In his book Age of Em: Work, Love and Life when Robots Rule the Earth, Robin Hanson briefly discusses software rot:
As software that was designed to match one set of tasks, tools, and situations is slowly changed to deal with a steady stream of new tasks, tools, and situations, such software becomes more complex, fragile, and more difficult to usefully change (Lehman and Belady 1985)1. Eventually it is better to start over and write whole new subsystems, and sometimes whole new systems, from scratch.
I’m pretty sure this is true. Adapting mature software to new circumstances tends to take more time and effort than writing new software from scratch. Software people don’t like to admit this, but the evidence is clear. Open source software has several high-profile examples.
When it was first written, Mozilla Firefox ran everything in a single process. After the release of Google Chrome, it was clear that a multi-process model allowed for better security and performance. Mozilla developers soon started planning to make Firefox multi-process. That was in 2007.
Almost a decade later, Mozilla finally began rollout of multi-process Firefox. This delay is not for want of trying. The teams at Mozilla are talented and driven. Still, Chrome was written from scratch in far less time than it has taken Firefox to change. There are two main reasons for this:
- Making a single process architecture multi-process means changing a lot of small things. Certain function calls have to be replaced with inter-process communication. Shared state must be wrapped in mutexes. Caches and local databases must handle concurrent access.
- Firefox needed to remain compatible with existing add-ons (or force devs to update their add-ons). Chrome got to create an extention API from scratch, avoiding such constraints.
It gets worse. These constraints are at odds with each other: Overhaul the internal architecture, but alter public-facing APIs as little as possible. It’s no wonder Mozilla needed 10 years to accomplish this feat.
When Apache httpd was first written, it used a process-per-connection model. One process would listen on port 80, then
fork(). The child process would then
write() on the socket. When the request was finished, the child would
close() the socket and
This architecture had the advantage of being simple, easy to implement on many platforms, and… not much else. It was absolutely terrible for performance, especially when handling long-lived connections. To be fair: this was 1995. And Apache soon moved to a threaded model, which did help performance. Still, it couldn’t handle 10,000 simultaneous connections. A connection-per-thread architecture takes 1,000 threads to service 1,000 concurrent connections. Each thread has its own stack and state, and must be scheduled by the operating system. It makes for a bad time.
Nginx was first released in 2007, and its performance advantage was apparent. Years before the release of Nginx, the Apache devs had begun re-architecting httpd to perform better. The event MPM shipped with Apache 2.2 in 2005. Still, there were teething issues. Most importantly, the event MPM broke compatibility with popular modules like mod_php. It wasn’t until 2012 that Apache 2.4 shipped with it as the default.2 While far better than the previous prefork and worker MPMs, the worker MPM didn’t acheive parity with Nginx. Instead, it used separate thread pools for listening/accepting connections and processing requests. The architecture is roughly equivalent to running a load balancer or reverse proxy in front of a worker MPM httpd.3
Python is a nice programming language. It’s expressive, easy to learn (at least as programming languages go), and it’s supported on a wide variety of platforms. But for the past two decades, the most popular implementation of Python has had one major problem: it can’t easily take advantage of multiple CPU cores.
The cause of Python’s lack of parallelism is its global interpreter lock, or GIL. From the Python wiki:
In CPython, the global interpreter lock, or GIL, is a mutex that prevents multiple native threads from executing Python bytecodes at once. This lock is necessary mainly because CPython’s memory management is not thread-safe. (However, since the GIL exists, other features have grown to depend on the guarantees that it enforces.)
Originally, the GIL wasn’t a big deal. When Python was created, multi-core systems were rare. And a GIL is simple to write and easy to reason about. But today, even wristwatches have multi-core CPUs. The GIL is an obvious and glaring defect in what is otherwise a pleasant language. Despite CPython’s popularity, despite the project’s capable developers, despite sponsors such as Google, Microsoft, and Intel, fixing the GIL isn’t even on the roadmap.
Even when given talented engineers, plenty of money, and clear vision, mature software can be extremely difficult to change. I tried to find cases that disproved software rot, but they don’t seem to exist. Robin Hanson asked for counterexamples and nobody came up with anything convincing. There are plenty of old software projects, but they haven’t had to adapt much. I’d love to find good counterexamples, as the current evidence paints a bleak picture for the long-term future of software.