Client-go: Switch to dep for dependency management

[lavalamp]

https://github.com/golang/dep

I.e., publish whatever dep needs instead of Godeps.json, and change documentation to tell people to download via dep ensure. (After testing to make sure it works--feel free to make a temporary branch if that helps testing.)

[caesarxuchao]

Switching to dep or glide will enable an alternative to this proposal on how to publish client-go.

We set up the manifest of dep (or glide) to always vendor the latest version of k8s.io/apimachinary. We set up the publishing robot to publish to k8s.io/apimachinary first, so that a real commit exists in the apimachinary repo; then the publishing robot copies staging/client-go, and run dep ensure k8s.io/apimachinary, which pulls down all the latest commits in k8s.io/apimachinary and overwrites the vendor/ folder, then the robot publishes to k8s.io/client-go.

Did I miss something? I'll experiment with this approach later tonight. @deads2k @lavalamp @stts.

[deads2k]

We set up the manifest of dep (or glide) to always vendor the latest version of k8s.io/apimachinary. We set up the publishing robot to publish to k8s.io/apimachinary first, so that a real commit exists in the apimachinary repo; then the publishing robot copies staging/client-go, and run dep ensure k8s.io/apimachinary, which pulls down all the latest commits in k8s.io/apimachinary and overwrites the vendor/ folder, then the robot publishes to k8s.io/client-go.

I think the idea is good. Assuming one pusher (the bot) or a coordinated pusher (one of us manually), this should guarantee a matched set assuming client-go has a similar sync script.

For branches, you'll need to restrict to a branch.

[caesarxuchao]

I don't think client-go having a similar sync script is a requirement for this idea to work. But apimachinery's sync script does some nice things like keeping the history, so I'll migrate client-go's script to yours.

I'll implement the necessary changes for the robot to implement the idea.

[sttts]

@caesarxuchao I like the idea. I don't see the connection though to golang/dep. We can do the same with Godeps now (if golang/dep support takes some time).

[caesarxuchao]

Yeah, I'm implementing it with Godeps now.

[caesarxuchao]

I manually experimented the idea, it's working. The robot is mostly done, but I stuck at running https://github.com/kubernetes/apimachinery/blob/master/hack/sync-from-kubernetes.sh, it keeps complaining fatal: Invalid revision range acda1e7401fbf79d89ba23df5e11d81f4a9d4e0d..HEAD at this line. git cannot find the old filter-branch-sha in the newly filtered branch. @sttts @deads2k could you help?

The only change I made to the script it fetching from https://github.com/kubernetes/kubernetes.git instead of [email protected]:kubernetes/kubernetes.git. [update] It also failed if fetching from the original ssh address. I ran the script from the repo root of k8s.io/apimacinery.

[deads2k]

I manually experimented the idea, it's working. The robot is mostly done, but I stuck at running https://github.com/kubernetes/apimachinery/blob/master/hack/sync-from-kubernetes.sh, it keeps complaining fatal: Invalid revision range acda1e7401fbf79d89ba23df5e11d81f4a9d4e0d..HEAD at this line. git cannot find the old filter-branch-sha in the newly filtered branch. @sttts @deads2k could you help?

The repo isn't a straight filter branch. The script as a for instance, changes the SHAs. So does the metadata about what it matches.

[caesarxuchao]

Sorry, I don't follow you. What should I do to run that script successfully?

[deads2k]

Sorry, I don't follow you. What should I do to run that script successfully?

Sorry, I misread what you were doing. My log on the kube-sync branch post-filter branch looks like this: https://gist.github.com/deads2k/7ce8f3a2dbcb564a8b369c782625800e What does yours look like?

[caesarxuchao]

Different: https://gist.github.com/caesarxuchao/9f7eec6d90c26698580a6eb1810ec5f4

[caesarxuchao]

The SHAs are different, commit messages are the same. Also, did you miss the lines after "d3c7cb5 Move first pkg/api/validation's into apimachinery"?

