[diegoveralli]

At my previous company there were several attempts to build declarative migration systems, one of which was more or less successful. They were all a lot more modest in scope and goals than Atlas.

The motivation is very straight forward: if you work in data management with different schemas, you often see similar migration patterns, so it makes sense to try to build abstractions representing these patterns.

If you couple this with the usual problem of the migration / schema duality, and some people who want to have a replica of the current state of the schema in the codebase in some way ("desired state"), you can see how we end up with schema-diff migration tools.

Whether these tools work well or not I think depends on the rate of schema changes vs size of the data store. The larger the dataset, the more specific you need to be with the migration strategies, so if you're not making schema changes very often, using a "clever", schema-diff migration tool adds risk for little benefit, and you might need to bypass the tool on a regular basis.

I don't work in a scenario in which such a tool would bring benefits, but there are projects that have frequent schema changes and data sizes small enough that you don't need to do anything special to maintain availability during migration.