A response-time percentile states the duration within which a specified share of requests finishes. P50 is the median, p95 describes the slower tail, and p99 exposes the slowest meaningful slice. A p95 of 200 ms means 95% of requests complete within 200 ms and 5% take longer. Percentiles make the user impact of latency visible in a way an average can hide. section 3.2.2
Table 3.4 Backend performance metrics
|
Metric
|
Measures
|
Example
|
|
Response time percentiles (e.g. p50, p95, p99)
|
How fast do requests complete across the distribution of requests
|
“p95 of 200ms” means 95% of requests complete within 200ms
|
|
Throughput
|
Requests handled per second
|
Service handles 500 requests per second at peak
|
|
Error rate
|
Percentage of requests that fail
|
0.05% of payment requests return errors
|
Compare percentiles with throughput and error rate when evaluating a backend service: speed alone does not establish capacity or reliability. section 3.2.2
Use percentiles for API goals and budgets because response time changes with load, data volume, and third-party dependencies. Choose the percentile and threshold according to the request’s role: interactive endpoints commonly need a tighter p95 than background work, while external APIs must also honor committed SLAs. Set internal budgets tighter than an SLA to leave room for variation. section 6.2.4
Checkout endpoint budgets
A checkout flow can assign different percentile budgets to different user actions: 200 ms for a cart summary, 2 s for payment processing, and 300 ms for order confirmation. The cart summary needs the tightest target because it blocks what the user sees; payment processing must allow for external-service latency. section 6.2.4
Build a usable percentile budget with an operational baseline:
- Measure the current p95 under realistic load rather than choosing an arbitrary number. section 6.2.4
- For a critical endpoint, pull p95 from an APM dashboard over a week of normal traffic; repeat measurements and use the median baseline where instructed. section 6.8
- Set an achievable threshold below current performance or use phased budgets when the gap to the goal is large. section 6.2.4 section 6.8
- Recheck field data after changes: field measurements reveal user experience, while lab measurements support rapid debugging. section 3.6
Do not treat a healthy average or p50 as proof that users are protected. Under load, an undersized database connection pool can create a queueing cascade: waiting requests hold threads, capacity falls, and p99 can climb from acceptable values to seconds or timeouts. Comparing p50 and p99 as traffic rises helps distinguish a scalability problem from a single-request speed problem. section 5.3.1 section 5.1
Percentile targets turn performance into a clear service promise: how many users receive a response within a stated time. Tie that promise to complete user flows, monitoring, ownership, and the definition of done rather than treating it as a one-time optimization metric. section 3.2.2 section 3.6