[deads2k]

The SHAs are different, commit messages are the same. Also, did you miss the lines after "d3c7cb5 Move first pkg/api/validation's into apimachinery"?

Yours handles the signed commits better. Do I have an old git maybe?

[caesarxuchao]

$ git version
git version 2.11.0.390.gc69c2f50cf-goog

And I'm on OSX 10.11.6

[deads2k]

Fedora

[deads@deads-dev-01 apimachinery]$ git version
git version 2.5.5

[deads2k]

well, this going to be a problem.

[deads2k]

Ok, found http://stackoverflow.com/questions/31552774/unexpected-behavior-with-git-filter-branch-and-signed-commits . Looks like mine is working correctly even if its ugly, since I changed the diff of the patch with filter-branch and logically the owner didn't sign it anymroe.

[caesarxuchao]

The robot is using git version 2.1.4, and it's producing the same git log output [edit] as my local git.

[sttts]

@caesarxuchao will you put most of the tooling about the godep update into the kube repo?

[deads2k]

So I have git version 2.5.5, chao has 2.11. 2.1 worked, 2.11 worked, 2.5.5 produces weird commits. We shouldn't rewrite history, but we can correct since the metadata is on disk. We should correct the filter-branch sha and use a working version of git moving forward.

[deads2k]

Alright, I've updated my git version, corrected the filter-branch-sha, and pushed a new sync. I did not rewrite history, so we've got some ugly commit messages we'll have to live with for all eternity.

[deads2k]

@lavalamp @smarterclayton @liggitt in the experiment you've created here: https://github.com/lavalamp/client-go-flat/, how would another library writer who generated his clients using our client generator against the apimachinery repo go about creating a compatible client? The signatures for anything that uses an apimachinery type like RESTClient https://github.com/kubernetes/kubernetes/blob/master/staging/src/k8s.io/client-go/rest/client.go#L61-L89 will be incompatible.

As we add aggregated API servers, we expect this to be more common, not less, correct?

[lavalamp]

Yes, my experiment is aimed at people who don't want to think about
vendoring. People who want to write an apiserver will have to vendor.

On Thu, Feb 2, 2017 at 7:11 AM, David Eads [e-mail] wrote:

@lavalamp https://github.com/lavalamp @smarterclayton
https://github.com/smarterclayton @liggitt https://github.com/liggitt
in the experiment you've created here: https://github.com/lavalamp/
client-go-flat/, how would another library writer who generated his
clients using our client generator against the apimachinery repo go about
creating a compatible client? The signatures for anything that uses an
apimachinery type like RESTClient https://github.com/kubernetes/
kubernetes/blob/master/staging/src/k8s.io/client-go/rest/client.go#L61-L89
will be incompatible.

As we add aggregated API servers, we expect this to be more common, not
less, correct?

—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
https://github.com/kubernetes/client-go/issues/78#issuecomment-276983744,
or mute the thread
https://github.com/notifications/unsubscribe-auth/AAnglpZFph6zIgxSmrhJmrHoqgCFPS-rks5rYfIggaJpZM4Ls_Zy
.

[deads2k]

Yes, my experiment is aimed at people who don't want to think about
vendoring. People who want to write an apiserver will have to vendor.

But at the same time, the experiment makes it impossible for someone writing their API server to provide a client of their own for people who don't to think about it. Basically, it means that no aggregated API server will be able to have my-aggregated-apiserver/client-go because the types we generate for them are incompatible.

[caesarxuchao]

Sent https://github.com/kubernetes/kubernetes/pull/40909 and https://github.com/kubernetes/test-infra/pull/1784 to implement the idea.

[ericchiang]

Some issues that may be blocking:

• [ ] Dep doesn't remove nested vendor directories https://github.com/golang/dep/issues/120
• [ ] Dep can't detect transitive dependencies of client-go. Need to either:
  
  
  
  • [ ] Publish a dep manifest once that format is stable https://github.com/golang/dep/issues/276
  
  
  • [ ] Wait for glide to support detecting other vendor tool manifests https://github.com/golang/dep/issues/380
  
  

Anything else anyone's hit?

