Backward compatibility
A change is backward compatible when everything built for the old version keeps working after the upgrade. An old mobile app can still call the new API, code written for v1 of a library still compiles against v1.4, and the previous release can still read the database after a migration. The opposite is a breaking change. Forward compatibility is the other direction: old code tolerating data written by a newer version, such as a v1 client ignoring a field it has never seen.
How it works
Adding things is usually safe and changing or removing them usually is not. In a JSON API you can add a response field, add an endpoint or add an optional request parameter. Renaming a field, changing its type, making an optional input required or removing an endpoint breaks somebody. Semantic versioning writes this down: a minor release may only add, and anything that breaks callers needs a new major version.
In a database the same rule becomes expand and contract. A schema migration first adds the new column and backfills it, the code moves over, and the old column is dropped in a later release.
For example, an orders REST API returns "total": 1499 and release 3.0 renames it to "amount". The app has 50,000 daily users, and 30% of them are still on version 2.x, which reads total. That is 15,000 people a day who see a blank order total until they update. The compatible version sends both fields and watches how many requests still come from 2.x. It removes total only when that share drops below 1%, which is 500 users a day at this size.
Why it matters in production
Most compatibility bugs come from old and new versions running at the same time. In production they always do: during a rolling deploy, while users delay app updates, and during a rollback, when the old release has to read whatever the new one already wrote.
Library upgrades raise the same question from the other side. When One2N moved its open-source cloudlens CLI from AWS SDK for Go v1 to v2 to support SSO credentials, the first question was whether v2 was backward compatible with v1. The team found that v2 builds clients from a config instead of a session, and each AWS service needs its own client, which meant code changes in every service cloudlens supports.
See also
On the One2N blog
References
