Rollback

Software is constantly being changed. Someone builds in a new feature, deploys the change to the servers, and suddenly the app crashes for thousands of users. A rollback means: the change is undone and the previous state is restored. The program then runs again just as it did before the update, with all the old bugs, but at least without the new one. The idea is comparable to the save function in a video game. Anyone who gets stuck loads the last save point and is immediately back in the game.

Why reverting is faster than repairing

When an online service goes down, every minute costs money and trust. Finding a bug in someone else’s code can take hours. A rollback often takes less than a minute. That’s why the rule in many tech teams is: roll back first, then calmly search for the cause.

A rollback is also the precondition for teams being able to change things often in the first place. Large providers roll out hundreds of changes every day. That’s only justifiable because each individual one can be undone in seconds in an emergency. Without this safety net, companies would update less often but in huge bundles instead, and it’s precisely such large-scale changes that tend to go wrong especially often.

A common misconception is that a rollback is a sign of sloppiness. In practice, the opposite is true. A team that can roll back without a fuss has its systems well under control. The situation only becomes dangerous once a rollback is no longer technically possible.

What happens technically during a rollback

For a rollback to work, the old state must still exist. Program code is therefore stored in a version control system, a kind of archive that keeps every previous version. During a rollback, the bug isn’t deleted—instead, the previous version is loaded onto the servers again. Often both versions even run in parallel for a short time, so that traffic can simply be switched back.

In databases, rollback means something slightly different. There, several changes are grouped into a transaction, a work package that must apply either entirely or not at all. In a bank transfer, the debit and the credit belong together. If the process is interrupted midway, the database performs a rollback and reverses both steps. Otherwise money would have vanished.

Things get difficult when an update has changed the structure of the stored data. If a customer table has been restructured and then filled with new entries, the old program version no longer understands this data. Such changes are therefore called hard to roll back. Good development teams plan them in small steps, so that each step remains reversible on its own.

Rollbacks in apps, games, and AI models

On your own phone, you rarely experience rollbacks, because app stores usually only offer the latest version. The process becomes visible in games. After a server error, it’s sometimes announced that the game state will be reset to a point three hours earlier. That means the items collected during those three hours are also gone. This is a rollback of the game data.

In news reports about tech companies, rollbacks turn up after outages. Typical phrasing: A faulty update was rolled out and reverted again a short time later. Operating systems know this too, for instance when Windows automatically restores the previous state after a failed update.

Rollbacks also exist for artificial intelligence. Providers regularly replace their language models with newer versions. If many users complain that the new version gives worse or stranger answers, the providers temporarily switch back to the older model version. Because models are so large, companies deliberately keep older versions operational and ready for exactly this purpose.