Response compression: where to do it and what to exclude

What it is

A client advertises acceptable codings with Accept-Encoding; the server or proxy may respond with Content-Encoding: gzip|br|zstd and must add Vary: Accept-Encoding so that caches keep encoded and unencoded variants apart. RFC 9110 treats the content coding as part of the representation, which has consequences for validators.

Why it matters

Text shrinks several-fold, which matters for JSON APIs and Markdown that agents fetch repeatedly. Compressing the wrong things wastes CPU (images, archives) or breaks streaming (event streams whose bytes are held back until a buffer fills).

How to apply

• Compress at one layer only, usually the reverse proxy; set a minimum size so tiny responses are left alone.
• Exclude binary and already-compressed media types and text/event-stream; make sure the proxy honours the exclusion.
• Because the proxy usually keeps the origin's ETag, use weak validators for content that may be delivered in more than one encoding, or ensure the proxy weakens them.
• Measure: response sizes with and without Accept-Encoding: gzip for representative endpoints.

Pitfalls

Double compression when both application and proxy compress. Compression of secrets alongside attacker-controlled input in the same response has been exploited (BREACH); avoid reflecting secrets in compressed responses. Length-based checks after compression compare the wrong number.