A service tier is a named range of non-functional expectations with similar architectural implications. Assign a tier for each relevant dimension rather than demanding a precise number for every requirement or building every system to the same specification. An off-peak marketing email queue and a peak-trading payment service carry different consequences of failure; treating them alike can mean either unnecessary expense or insufficient resilience. section 2.2
A tier matters when it changes a decision:
- Availability targets can determine whether a single region is sufficient or multiple regions are needed. section 2.2
- Latency targets can determine whether work fits on a synchronous path. For requests, a percentile target such as p95 < 300 milliseconds is more practical than an average or a guarantee for every request. section 2.2 section 2.2.3
- Durability expectations can affect storage, backups, and sometimes database choice. Throughput and scalability, and change volatility are also dimensions to assess. section 2.2
Table 2.3 Latency tiers
|
Tier
|
Latency, p95
|
Architecture implications
|
|
Instant
|
<100ms
|
Local state or precomputed data. No unindexed or heavy
backend work on the critical path.
|
|
Responsive
|
100–300ms
|
Well-designed backend queries with indexing and caching
when required.
|
|
Acceptable
|
300ms–1s
|
Synchronous upper bound for core user journeys.
Approaching the edge of comfortable. Think of performance optimisation to get
to Responsive.
|
|
Deferred
|
1–10s
|
Async processing with some progress or loading indicators
essential. Not suitable for interactive flows
|
|
Background
|
>10s
|
Background job with explicit completion signals. Better
not to be designed as a blocking interaction
|
Latency tiers link user-visible response ranges to different approaches, from local or precomputed responses through synchronous queries to asynchronous work and background jobs. The ranges make trade-offs discussable; they are not a universal performance prescription. section 2.2.3
Customer-record search
A search requirement can carry different tiers on different dimensions: 99.9% availability for core search calls for planned multi-node failover, while moderate durability is acceptable because the index can be rebuilt from source. An acceptable latency tier calls for responses under one second, with indexing and sound query design but without excessive optimisation. Expected load—under 200 concurrent users and 50 requests per second—adds context for provisioning. These expectations inform the solution discussion without prematurely choosing the entire design. section 2.4

