What exactly should I put in .npmignore?
Tests? Stuff like .travis.yml, .jshintrc? Anything that isn't needed when running the module (except the readme)?
I can't find any guidance on this.
As you probably found, NPM doesn't really state specifically what should go in there, rather they have a list of ignored-by-default files. Many people don't even use it as everything in your .gitignore is ignored in npm by default if .npmignore doesn't exist. Additionally, many files are already ignored by default regardless of settings and some files are always excluded from being ignored, as outlined in the link above.
There is not much official on what always should be there because it is basically a subset of .gitignore, but from what I gather from using node for 5-ish years, here's what I've come up with.
Note: By production I mean any time where your module is used by someone and not to develop on the module itself.
Pre-release cross-compiled sources
Pros: If you are using a language that cross-compiles into JavaScript, you can precompile before release and not include .coffee files in your package but keep tracking them in your git repository.
Build file leftovers
Pros: People using things like node-gyp might have object files that get generated during a build that never should go into the package.
Cons: This should always go into the .gitignore anyway. You must place these things inside here if you are using a .npmignore file already as it overrides .gitignore from npm's point of view.
Tests
Pros: Less baggage in your production code.
Cons: You cannot run tests on live environments in the slim chance there is a system-specific failure, such as an out of date version of node running that causes a test to fail.
Continuous integration settings/Meta files
Pros: Again, less baggage. Things such as .travis.yml are not required for using, testing, or viewing the code.
Non-readme docs and code examples
Pros: Less baggage. Some people exist in the school-of-thought where if you cannot express at least minimum viable functionality in your Readme, your module is too big.
Cons: People cannot see exhaustive documentation and code examples on their own file system. They would have to visit the repository (which also requires an internet connection).
Github-pages objects
Pros: You certainly don't need to litter your releases with CNAME files or placeholder index.htmls if you use your module serves double-duty as a gh-pages repository as well.
bower.json and friends
Pros: If you decide to build in your dependencies prior to release, you don't need the end-user to install bower then install more things with that. I would, personally, keep that stuff in the package. When I do an npm install, I should only be relying on npm and no other external sources.
Basically, you should ever use it if there is something you wish to keep out of your npm package but checked-in to your module's repo. It's not a long list of items, but npm would rather build in the functionality than having people stuck with irrelevant objects in their package.
I agree with lante's short and syntetic answer and SamT's big answer:
You should not include your tests in your package.
Your package should only contains production runtime files.
That will make your package more straightforward and faster to be dowloaded.
My contribution to those answers:
.npmignore is the blacklist way to achieve package file selection. But in a more practical way, you can whitelist files you need to include in your package using the files field in your package.json:
{
"files": [
"lib/",
"index.js"
]
}
I think that's simpler, future proof and have better semantics ;)
Just to clarify, anytime someone do npm install your-library, npm will download all source files that the package includes. Those files that were included in the .npmignore file in the source code of the package your-library will be excluded when publishing the lib, so users of your-library won't download them.
Know that people installing your library will need just your library running, anything else will be not necessary.
For example, when someone installs a library, its probably that he/she doesn't care about your .travis.yml or your .jshintrc files, or even some images, Grunt files, documentation, etc.
.npmignore could let your npm package to have less files, and faster to be downloaded
Don't include your tests. Oftentimes tests are like 5x the size of the actual codebase. As long as your tests are on Github, etc, that's good enough.
But what you absolutely should do is test your NPM package in its published format. Create some smoke tests that reside in the actual codebase, but are not part of the test suite.
You can read about testing your package after tarballing it, here:
https://github.com/ORESoftware/r2g
How to test an `npm publish` result, without actually publishing to NPM?
Related
I lately helped out on a project, where I added a really small dependency - in fact, it only contained a regular expression (https://www.npmjs.com/package/is-unc-path).
The feedback I got from the developer of the project was that he tries to minimize third-party dependencies if they can be implemented easily - whereby he - if I understand it correctly - asks me to just copy the code instead of adding another dependency.
