Glosarium
Contents

Architecture & System Design › Cloud Design Patterns

Busy Database

Pushing too much processing into the database.

Backend Engineer · Senior

Also known as: busy database, database as integration, shared database bottleneck

The busy database anti-pattern funnels every integration through one shared store: services read each other’s tables, batch jobs hammer production, analytics query live — until the database (and every team sharing it) contends on locks, I/O and schema changes. The database becomes the coupling point, the bottleneck and the blast radius simultaneously.

services + jobs + analytics → ONE database (locks, slow queries, migration freezes)

Symptoms are unmistakable: schema changes needing cross-team sign-off, one workload’s queries starving others, “don’t touch the database” deployment freezes, and scaling debates that are really about sharing. The cures separate concerns: private databases per service, read replicas/warehouses for analytics, queues instead of polling tables.

The classic mistakes:

  • Integration via shared tables. Services reading/writing each other’s tables couples schemas invisibly — every column change breaks unknown readers. Integrate through APIs/events, never tables.
  • Analytics on production. Ad-hoc analytical queries on OLTP databases contend with live traffic unpredictably. Replicate to warehouses; analyse there.
  • Batch jobs inline. Nightly jobs locking live tables stall users; schedule, isolate (replicas), or stream instead of batch-scanning production.
  • Schema change paralysis. Shared ownership making migrations multi-team negotiations signals the database has become organisational coupling. Split ownership with the data.
  • Scaling the shared box. Bigger hardware postpones the contention without removing coupling. Partition workloads (and data) instead of upsizing shared fate.
  • Missing the people problem. “Busy database” is often Conway’s law in SQL — teams sharing stores because sharing services felt harder. Fix ownership boundaries, not just queries.
  • Monitoring queries, not contention. Slow-query logs miss lock waits and I/O contention from other workloads. Measure contention (locks, buffer pressure), not just speed.

The escape: private databases, analytics off production, queues over polling, ownership with data. Shared databases integrate teams by accident — integrate deliberately instead.