Best tip for docs in small teams: Have a server running in the local network that pulls the latest develop branch, renders it and serves it on the network for each project. For microservices I would recommend API Blueprint, for normal libraries, whatever is the language standard. Really cuts down the friction of getting to the docs, since you don't have to do all those steps on you local machine anymore whenever you need to access the documentation of an internal project.
Next best thing would be building Dash docfeeds for internal projects.
in addition to tools for documenting interfaces (i.e. APIs), I'm interested in tools that document behavior (both desired & observed) and integrate it back to codebases.
In type safety terms, this is like the difference between function signatures that verify inputs & outputs vs something like coq/TLA that can verify more advanced properties of your program.
Configuration is particularly difficult to document in 2017 -- if you solve that problem you fix a lot of migraines.
Some of my colleagues in Pivotal Labs swear by Spring Cloud Contract[0], but I haven't used it in anger myself. It seems to be what you want -- an independent statement of protocol that's shared between server and client codebases, with code-generation to help each.
> I'm interested in tools that document behavior (both desired & observed) and integrate it back to codebases.
> Configuration is particularly difficult to document in 2017 -- if you solve that problem you fix a lot of migraines.
Could you expand on those points a bit more? As for the first one, it sounds vaguely like https://deckard.ai, but I feel you could also mean something entirely different. (Disclamier: I know the founders of Deckard very well.)
(1) contractual behavior (what you want/expect) can be marked up in comments that generate docs but are also readable inline in the code.
And (2) actual behavior in the wild from monitoring, tracebacks, profiling. Would be cool if IDEs could drop that next to a function so I can ask 'is this slow', 'does this ever get called'.
Re: configuration I'm saying it's not always clear what a config setting does, and esp. how related config settings combine. Having standard types for configuration would let us generate docs for which part of the program are affected by a config setting; at least provide a starting point for understanding the flag's behavior.
I used to feel that way, but now generally don't. Most documentation is effectively duplication of something already expressed in the code. That means that whenever you add docs, you are multiplying the effort needed to change things in the future. You also create opportunities for the various expressive duplications to diverge, causing untold confusion.
Now documentation is always a last resort for me. That's a shame, as I enjoy writing. Instead I try to put knowledge in existing places, like unit tests, acceptance tests, method names, object names, code structure, file layout, in-app text, and all of the other little things I'm already doing along the way. Only if I can't put something there will I write docs, and then reluctantly and with a feeling I probably missed a chance to minimize duplication.
The only documentation I still write without reservation is stuff that obviously doesn't have to keep up with the project as it changes. E.g., personal journals and project blogs are great, in that everybody understands those are out of date right after they're written.
I don't think the things you listed, even when combined, are sufficient substitutes for well-written documentation.
The fact of the matter is that, for humans, plain English is always much easier to understand than computer code. The more complex the code, the more true that is going to be. You can use all the friendly function names and object names and file layouts you want. At the end of the day, it is not going to be enough for someone who is not familiar with that part of the system to understand it quickly and to the appropriate level of detail. This is especially true if you have junior-level or intermediate-level people on your team. It is also true if you ever end up coming back to some piece of code you yourself wrote a year ago.
Furthermore, documentation can provide additional value by including examples of code usage in different settings that you may not have covered in your test cases due to not being immediately relevant. It can contain links to other websites and have embedded images of system diagrams or data flow. It can also describe third-party systems your code has to interact with and various bugs and unconventional behavior that may exist in those systems that your code has to account for.
Yes, documentation involves more upfront effort. IMO the correct way to deal with that is to include that in project estimates, and to not mark tasks and features to be "done" until their documentation is written. Same with revisions and bug fixes. I understand the feeling that it duplicates logic already expressed in the code. But look at it this way: code is for computers. Documentation is for humans.
I should be clear that I like good documentation, and will happily write it when necessary. It's just that over the years, I've a) found a lot of ways to make it unnecessary, b) grown to despise useless or out of date docs, which is the bulk of what I see, and c) grown very tired of writing documentation that people end up not reading.
> [...] plain English is always much easier to understand than computer code [...] This is especially true if you have junior-level or intermediate-level people on your team.