To me, adding a new dependency looks just like putting some lines of code into an extra file in the repo. In addition, the developers will get informed by an update if the code needs a change.
Is it just a religious thought that drives a developer to do this? Are there maybe any costs (performance- or space-wise, etc) when adding a dependency?
I also had some disputes with my managers once concerning the third party libraries, the problem was even greater he got into believing that you should version the node_modules folder.
The source of any conflict usually is the ignorance.
His arguments were:
you should deliver to the client a working product not needing for him to do any other jobs like npm install
if github, npm is down in the moment when you run npm install on the server what you will do ?
if the library that you install has a bug who will be responsible
My arguments were:
versioning node_modules is not going to work due to how package dependencies work, each library will download his own node_modules dependencies and then your git repository will grow rapidly to hundreds of mb. Deploy will become more and more slow, downloading each time half a gb of code take time. NPM does use a module caching mechanism if there are no changes it will not download code uselessly.
the problem with left-pad was painfull but after that npm implemented a locking system and now for each package you just lock to a specific commit hash.
And Github, and npm does not have just a single instance service, they run in cloud.
When installing a dependency you always have some ideas behind and there are community best practices, usually they resume to: 1. Does the repo has unit tests. 2. The download number 3. When was the latest update.
Node.js ecosystem is built on modularity, it is not that node is so popular cause of some luck, but cause of how it was designed to create modules and reuse them. Sometimes working in node.js environment feels like putting lego pieces together and building your toy. This is the main cause of super fast development in node.js. People just reuse stuff.
Finally he stayed on his own ideas, and I left the project :D.
I assume that when developing an NPM project, that every git branch (or whatever version control system you use) probably points to a different set of node_modules on the filesystem. Is that true? How does that work? Does it pose any problems for diskspace etc?
Or perhaps, since node_modules is most commonly .gitignore'd, then the node_modules files are shared between Git branches? Again, how would/does that work?
*Note that Node.js / NPM is fundamentally different than other platforms/languages since dependencies are typically stored locally to a proejct rather than in some central location on a machine.
By convention, one should not add any files, libraries or binaries which can be generated or pulled in from an external source. This includes things like node_modules; since that is made readily available* once you do npm install, there's no reason or incentive** to want to put that into your source control. At worst, it will also bloat your repository, filling your diffs with things you simply don't control and don't necessarily want to review.
I would not expect different Git branches of an NPM project to contain different node_modules folders. I'd only expect the one node_modules folder, and if a branch gave me fits about dependencies, I'd look to reinstall the dependencies (and note it down to be sure that something else hadn't gone awry).
As an addendum, any files or folders in .gitignore are simply not indexed or tracked by Git. If the contents of those files or folders change, Git is none the wiser. This also means, when switching between branches, the contents of the files or folders in .gitignore remain the same.
*: Provided that the library you're using isn't suddenly yanked. Or the repository is not impacted by a colossal DDoS.
**: There may be some incentive to do this given that the reliability of certain NPM packages hasn't been 100% this year, but that's a team and architecture-driven decision, and I doubt that placing it into source control is the most ideal and convenient way to deal with it.
There are two schools of thought, and both have merit.
1) Never check in node_modules and rebuild on deploy/install
The approach relies heavily on NPM and the connectivity of your deploy environment. node_modules are downloaded and installed (and/or compiled) each time the deploy is run.
Positives:
Your repository is much smaller.
NPM modules are installed in the environment they will run on.
Concerns:
Tied to 3rd party for sources - Go read about that whole left-pad thing. If one dependency cannot be downloaded, your entire build system is hung out to dry. "Cranky and paranoid old timers" will cite this as the reason to check everything in (or run your own private NPM somewhere).
