Configuration is an API, not an SDK
Published on , under Programming, tagged with python, best-practices, tools, configuration and architecture.
Configuration is every setting that needs to change based on the environment where the app is executed. It helps you preset the state of your app, without having to change the code. This is why it is important to provide a clear separation of configuration and code.
Why are executable config files a bad idea?¶
Configuration is just another API of your app.
It might be used by people with or without technical skills, that install or run your app.
Executable files can be used as config sources like
.vimrc for Vim,
Vagrantfile for Vagrant, etc, but that approach has some drawbacks.
First, your users now need to learn a new programming language, just to configure your application. Some apps (like the suckless bundle) go as far as requiring you to patch and compile your app to change it’s configuration.
And second, your configuration is no longer hierarchical. Your application cannot extract configuration from different sources by executing different files, because you cannot know in advance what is being executed. A programming language has control structures, can make calls to the internet, etc, so you can't known in advance the output you will get. Your configuration is no longer deterministic.
If you need to let users alter the behavior of your program, that is not strictly configuration, there is a better solution: a plugin system.
The best way to think of configuration is as a set of key/value dicts that need to be merged into a single config dict. No need to get fancy with yet another DSL.
What configuration format/sources should I use?¶
The short answer is: it depends, but probably more than one.
For example command line args are great to explore an app from the shell, and
tinker around it's possibilities. Many apps allow for both
-s short and long
--more-verbose arguments, for those who don't want to be guessing later.
When you already know what you want it would be great to set some defaults in a configuration file somewhere. Yes, you can always set an alias, like many distros do:
$ type ll ll is aliased to `ls -alF`
Alternatively, config files are naturally better documenting and declarative. Some file formats allow for comments and are great as starter templates to build upon. With files it's easier to make diffs and track changes over time.
Environment variables are the simplest way to configure apps/scripts. You just need to populate a global dictionary before executing it. There is no need to parse files or command line arguments. This practice is very common in cloud providers to inject credentials or connection strings for your app. Another example is qutebrowser; it allows the user to write userscripts and passes many environment variables to share the browser state before executing them.
On the other hand, environment variables shouldn’t hold sensitive data, there are potential security issues regarding accidental leaks via logging and error reporting services or child process inheritance.
Well designed applications allow different ways to be configured. Each having it's pros and cons.
If your app is a long running process, like a webserver, you can issue a
SIGHUP signal so that it reloads it's config from files. Env vars and
command line arguments cannot be easily changed from the outside after the
What matters here is that env vars, command line args and files (
.toml) are the most standard formats to configure apps. You should be able
to configure any app with standard unix tools and a text editor. Stick to these
formats when possible.
But what happens if a setting is passed as command line argument but also exist in a config file? Which source is more important?
A proper settings-discoverability chain goes as follows:
- First command line args are checked.
- Then environment variables.
- Third, config files in different directories, that also imply some hierarchy. For example: config files in
/etc/myapp/settings.iniare applied system-wide, while
~/.config/myapp/settings.initake precedence and are user-specific.
- Finally you fallback to hardcoded constants as defaults.
Some of these sources may not be present or relevant to your app. But each one of these sources of configuration needs to be properly collected and overwritten with an explicit level of hierarchy. This gives more flexibility to your users, so they can run your app/script in the cloud or in their multi-user computers, using systemd or docker, etc.
Using the right tool for the job¶
Your application will require other tools, like compilers, installers, package managers, process supervisors, etc. These tools solve different problems of the architecture of your software.
Configuration is also part of that architecture. Along with your program you will have to ship the configuration artifact. But this is not an issue that only comes up when you make a new release. When you are developing you are also making tiny releases on your laptop.
So this raises the need for some tool to provide your code with the right configuration, in all its stages. Some of these tools are only used when developing or in production, ideally both envs match.
There are many tools for managing configuration. For example, direnv and envdir load environment variables from directories and files. Systemd units have a section to list them or point to an environment file populating the environment. Ansible includes a templating language to generate configuration files and place them anywhere in the system. Ansible Vault can be used to provide the app with encrypted secrets.
Some tools like consul-template or envconsul can also listen to
changes in the configuration and issue a
SIGTERM or (even better) a
signal to your long running process, so that it can pick up new config values
No matter which tool you choose to manage, generate and populate the configuration artifact, the code of your app should only care about reading files, env vars and/or cli args.
Naming conventions and namespaces for settings¶
There happen to be some formatting conventions for configuration parameters based on where they are set. For example, it is common to declare environment variables in uppercase:
$ DEBUG=yes OTHER_CONFIG=10 myapp
Since the environment is a global and shared dictionary, it is a good practice
to also apply some prefix to each setting to avoid collisions with other known
TZ, etc. This prefix works as a namespace for your
$ MY_APP_DEBUG=yes MY_APP_OTHER_CONFIG=10 myapp
but if you were to set this config in an .ini file, each setting should probably
be in lower case, the namespace is implicit in the file path, i.e:
[settings] debug=yes other_config=10
Command line arguments have yet another conventions:
$ myapp --debug=yes --another-config=10
You probably noticed that the debug setting is a boolean value. These flags
should accept different inputs like
It is important to be consistent in naming these variables, but to respect the conventions too.
A solution for the working dev¶
If your app is written in Python, it's your lucky day.
Classyconf is a library that helps you with a configuration architecture for perfectionists with deadlines™, or for the working dev if it sits you better.
The good practices that it suggests have an agnostic approach to configure applications, no matter if they are web, CLI or GUI apps, hosted on the cloud or running in your desktop.
You can find out more documentation at Read the Docs website, but here is a preview of how to use it.
from classyconf import Configuration, Value, Environment, IniFile, as_boolean, EnvPrefix class AppConfig(Configuration): class Meta: loaders = [ Environment(keyfmt=EnvPrefix("MY_APP_")), IniFile("config.ini", section="myapp") ] DEBUG = Value(default=False, cast=as_boolean, help="Toggle debugging mode.") DATABASE = Value(default="postgres://localhost:5432/mydb", help="Database connection.")
As you can see is very declarative. It uses the concept of loaders, which collect settings from different sources and merges them in the right order.
This class can be extended according to different environments or needs.
class TestConfig(AppConfig): class Meta: loaders = [IniFile("test_settings.ini"), IniFile("config.ini")]
overridden at runtime
>>> dev_config = AppConfig(loaders=[IniFile("dev_settings.ini")]) >>> dev_config.DEBUG True
accessed (and lazily evaluated) as dict or object
>>> config.DEBUG False >>> config["DEBUG"] False
introspected and iterated
>>> for setting in config: ... print(setting) ... ('DEBUG', Value(key="DEBUG", help="Toggle debugging on/off.")) ('DATABASE', Value(key="DATABASE", help="Database connection."))
The idea of this blog post was to highlight the importance of thinking configuration as an API, to follow best practices and conventions and the need to be flexible for different sources of configuration.
I've shamelessly introduced classyconf as a way to address all this topics, but I'm sure that similar libraries can be found for other languages as well.