Set-based design delays architectural commitment by exploring several options in parallel. In software, teams use short, time-boxed spikes to answer shared questions about competing approaches, then let evidence guide the decision. section 11.1 section 11.10
Shape the exploration around the system’s needs:
- Choose candidates using team knowledge, published benchmarks, and operational constraints. section 11.1
- Ask what must be true and what could go wrong; evaluate every option against the same requirements. section 11.1
- Commit at the last responsible moment: before delay eliminates an important option, keep learning; afterward, decide rather than block progress. section 11.10
Choosing notification delivery
A team compares Redis Pub/Sub, Kafka, and a PostgreSQL outbox in three-day spikes. The options differ in throughput, offline-message retention, operational complexity, team familiarity, and p99 latency—trade-offs that an early commitment could hide. section 11.1
Table 11.1 Notification delivery spike comparison
|
Question
|
Spike A (Redis Pub/Sub)
|
Spike B (Kafka)
|
Spike C (PostgreSQL Outbox)
|
|
Can it deliver 5,000 notifications/sec?
|
Yes, very fast in-memory
|
Yes, with partitioning
|
Moderate, needs polling or NOTIFY
|
|
What happens when a consumer is offline?
|
Messages lost
|
Messages retained on disk
|
Messages retained in the outbox table
|
|
How complex is the operational story?
|
Simple but fragile
|
Cluster management, ZooKeeper/KRaft
|
Familiar, same DB the team already runs
|
|
What does the team already know?
|
Some experience
|
No experience
|
Deep experience
|
|
What's the delivery latency at p99?
|
< 5ms
|
10-50ms
|
50-200ms
|
Compare the notification options across the questions the team uses to evaluate them. section 11.1
