What are the major challenges of blue/green deployments with relational databases?

WHAT THIS TESTS: This question tests whether you understand that relational databases are shared infrastructure across blue and green environments rather than per-environment state. It probes your ability to manage schema evolution safely when two application versions must coexist during a deployment window, and whether you know how to decouple schema changes from code changes to preserve rollback capability.

A GOOD ANSWER COVERS: A strong candidate first states that both environments must share the same data store so that green can take over production and blue remains rollback-ready with current data. Next, they explain the decoupling pattern: schema changes should happen separately from code deployments. The first phase uses additive changes before the cutover, such as adding new tables, columns, or relationships, while ensuring the old application version can still read and write safely. The candidate should mention that applications must ignore unknown fields during deserialization to avoid runtime errors, and that triggers or backfill jobs may populate new structures from old-version writes. The second phase applies deletive changes only after the green environment is stable and the old version is no longer needed, removing obsolete columns or merging entities. They should emphasize that both phases require backward compatibility: the old code against the new schema for additive changes, and the new code against the old schema if schema lags behind code.

COMMON WRONG ANSWERS: A major red flag is proposing completely separate databases for blue and green without a viable synchronization or rollback strategy. Another weak answer treats schema and code as a single atomic deployment, which eliminates safe rollback because reverting the code leaves the database in an incompatible state. Candidates who suggest destructive schema changes before cutover, such as dropping columns the old application still needs, demonstrate a lack of appreciation for the overlap period. Ignoring the need for application-level tolerance of extra fields is also a subtle but important miss.

LIKELY FOLLOW-UPS: An interviewer may ask how you would handle a long-running migration on a large table without locking, or how you would verify that the old application truly has no latent dependency on a column you plan to drop. They might also ask about data consistency requirements when sharing NoSQL stores versus relational databases, or how to automate the additive and deletive phases in a CI pipeline.

ONE CONCRETE EXAMPLE: Suppose you need to replace a monolithic user profile table with normalized contact and preference tables. In the additive phase before cutover, you create the new tables and add foreign key columns to the original table, then deploy a trigger that populates the new tables on every update from the current application. The green environment is deployed and validated while the old application continues to run, ignoring the new columns. Once you confirm no rollback is needed, you enter the deletive phase: remove the old columns and drop the trigger, leaving only the normalized schema for the new application version.

Source: docs.aws.amazon.com