I personally found this write up extremely interesting and exciting.
I have always been interested in documentation and its order in the priority list of tasks which a development team has to tackle. It is not an original observation that documentation is critically important to the success of a project/code-base and yet it is often the last artifact produced (and many skip it altogether). I have recently been extremely interested in the idea that documentation should be moved to the top of the priority list and, rather than being a duplicative post-processing step, should be the "ground-truth" for generating lots of the follow on artifacts. For example, write API documentation first and use that to generate client side libraries, an API test suite, and server boiler plate code/skeleton.
In my search for existing projects and approaches, I came across many interesting things.
It was very interesting to read this GitHub post because they presented yet another approach to treating documentation as a first class citizen with different methods to write docs, host docs, and keep the docs updated.
I recently updated the API docs at my workplace to use the Slate tool I referenced above. We manually write docs in a Markdown file, manually use Slate to compile the MD file into HTML, and then manually deploy it to our host. This is approach is incredibly basic and non-scalable, but is light years better than what we had previously, which was API docs directly in the repo's README file.
I hope to learn more about the projects listed above (and many others!) as I explore different approaches for treating docs as a first class citizen and pick the approach which meets the requirements of my current team.
I always thought it would be a good idea to generate API documentation from tests. Kind of the same idea you have but the other way around.
The problem I used to have is that if you write your documentation by hand it tends to get out of sync with the code. You make a quick change to the code and forget to update the docs. After a while it's a mess unless you stay vigilant.
But if you generate documentation from tests they can't get out of sync. The example output that gets written to the docs comes from the application so it can't be wrong. And if a test fails, documentation doesn't get written. It also forces you to write tests which is a good thing. If you don't, you don't have docs.
I don't know if there are tools like this. I created one for Ruby/Rack apps that I have used in some of my projects. I think this approach works pretty well.
I wholeheartedly agree with everything above. Two additional and related thoughts:
1. The original documentation tool I wrote for Atom, Biscotto, kept track of the number of undocumented classes / methods: https://github.com/gjtorikian/biscotto/blob/59f48ba2621a92ae... . It was hooked up to CI, and if the count fell below a certain threshold, the test failed.
2. Right now, we're exploring into working with JSON schema as a means of providing both testing validation and accurate documentation. If the schema says "This REST method expects a parameter of this type," it becomes very easy to write a test to enforce that behavior; documentation can be easily generated from it; and of course your production code is safer for it.
I'm a huge fan of introducing more cross-overs between testing and documentation. I think a lot of time is spent on "clever" (and subjective) validations like http://www.hemingwayapp.com/, but not enough time is spent on basic content checks. It's very easy to drift code, tests, and docs apart. We need to start thinking about all three of them working together.
I could not agree more strongly that working on code, docs, and tests as a single unit is an excellent approach to development. I think the best way to convince most developers of doing this is to create tools which make it as automated as possible. If using a tool actually makes their lives easier to accomplish something they already do (write tests? write docs?) and they get the other for free, then that is a huge selling point.
I had a similar idea about using JSON Schema to validate my API. However, a few issues that I thought of and a few I discovered while researching:
1) I am not aware of a decent JSON Schema editor. It has been on my wishlist/TODO list for a looong time to write one myself which has similar capability to something like Oxygen for XML [1], but I have not had the time. Do you plan to write and update the JSON Schema by hand?
2) I have not researched them in great detail yet, but Swagger [2] and API Doc [3] seem to have already defined something similar to what I had in mind with JSON Schema. Have you looked into those tools yet?
> Do you plan to write and update the JSON Schema by hand?
Currently, yes, unfortunately. ;____; We've talked about doing it in an intermediate format, like YAML, which is slightly less painful, but....ugh. Years ago I wrote a very badly-coded-but-essentially-functional tool called Panino, which converted Markdown-to-JSON: https://github.com/gjtorikian/panino-docs/tree/master/lib/pa....
I see that some people have picked up on that and ran along with it, too: https://github.com/apiaryio/mson. So there might be something there to explore.