Branch management - Like you mentioned in the question, some branches might not have the same dependencies. Dev1 adds a new features and used a new package. Now Dev2 runs the dev branch or whatever, and everything is broken and they need to know to npm install the new package. More subtle is the case where a npm package is version changed (now you need npm update as npm install will say nothing has changed), or where their node_modules are upgraded to work on "new feature 10" but they need to clear everything out to "downgrade" to go fix "prior bug 43". If you are in active development with a team of more than 2-3, watch out for this one.
Build Time - If it is a concern, it takes a little longer to download and install everything. Or a lot longer.
2) Always check in everything you can
This approach includes node_modules as part of the repo.
Positives:
Not dependent on 3rd party sources. You have what you need to run. You code can live on its own forever, and it does not matter if npm is down or a repo is deleted.
Branches are independent, so new features from Dev1 are auto included when Dev2 switches to that branch
Deploy time is shorter because not much needs to be installed.
Concerns:
Repository is much larger. Clones of code take longer as there are many more files.
Pull Requests need extra care. If a package is updated (or installed) along with core code, the PR is a mess and sometimes unintelligible. "500 files changed", but really you updated a package and changed two lines of core code. It can help to break down into two PRs - one that is is a mess (the package update) and one that is actually reviewable (the core code change). Again, be prepared for this one. The packages will not change too often, but your code review takes a little longer (or a little more care) when they do.
OS Dependent Packages can break. Basically anything that is installed/compiled with gyp can be OS dependent (among others). Most packages are "pure JS" and, being just scripts, run everywhere. Imagine all your devs run and test on OSX while you deploy to Linux, you cannot check in those packages that were compiled on a MAC because they will not run on Linux. An odd workaround for this is to define most packages as "dev dependencies" (--save-dev) and the ones that need compiled as normal ("production", --save), then you run npm install --production so the dev dependencies are not installed (and are already present), but the others are.
Conclusions
It depends. (Don't you hate hearing that all the time? : )
Depending on your team and your concerns, you might go either approach. Both have their merits, and you will decide which is more beneficial to you. Both have drawbacks as well, so be aware of those before you get bit!
Personally I ignore .node_modules but I have different package.json in different branch and when I switch I reinstall the dependencies
Two branches having different set of node modules is in scenario where one branch is in development phase and other is your production branch. In such cases development branch will have more node modules than production. If I am not wrong any other scenario might get you in trouble.
Pushing node_modules to remote version control repository is bad practice hence just rely on npm install whenever you clone a branch or pull the code to download any new node module added to package.json.
Apparently, since you don't have your node_modules in your actual repository, you need to install node modules again and each branch might have its own requirement, as you might update your server.js with new dependency and you also need to make sure you have these newly added node dependencies in your production server as well.
I have a repo which consists of several "micro-services" which I upload to AWS's Lambda. In addition I have a few shared libraries that I'd like to package up when sending to AWS.
Therefore my directory structure looks like:
/micro-service-1
/dist
package.json
index.js
/micro-service-2
/dist
package.json
index.js
/shared-component-1
/dist
package.json
component-name-1.js
/shared-component-2
/dist
package.json
component-name-2.js
The basic deployment leverages the handy node-lambda npm module but when I reference a local shared component with a statement like:
var sharedService = require('../../shared-component-1/dist/index');
This works just fine with the node-lambda run command but node-lambda deploy drops this local dependency. Probably makes sense because I'm going below the "root" directory in my dependency so I thought maybe I'd leverage gulp to make this work but I'm pretty darn new to it so I may be doing something dumb. My strategy was to:
Have gulp deploy depend on a local-deps task
the local-deps task would:
npm build --production to a directory
then pipe this directory over to the micro-service under the /local directory
clean up the install in the shared
I would then refer to all shared components like so:
var sharedService = require('local/component-name-1');
Hopefully this makes what I'm trying to achieve. Does this strategy make sense? Is there a simpler way I should be considering? Does anyone have any examples of anything like this in "gulp speak"?
I have an answer to this! :D
TL;DR - Use npm link to link create a symbolic link between your common component and the dependent component.
