There are two compression features, both bounded by construction, and a larger territory zgix deliberately leaves to a CDN or the origin.

Precompressed static answers

Every configured body — a respond action’s page, an error_pages entry — is gzipped once at load. When the request’s Accept-Encoding accepts gzip (an explicit gzip or *; q=0 declines; q-values are parsed), the answer is sent with Content-Encoding: gzip, Vary: Accept-Encoding and the compressed length; otherwise the plain body is sent unchanged. HEAD gets the selected variant’s headers without a body.

"bodies": {
    "maintenance": { "file": "/etc/zgix/503.html", "content_type": "text/html; charset=utf-8" },
    "robots":      { "inline": "User-agent: *\nDisallow: /private\n",
                     "content_type": "text/plain" }
},
"error_pages": { "503": "maintenance" }

Bodies live in memory — a 503 is raised because something ran out, so an error page whose delivery needed a resource would fail exactly when it is needed. The cost is capped per body, counted in the startup banner, and paid once per process. Listing a status is the opt-in, including for sheds: leave 503 out and overload answers an empty 503 as before.

Compressing proxied responses

A listener may opt in to gzip for proxied answers as well:

"http": {
    "cluster": "api",
    "compress": { "min_bytes": 1024 }
}

• min_bytes (default 1024) is the floor; smaller bodies are sent unchanged, since the gzip framing alone would cost more than it saves.
• Compression applies only when the whole body arrived in the same read as the response head (a small response from a fast origin), the status is 200, Content-Length is known, Content-Type is text-like (text/*, JSON, JavaScript, XML, SVG, NDJSON), no Content-Encoding is already set, and Cache-Control does not say no-transform.
• The ceiling is 64 KiB. Anything larger, chunked, until_close, an interim response, a HEAD, or a body that took more than one read is streamed through untouched.
• A compressed answer carries a recomputed Content-Length, Content-Encoding: gzip and Vary: Accept-Encoding; a client that did not offer gzip gets the original bytes.

The compressor is one buffer allocated at startup and counted in the banner — there is no per-request allocation.

What is not compressed

This is not full streaming compression: the deflate pump zgix ships cannot emit a valid gzip stream incrementally, so streaming responses — SSE, large bodies, chunked replies — pass through uncompressed. Front zgix with a CDN, or compress at the origin, when those need it.