APIs That Respond in Milliseconds, Not Seconds

API performance is the critical path between your infrastructure and your users’ experience. A P99 latency of 2 seconds at the API layer means that 1 in 100 user interactions takes 2 seconds — and at scale, that 1% becomes thousands of users per hour. Distributed tracing is the tool that makes API latency tractable: by tracing every step of a request from entry point through every service call and database query, we identify exactly where time is spent.

GraphQL performance deserves specific attention because the resolver pattern creates a structural temptation toward N+1 problems: each field resolver may independently query the database, producing O(n) database calls for a query returning n records. DataLoader batching collapses those into O(1) batched calls, often reducing database load by an order of magnitude. We find N+1 resolver patterns in the majority of GraphQL APIs we audit.

Payload optimisation is the other major lever: REST APIs that return every field regardless of client need transfer unnecessary data, slowing responses, increasing bandwidth costs, and forcing clients to parse data they discard. Response compression, field selection (sparse fieldsets), and pagination design each reduce payload size. For mobile clients in particular, payload reduction from 48MB to 2MB represents a 96% reduction in parsing time and data transfer cost.