\Drupal\strata\Codec CodecCatalog

The shipped reference figures for every codec, and how to install the ones that are missing.

Where the numbers come from: a fixed 6,357,223-byte corpus of Drupal-shaped data - 1,000 node images in toArray() shape, 1,000 user images carrying JSON field values, and 4,000 custom-table rows including hashed API tokens. Split into frames, each frame compressed independently - which is what Strata actually does - with a zstd dictionary trained on 5% of the frames. Measured on PHP 8.5.7, arm64, with ext-zstd 0.18.0 and ext-brotli 0.21.0.

What the numbers say, at 16 KiB frames, which is the shipped default:

Three things follow, and they are why the defaults are what they are:

• zstd level 1 is the flush-path choice. 4.28x at 422.9 MB/s against gzip level 9's 4.40x at 67.0 MB/s: a fraction worse on size for more than six times the throughput, on a path that runs inside a web request.
• A dictionary costs throughput, not just disk. At level 1 it takes 422.9 MB/s down to 105.1 MB/s for 4.28x to 4.67x. Whole-batch measurement hides this, because the dictionary is loaded once there and per frame here. It is used during compaction, where 5.86x against 4.87x is worth 2.4 MB/s, and by delta coding, which cannot work without it.
• brotli wins on size and loses on time. 6.07x is the best figure here, at 0.9 MB/s. That is a compaction codec on a large site and nothing else.

What they do not say: ratio depends on content. This corpus includes incompressible hashed tokens, so it runs low against a prose-heavy site and high against one storing mostly media. These figures exist so two hosts can be compared and so a size estimate has somewhere to start; Calibrator measures the real thing on real site data and the settings form shows both, marking which is which.

Methods