So, I have a a project with only two modules:
- main-module
- referenced-module
Each of these is a node module. If I cd into referenced-module and run npm link, then cd into main-module and npm link referenced-module, npm will 'install' my referenced-module into my main-module and store it in my node_modules folder. NOTE: When running the second npm link, the name of the project is the one you find in your package.json, not the name of the directory (see npm link documentation, previously linked).
Now, in my main-module all I need to do is var test = require('referenced-module') and I can use that to my hearts content. Be sure to module.exports your code from your referenced-module!
Now, when you zip up main-module to deploy it to AWS Lambda, the links are resolved and the real modules are put in their place! I've tested this and it works, though not with node-lambda yet, though I don't see why this should be a problem (unless it does something different with the package restores).
What's nice about this approach as well is that any changes I make to my referenced-module are automatically picked up by my main-module during development, so I don't have to run any gulp tasks or anything to sync them.
I find this is quite a nice, clean solution and I was able to get it working within a few minutes. If anything I've described above doesn't make any sense (as I've only just discovered this solution myself!), please leave a comment and I'll try and clarify for you.
UPDATE FEB 2016
Depending on your requirements and how large your application is, there may be an interesting alternative that solves this problem even more elegantly than using symlinking. Take a look at Serverless. It's quite a neat way of structuring serverless applications and includes useful features like being able to assign API Gateway endpoints that trigger the Lambda function you are writing. It even allows you to script CloudFormation configurations, so if you have other resources to deploy then you could do so here. Need a 'beta' or 'prod' stage? This can do it for you too. I've been using it for just over a week and while there is a bit of setup to do and things aren't always as clear as you'd like, it is quite flexible and the support community is good!
While using serverless we faced a similar issue, when having the need to share code between AWS Lambdas. Initially we used to duplication the code, across each microservice, but later as always it became difficult to manage.
Since the development done in Windows Environment, using symbolic links was not an option for us.
Then we came up with a solution to use a shared folder to keep the local dependencies and use a custom written gulp task to copy these dependencies across each of the microservice endpoints so that the dependency can be required similar to npm package.
One of the decisions we made is not to keep two places to define the dependencies for microservices, so we used the same package.json to define the local shared dependencies, where gulp task passes this file and copy the shared dependencies accordingly also installing the npm dependencies with a single command.
Later we made the code open source as npm modules serverless-dependency-install and gulp-dependency-install.
How should we deal with local packages that are a dependency in other local packages?
For simplicities sake, say we have the follow packages
api - express application
people - a package to deal with people
data-access - a package that deals with data access
And then the dependencies are
api depends on people
people depends on data-access
Currently we have these dependencies setup as file dependencies.
I.e. api package.json would have
"dependencies": {
"people": "file:../people"
}
Trouble with this is that we're finding it a PITA when we make updates to one package and want those changes in the other packages that depend on it.
The options we have thought of are:
npm install - but this won't overwrite previously installed packages if changes are made, so we have to delete the old one from the node_modules directory and re-run npm install... which can be niggly if the package dependency is deep.
npm link - we're not sold on the idea because it doesn't survive version control... Just thinking about it now, maybe we have some kind of local build script that would run the npm link commands for us... this way it could survive version control. Would that be a grunt job?
grunt - we haven't dived too deep into this one yet, but it feels like a good direction. A little bit of googling we came accross this: https://github.com/ahutchings/grunt-install-dependencies
So, what option would work best for our situation?
Are there other options that we haven't thought of yet?
Ps. we're a .NET shop doing a PoC in node, so assume we know nothing!
Pps. if you strongly believe we're setting up our project incorrectly and we shouldn't have smaller individual packages, let me know in the comments with a link to some reading on the subject.
So, I agree that going with 'many small packages' is usually a good idea. Check out 12factor.net if you haven't already.
That said, in specific answer to your question I'd say your best bet is to consider mainly how you want to maintain them.
