Software That Heals Itself

While you are reading this line, your heart has adjusted its rhythm to the fact that you are sitting down, your pupils have adjusted to the lamp, and your kidneys have filtered another portion of blood, and not one of these decisions has disturbed your consciousness. The autonomic nervous system works your whole life without days off and troubles its owner only when it cannot cope on its own: then he feels pain, thirst or fear and starts doing something.

In 2001 IBM’s head of research, Paul Horn, proposed building computer systems the same way and borrowed the very word from the physiologists: autonomic computing. Systems were supposed to configure, protect and heal themselves. Over the next twenty years computers did learn a great deal, including how to restart after a crash, and almost nothing beyond that. To find the bug in the code that brought a server down and fix it, you had to understand the code, and only people understood it, and they were woken in the middle of the night for the purpose.

---

If you recall Reznik’s method from the first chapter, you can see what was missing here. Servers had had selection for a long time: monitoring saw that something had broken, and tests could say whether it had been fixed. What was missing was a generator capable of proposing any kind of fix at all for an error nobody had described in advance.

Facebook built one first. In 2018 the company ran a system called Sapienz, which tirelessly poked at the company’s mobile apps looking for actions that made them crash. A second system, SapFix, was attached to it: on receiving a crash report, it wrote a patch by itself, checked by itself that the crash was gone and nothing else had broken, and put the finished fix on an engineer’s desk. It still wrote in a simple way, from templates and small edits to existing code, but during a ninety-day trial period it proposed 165 patches for 57 crashes, and on average 69 minutes passed from the detection of a crash to a ready fix. This was the first night shift of Reznik’s monkey: it typed out variants, and the tests decided which of them to release into the world.