Skip to content

Latest commit

 

History

History
346 lines (258 loc) · 20.4 KB

FAQ.md

File metadata and controls

346 lines (258 loc) · 20.4 KB
Raw

FAQ

The first rule of FAQ is don't bikeshed the FAQ, leave that for Create structure for managing docs.

Please contribute to the FAQ! Found an explanation in an issue or pull request helpful? Summarize the question and quote the reply, linking back to the original comment.

Concepts

Configuration

Behavior

Best Practices

Concepts

Does dep replace go get?

No, dep is an experiment and is still in its infancy. Depending on how this experiment goes, it may be considered for inclusion in the go project in some form or another in the future but that is not guaranteed.

Here are some suggestions for when you could use dep or go get:

I would say that dep doesn't replace go get, but they both can do similar things. Here's how I use them:

go get: I want to download the source code for a go project so that I can work on it myself, or to install a tool. This clones the repo under GOPATH for all to use.

dep ensure: I have imported a new dependency in my code and want to download the dependency so I can start using it. My workflow is "add the import to the code, and then run dep ensure so that the manifest/lock/vendor are updated". This clones the repo under my project's vendor directory, and remembers the revision used so that everyone who works on my project is guaranteed to be using the same version of dependencies. -@carolynvs in #376

The long term vision is a sane, overall-consistent go tool. My general take is that go get is for people consuming Go code, and dep-family commands are for people developing it. -@sdboyer in #376

Why is it dep ensure instead of dep install?

Yeah, we went round and round on names. A lot.

The idea of "ensure" is roughly, "ensure that all my local states - code tree, manifest, lock, and vendor - are in sync with each other." When arguments are passed, it becomes "ensure this argument is satisfied, along with synchronization between all my local states."

We opted for this approach because we came to the conclusion that allowing the tool to perform partial work/exit in intermediate states ended up creating a tool that had more commands, had far more possible valid exit and input states, and was generally full of footguns. In this approach, the user has most of the same ultimate control, but exercises it differently (by modifying the code/manifest and re-running dep ensure). -@sdboyer in #371

What is a direct or transitive dependency?

  • Direct dependencies are dependencies that are imported directly by your project: they appear in at least one import statement from your project.
  • Transitive dependencies are the dependencies of your dependencies. Necessary to compile but are not directly used by your code.

Configuration

What is the difference between Gopkg.toml (the "manifest") and Gopkg.lock (the "lock")?

The manifest describes user intent, and the lock describes computed outputs. There's flexibility in manifests that isn't present in locks..., as the "branch": "master" constraint will match whatever revision master HAPPENS to be at right now, whereas the lock is nailed down to a specific revision.

This flexibility is important because it allows us to provide easy commands (e.g. dep ensure -update) that can manage an update process for you, within the constraints you specify, AND because it allows your project, when imported by someone else, to collaboratively specify the constraints for your own dependencies. -@sdboyer in #281

When should I use constraint, override, required, or ignored in Gopkg.toml?

  • Use constraint to constrain a direct dependency to a specific branch, version range, revision, or specify an alternate source such as a fork.
  • Use override to constrain a transitive dependency. See How do I constrain a transitive dependency's version? for more details on how overrides differ from constraints. Overrides should be used cautiously, sparingly, and temporarily.
  • Use required to explicitly add a dependency that is not imported directly or transitively, for example a development package used for code generation.
  • Use ignored to ignore a package and any of that package's unique dependencies.

Refer Gopkg.toml for detailed definitions of the above terms.

How do I constrain a transitive dependency's version?

First, if you're wondering about this because you're trying to keep the version of the transitive dependency from changing, then you're working against dep's design. The lock file, Gopkg.lock, will keep the selected version of the transitive dependency stable, unless you explicitly request an upgrade or it's impossible to find a solution without changing that version.

If that isn't your use case and you still need to constrain a transitive dependency, you have a couple of options:

  1. Make the transitive dependency a direct one, either with a dummy import or an entry in the required list in Gopkg.toml.
  2. Use an override.

Overrides are a sledgehammer, and should only be used as a last resort. While constraints and overrides are declared in the same way in Gopkg.toml, they behave differently:

  • Constraints:
    1. Can be declared by any project's manifest, yours or a dependency
    2. Apply only to direct dependencies of the project declaring the constraint
    3. Must not conflict with the constraint entries declared in any other project's manifest
  • Overrides:
    1. Are only utilized from the current/your project's manifest
    2. Apply globally, to direct and transitive dependencies
    3. Supersede constraints declared in all manifests, yours or a dependency's

Overrides are also discussed with some visuals in the gps docs.

