If a developer tried to `meteor update` a project whose `.meteor/release`
file corresponded to a version of `meteor-tool` that no longer exists in
`~/.meteor/packages/meteor-tool`, this code would throw an ENOENT error.
This could be fixed by running `meteor update --release <old release>`
first, but that kind of workaround shouldn't be necessary when updating
Meteor to the latest version.
As a refinement of my previous commit, we will continue setting the
$GYP_MSVS_VERSION variable to "2015" except in cases where the Python
library behind node-gyp succeeds in inferring the correct version.
Note that the Python library will never infer "2015" by itself, even
though that's the version recommended by the node-gyp setup instructions:
https://github.com/nodejs/node-gyp#installation.
Note also that the Windows machines used by `meteor admin get-machine` and
build.meteor.com have version "2010e", which can be inferred.
In other words, this commit allows the following groups of developers to
avoid setting $GYP_MSVS_VERSION manually:
1. Developers who followed the node-gyp setup instructions.
2. Package authors who run `meteor publish-for-arch` on build machines.
Setting GYP_MSVS_VERSION to 2015 when the Visual C++ 2015 build tools are
not (and/or cannot be) installed is (in my experience) the leading cause
of `npm rebuild` failures on older Windows machines.
Auto-detecting the appropriate version is something node-gyp ought to be
able to do as well as Meteor possibly could, at least according to
https://github.com/nodejs/node-gyp/blob/1dcf356ca7/gyp/pylib/gyp/MSVSVersion.py#L222-L223
When selecting a different version is important, you can always set the
GYP_MSVS_VERSION environment variable yourself. Specifically, if you're a
package author, and your package depends on npm packages that can only be
compiled with a newer version of MSBuild.exe, then you might need to run
$env:GYP_MSVS_VERSION = "2015"
meteor publish-for-arch name:package@x.y.z
I fully trust that the folks in that small category can handle that
additional burden.
We now use a smaller file (the `babel-compiler` archive instead of
`meteor-tool`), and we interrupt every 500kB, which leads to three total
interruptions for that file. Also, we only delay one second (instead of
five) between attempts.
This is a test that really needs to run and pass every time we run the
test suite, so I decided it shouldn't be --slow. If it takes too long, we
can always download a smaller test file.
Hard-coding the download length was a recipe for brittleness, so now I'm
downloading the file without interruptions in parallel with the
interrupted/resumed download, so that we can compare the two files when
both have finished downloading.
Follow-up to #7399.
I'm deliberately leaving in the code to ensure .meteor/.gitignore contains
dev_bundle for now, so that the old file won't get accidentally checked
into anyone's app repository.
