So, having spoken to Tim about this, and coming to understand the motivations a bit more, I think the "env var propagated everywhere, no exceptions" approach is the winner.
The deciding factor is that environ-config requires an API connection, and that (one of) the motivating use case(s) involves exposing incomplete CLI commands via feature flag (and I'm not aware of ones that can't be fulfilled by env vars). I understand that environ-config is superior in some other respects, but it's also more complex; and furthermore I'd like to avoid having multiple mechanisms for the same thing. The biggest risk is that people will start to use and depend upon feature flag behaviour, so I require that anyone using a feature flag make it abundantly clear that it's unsupported behaviour, by (1) including something like "DEV" in the var name and (2) logging its usage promiscuously, and making clear in those messages that the behaviour is not supported. Cheers William On Wed, Nov 26, 2014 at 4:50 PM, Casey Marshall <casey.marsh...@canonical.com> wrote: > +1 for feature flags in general and +1 for using environment variables > in upstart to get them to the servers & agents. > > I think it'd be nice to have an environment variable per flag, with a > common prefix JUJU_FEATURE_. That way, if you need to check one in a > package init(), you don't have to parse the whole list of flags to find > the one you care about -- or depend on a globally initialized parsing of > that list. > > env config seems reasonable, but dealing with parsing, errors and then > making that config globally available at init seems complex and not > always feasible. > > How about defining them "free form" in an env config field, which is > then used to emit the env vars as described above to the upstart config > during bootstrap? > > -Casey > > On 11/25/2014 10:16 PM, Ian Booth wrote: >> I like feature flags so am +1 to the overall proposal. I also agree with the >> approach to keep them immutable, given the stated goals and complexity >> associated with making them not so. >> >> I think the env variable implementation is ok too - this keeps everything >> very >> loosely coupled and avoids polluting a juju environment with an extra config >> attribute. >> >> On 26/11/14 08:47, Tim Penhey wrote: >>> Hi all, >>> >>> There are often times when we want to hook up features for testing that >>> we don't want exposed to the general user community. >>> >>> In the past we have hooked things up in master, then when the release >>> branch is made, we have had to go and change things there. This is a >>> terrible way to do it. >>> >>> Here is my proposal: >>> >>> http://reviews.vapour.ws/r/531/diff/# >>> >>> We have an environment variable called JUJU_FEATURE_FLAGS. It contains >>> comma delimited strings that are used as flags. >>> >>> The value is read when the program initializes and is not mutable. >>> >>> Simple checks can be used in the code: >>> >>> if featureflag.Enabled("foo") { >>> // do foo like things >>> } >>> >>> Thoughts and suggestions appreciated, but I don't want to have the >>> bike-shedding go on too long. >>> >>> Tim >>> >> > > > -- > Juju-dev mailing list > Juju-dev@lists.ubuntu.com > Modify settings or unsubscribe at: > https://lists.ubuntu.com/mailman/listinfo/juju-dev > -- Juju-dev mailing list Juju-dev@lists.ubuntu.com Modify settings or unsubscribe at: https://lists.ubuntu.com/mailman/listinfo/juju-dev
