Break generational cycle

[squaremo]

Adapted from commit msg:

For fluxctl install, we embed a set of templates in the fluxctl binary by generating a fake filesystem as Go code (pkg/install/generated_templates.gogen.go). But the code that generates the fake filesystem also depends on the fake filesystem, since it does double duty as a program for generating example manifests from those templates.

This leads to a problem: if anything has upset the generation process, it's not possible to regenerate the files, since the generation tool won't build.

A related problem, worked around elsewhere (#2473), is that it's easy to introduce spurious differences in the generated code -- which, since it's checked in -- means painful merges.

In general we don't care about the generated Go code, only about what it produces. This PR removes it from git and from the uncommitted change detection (make check-generated); and uses a new capability of fluxctl install to write to individual files to generate the example files under deploy/, removing the need for the code generating program to depend on the code it's generating.

[squaremo]

NB: there is one change to the example deployments in deploy/ as a result of this: the git label, email and user are set explicitly to example values rather than left to default, since we require them to be supplied to fluxctl install.

I think this is a good thing, but I have little idea of how deploy/ gets used nowadays -- @stefanprodan, do you think it is OK?

[stefanprodan]

@squaremo I see no problem with setting git label, email and user, but I would use flux-sync as the label default or better yet use sync-state=secret.

[squaremo]

@stefanprodan Thanks -- I'll change the label. I am reluctant to change too much about either fluxctl install or the example manifests here, so I'll leave the "better yet" to another PR.

[hiddeco]

The change is a great improvement for us, both in complexity and code reduction, and improved maintainability.

I am however wondering if there is any impact on eksctl and how they currently make use of the install package, as iirc, they make use of the embedded file that is no longer there.

[stefanprodan]

I think this breaks eksctl, wksctl, firekube and any other project that depends on the install package.

[squaremo]

any other project that depends on the install package

WHYYYYYYY
(did they do that)

[squaremo]

any other project that depends on the install package

WHYYYYYYY
(did they do that)

Oh yes, it was the motivation for the install package in the first place.

[2opremio]

I think it's OK to break the API (eksctl vendors it anyways) but Flux should still supply a way for eksctl to generate the files (ideally in memory and including the filenames)

[hiddeco]

@2opremio would it be possible for them to generate their own embedded file based on the templates they receive from vendoring the package?

[squaremo]

I can commit the generated file, and restore the change detection, while still breaking the dependency that was causing problems. (It's just a shame go doesn't let us avoid it)

[2opremio]

Yes, I guess that would be possible, but in practice that's pushing the problem to the consumer which I don't think it's ideal (particularly if it's not just eksctl which uses this functionality)

[squaremo]

that's pushing the problem to the consumer

Not really -- the consumer wouldn't have the circular dependency problem that's being fixed here.

However, it is fiddly to call go generate on an imported module (in fact, go will refuse to do so). So: I have pinched my nose and reinstated the generated file.

[hiddeco]

Thanks for double checking the timestamp issues @squaremo, LGTM 🥇