| title | Reduce the environment variables provided to lifecycle scripts |
|---|---|
| number | 21 |
| status | implemented |
| created | |
| accepted_at | |
| implemented_at | |
| withdrawn_at | |
| implementation |
In versions of npm up to v6, the following items are all placed into the
environment of scripts run for various lifecycle events (install, prepare,
etc., as well as explicit scripts such as test and start).
PATHConfigured to include.../node_modules/.binfor current and all parentnode_modulesdirectories.npm_package_*for allpackage.jsonvalues in the current package for which the lifecycle event is running.npm_config_*for all npm configuration values that do not start with a_character.npm_lifecycle_eventthe current lifecycle event.npm_lifecycle_scriptthe command being run.npm_node_execpaththe path to the Node.js executable npm is using.npm_execpaththe path to the npm executable being run.
The suggestion presented here is to remove (or vastly reduce) the
npm_config_* and npm_package_* environment variables from the context
of lifecycle scripts, and potentially also add new fields that may be more
useful to more users.
Lifecycle scripts are run in many different contexts throughout the npm codebase.
- Explicit scripts are run directly from the
lib/run-script.jscommand implementation. - Build scripts are run from the context of the tree building logic, which
is moving to a new implementation with
@npmcli/arboristin v7. - Prepare scripts are run by
pacotewhen it creates a tarball for publication or when it installsgitdependencies.
All of this necessitates passing around a single configuration object, which has some problems.
- It is tedious and error-prone, and has led to a more complicated codebase
- While we have not had security issues with it in the past, it runs the risk of exposing something sensitive in a context where it should not be exposed.
- It invites users to fork package behavior based on npm configuration, which should be a contract between the user and npm, and not between the user, npm, and the publisher.
- While the package.json data does not have as many of these problems, it
is also largely unnecessary (and not widely used). The
package.jsonfile is readily available and easily parsed, and most scripts that would depend on package data simply read it directly. - The environment is created anew for every script that's run. This could be optimized further, but as it currently stands, it's pretty inefficient.
- Lastly, exposing the full configuration and
package.jsonmakes the environment significantly larger, and can lead to problems on memory-constrained systems.
The advantage of including npm_config_* values in the lifecycle
environment is that npm commands run from within lifecycle events will have
the same config values as the process that spawned them, since env values
will override any other values except explicit command line flags.
For example, a script named release may run tests, update the changelog,
and then publishe the package. Running npm run release --otp=123456 will
put the two-factor auth one-time password into the npm_config_otp
environment variable, so that the subsequent npm publish command will
have the one-time password provided in the config.
- Remove
npm_package*values from the script lifecycle environment. - Provide a new field,
npm_package_json, with the path to thepackage.jsonfile. - Provide a new field,
npm_command, with the canonical name of the command being run. - Remove all
npm_config_*values from the script lifecycle environment except:npm_config_userconfignpm_config_globalconfig- Environment variables corresponding to any non-default config values.
- Add
npm_package_from,npm_package_resolved, andnpm_package_integrityfor the package whose lifecycle event is running, if it's part of an install. (This addresses the needs of build tools, as discussed in #38.) PATHwill continue to be provided as it currently is, so that scripts find their dependencies' executables first.
This makes it easier to find and rely on package.json data, while ensuring
that config defaults are maintained, without blowing up the size of the
environment for lifecycle processes, or requiring access to the npm config
subsystem in every npm CLI dependency.
In addition to these, the following environment variables will be preserved from npm v6's set, in order to provide an easier upgrade path:
npm_package_namenpm_package_versionnpm_package_config_*Each key in theconfigobject will be included, but npm will not override values with a<pkgname>:<keyname>config value if one exists.npm_package_engines_*npm_package_mainnpm_package_bin_*
Possible alternatives:
This is not ideal for the reasons mentioned above, but also, it makes it virtually assured that Arborist remains tightly coupled to the npm cli. While some degree of coupling is unavoidable, having to provide a valid npm config object would make this coupling much tighter than necessary.
Rather than provide a config object matching a given interface, provide
npm-lifecycle with a method that can build up and return the environment
object.
This approach would address the tight coupling between cli and arborist, but it doesn't address the other problems with having a giant config object dumped into the environment.
The npm CLI will set all non-default config flags in the environment so that scripts and sub-scripts will have them set in their configs by default at the env level.
Replace npm-lifecycle with a lighter-weight approach, published as
@npmcli/run-script.
Instead of building the environment up from the config and package data,
@npmcli/run-script will only set npm_package_json to the path to the
package.json file for the package being run, take an object to define
additional environment variables, and always inherit the environment in the
child process being run.
Because the npm CLI sets the relevant config fields, they'll be inherited
to the child processes automatically. Arborist will use the environment
option to pass in the npm_package_from, npm_package_resolved, and
npm_package_integrity values.
npm v6 and yarn both do roughly the same thing, though they have different config values.
Some modules today use npm_config_argv. These will have to be updated to
use other means to get this information.
Where the argv is being parsed in order to determine the command being run,
the npm_command environ provides a safer approach.