This is why I'm a big fan of collective code ownership, pair programming, and frequent pair rotation. That has a variety of benefits. It also has substantial advantages over documentation. A big piece of which is that the need for documentation is always speculative, whereas answering a question always works from a confirmed need.
> It is also true if you ever end up coming back to some piece of code you yourself wrote a year ago.
This is what I use tests and journals for. Well-written tests are essentially machine-verifiable documentation. They tell me what the intent of the system is. If I want to know the history of the intent, then I turn to my journals.
> But look at it this way: code is for computers. Documentation is for humans.
I couldn't disagree more. As Martin Fowler says: “Any fool can write code that a computer can understand. Good programmers write code that humans can understand.”
Code and tests are primarily means of communication between developers. They have the constraint that they should result in a running system. But we stuff most of the work of that into compilers, optimizers, and run times. Our tools are -- and should be -- optimized for human productivity. And since programming is mostly a team sport, the biggest value of our tools is in how they enable that teamwork.
The biggest value in documentation to me has always been efficiency in understanding: that usually means documenting at the architecture / "components as black boxes" level.
I could trace through to see how someone glued something together without having a hint what I'm looking for, or I could get the architecture gist and then generally know where the thing I should be looking at is.
To me, this delivers two benefits. (1) Reduced redundancy / potential for divergence (as you mentioned above), as you're not re-documenting code but rather documenting at a higher level. (2) Reduced churn, as architecture changes more slowly than code, and generally an arch change will have some organizational process before being made (where docs can be updated).
Disclaimer: This is coming from a place of diving into a lot of client systems with really poor or utterly lacking documentation.
Sure, but no documentation here will be as efficient as an expert having a conversation with you and answering your questions. I'd much rather have a few whiteboard conversations and maybe some pair programming time, and that is cheaper and easier for everybody then speculative production of documentation.
If the choices are only "shitty code with no docs" and "shitty code with good docs", I'll definitely take the latter. And if I'm rolling a team off a project, I'm happy to take some time to write some high-level docs for the people who will next pick it up. That's a one-time expense with documentation written for a defined audience.
But when I'm picking up a system and the original team is gone, I'd much rather have good code and good tests over good docs.
And honestly, I'm not sure I've even inherited a system with good docs. I have seen things that managers probably thought were good docs. And I've seen things that the developers intended to be good docs, and might have even been good once upon a time. But by the time I get to them, it's either the "we were told to write docs" garbage that is useless, or stuff that's so far out of date that it's just as often misleading as helpful.
Docs/code impedance mismatch is a problem that a docs tool needs to solve to stay relevant. Good docs will resemble or contain a test suite.
I agree with your point about blogposts -- publish date is like a sell-by date, makes it clear when the post has gone stale.
Getting insiders to read docs is also hard. Not every piece of code is worth documenting. But I think highly critical and highly reused code is worth documenting, in part for programmers and in part so that non-coder stakeholders can understand the details of the product (i.e. 'sell what you make' if they're on the sales team).
> Docs/code impedance mismatch is a problem that a docs tool needs to solve to stay relevant.
It's not an impedance mismatch (which results from difference in the structural representations supported by different systems), it's just a content mismatch. And it's not a tooling problem, it's a process problem. Documentation is the single source of truth for intent, and development process must respect that.
This becomes more and more important the more people who aren't developers need to understand the logical function of the system.
We do the same - it was surprisingly simple to implement -
- CircleCI listens across all GIT branches and injects $CIRCLE_BRANCH into ENV
- A build step populates a vhost template to point "$CIRCLE_BRANCH.qa.bofh.com" to "/var/www/$CIRCLE_BRANCH/current"
- Another build step creates "manifest.json" with $CIRCLE_BRANCH and $CIRCLE_BUILD_NUM
- The vhost config and the manifest are added to the release tarball/Docker image
- The tarball/image gets shipped to the staging server
- (we ship to S3/Docker, then pull because of security concerns. But can ship inside the build process)
- In staging, a script reads manifest.json and moves content to /var/www/$CIRCLE_BRANCH/$CIRCLE_BRANCH
- The provided nginx vhost is copied into a standard dir with other vhosts
- A symlink of "/var/www/$CIRCLE_BRANCH/current" is forced to point to the fresh release
- nginx -s reload && profit
You can use a wildcard SSL to cover subdomains or add a step to trigger LetsEncrypt certbot as required.
