154 KiB
v2.8.3 (2015-04-15):
TWO SMALL GIT TWEAKS
This is the last of a set of releases intended to ensure npm's git support is robust enough that we can stop working on it for a while. These fixes are small, but prevent a common crasher and clear up one of the more confusing error messages coming out of npm when working with repositories hosted on git.
387f889
#7961 Ensure that hosted git SSH URLs always have a valid protocol when stored inresolved
fields innpm-shrinkwrap.json
. (@othiym23)394c2f5
Switch the order in which hosted Git providers are checked togit:
,git+https:
, thengit+ssh:
(fromgit:
,git+ssh:
, thengit+https:
) in an effort to go from most to least likely to succeed, to make for less confusing error message. (@othiym23)
v2.8.2 (2015-04-14):
PEACE IN OUR TIME
npm has been having an issue with CouchDB's web server since the release
of io.js and Node.js 0.12.0 that has consumed a huge amount of my time
to little visible effect. Sam Mikes picked up the thread from me, and
after a lot of effort
figured out that ultimately there are probably a couple problems with
the new HTTP Agent keep-alive handling in new versions of Node. In
addition, npm-registry-client
was gratuitously sending a body along
with a GET request which was triggering the bugs. Sam removed about 10 bytes from
one file in npm-registry-client
, and this problem, which has been bugging us for months,
completely went away.
In conclusion, Sam Mikes is great, and anybody using a private registry hosted on CouchDB should thank him for his hard work. Also, thanks to the community at large for pitching in on this bug, which has been around for months now.
431c3bf
#7699npm-registry-client@6.3.2
: Don't send body with HTTP GET requests when logging in. (@smikes)
v2.8.1 (2015-04-12):
CORRECTION: NPM'S GIT INTEGRATION IS DOING OKAY
A helpful bug report
led to another round of changes to
hosted-git-info
,
some additional test-writing, and a bunch of hands-on testing against actual
private repositories. While the complexity of npm's git dependency handling is
nearly fractal (because npm is very complex, and git is even more complex),
it's feeling way more solid than it has for a while. We think this is a
substantial improvement over what we had before, so give npm@2.8.1
a shot if
you have particularly complex git use cases and
let us know how it goes.
(NOTE: These changes mostly affect cloning and saving references to packages hosted in git repositories, and don't address some known issues with things like lifecycle scripts not being run on npm dependencies. Work continues on other issues that affect parity between git and npm registry packages.)
66377c6
#7872hosted-git-info@2.1.2
: Pass through credentials embedded in SSH and HTTPs git URLs. (@othiym23)15efe12
#7872 Use the new version ofhosted-git-info
to pass along credentials embedded in git URLs. Test it. Test it a lot. (@othiym23)
SCOPED DEPENDENCIES AND PEER DEPENDENCIES: NOT QUITE REESE'S
Big thanks to @ewie for identifying an issue with
how npm was handling peerDependencies
that were implicitly installed from the
package.json
files of scoped dependencies. This
will be a moot point
with the release of npm@3
, but until then, it's important that
peerDependency
auto-installation work as expected.
b027319
#7920 Scoped packages withpeerDependencies
were installing thepeerDependencies
into the wrong directory. (@ewie)649e31a
#7920 TestpeerDependency
installs involving scoped packages usingnpm-package-arg
instead of simple path tests, for consistency. (@othiym23)
MAKING IT EASIER TO WRITE NPM TESTS, VERSION 0.0.1
@iarna and I
(@othiym23) have been discussing a
candidate plan
for improving npm's test suite, with the goal of making it easier for new
contributors to get involved with npm by reducing the learning curve
necessary to be able to write good tests for proposed changes. This is the
first substantial piece of that effort. Here's what the commit message for
ed7e249
had to say about this work:
It's too difficult for npm contributors to figure out what the conventional style is for tests. Part of the problem is that the documentation in CONTRIBUTING.md is inadequate, but another important factor is that the tests themselves are written in a variety of styles. One of the most notable examples of this is the fact that many tests use fixture directories to store precooked test scenarios and package.json files.
This had some negative consequences:
- tests weren't idempotent
- subtle dependencies between tests existed
- new tests get written in this deprecated style because it's not obvious that the style is out of favor
- it's hard to figure out why a lot of those directories existed, because they served a variety of purposes, so it was difficult to tell when it was safe to remove them
All in all, the fixture directories were a major source of technical debt, and cleaning them up, while time-consuming, makes the whole test suite much more approachable, and makes it more likely that new tests written by outside contributors will follow a conventional style. To support that, all of the tests touched by this changed were cleaned up to pass the
standard
style checker.
