state is a proxy exposing flags from different data sources. And this will require flagd to be a core component of each SDK and write convertors from other providers to flagd
Client libraries heavily rely on caching to give flag evaluation performance. Wire protocol may have potential performance issues as it requires a proxy/server request. Further, caching must not be a concern of wire protocol implementation as it could lead to other issues, hence it should be minimal.
Besides, we might also need to consider security for the SDK to proxy/server communication. This could be outsourced to a proxy but may be protocol will need to define a recommendation.
[1] - https://docs.google.com/document/d/1pp6t2giTcdEdVAri_2B1Z6Mv8mHhvtZT1AmkPV9K7xQ/edit#bookmark=id.9ltvj8qmp30v [2] - https://docs.google.com/document/d/1pp6t2giTcdEdVAri_2B1Z6Mv8mHhvtZT1AmkPV9K7xQ/edit#bookmark=id.le89k8eo7mrq
Sorry for my delay. I'm getting lost on this proposal w/r/t jenkins and how they need it. It seems like we're talking about replacing jenkins' existing configuration mechanism with open feature. I'm not convinced that's a targeted use-case.
That said, what the proposal is reading to me as is: "Hey, it sucks to have to write N different providers where N is the number of languages that I, $vendor, would like to support." Instead, we're offering them a simpler "implement this interface over http and we'll support you in basically every language that we write SDKs in. That feels like a valuable thing. Thoughts @moredip ?
Hey @justinabrahms, yes, I agree that's the immediate value. At the very least, it would consolidate the dev effort for Go Feature Flagd and flagd providers.
I included the Jenkins parts because @oleg-nenashev, a Jenkins maintainer, expressed interest in this functionality.
I am yet to review it in details, but I definitely express interest. Less for Jenkins, more for the ecosystem of small Golang/Rust projects that need to bundle something vendor neutral
This PR is stale because it has been open 30 days with no activity. Remove stale label or comment or this will be closed in next 60 days.
This PR is stale because it has been open 30 days with no activity. Remove stale label or comment or this will be closed in next 60 days.
This PR is stale because it has been open 30 days with no activity. Remove stale label or comment or this will be closed in next 60 days.
This PR is stale because it has been open 30 days with no activity. Remove stale label or comment or this will be closed in next 60 days.
@beeme1mr should we close this with the addition of the OFREP to the spec repo and it's own repo? Is the OFEP still needed for record-keeping purposes you think?
I think https://github.com/open-feature/protocol/blob/main/README.md may accomplish the same goals :shrug:
I think we can close this one, since open-feature/protocol is the place where things are happening.
To follow the evolution of the project please follow to the protocol repository.
This PR
Related Issues
Fixes #65