WRITE_PREFERENCE
WRITE_PREFERENCE = ['zstd', 'brotli', 'gzip', 'none']
Write preference, best first.
Ordered by measured ratio with dictionary support weighted above it, because a codec that cannot take a dictionary cannot produce a delta.
Chooses which codec writes a frame, and finds the one that can read an existing frame.
Two separate jobs, and conflating them is how a store becomes unreadable. Writing wants the best
codec this host can run. Reading wants whichever codec understands the id already recorded in a
frame header, even when that is not the one this host would choose - a bucket written on a box
with ext-zstd must stay readable on a box that only has the zstd binary, and a bucket written
years ago on gzip must stay readable forever.
Write preference follows the figures in CodecCatalog. At 16 KiB frames zstd level 1 reaches 4.28x at 422.9 MB/s against gzip level 9's 4.40x at 67.0 MB/s, so zstd wins the flush path on throughput. Dictionary support outranks raw ratio, because DeltaCodec cannot produce a delta without it.
A PipeCodec is never the hot-path choice. A process spawn dominates the work at frame sizes -
1,000 invocations measured 5.88 seconds - so zstd and brotli on disk are registered for
reading and for batch work, and a host with neither extension gets deflate on the flush path
instead. What they buy is that a frame stays readable when a site moves to a host whose PHP was
built without the extension that wrote it.
$writers : array
Codecs able to write a single frame, keyed by id. First registration for an id wins.
$bulkWriters : array
Codecs able to write in bulk but not per frame, keyed by id.