This method works equally well with Docker images - just requires an interim step to launch the container. In staging, we use a naming convention to launch on predictable ports for every app. First 2 numbers are app specific, second 3 numbers are build-specific. For example, "25$CIRCLE_BUILD_NUM" (where $CIRCLE_BUILD_NUM is just the last 3 numbers of the build).
How do you handle the database? Do you reuse the one on staging or create a new instance with prepopulated data? If the former how do you deal with migrations and schema edits?
Yes, the db has been the most challenging aspect for us. We have 3 situations -
1. "Common baseline". With a relatively stable product, most branches (as in ~51%) do not impact the schema. For testing / QA purposes, these share one central QA db and pollute each out. Turns out, a lot of the times this is quite ok because the PR is about how the data is displayed, or improved logging, or UX change, or a security layer or anything else other than core domain knowledge - they don't care for the data that much.
2. "I'm special". Some branches do modify data (whether the format or the structure). To handle these, the manifest.json file has an option to request a separate database. If present, the rollout script will do "pg_dump + copy" of the shared staging DB, and duplicate it into "qa_$BRANCH", then update the config file (or .env for Docker) with the appropriate connection value. Additionally, it will all *sql files in a dir specified in manifest.json against the clone DB. This is done on every release, which does get annoying by resetting the qa data (we could add another manifest switch here I). On the upside, it forces you to codify all data migration rules from the start.
3. "I am very special". Some changes transform data in a way that requires business processing and cannot be done with easy SQL. Sorry, out of luck - we don't automate special cases yet. The developer has to pull the QA database to localhost, do his magic, and push it back. Not ideal, but hasn't caused any problems yet. If ain't broke...
We have a job every morning that clones the current production db instance and then deploys it onto a staging instance. All PRs with schema changes will be applied to this instance when merged. If something goes wrong, we just redeploy the saved image.
In a previous job, I built (what sounds like) a similar system. From a web UI, coworkers would see a list of recent commits, from which they could fire up an instance of the web app on demand. This created a new Docker container running the server that they could navigate to. A unique port distinguished each instance. Git tags indicated particularly important commits, e.g. new features that required more extensive testing.
Docker worked great for this task, allowing instances to be created and destroyed quickly. In a fresh instance of the web app you could create new accounts and import data then throw away the container when you were finished or wanted a fresh start. It worked especially well for non-technical people to see the latest changes as soon as they were pushed.
We're doing the same for microservice deployments to kubernetes, using Jenkins' multibranch pipeline - Jenkins just watches the git repo, and any commit to a branch that contains a Jenkinsfile gets built. At the end of the build process, we publish a docker image tagged with the branch and build number, and then use helm to upgrade or create a deployment in a kubernetes cluster, tagged to run the just-published docker image. Our cluster's running in AWS and has a package deployed to it called 'area53' which can set up route 53 DNS records for kubernetes services automatically.
Upshot of all that is that if you create and push a new branch called 'foo', a few minutes later that branch's code is up and running in AWS, under the name 'service-foo.dev.example.com'. We can then automatically run integration tests against that endpoint.
Best part is that the build process defined by the jenkinsfile, and the deployment defined by the helm chart, all live in the source, so you can play with them on a branch. Want to add another testing step to the build process? Branch, add it to the jenkinsfile, and your build pipeline change gets executed just for your branch, without stopping other branches from building. Once you're happy with the addition, merge the change and now every future branch gets that extra deployment step. Likewise if you modify the deployment chart, you can test it in a branch.
Red Pen for annotations. (Does't integrate with anything and that pisses me off -- how freakin' hard would it be to build web hooks so it could tie into Slack?! But on the whole it's got an easy to use interface.)
Thanks for the list, it’s quite good! However, our point is to not use third-parties for things where we don’t need it.
