Developer experiences from the trenches
Sun 09 June 2024 by Michael Labbe
tags code
Modern web applications are a façade consisting of many smaller programs and libraries that are configured to run in concert to produce a result. To developers outside of games, this has been acutely obvious for a long time. Games have largely been spared the configuration needs this brings due to a focus on producing a monolithic runtime. However, many modern games ship proprietary logic outside of the code that runs on the disc, such as backend services, so has been affecting games for some time, as well.
At the heart of all this is the need for configuration. Having personally experienced professional devops roles, there seems to be a lack of deep thinking about configuration. This article hopes to inspire deeper thinking about configuration design for programs.
Application configuration is our opportunity to affect runtime state before a program begins its main execution. Static declarations are easily definable, immutable, loggable, can be stored in revision control and can be easily reviewed by a team. Runtime state, on the other hand, is ephemeral and mutable. Through configuration, we have the opportunity wield the runtime state of large, distributed applications in predictable, effcient ways. Most programs do not seize this opportunity.
We treat configuration like it is simple and easy. It is time to start respecting configuration in application design and maintenance.
What is the ground truth configuration for a program? Is it the config file? Not even remotely close. It is the portion of in-memory state that is necessary to cause an (approximately) deterministic, repeatable execution of the program. This is what I call the “ground truth” of an application’s configuration. It usually includes:
Commonly, programs read configuration from many sources. A bespoke search path for configuration, starting from system-wide, and moving in to home directories. Environment variables as an override. Then, command line arguments.
This process differs for each program which is why you’ll see each program document it. Even specifying the system hostname requires addressing multiple files, deprecations and symlinks on Linux.
What happens if there is a system-wide config file but it is not readable because of the permissions of the current user? Pass over it? Throw an error because it exists? This, too, is ambiguous and varies from program to program.
The bottom line is that most programs accumulate a ground truth configuration haphazardly, and then begin executing, perhaps destructively, with no means to review the configuration before it starts.
Writing code is commonly less time consuming than maintaining and debugging the same code. The same is true of configuring software versus troubleshooting it. A misconfigured application produces errors for end users. Many of the configuration formats that are commonly in use (JSON, YAML, TOML) prioritize convenient authorship over unambiguous runtime states. This allows for rapid configuration in exchange for potential risks involving:
Implicit defaults are exceptionally bad when ground truth configuration is not reviewable. You may not even know that you are operating on a bad default, or that an option exists.
Consider:
secrue=true
An insufficiently rigorous program can be misconfigured to breach security without error due to these two aforementioned properties.
YAML, in particular, has a lot of known pitfalls. The point of this article is not to debate popular config file formats. A good developer can overcome YAML’s problems with knowledge and practice, but the problem of contending with underspecified ground truth configuration state is a lifelong drag which can only be overcome through good program design.
JSON, YAML and TOML all have versioned file format specs, but those specifications have no details about how they should affect on-disk performance. Some examples of ambiguities:
Every program behaves differently as a result of this underspecification.
When folks debug a program, they have a mental model of its execution in their heads. Consider:
b = 1;
if (cfg.a)
b += do_optional_thing()
// code continues to do complex things with b
When a developer reads this code, they will either consider b to be augmented by config option a or not. Their mental model of the code necessarily includes this mutating state. Therefore, in removing as much uncertainty as to what the state of a is, is important to someone attempting to ascertain why they are seeing the result of b on their screen.
The rest of this article’s solutions emphasize the need for reducing the size of the mental model necessary for proper configuration troubleshooting.
Which one is right for your application depends on your context. Declarative configuration is a turing-complete program that configures a program. Keeping a mental model of config state requires mentally interpolating variables, simulating loops in your head and jumping through nested function calls.
Imperative configuration lays it all out flat, which lets you see what things are. However, almost everything imperative ends up becoming awkwardly complex when it layers in declarative concepts. See: HCL for_each loops or Ansible adding Python dictionary lookups to YAML files.
A better approach is to think of imperative configuration as a funnel. A data table, perhaps nested, of configuration values can be derived from all sources and fed as input to the ground truth configuration. This table could be declared, or imperatively derived.
The healthy thing is to arrive at a data table of explicit program configuration before core execution of the program starts — an imperative funnel which can be arrived at declaratively.
Schemas are for constraining config file formats, not for constraining ground truth configuration. Ground truth configuration is subjected to underspecified parsers, config file search orders, environment variable and command line overrides and more. Therefore, a schema for a config file does not solve the larger program configuration problem by itself. It doesn’t necessarily hurt it, either, though.
When someone says “we need schemas”, it is useful to explore the root reasoning of that statement before jumping in.
In structured languages, a ground truth configuration can be typed and could be used to produce a schema. The right choice is to keep as much ground truth about the program’s configuration in one reviewable structure.
Most importantly, provide the best tooling for your in-context situation to edit and review the program’s ground truth.
Configuration has a way of becoming layered, especially in devops. For example:
values.yaml file overrides the Helm Chart for a forked Docker imageIn this case, we reap the benefits of a highly-available program that is configured to our specification, compiled by a program provider else and made to work for our purposes. This reduces a large one-time up-front cost. However, we incur a cost of five configuration files, implicitly depending on values from each other to derive whole program state. This has a drag on efficiency for the lifetime of the product. This is an important tradeoff in where you spend effort — one to commit to consciously.
Each small program comes with its own configuration files and state. Since your application consists of multiple programs, you end up producing configuration files that require values that are similar between them. This is brittle when changed.
Further, if there are multiple versions of an application (eg: test and production), there is an n by m problem, where each dependent configuration must exist for each version of the application.
This can be addressed by having a single source of truth for each application configuration, used to produce the smaller configurations for each program.
For the remainder of my life I will depend on large applications that are made up of many small, configured programs operating in concert. Making configuration correct, safe and expressive is an opportunity to wield large numbers of these programs with minimal cost and overhead.
Many of these smaller programs came from programming cultures that emphasized getting something up and running over long term maintainability, loose coupling and quick-and-dirty scripting. As computing complexity increases, it is my hope that the sort of rigorous values that spurred the creation of languages like Rust are applied to configuration management.