2. I've never looked into either. I know Swagger is used by several companies, like Twitter, but API Doc is new to me.
I do like the idea of starting by writing docs because they are very lightweight and it forces you to think through the design of the API. For example, "the params on that route don't look correct" or "this is really difficult to explain, so I think we just over complicated this end point. Let's make it simpler".
One could make the TDD argument that writing tests first would accomplish the same thing, but I think tests are much heavier than docs because certain routes could have dependencies and other similar gotchas. For example, a route may require authentication so you either need to have a fake auth token in the test DB or you need to have your test first obtain an auth token via the API and then hit the end point you really want to test.
If you start by writing docs in something like Swagger, then it should be straight forward to generate API tests from those docs (given, you will still have some gotchas like in the authentication scenario, but if you are generating the tests, then you solve that issue once in your test generator and then the boiler plate is created for you). For example, if you add a required param to a route, you first update your docs to define this new require param and then re-generate the automated test suite. A generated test suite like that could be used for TDD purposes to develop against.
Can you expand on your idea of generating docs from tests? Your tool sounds very interesting. Is it open source for me to look at? If not, how did you have to format/structure your tests so that you knew what information to include in the docs? How did the generation work?
It isn't something I would encourage others to use in its current state, but maybe it will give you some ideas. It uses Ruby's Minitest and adds some syntax to it. When the tests run it captures some things and writes those to an html template.
Another option is Read the Docs, which is how we host the Write the Docs docs: http://docs.readthedocs.org/en/latest/ -- It's basically the above workflow that GitHub describes, but with Sphinx (http://sphinx-doc.org/) -- which is a mature and powerful documentation system.
I forgot to mention that, but it is already on my list of things to investigate. I came across readthedocs.com in the last few months (probably on HN) and sent a passionate email requesting access to the private beta. I haven't been selected yet, but I am anxiously waiting to try it out.
How would you say that readthedocs.com differs from the tool mentioned in a sibling comment (Readme: https://readme.io/)?
>I have always been interested in documentation and its order in the priority list of tasks which a development team has to tackle. It is not an original observation that documentation is critically important to the success of a project/code-base and yet it is often the last artifact produced (and many skip it altogether). I have recently been extremely interested in the idea that documentation should be moved to the top of the priority list and, rather than being a duplicative post-processing step, should be the "ground-truth" for generating lots of the follow on artifacts. For example, write API documentation first and use that to generate client side libraries, an API test suite, and server boiler plate code/skeleton.
I wish people would stop viewing documentation as an additional formal step needed to 'package' the software (a nasty chore) and instead viewed it as an exercise in communicating as clearly and concisely as possible what the project is about and how it works.
The number one thing on my wishlist for any project's documentation is simply a glossary, which is trivial to create but is almost never done. Virtually every project has special terms for artefacts, usage, features, etc. which are burned into developers' skulls so deeply that they often forget that outsiders do not use this special terminology.
It gets worse when multiple terms refer to the same thing and the same terms refer to multiple (often ever so slightly different) things.
No fancy technology needed to write one of these things either. A simple text file will do.
The other thing that bugs me is how much documentation is simply describing code that should exist but doesn't. Scripts to build a project or deploy it, for instance. Or mindless test scripts.
> I wish people would stop viewing documentation as an additional formal step needed to 'package' the software (a nasty chore) and instead viewed it as an exercise in communicating as clearly and concisely as possible what the project is about and how it works.
Yes 1000x. One of the first things I do when evaluating a new project is read the docs and tutorials.
A glossary for each project would also be super useful. I have also run into the issues that you outline (one term to many definitions and vice versa). Do you think the glossary would help solve that by forcing folks to rename things to have clearer terms (if one is already used in the glossary, for example) or do you think that it is just a barrier for some developers to even create the glossary in the first place (out of sight, out of mind mentality).
