commit b22f412dffb72c08f804418a8348facd8f0a0f1b
Author: James Halliday
Date: Thu Mar 14 19:16:19 2013 -0700

so it is

As a true believer in the UNIX way of tiny self-contained modules that each do just one thing, I find myself at a fundamental disagreement when talking with IDE-adherents or framework-oriented programmers.

When abstractions are kept to the minimum necessary, it's much easier to see which ones are good ideas and which ones are mistaken ideas. With larger collections of ideas branded into cohesive framework ideologies, it's very easy for unjustified abstractions to sneak in under the cover of the more useful pieces.

So it is with many things.

Do one thing well even if that one thing turns out to be a bad idea. Publish your mistakes to save the next person who gets the same idea about how things could be the trouble of re-implementing the same mistakes.

monopoly

It used to be the languages had a monopoly on abstractions. The standard distribution would bundle some handy libraries to do common things like option parsing, making http requests, and sending email. This made sense at the time. When you installed a distribution of UNIX or Linux you get the BSD or GNU toolchain of handy utilities so the same logic should apply.

But these standard libraries grew rotten. Nobody could give them the love and care they needed because the core team was too busy evolving the language, iterating on the runtime, and improving performance. The core team forgot about poor little distutils.

It's actually far worse than merely forgetting about a neglected expanded core: the versioning on core modules is intrinsically linked to the distribution release cycle as a whole so making any changes becomes a beaurocratic protracted struggle that few potential contributors have the patience for. Warts accumulate. Code calcifies around bugs that make it impossible to make necessary changes without breaking everything.

pulling from the outside

The reason bundles can work for system distributions is that the people working on the distributions are just pulling in from external resources, not pushing the new iterations themselves. The work happens outside from the community at large and is pulled in. This independence makes it much less daunting to fix bugs, improve APIs, and experiment.

Need to make a breaking change? No problem, just bump the major version. Trust that any packages downstream will be sufficiently careful with their dependency semvers.

welded shut

Frameworks side-step the versioning problem that core platforms have by adopting the system distribution model for bundling up handy, common utilities in a convenient way.

However, when you choose a framework, you trade away some temporary convenience for long-term flexibility. This may be an appealing bargain, particularly for short-lived applications as it imposes an immediate sense of structure on an otherwise blank slate that might very indeed get software built before a deadline.

However, it's much harder to change your mind about a choice that you never made yourself. When frameworks make your decisions for you, you very often won't even realize that a decision has been made at all so it's much harder to identify problems when the assumptions grounded in that technology choice no longer apply.

Another problem with short-sighted calculations is that it's very hard to know which applications really are one-offs and which ones will have a surprisingly enduring lifespan. And when a project "ends", it doesn't usually "end". It's just that the current contract is up, so any improvements will need to be made in a separate contract with perhaps a different team at the helm. Extend empathy to next person to work on some code, because that next person could be you!

An argument I often hear in favor of frameworks is that it helps a big team rally around some common tools that everybody is familiar with. This is true, but I often observe the same benefits collaborating with people who use some of my favorite low-level stream modules like through and duplexer or when authors use interfaces from node core like .pipe().

To be fair, there is a distinction here between frameworks that are sheer mudballs of compounded assumptions and frameworks that take the time to carefully factor out themselves into lots of tiny pieces. I think an important standard to judge frameworks is how many of the framework's core abstractions can be used ala-carte without using the framework itself. If your framework melts away into an informal collection of modules that happen to work well together but can be easily repurposed by people who don't use the framework, then you have built something very sublime.

cooking up some modules

If some problems naturally turn out to naturally trend towards monopoly then that is something that we should notice and accept. However, I expect that the number of these natural monopolies which actually exist is far lower than many people assume right now.

Tools like github and free-for-all package managers like npm continue to remove barriers to entry to such a point that the fitness landscape of software is a fundamentally different place than just 10 years ago.

What we need right now is not more frameworks to give a collection of modules a cohesive "brand" with divisive opinions about what "belongs". Instead, we need more recipes, articles, talks, and screencasts about how all these pieces fit together.

We need to be better about how to go from problem A to solution B by way of modules X, Y, and Z.

There should be more public discussion about what modules to use for which problems and when.

There should be more meta-analysis too about different fundamental design axioms at work behind certain clusters of libraries versus other clusters.

Keep publishing modules and learning new facts by experiment of course, but if we are going to obtain this lofty goal set forth by ryan dahl

I want programming computers to be like coloring with crayons and playing with
duplo blocks.

then we'll need more than code. We'll need documentation and community!

more
git clone http://substack.net/blog.git