The point of the article is to show that you can use Github solely for project management without a third-party service.
But I completely agree with you: If you don’t like the little extra-work we do to achieve that with just github.com, it’s probably right to add another tool like ZenHub or waffle.io on top of it.
Making your toolset work for your team is the most important thing. For us, it’s enough — and maybe for a few others as well. That’s why we shared our approach :)
Gonna rant a bit... GitHub boards cost time. No one should use them. They're just an inferior option to ZenHub (or even Waffle or Asana or any number of other "board" interfaces you can tack on to GitHub). I've wasted my team's time on GitHub boards... everyone quickly demanded we go back to ZenHub.
Most projects have multiple repositories, right? But GitHub boards have it so that each board is based on one repo. Stupid. You want to see a project view of all your repos at once... front-end, back-end, deployment, whatever... not 4 or 5 separate views on it.
I feel like GitHub REALLY dropped the ball on not having a board forever, then putting out such a "beta-feeling" board. They should have just bought ZenHub -- still time... but literally any tool out there makes the boards work better than default GitHub.
ZenHub is free for small teams, and I'd argue that anyone can afford $5 / month / user. Budget $150 / month / user for any team for tools -- seems about right.
Time tracking seems very microcontrolish for such a small team, and doesn't really seem worth the overhead. If the team includes freelancers, that's of course a different thing.
You track time so you have metrics and can back up your opinions with facts.
"We need more developers..."
Ok, show me how much work we are doing... and where the hours are being spent... so we can figure out how to optimize and recruit more folks.
There isn't a team out there that can afford not to track how they spend their time. Not tracking hours is like not brushing your teeth... It's such a simple thing, takes 3 minutes a day.
We've use Pivotal Tracker before, and it always felt challenging to have more than a basic description/discussion (maybe that's the point), but most tools like Jira, Github Issues, Trello, etc, facilitate a larger view.
One of the basic notions in the school of thought that Tracker comes out of is that team interaction is valuable and to be encouraged. So making a tool that makes it easy for people to not talk undermines the broader goal.
One of the basic insights of the Agile movement (R.I.P.) about Waterfall is that processes structured around documentation rather than collaboration have a lot of subtle bad effects that gradually destroy characteristics you'd like your teams to have. E.g., responsiveness to change, systemic effectiveness, resilience to failure, low overhead, ability to ship frequently, ability to deliver customer value.
That's why when I set things up I generally drive things off of index cards. [1] Those are obviously insufficient, which forces people to discuss and collaborate. (That's in contrast to more voluminous documentation, which is subtly insufficient.)
Tracker is very opinionated and is deliberately trying to not be the be-all and end-all for software projects. It's really closely tuned to how Pivotal works: XP with Lean trimmings in small teams with a product manager, designer and engineers.
The downside is that because it is deliberately limited, now and then you will find that you sorely miss something. Often those missing features are recreated with conventions around tagging or release markers. But tagging conventions are not the same as a first-class feature, for good or ill.
JIRA is massively more featuresome and flexible. These days it's grown into a workflow middleware which comes with a bug tracker as the first application installed. For some organisations that will make more sense.
Disclosure: I work for Pivotal, but not on Tracker.
It’s a combination of a few things. First, when we started doing calls together Slack calls weren’t available yet for multiple people.
In our experience, Slack calls often fail for a variety of reasons. Other than that, we don’t want to shift too much data to Slack and since it’s pretty unclear how their calling technology works and appear.in uses WebRTC, we rather use another service here than a service where we don’t know much about it but only that calls aren’t too reliable. But as said, we haven’t bothered to check the group-video calling yet there. We’re pragmatic. Appear.in works in any browser (Slack calls don’t) and therefore… why should we change a well working system?
Slack’s audio/video conferencing feature sometimes double-connects an user and frequently connects without sending audio traffic either way.
My team has been using ScreenHero prior to Slack buying them out. ScreenHero also supports audio calls without screen-sharing, and its audio calls have not been problematic for us.
Making docs sexy would be a big quality of life boost for programmers and the people who love them (i.e. project managers and users of software).