Would you keep the glossary in git (or similar) so that it versions along with the code? It could be a required part of a code review to make sure that reasonable terms and/or acronyms introduced in a PR are added to the glossary.
Thanks for suggesting another tool to look into! I only glanced at their homepage, but it looks like they may be using Swagger under the hood because some visual elements are similar and the functionality is also similar. It does appear to have many of the features that I am looking for, so it definitely warrants further investigation.
We existed (much) after Swagger. I've never used Swagger myself, although I briefly played around with the client. Not really that inspired by them.
We use APIDoc (and eventually want to add more) as our "schema", and you can auto-sync from GitHub. Or, you can add it manually.
I am familiar with RTD (moreso the .org version than the .com); they're great. However, they mostly just host Sphinx/markdown/etc files. ReadMe does much more (support section, API explorer, "Suggest Edits", show user keys in the code, etc). If you want to write docs locally and deploy, use RTD. If you want something more full-featured, check out ReadMe.
Thanks for comparison. I have not played with either your tool nor RTD yet, but I look forward to trying both.
I came across the APIDoc project when watching an excellent presentation by the Gilt CTO titled "Immutable Infrastructure with Docker and EC2" [1]. It is worth a watch and he mentions that APIDoc allows them to more easily create new micro services in their cluster (of which they have over 300!).
I was reading around the APIDoc site and saw they had a note to investigate Swagger 2.0 to see if it could be useful [2].
I would love to see a few projects explore this concept of generating docs, client code, and server code from a single "ground truth" file and eventually have them all merge into a single defacto standard. That would be exciting indeed (though it seems Swagger may already be approaching that point in some circles).
Thanks for the info! I have many questions about your tool, but if the answer is "go check it out via a free account!" then just tell me that and I'll get around to it :)
- Were you influenced by Swagger at all? Did you exist before/after?
- Do you expose any schema that developers can modify via an API or similar? Do you have an internal schema that all docs follow?
- I edited my original comment to include Read the Docs [1]. Are you familiar with that service (it is in private beta)? What are some advantages of using ReadMe instead of Read the Docs?
I wasn't familiar with the term, but a quick search reveals Literate Programming on Wikipedia [1]. I assume that is what you were referencing. I skimmed the summary of the article and it sounds interesting (and similar to what I'm describing).
Thanks for suggesting this. I glanced over the home page and I am definitely adding it to my list of projects to explore. It claims to support many of the features that I am looking for and it will be fun to kick the tires!
I'm curious, how they write internal-facing documentation and how that effects the development experience for new github engineers.
Source diving through open source libraries, I've often wished for a "spelunker's guide": a text file laying out where things were and what I should read first to build a mental model I could use in understanding the rest of the source. I'm currently trying to figure out what the best way is for someone to write a spelunker's guide, especially if they've forgotten what it's like to be a beginner.
> I'm curious, how they write internal-facing documentation and how that effects the development experience for new github engineers.
I wish I could show up a sample, but I can't, because it's internal. ;)
Honestly, I think a lot of the engineering documentation started organically. When you have a small team working on a feature, it's difficult to scale explanations to the rest of the company. One day someone sits down and starts writing all their thoughts out in Markdown, and just checks it into a docs folder. That's it. It's easy-to-read, short on code, and usually full of ASCII, like this: https://i.imgur.com/KTbyhyq.png
Writing documentation is the best way to get outside contributors involved with minimal investment on your part. It also forces you to try and explain what you've built.
If you can't pretend to go back and look at things like a beginner, grab someone unfamiliar with the project, and have them describe to you what they would expect, and how they think they should proceed. They may be able to provide you with insights on what needs to be described.
I'm planning to build a stack for internal company domain knowledge, and I've been thinking about middleman (http://middlemanapp.com) instead of Jekyll.
Middleman has impressive workflows and markdown processing (I'm guessing parallel to that of the Github/Jekyll solution or better).
