Schema drift doesn't announce itself. It accumulates through dozens of small migrations, each of which seemed safe at the time. A column added here, a type changed there, a foreign key quietly dropped because it was slowing down a batch job. Then one day, a query that worked perfectly for three years starts returning wrong results, and the investigation begins.
How drift starts
The earliest stage of schema drift is additive. Teams add columns to accommodate new features, often without communicating the change to every consumer of that table. A data warehouse sync that reads all columns starts pulling in nulls it doesn't know how to handle. An ORM that maps results to a struct gets a field it doesn't know about. These failures are usually silent — the data just disappears into a default value somewhere downstream.
The destructive phase
The more dangerous phase is when drift involves removals or type changes. A column gets renamed as part of a refactoring project. Three services that referenced the old name stop working. A varchar column gets extended from 100 to 500 characters to accommodate a new market — and a downstream system that was assuming max 100 breaks an index it uses for substring matching.
The worst incidents come from cascade effects: a foreign key gets dropped to unblock a migration, and three months later someone runs a report that relied on referential integrity guarantees the database no longer enforces. The data looks fine. The results are wrong.
Detection is the hard part
Most teams detect schema drift after it's already caused a problem. The right approach is continuous schema comparison: snapshot your schema on every deployment, diff it against the previous state, and surface any change that wasn't explicitly declared in a migration file.
What Dreambase adds is the downstream impact analysis: it doesn't just tell you that a column was added — it tells you which queries are affected by the change, which indexes might need updating, and whether any existing queries will behave differently against the new schema. Schema drift, caught early, is a diff. Caught late, it's an incident.