And here's a little extra context from a comment I left on #7929:
One of the other things that encouraged me was looking at this presentation on technical debt from Pycon 2015, especially slide 53, which I interpreted in terms of difficulty getting new contributors to submit patches to an OSS project like npm. npm has a long ways to go, but I feel good about this change.
ed7e249
#7929 Eliminate fixture directories fromtest/tap
, leaving each test self-contained. (@othiym23)4928d30
#7929 Move fixture files fromtest/tap/*
totest/fixtures
. (@othiym23)e925deb
#7929 Tweak the run scripts to stop slaughtering the CPU on doc rebuild. (@othiym23)65bf7cf
#7923 Use an alias of scripts and run-scripts innpm run test-all
(@watilde)756a3fb
#7923 Sync timeout time ofnpm run-script test-all
to be the same astest
andtap
scripts. (@watilde)8299b5f
Set a timeout for tap tests fornpm run-script test-all
. (@othiym23)
THE EVER-BEATING DRUM OF DEPENDENCY UPDATES
d90d0b9
#7924 Removechild-process-close
, as it was included for Node 0.6 compatibility, and npm no longer supports 0.6. (@robertkowalski)16427c1
lru-cache@2.5.2
: More accurate updating of expiry times whenmaxAge
is set. (@isaacs)03cce83
nock@1.6.0
: Mocked network error handling. (@pgte)f93b1f0
glob@5.0.5
: Usepath-is-absolute
polyfill, allowing newer Node.js and io.js versions to usepath.isAbsolute()
. (@sindresorhus)a70d694
request@2.55.0
: Bug fixes and simplification. (@simov)2aecc6f
columnify@1.5.1
: Switch to using babel from 6to5. (@timoxley)
v2.8.0 (2015-04-09):
WE WILL NEVER BE DONE FIXING NPM'S GIT SUPPORT
If you look at the last release's release
notes,
you will note that they confidently assert that it's perfectly OK to force all
GitHub URLs through the same git:
-> git+ssh:
fallback flow for cloning. It
turns out that many users depend on git+https:
URLs in their build
environments because they use GitHub auth tokens instead of SSH keys. Also, in
some cases you just want to be able to explicitly say how a given dependency
should be cloned from GitHub.
Because of the way we resolved the inconsistency in GitHub shorthand handling
before, this
turned out to be difficult to work around. So instead of hacking around it, we
completely redid how git is handled within npm and its attendant packages.
Again. This time, we changed things so that normalize-package-data
and
read-package-json
leave more of the git logic to npm itself, which makes
handling shorthand syntax consistently much easier, and also allows users to
resume using explicit, fully-qualified git URLs without npm messing with them.
Here's a summary of what's changed:
- Instead of converting the GitHub shorthand syntax to a
git+ssh:
,git:
, orgit+https:
URL and saving that, save the shorthand itself topackage.json
. - If presented with shortcuts, try cloning via the git protocol, SSH, and HTTPS (in that order).
- No longer prompt for credentials -- it didn't work right with the spinner,
and wasn't guaranteed to work anyway. We may experiment with doing this a
better way in the future. Users can override this by setting
GIT_ASKPASS
in their environment if they want to experiment with interactive cloning, but should also set--no-spin
on the npm command line (or runnpm config set spin=false
). - EXPERIMENTAL FEATURE: Add support for
github:
,gist:
,bitbucket:
, andgitlab:
shorthand prefixes. GitHub shortcuts will continue to be normalized toorg/repo
instead of being saved asgithub:org/repo
, butgitlab:
,gist:
, andbitbucket:
prefixes will be used on the command line and frompackage.json
. BE CAREFUL WITH THIS.package.json
files published with the new shorthand syntax can only be read bynpm@2.8.0
and later, and this feature is mostly meant for playing around with it. If you want to save git dependencies in a form that older versions of npm can read, use--save-exact
, which will save the git URL and resolved commit hash of the head of the branch in a manner similar to the way that--save-exact
pins versions for registry dependencies. This is documented (so checknpm help install
for details), but we're not going to make a lot of noise about it until it has a chance to bake in a little more.
It is @othiym23's sincere hope that this will resolve all of the inconsistencies users were seeing with GitHub and git-hosted packages, but given the level of change here, that may just be a fond wish. Extra testing of this change is requested.
