The vendor dir (here called "dir") is already joined with the CWD
in installCommand(). it should not be joined again here.
Also adds a logging message to show that a local package was
installed.
Allows `installing` a package without pulling dependencies of that
package. This comes in handy when having multiple directories on
`JPATH`, which overlay each other.
Because dependencies may be already present at lesser ranked import
paths, it's not required to pull them, when only needing to bump a
single one.
* feat: go-like import style
jb now creates a directory structure inside of vendor/ that is similar to how go
does (github.com/grafana/jsonnet-libs). This is reflected in the final import
paths, which means they will be go-like
* refactor(spec/deps): named regexs
* feat: make goImportStyle configurable
Defaults to off, can be enabled in `jsonnetfile.json`
* fix: integration test
* doc: license headers
* fix(deps): remove GO_IMPORT_STYLE
not an option anymore, will always do so and symlink
* feat: symlink to legacy location
* feat: allow to disable legacy links
* fix(test): legacyImports in integration tests
* fix(spec): test
* fix: respect legacyName aliases
It was possible to alias packages by changing `name` previously.
While names are now absolute (and computed), legacy links should still respect
old aliases to avoid breaking code.
* fix(test): integration
* fix(init): keep legacyImports enabled for now
* feat: rewrite imports
adds a command to automatically rewrite imports from legacy to absolute style
* fix(tool): rewrite confused by prefixing packages
When a package was a prefix of another one, it broke.
Fixed that by using a proper regular expression. Added a test to make sure it
works as expected
* Update cmd/jb/init.go
* fix: exclude local packages from legacy linking
They actually still use the old style, which is fine. LegacyLinking
messed them up, but from now on it just ignores symlinks that match a localPackage.
Adding a dep, or updating a dependency version makes writes to the
jsonnet files.
We evaluate the changes on each of the files. An empty jsonnetfile.json
does not create a corresponding lockfile, as a missing lockfile is not
different from its previous (non existent).
- allow to install pkg that is already locked
- clean unknown files from vendor
- correctly handle checksums and locked versions (was accidentally ignoring
- these before)
Packages are unique anyways so it makes sense to use a map to avoid having
duplicates.
For compatibility reasons custom json (un)marshallers hide this change from the
end user
rewrites the installation of packages from scratch to solve several issues with
the existing implementation:
- does not need to choose between lockfile and jsonnetfile anymore. The
jsonnetfile what to be installed, while the lockfile also has versions and
checksums of all packages, even nested ones.
- the lockfile is regenerated on every run, preserving the locked values
- downloaded packages are hashed using sha256 to make sure we receive what we
expect. If files on the local disk are modified, they are downloaded again.
So far, `pkg` and `pkg/jsonnetfile` had overlapping functionality when it came
to choosing and loading jsonnetfiles.
This fully switches to the separate package `pkg/jsonnetfile` that seems to be
created for exactly this purpose
Not using an undocumented ETag header from the GitHub archive API is
probably for the best. This is slightly slower due to extra round-trip,
but it is still much faster than cloning the repository to resolve the
ref.
Signed-off-by: Benoit Gagnon <benoit.gagnon@ubisoft.com>
If the server supports it, fetch a specific
revision with --depth 1. Otherwise, fall back
to the normal fetch.
This replaces the previous "clone" operation. The bandwidth and time savings
can be significant depending on the history
of the repository (number of commits).