If the 'subcomponents' are all just parts of your app (as, for example, data-access implies), then I'd keep them in the same folder structure, not map them in package.json at all, and just require them where you need them. In this case, everything versions together and is part of the same git repository.
If you really want to or need to keep them all in separate git repositories, then you can do npm link, but to be honest I've found it more useful to just use the URL syntax in package.json:
dependencies: {
"people" : "git://path.to.git:repo#version.number"
}
Then, when you want to explicitly update one of your dependencies, you just have to bump the version number in your package.json and run npm install again.
In doing research on whether Node's node_modules should be checked into your version control repository, the most recent consensus seems to be that you should include it for applications you deploy.
Sources:
Check in node_modules vs. shrinkwrap
Should I check in node_modules to git when creating a node.js app on Heroku?
https://www.npmjs.org/doc/misc/npm-faq.html#Should-I-check-my-node_modules-folder-into-git
In reading these arguments, it lead me to question whether Composer's /vendors directory should also be checked into version control. Composer's documentation suggests that you don't:
Should I commit the dependencies in my vendor directory?
The general recommendation is no. The vendor directory [...] should be added to .gitignore/svn:ignore/etc.
The best practice is to then have all the developers use Composer to install the dependencies. Similarly, the build server, CI, deployment tools etc should be adapted to run Composer as part of their project bootstrapping.
While it can be tempting to commit it in some environment, it leads to a few problems:
Large VCS repository size and diffs when you update code.
Duplication of the history of all your dependencies in your own VCS.
[...]
Contrasting that argument is this one (source):
Doesn’t checking in node_modules create a lot of noise in the source tree that isn’t related to my app?
No, you’re wrong, this code is used by your app, it’s part of your app, pretending it’s not will get you in to trouble. You depend on other people’s code and they are just as likely to write new bugs as you are, probably more so. Checking all of that code in to source control gives you a way to audit every line that ever changed in your application. It allows you to use $ git bisect locally and be ensured that it’s the same as in production and that every machine in production is identical. No more tracking down unknown changes in dependencies, all the changes, in every line, are viewable in source control.
In summary, the question is this: Why would one gitignore (i.e. not version control) node_modules but not do the same for Composer's vendor/ directory?
The reason to commit external dependencies is
it's easier to deploy with git pull
the code used is directly included in the commit anyone checks out
Arguments against this are
Git is no deployment tool
all dependency managers do have a way to make exactly sure the code used can be fetched
I don't know anything about npm, but for Composer that last point is implemented by committing composer.lock.
I don't think the "audit code" argument is a valid one in every case. If you do write software that get's audited by itself, and subsequently needs all libraries to be audited, then probably all code changes need to be conserved. This isn't true for the general case.
git bisect works still as well with a committed composer.lock. It does require installing the dependencies with every bisect step, but this is just one simple step, probably already done in the automatic test suite run.
The last thing to worry about is when used packages go offline. This really is more of a problem with Composer, because there is no central hosting of the downloadable releases (npm probably does this). If this is a problem, either commit the code (and try to figure out how to update that missing package in the future), or setup an instance of Satis to create a local copy of the code you use.
Putting all your modules in you VCS makes it really heavy to download or upload. For example, I work on two node.js projects and the total node_modules directories size is between 250MB and 500MB whereas the whole code with assets is generally less than 40MB. Every Node.js developer likes Node lightness, so the code must stay easy to download and share.
For the second point, an alternative to avoid regressions is to be more restrictive in your package.json dependencies versions. You will find more information here.
Finally the best argument is to take a look on the famous modules everybody know :
express ignores
node_modules
mocha ignores
node_modules
q ignores node_modules
...
The more I research this the more I'm starting to think that there is no one correct answer to this, just different opinions as well as pros and cons of each method.
This blog, looking from the context of Bower, does a good job of weighing the pros and cons of each: http://addyosmani.com/blog/checking-in-front-end-dependencies/.
In short: At least for right now, weigh the pros and cons and decide what best fits your situation.