Semantics versus representation
HTTP/2 preserves familiar HTTP semantics such as methods, status codes and fields while changing their wire representation into frames. Multiple request-response exchanges are identified as streams within a connection and can progress concurrently.
Instead of treating traffic as HTTP/1.1 text lines, follow frames such as HEADERS and DATA together with their stream identifiers. This identification lets pieces of different responses be interleaved. Multiplexing describes how exchanges share a connection; it does not change the application meaning of a GET or a 404 response.
Interleave two responses
Consider a small JSON response and a large image requested over the same connection. Frames for their separate streams can be interleaved, reducing the need to finish transmitting one entire response before beginning another. This can improve utilization when several independent exchanges are ready.
The observed order still depends on scheduling, server work, buffering and available network capacity. Multiplexing does not by itself promise that the most important response always arrives first. Inspect actual response timing rather than inferring a performance result solely from the negotiated protocol version.
TCP still orders the connection
When HTTP/2 runs over TCP, a missing byte range can block delivery of later bytes on the connection until recovery. The gap might belong to an image frame, while a later JSON frame on another HTTP stream waits behind it.
Separate HTTP stream identifiers do not create separate TCP ordering domains. This distinction is central when comparing HTTP/2 with HTTP/3 over QUIC. It also explains why application-level concurrency and transport-level loss behavior should be measured separately when diagnosing latency on a lossy path.
Compression and resource limits
HPACK compresses HTTP/2 header fields; it does not mean every response body is compressed with that format. Body Content-Encoding remains a separate concern. Header overhead can matter particularly when requests are small but repeat substantial fields.
Concurrent-stream limits, flow control and connection shutdown still constrain progress. Packing more requests onto one connection cannot create unlimited server capacity. Diagnose application execution time, stream scheduling and TCP recovery independently before choosing a different number of connections or claiming that a protocol upgrade solved the bottleneck.