The ATLAS deprecation policy is short: no breaking changes within a major version, and 36 months of support after a successor major version ships. That is the whole thing.
It is deliberately dull. An API deprecation policy is a promise about the future made to people who are deciding whether to build something expensive on top of you, and the only useful property of such a promise is that it is simple enough to be verifiable.
The cost is real. v2 is still supported and will be until September 2026, three years after v3 shipped. Maintaining it constrains what we can do in the shared layers underneath, and there is a recurring internal argument about accelerating the end date.
We have not, and I do not expect to. The argument that ends it is always the same: a customer who built an integration in 2023 on the strength of a written commitment should not discover that commitments are negotiable when they become inconvenient for us.
What we do instead is make migration cheap. The v2-to-v3 migration guide is thorough, the SDKs support both, and there is a compatibility shim for the three endpoints that changed shape most.