Up to Interpretation

What Do You Want (Instead)?

A lot of virtual ink has been spilled about dependency hell and left-pad and how NPM is hell incarnate. Medium posts moan about how NPM is insecure, is full of bad packages and is emblematic of developers' laziness. They’re not wrong. However, when one raises an issue with a service, there is always the inevitable follow up: what do you want instead? There’s no clear answer to this question for package management. Do we want to go to the bad old days of adding script tags and polluting our global namespace? Do we want to switch to large catch-all libraries a la Boost or jQuery? What do you want?

One common answer to this question is to point to other package distributions that appear to not have these problems. Take your pick between Bundler, Cargo, Composer, Pip, etc. Except, there’s a few flaws with this. First, these registries and respective package managers have their own problems. Pip has the inane global installation by default and non determinism. Cargo is beginning to have the same problems as NPM in dependency size, especially compounded by Rust’s compile times. I love Rust, but downloading and compiling packages can take a bit. Second, there’s the issue of scale. While it is nice to compare to other languages, NPM is a little unique in the sheer size of the ecosystem. Due to several factors which quite frankly I don’t feel like explaining (fragmentation, churn, diversity of usage), the ecosystem has the perfect storm of repetition and scale. Granted, Python is also rather diverse in its usages, but ask anybody who has had to install Tensorflow and they’ll tell you Python dependency management needs work as well.

Which brings me to my main point: the problems of NPM are problems that every package manager either faces or will face. I.e. NPM is not unique. The reason NPM gets all the coverage is because it’s reached these problems faster than any other ecosystem. But the fundamental problems of massive dependencies; tiny, yet powerful packages and massive attack surfaces will be problems that everybody will need to address.

Why? Well first, because packages work dammit. They’re an effective way of reusing software at a much more granular level than the massive swiss army knife libraries of yore. Try to find a newish language that isn’t moving to the package model. Even Go, the grumpy grampa yelling “get off my lawn” of 21st century languages, is moving to packages. This isn’t surprising at all. When we design our own projects, do we rely on monolithic objects that provide all the functions we need? No, we split them up into granular, modular abstractions.

Now, there’s always the straw man argument of “what about the is-odd package??” Well of course we shouldn’t make packages for checking the parity of a number. That shouldn’t have to be said. But there’s a lot of useful cases in between is-odd and jQuery in which packages solve a real problem.

The next point that people often make is, why not write the code yourself? And sure, in the trivial cases that’s true. You can write a decent set of array/string/object primitives instead of using lodash. You can write a delay Promise function yourself. But people don’t just avoid writing code due to laziness. They avoid it because of specialization. A commonly used package will probably have less bugs (cf Linus' Law), be faster than the naive implementation, and handle more cases. Not to mention, they’ll get bugfixes for free. I can’t count the number of times I’ve fixed a bug by bumping a version number.

Another interesting point people make is that the reason behind NPM’s issues is JavaScript’s lack of standard library. I don’t disregard this completely, there’s probably some amount of packages (cough Lodash cough) which depend on JS’s poor standard library. But if you look at the most depended upon packages, beside Lodash most of the packages wouldn’t fall under a standard library. Request, chalk, react, express, commander, or moment don’t really appear to be standard library material (okay maybe moment/request, but Rust for instance does not provide either functionality in its stdlib).

And besides, large standard libraries don’t exactly seem great either. They’re tied to the release cycle of the language, meaning if there’s a bug in the standard library, or worse, a security hole, you need to beg and plead users to upgrade their language. Meanwhile, a package means that even if a user doesn’t upgrade their existing projects, future projects will automatically use the newer package. And tools like NPM and GitHub are starting to warn people automatically about their dependency security holes.

Besides, as Python is finding out, if you bundle certain things in your standard library, but newer, better alternatives come out, you’re going to be stuck with a bunch of outdated packages in your standard library. Imagine if jQuery was JavaScript’s standard library. Hell, for the longest time it basically was.

This isn’t to say packages are perfect or that we should just ignore the issues with packages. I’m just sick of people hating on NPM as if there’s something unique about it that makes it bad and I’m sick of people hating on packages without proposing a better option. If you want to go back to script tags and globals be my guest. I’ll be here with my million pound node_modules.