[sdboyer]

Dep doesn't remove nested vendor directories

It does. (It doesn't provide an option not to, as otherwise it couldn't guarantee correct compilation).

golang/dep#120 is about pruning _additional_ directories - e.g., unused packages.

Dep can't detect transitive dependencies of client-go. Need to either:

This shouldn't be the case - if it's in the import graph, direct or transitive, then it should be picked up without issue. If that's not the case, please please, open an issue against dep!

If your goal here is applying constraints to transitive dependencies, we've got a Q about that in the FAQ. 😄

[caesarxuchao]

This shouldn't be the case - if it's in the import graph, direct or transitive, then it should be picked up without issue. If that's not the case, please please, open an issue against dep!

client-go only have godep.json, which is not recognized by dep, so dep just downloads the HEAD of the transitive dependencies of client-go.

[sdboyer]

client-go only have godep.json, which is not recognized by dep

Ah! I see what you meant now. There's still a few things this could mean, but if the concern is just the initial setup of dep, then you can achieve the same effect by godep restore-ing onto your GOPATH: dep init reads the state of your GOPATH and tries to construct a constraint set from there.

For more direct integration, I see you already found golang/dep#222; the main issue for it is golang/dep#186. Honestly, godep integration would be pretty easy to write 😄

dep just downloads the HEAD of the transitive dependencies of client-go.

A bit more sophisticated than that - it tries a versions in this order.

[caesarxuchao]

For more direct integration, I see you already found golang/dep#222; the main issue for it is golang/dep#186. Honestly, godep integration would be pretty easy to write :smile:

Cool. Is there an ETA?

A bit more sophisticated than that - it tries a versions in this order.

Thanks! Good to know.

[sdboyer]

Cool. Is there an ETA?

The main priority right now is stabilizing the file formats, so I haven't put a timeframe on it yet. But stabilizing the CLI interface - golang/dep#277 - is the next major priority in the roadmap, and entails doing this. If current plans hold, we'll turn our focus to that mid-May. I suspect that godep support will be high on the priority list. So, maybe...a month? That's totally not firm, of course.

That said, progress is already happening on stabilizing the commands. A contributor has just started in on golang/dep#496, which lays some necessary foundation in dep init for this. So, it could also be sooner.

[ahmetb]

@sttts do you have any updates from https://github.com/golang/dep/issues/1124? It would be awesome to be able to use client-go with dep.

[sttts]

@ahmetb no update. But we can archieve 90% of that without transitive dependencies. That's the plan right now, targetting to have that for 1.9.

[ahmetb]

dep ensure -v -add k8s.io/client-go@~5.0.1 just worked for me. I don’t know if it's a point-in-time coincidence.

[fejta-bot]

Issues go stale after 90d of inactivity.
Mark the issue as fresh with /remove-lifecycle stale.
Stale issues rot after an additional 30d of inactivity and eventually close.

If this issue is safe to close now please do so with /close.

Send feedback to sig-testing, kubernetes/test-infra and/or fejta.
/lifecycle stale

[sathieu]

Hi,
Any news on this?

[sttts]

Any news on this?

Yes, we are actively working on creating a Gopkg.toml inside of client-go during publishing. Some more patience please. We finally make progress.

[fejta-bot]

Stale issues rot after 30d of inactivity.
Mark the issue as fresh with /remove-lifecycle rotten.
Rotten issues close after an additional 30d of inactivity.

If this issue is safe to close now please do so with /close.

Send feedback to sig-testing, kubernetes/test-infra and/or fejta.
/lifecycle rotten
/remove-lifecycle stale

[sttts]

/remove-lifecycle rotten

[AkihiroSuda]

Any update?

[sttts]

Yes, the publishing-bot can publish Gopkg.toml (https://github.com/kubernetes/publishing-bot/pull/55), but we are blocked on dep.

