The Database Architecture Decision That Took Us 18 Months to Undo

Three years ago, we made a database architecture decision based on what was popular rather than what fit our use case. We chose a document database for data that was fundamentally relational. We paid for that decision for eighteen months before we finally migrated.

Why We Made the Wrong Choice

The team was excited about the flexibility of schema-less storage. We were tired of migrations. We believed — incorrectly — that our data was less relational than it actually was. We underestimated how often we’d need to query across what turned out to be relationships.

The Symptoms That Should Have Told Us Sooner

We started denormalizing aggressively. We were duplicating data across documents to avoid “joins” that the document model doesn’t support natively. Query code became increasingly complex as we worked around the model instead of with it. Performance degraded predictably as data volume grew.

The Migration

The migration to PostgreSQL took three months. We ran both databases in parallel for six weeks. Performance improved immediately on read-heavy queries. Write performance was equivalent. The query code became dramatically simpler. We should have made this choice initially — the document database was solving a problem we didn’t actually have.