Compression Dictionary Transport cut a 272 KB bundle to a 2.6 KB delta in Cloudflare's lab test

Gzip and Brotli compress each response in isolation ^[5]. When a one-line fix gets a new hashed filename, a returning visitor downloads the whole file again. In the source's example that file is about 270 KB, and the browser already holds 99 percent of those bytes under the old hash ^[6]. Dictionary transport lets the server compress the new file against the copy the client says it has cached ^[4].

The 2.6 KB delta in Cloudflare's test is about 2.8 percent of the gzip size, or roughly 35 times smaller ^[10]. That result comes from one asset in one vendor's lab. According to the source's author, results depend entirely on how much changed between builds, and frequent small deploys are the case the ratio is meant for ^[11]. For the figure to transfer to another codebase, each release has to change only a small part of the bundle ^[11].

The build step is short. The source's pattern compresses each new bundle against the previous release's copy of the same file and stores the result next to it ^[12]. Its example command ^[13]:

``` zstd -19 --patch-from=dist-prev/app.8f3c1a.js dist/app.2d9e44.js -o dist/app.2d9e44.js.dcz-raw ```

The -raw suffix is the most useful documentation in that line. The dcz and dcb formats wrap the compressed stream in a short header that embeds the dictionary hash, so the browser can confirm it decoded against the right file ^[14]. The zstd output still needs that header before it is a valid dcz body ^[19]. Cloudflare's implementation adds the framing. For anyone doing it by hand, the source's advice is to follow the RFC's format section ^[15]. Because of that framing step, I would start with a CDN implementation.

On the server, the source's Express handler looks up a delta by filename and dictionary hash ^[16]. It serves the delta with Content-Encoding: dcz only when one exists and the request advertised dcz. Every other request falls through to the ordinary Brotli or gzip handler ^[16]. Browsers without support get the same response they got before ^[16].

The part of the design I like is the chaining. The new bundle's response carries its own Use-As-Dictionary header. Each release becomes the dictionary for the next, so a returning visitor keeps getting deltas across deploys ^[18]. The match value is a URL pattern for the future requests the stored file can help with: `Use-As-Dictionary: match="/js/app.*.js", id="app-v1"` ^[3].