[ryanwalls]

If you get to this issue and want to follow the workaround of specifying overrides for all transitive dependencies, here are the overrides for release 7.0. https://gist.github.com/ryanwalls/96ffc7a218c05cfdfcdecdaf413b9e13

• Update: dep doesn't like the list having non-root projects... wait one while I fix it.
• Update 2: Okay, the gist has the dependency overrides that actually worked.

[lvangool]

Hmm - that gist is not working for me - I'm seeing:

$ dep ensure
Solving failure: No versions of k8s.io/client-go met constraints:
v7.0.0: Could not introduce k8s.io/[email protected] due to multiple problematic subpackages:
Subpackage k8s.io/client-go/pkg/api/v1 is missing. (Package is required by (root).) 
Subpackage k8s.io/client-go/pkg/apis/batch/v1 is missing. (Package is required by (root).)
Subpackage k8s.io/client-go/pkg/apis/extensions/v1beta1 is missing. (Package is required by (root).)
v6.0.0: Could not introduce k8s.io/[email protected], as it is not allowed by constraint ^7.0.0 from project ...
... 

Having so much pain with go-client deps!

[lavalamp]

So, it seems the go team mostly considers dep a failed experiment. We'll have to continue to live in a world of pain until https://github.com/golang/go/issues/24301 rolls out and we figure out how to start publishing packages that play nice.

[lavalamp]

I filed https://github.com/kubernetes/kubernetes/issues/63607 to call more attention to this.

[briansmith]

So, it seems the go team mostly considers dep a failed experiment.

I agree.

We'll have to continue to live in a world of pain until golang/go#24301 rolls out and we figure out how to start publishing packages that play nice.

I don't agree with this. It would be very good for client-go to be made to work well with dep-based build systems in the interim. We shouldn't underestimate how long it will take for vgo to become a real thing; it seems very likely it will take a long time to happen, and it seems pretty likely it will be abandoned as another "failed experiment."

It is seriously painful to upgrade client-go in dep-based projects so projects like the ones I work on avoid doing so unless/until absolutely necessary. This is bad and is worth fixing.

[briansmith]

At the very least, in the event of a security vulnerability that gets fixed by releasing new version(s) of client-go, it is important to provide the workarounds needed to get dep-based software to be able to do the upgrade at the same time the new release is made. Otherwise lots of software won't be able to consume the fix in a reasonable amount of time.

[sdboyer]

that is indeed Russ' opinion. I have a number of thoughts about what k8s can do that will roll sanely into the future. unfortunately, all my time is right now is taken up by working on vgo response blog posts 😢

imma be publishing the first of those this week, though, so my time will start unblocking soon.

[fejta-bot]

Issues go stale after 90d of inactivity.
Mark the issue as fresh with /remove-lifecycle stale.
Stale issues rot after an additional 30d of inactivity and eventually close.

If this issue is safe to close now please do so with /close.

Send feedback to sig-testing, kubernetes/test-infra and/or fejta.
/lifecycle stale

[fejta-bot]

Stale issues rot after 30d of inactivity.
Mark the issue as fresh with /remove-lifecycle rotten.
Rotten issues close after an additional 30d of inactivity.

If this issue is safe to close now please do so with /close.

Send feedback to sig-testing, kubernetes/test-infra and/or fejta.
/lifecycle rotten

[m1o1]

Any update here? Has client-go abandoned dep (even in the interim until vgo is stable)?

[fejta-bot]

Rotten issues close after 30d of inactivity.
Reopen the issue with /reopen.
Mark the issue as fresh with /remove-lifecycle rotten.

Send feedback to sig-testing, kubernetes/test-infra and/or fejta.
/close

[k8s-ci-robot]

@fejta-bot: Closing this issue.

In response to this:

Rotten issues close after 30d of inactivity.
Reopen the issue with /reopen.
Mark the issue as fresh with /remove-lifecycle rotten.

Send feedback to sig-testing, kubernetes/test-infra and/or fejta.
/close

Instructions for interacting with me using PR comments are available here. If you have questions or suggestions related to my behavior, please file an issue against the kubernetes/test-infra repository.