Key points
- REST is usually the simplest option for browsers, serverless functions, administration tools, and low-to-moderate request rates.
- gRPC is a strong fit for persistent service-to-service traffic and typed clients with controlled connection reuse.
- Measure end-to-end latency from the customer region; an in-region target latency is not the same as public internet latency.
When REST is the practical default
REST works with standard HTTP tooling, is easy to inspect, and passes through common gateways and corporate networks. It is appropriate when requests are occasional, clients are diverse, or a team wants to start with curl and a conventional SDK.
Connection reuse still matters. A benchmark that creates a new TLS connection for every request measures handshake cost as much as API processing. Production clients should use keep-alive, bounded connection pools, deadlines, and structured error handling.
When gRPC is worth the operational cost
gRPC provides a typed contract, compact binary framing, HTTP/2 multiplexing, and efficient connection reuse. These properties suit backend services that make continuous requests and can run a supported gRPC client.
The application still needs reconnect behavior, deadlines, health checking, and load-balancer compatibility. Long-lived channels should be reused rather than created per request. A low core processing time does not guarantee the same p95 after network routing, authentication, rate limiting, and scaling events are included.
Retries can change the business outcome
A timeout does not prove that the server did no work. Retrying a draw or shuffle without an application idempotency rule can create a second valid random result. The client should distinguish transport retries from a new business operation and retain its own operation identifier where one logical action must produce one accepted outcome.
A useful pre-launch test matrix
Run clients in the intended AWS Region or customer network and keep the transport configuration fixed. Increase request rate and API-key count independently, then repeat the test with different receipt ratios. Include a sustained period, a short burst, and a scaling event so capacity and recovery can be observed rather than inferred.