gRPC vs HTTP, and HTTP/1.1 vs HTTP/2 vs HTTP/3 compared

The same infrastructure loop asked protocol fundamentals: compare HTTP/1.1, HTTP/2 and HTTP/3, and compare gRPC with plain HTTP. Answer as an engineer who has to choose these protocols and run them in production. ### Clarifying Questions - Is the comparison about browser-facing traffic, service-to-service calls inside a data center, or both? - Do the clients and everything on the network path (proxies, load balancers, firewalls) support HTTP/2 and allow UDP? ### Part 1 — HTTP/1.1 vs HTTP/2 vs HTTP/3 How do HTTP/1.1, HTTP/2 and HTTP/3 differ? For each, explain how requests share a connection, where head-of-line blocking happens, and what the transport and the connection handshake look like. ```hint Follow the blocking Track where head-of-line blocking happens in each version, and which layer each new version changed to get rid of it. ``` #### What This Part Should Cover - The connection and multiplexing model of each version, and its framing and header compression - Head-of-line blocking at the HTTP layer versus the transport layer - Transport and handshake differences, including what QUIC changes - The operational costs of adopting each version ### Part 2 — gRPC vs HTTP How does gRPC differ from calling a plain HTTP API with JSON bodies? When would you choose each? ```hint Inherited versus added gRPC runs on top of HTTP/2. Separate what it adds (contract, encoding, call types) from what it inherits, including how its connections behave behind a load balancer. ``` #### What This Part Should Cover - The interface definition, code generation and binary encoding versus JSON - Call types, deadlines, status codes and metadata - Load balancing, browser support and debuggability trade-offs, and a reasoned choice ### What a Strong Answer Covers - Correct mechanics of multiplexing and head-of-line blocking in each HTTP version - Transport and handshake differences, including QUIC's built-in TLS 1.3 and connection migration - What gRPC adds over HTTP/2 and what it costs - Operational consequences: long-lived connections and load balancing, networks that block UDP, observability - A choice for a concrete case, with the reasoning ### Follow-up Questions - Your gRPC service runs five replicas behind a Kubernetes Service, but one replica gets almost all of the traffic. Why, and how do you fix it? - HTTP/2 server push was largely abandoned. Why? - A corporate network blocks UDP. What happens to an HTTP/3 client? - How would you expose a gRPC service to browsers?