Can I put the manifest and lock in the vendor directory?

No.

Placing these files inside vendor/ would concretely bind us to vendor/ in the long term. We prefer to treat the vendor/ as an implementation detail. -@sdboyer on go package management list

Behavior

How does dep decide what version of a dependency to use?

The full algorithm is complex, but the most important thing to understand is that dep tries versions in a certain order, checking to see a version is acceptable according to specified constraints.

  • All semver versions come first, and sort mostly according to the semver 2.0 spec, with one exception:
    • Semver versions with a prerelease are sorted after all non-prerelease semver. Within this subset they are sorted first by their numerical component, then lexicographically by their prerelease version.
  • The default branch(es) are next; the semantics of what "default branch" means are specific to the underlying source type, but this is generally what you'd get from a go get.
  • All other branches come next, sorted lexicographically.
  • All non-semver versions (tags) are next, sorted lexicographically.
  • Revisions, if any, are last, sorted lexicographically. Revisions do not typically appear in version lists, so the only invariant we maintain is determinism - deeper semantics, like chronology or topology, do not matter.

So, given a slice of the following versions:

  • Branch: master devel
  • Semver tags: v1.0.0 v1.1.0 v1.1.0-alpha1
  • Non-semver tags: footag
  • Revision: f6e74e8d Sorting for upgrade will result in the following slice.

[v1.1.0 v1.0.0 v1.1.0-alpha1 footag devel master f6e74e8d]

There are a number of factors that can eliminate a version from consideration, the simplest of which is that it doesn't match a constraint. But if you're trying to figure out why dep is doing what it does, understanding that its basic action is to attempt versions in this order should help you to reason about what's going on.

What external tools are supported?

During dep init configuration from other dependency managers is detected and imported, unless -skip-tools is specified.

The following tools are supported: glide and godep.

See #186 for how to add support for another tool.

Why is dep ignoring a version constraint in the manifest?

Only your project's directly imported dependencies are affected by a constraint entry in the manifest. Transitive dependencies are unaffected.

Use an override entry for transitive dependencies.

[[override]]
  name = "github.com/pkg/errors"
  version = "=0.8.0"

Why did dep use a different revision for package X instead of the revision in the lock file?

Sometimes the revision specified in the lock file is no longer valid. There are a few ways this can occur:

  • When you generated the lock file, you had an unpushed commit in your local copy of package X's repository in your GOPATH. (This case will be going away soon)
  • After generating the lock file, new commits were force pushed to package X's repository, causing the commit revision in your lock file to no longer exist.

To troubleshoot, you can revert dep's changes to your lock, and then run dep ensure -v -n. This retries the command in dry-run mode with verbose logs enabled. Check the output for a warning like the one below, indicating that a commit in the lock is no longer valid.

Unable to update checked out version: fatal: reference is not a tree: 4dfc6a8a7e15229398c0a018b6d7a078cccae9c8

The lock file represents a set of precise, typically immutable versions for the entire transitive closure of dependencies for a project. But "the project" can be, and is, decomposed into just a bunch of arguments to an algorithm. When those inputs change, the lock may need to change as well.

Under most circumstances, if those arguments don't change, then the lock remains fine and correct. You've hit one one of the few cases where that guarantee doesn't apply. The fact that you ran dep ensure and it DID a solve is a product of some arguments changing; that solving failed because this particular commit had become stale is a separate problem. -@sdboyer in #405

Why is dep slow?

There are two things that really slow dep down. One is unavoidable; for the other, we have a plan.

The unavoidable part is the initial clone. dep relies on a cache of local repositories (stored under $GOPATH/pkg/dep), which is populated on demand. Unfortunately, the first dep run, especially for a large project, may take a while, as all dependencies are cloned into the cache.

Fortunately, this is just an initial clone - pay it once, and you're done. The problem repeats itself a bit when you're running dep for the first time in a while and there's new changesets to fetch, but even then, these costs are only paid once per changeset.

The other part is the work of retrieving information about dependencies. There are three parts to this:

  1. Getting an up-to-date list of versions from the upstream source
  2. Reading the Gopkg.toml for a particular version out of the local cache
  3. Parsing the tree of packages for import statements at a particular version

The first requires one or more network calls; the second two usually mean something like a git checkout, and the third is a filesystem walk, plus loading and parsing .go files. All of these are expensive operations.

Fortunately, we can cache the second and third. And that cache can be permanent when keyed on an immutable identifier for the version - like a git commit SHA1 hash. The first is a bit trickier, but there are reasonable staleness tradeoffs we can consider to avoid the network entirely. There's an issue to implement persistent caching that's the gateway to all of these improvements.