Also conrefs can be implemented by simple partials (which makes less contention for the probably huge conref file)
Though I have to be convinced by trying the Github/Jekyll stack, this does open my mind regarding Jekyll 2.0. I'm happy to see Github tell us their Jekyll story :)
I think the main take-away with this post is that static sites for docs are awesome. GitHub uses [nanoc](http://nanoc.ws/) for the [API documentation](https://developer.github.com/).
The only downside of hosting static text on GitHub Pages without Jekyll is that you have to push the generated HTML too.
> which makes less contention for the probably huge conref file
We're planning on splitting up the conref files by section. So for example, we'd get a separate conref file for Pages, one for UI stuff, one for Enterprise, etc. My only compliant with partials is that it's one piece of content per file, but it's a trade-off vs one file with several conrefs (gotta CTRL-F for that text you want to change).
Ironcially, just earlier today my co-worker and I were also discussing plans to setup a domain knowledge service internally. He came across a tool called Raneto [1] which looks very promising, but we have not had time to play with it extensively yet. Perhaps it could be useful for your use-cases too.
> Github's convention that web pages for a project are in a different branch of the same project is kind of strange.
It does have the advantage of 0-configuration 0-conflict.
But because it namespaces through the branch, if you're using the repository for something other than just the pages you can't have the gh-pages simply follow/trail master unless you want a bunch of site crap at the root of your repository, and interacting with both code and documentation at the same time is more painful than it needs be.
> Also, those things they call "conrefs" are just "macros".
Macros have a wider implied range of behavior, possibly completely arbitrary.
A content reference attribute is just a placeholder or a very small textual include[0] (XML calls them "named entities", rST calls them "substitutions")
> I think a macro implies something that can be executed
That isn't true... Traditionally a macro just refers to a substitution, maybe (but not necessarily) with parameter replacement, rescanning, etc. I'd say that lisp-style macros which can execute arbitrary code are actually rather rare historically....
I wasn't super thrilled with it either, until I dug in and discovered that it's the global average. We have a ton of international and mobile traffic, which factors into this sum:
1. United States (average load: 1.97 s)
2. United Kingdom (average load: 2.29 s)
3. India (average load: 7.48 s)
4. China (average load: 12.11 s)
Average load of these countries is 6 seconds, which seems absolutely horrid...until I tell you that the US has about seven times more traffic than the UK.
I didn't want to fudge the graph and take out those slow outliers--the truth's the truth.
GitHub Pages is one of the most shocking hacks I've ever come across. Not the worst mind you, just the most shocking. Most of GitHub is clean and good. But making a magic gh-pages branches is simply horrific. I'm still somewhat dumb founded that's the best method they could come up with.
I have always been interested in documentation and its order in the priority list of tasks which a development team has to tackle. It is not an original observation that documentation is critically important to the success of a project/code-base and yet it is often the last artifact produced (and many skip it altogether). I have recently been extremely interested in the idea that documentation should be moved to the top of the priority list and, rather than being a duplicative post-processing step, should be the "ground-truth" for generating lots of the follow on artifacts. For example, write API documentation first and use that to generate client side libraries, an API test suite, and server boiler plate code/skeleton.
In my search for existing projects and approaches, I came across many interesting things.
Swagger: https://helloreverb.com/developers/swagger
API Doc: http://apidoc.me/doc/gettingStarted
Slate: https://github.com/tripit/slate
Write the Docs: http://docs.writethedocs.org/
It was very interesting to read this GitHub post because they presented yet another approach to treating documentation as a first class citizen with different methods to write docs, host docs, and keep the docs updated.
I recently updated the API docs at my workplace to use the Slate tool I referenced above. We manually write docs in a Markdown file, manually use Slate to compile the MD file into HTML, and then manually deploy it to our host. This is approach is incredibly basic and non-scalable, but is light years better than what we had previously, which was API docs directly in the repo's README file.
I hope to learn more about the projects listed above (and many others!) as I explore different approaches for treating docs as a first class citizen and pick the approach which meets the requirements of my current team.
[EDIT] I am also anxiously awaiting a beta invite for http://readthedocs.com