6b0f588
#7867 Use git shorthand and git URLs as presented by user. Support newhosted-git-info
shortcut syntax. Save shorthand inpackage.json
. Try cloning viagit:
,git+ssh:
, andgit+https:
, in that order, when supported by the underlying hosting provider. (@othiym23)75d4267
#7867 Document new GitHub, GitHub gist, Bitbucket, and GitLab shorthand syntax. (@othiym23)7d92c75
#7867 When--save-exact
is used with git shorthand or URLs, save the fully-resolved URL, with branch name resolved to the exact hash for the commit checked out. (@othiym23)9220e59
#7867 Ensure that non-prefixed and non-normalized GitHub shortcuts are saved topackage.json
. (@othiym23)dd398e9
#7867hosted-git-info@2.1.1
: Ensure thatgist:
shorthand survives being round-tripped throughpackage.json
. (@othiym23)33d1420
#7867hosted-git-info@2.1.0
: Add support for auth embedded directly in git URLs. (@othiym23)23a1d5a
#7867hosted-git-info@2.0.2
: Make it possible to determine in which form a hosted git URL was passed. (@iarna)eaf75ac
#7867normalize-package-data@2.0.0
: Normalize GitHub specifiers so they pass through shortcut syntax and preserve explicit URLs. (@iarna)95e0535
#7867npm-package-arg@4.0.0
: Add git URL and shortcut to hosted git spec and usehosted-git-info@2.0.2
. (@iarna)a808926
#7867realize-package-specifier@3.0.0
: Usenpm-package-arg@4.0.0
and test shortcut specifier behavior. (@iarna)6dd1e03
#7867init-package-json@1.4.0
: Allow dependency onread-package-json@2.0.0
. (@iarna)63254bb
#7867read-installed@4.0.0
: Useread-package-json@2.0.0
. (@iarna)254b887
#7867read-package-json@2.0.0
: Usenormalize-package-data@2.0.0
. (@iarna)0b9f8be
#7867npm-registry-client@6.3.0
: Mark compatibility withnormalize-package-data@2.0.0
andnpm-package-arg@4.0.0
. (@iarna)f40ecaa
#7867 Extract a common method to use when cloning git repos for testing. (@othiym23)
TEST FIXES FOR NODE 0.8
npm continues to get closer to being completely green on Travis for Node 0.8.
26d36e9
#7842 When spawning child processes, map exit code 127 to ENOENT so Node 0.8 handles child process failures the same as later versions. (@SonicHedgehog)54cd895
#7842 Node 0.8 requires -e with -p when evaluating snippets; fix test. (@SonicHedgehog)
SMALL FIX AND DOC TWEAK
20e9003
tar@2.0.1
: Fix regression where relative symbolic links within an extraction root that pointed within an extraction root would get normalized to absolute symbolic links. (@isaacs)2ef8898
#7879 Better document thatnpm publish --tag=foo
will not setlatest
to that version. (@linclark)
v2.7.6 (2015-04-02):
GIT MEAN, GIT TUFF, GIT ALL THE WAY AWAY FROM MY STUFF
Part of the reason that we're reluctant to take patches to how npm deals with
git dependencies is that every time we touch the git support, something breaks.
The last few releases are a case in point. npm@2.7.4
completely broke
installing private modules from GitHub, and npm@2.7.5
fixed them at the cost
of logging a misleading error message that caused many people to believe that
their dependencies hadn't been successfully installed when they actually had
been.
This all started from a desire to ensure that GitHub shortcut syntax is being
handled correctly. The correct behavior is for npm to try to clone all
dependencies on GitHub (whether they're specified with the GitHub
organization/repository
shortcut syntax or not) via the plain git:
protocol
first, and to fall back to using git+ssh:
if git:
doesn't work. Previously,
sometimes npm would use git:
and git+ssh:
in some cases (most notably when
using GitHub shortcut syntax on the command line), and use git+https:
in
others (when the GitHub shortcut syntax was present in package.json
). This
led to subtle and hard-to-understand inconsistencies, and we're glad that as of
npm@2.7.6
, we've finally gotten things to where they were before we started,
only slightly more consistent overall.
We are now going to go back to our policy of being extremely reluctant to touch the code that handles Git dependencies.
b747593
#7630 Don't automatically log all git failures as errors.maybeGithub
needs to be able to fail without logging to support its fallback logic. (@othiym23)cd67a0d
#7829 When fetching a git remote URL, handle failures gracefully (without assuming standard output exists). (@othiym23)637c7d1
#7829 When fetching a git remote URL, handle failures gracefully (without assuming standard error exists). (@othiym23)