There's another major performance issue that's much harder - the process of picking versions itself is an NP-complete problem in dep's current design. This is a much trickier problem 😜

How does dep handle symbolic links?

because we're not crazy people who delight in inviting chaos into our lives, we need to work within one GOPATH at a time. -@sdboyer in #247

Out of convenience, one might create a symlink to a directory within their GOPATH/src, e.g. ln -s ~/go/src/github.com/user/awesome-project ~/Code/awesome-project.

When dep is invoked with a project root that is a symlink, it will be resolved according to the following rules:

  • If the symlink is outside GOPATH and links to a directory within a GOPATH, or vice versa, then dep will choose whichever path is within GOPATH.
  • If the symlink is within a GOPATH and the resolved path is within a different GOPATH, then an error is thrown.
  • If both the symlink and the resolved path are in the same GOPATH, then an error is thrown.
  • If neither the symlink nor the resolved path are in a GOPATH, then an error is thrown.

This is the only symbolic link support that dep really intends to provide. In keeping with the general practices of the go tool, dep tends to either ignore symlinks (when walking) or copy the symlink itself, depending on the filesystem operation being performed.

dep deleted my files in the vendor directory!

If you just ran dep init, there should be a copy of your original vendor directory named _vendor-TIMESTAMP in your project root. The other commands do not make a backup before modifying the vendor directory.

dep assumes complete control of vendor/, and may indeed blow things away if it feels like it. -@peterbourgon in #206

Best Practices

Should I commit my vendor directory?

It's up to you:

Pros

  • it's the only way to get truly reproducible builds, as it guards against upstream renames and deletes
  • you don't need an extra dep ensure step (to fetch dependencies) on fresh clones to build your repo

Cons

  • your repo will be bigger, potentially a lot bigger
  • PR diffs are more annoying

How can I test changes to a dependency?

I would recommend against ever working in your vendor directory since dep will overwrite any changes. It’s too easy to lose work that way. -@carolynvs in #706

If you have a fork, add a [[constraint]] entry for the project in Gopkg.toml and set source to the fork source. This will ensure that dep will fetch the project from the fork instead of the original source. Otherwise, if you want to test changes locally, you can delete the package from vendor/ and make changes directly in GOPATH/src/*package* so that your changes are picked up by the go tool chain.

How do I roll releases that dep will be able to use?

In short: make sure you've committed your Gopkg.toml and Gopkg.lock, then just create a tag in your version control system and push it to the canonical location. dep is designed to work automatically with this sort of metadata from git, bzr, and hg.

It's strongly preferred that you use semver-compliant tag names. We hope to develop documentation soon that describes this more precisely, but in the meantime, the npm docs match our patterns pretty well.

What semver version should I use?

This can be a nuanced question, and the community is going to have to work out some accepted standards for how semver should be applied to Go projects. At the highest level, though, these are the rules:

  • Below v1.0.0, anything goes. Use these releases to figure out what you want your API to be.
  • Above v1.0.0, the general Go best practices continue to apply - don't make backwards-incompatible changes - exported identifiers can be added to, but not changed or removed.
  • If you must make a backwards-incompatible change, then bump the major version.

It's important to note that having a v1.0.0 does not preclude you from having alpha/beta/etc releases. The semver spec allows for prerelease versions, and dep is careful to not allow such versions unless Gopkg.toml contains a range constraint that explicitly includes prereleases: if there exists a version v1.0.1-alpha4, then the constraint >=1.0.0 will not match it, but >=1.0.1-alpha1 will.

Some work has been done towards a tool to that will analyze and compare your code with the last release, and suggest the next version you should use.

Is it OK to make backwards-incompatible changes now?

Yes. But.

dep will make it possible for the Go ecosystem to handle backwards-incompatible changes more gracefully. However, dep is not some magical panacea. Version and dependency management is hard, and dependency hell is real. The longstanding community wisdom about avoiding breaking changes remains important. Any v1.0.0 release should be accompanied by a plan for how to avoid future breaking API changes.

One good strategy may be to add to your API instead of changing it, deprecating old versions as you progress. Then, when the time is right, you can roll a new major version and clean out a bunch of deprecated symbols all at once.

Note that providing an incremental migration path across breaking changes (i.e., shims) is tricky, and something we don't have a good answer for yet.

My dependers don't use dep yet. What should I do?

For the most part, you needn't do anything differently.

The only possible issue is if your project is ever consumed as a library. If so, then you may want to be wary about committing your vendor/ directory, as it can cause problems. If your dependers are using dep, this is not a concern, as dep takes care of stripping out nested vendor directories.