Conversation
A single large response cannot demonstrate a dictionary: it only contains within-response repetition, which deflate already handles inside the codec. Dictionaries exist for repetition ACROSS messages, so the benchmark now models what a route actually serves — many independent responses over a stable entity universe, at sizes APIs really return. - candles, devices and feed become 500-message corpora with 20-50 records per message (a chart window, a page of telemetry, a page of posts) instead of one 1000-row dump, each drawn from a fixed universe of instruments, devices and authors with Zipf-skewed selection. - orders is a new single-entity route (GET /v1/orders/:id, ~780 B of JSON) with a fixed 120-product catalogue and 200-customer base. Small responses are the case where a general compressor has almost no context to work with. - events is a new audit-log page (20-50 records) whose recurring user agents, actor emails and resource ids are exactly the shape dictionaries exist for. The harness reports per-message averages plus the dictionary's own size and how many requests it takes to pay for itself, because a dictionary is an out-of-band artifact and a reader is entitled to see what it costs: route json gzip br4 col prof full dictionary candles 3225 928 842 496 496 372 none devices 7994 1473 1422 896 705 638 5.2KB, 28 requests feed 6863 2307 2294 1908 1536 1535 4.9KB, 14 requests events 12687 2503 2512 2109 896 823 12KB, 10 requests orders 782 423 408 271 184 188 3.0KB, 35 requests Two results worth keeping honest about: profiles do nothing for candles, whose only string sits outside the array and so has no column to key on; and on orders the brotli pass makes the output slightly larger, because at that size its framing costs more than it saves. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_hf1
The benchmark now reports bytes per message over 500-message corpora rather than the size of one large response, because that is the setting in which a trained dictionary can be measured at all. Each route shows columnar, profiled and full so the contribution of each stage is visible rather than blended. Tabs lead with events, where recurring user agents and actor emails make the profile worth 57%, and end with candles, where it is worth nothing — that spread is the honest shape of the feature, and the candles row is left in precisely because it is flat. The footnote states that the profile is trained on the route's own traffic, that the dictionary is an out-of-band artifact, and how many requests it takes to pay for itself. Orders keeps the result where full is larger than profiled: at 184 bytes Brotli's framing costs more than it saves. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_hf1
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Reworks the benchmark around what a route actually serves, and puts the three Hyperfly stages on the site.
Why the shape changed. A single large response cannot demonstrate a dictionary — it only contains within-response repetition, which deflate already handles inside the codec. Dictionaries exist for repetition across messages. The old corpora were one response of 500–1000 records, so the profile row was measuring nothing; training on it gave 0–2%, and forcing a win meant putting the payload in the dictionary (7.5 KB of out-of-band data to "save" 2.3 KB on the wire).
Now: 500 independent messages per route, 20–50 records each (or one entity), drawn from a stable universe — a fixed 400-device fleet, a 120-product catalogue, a recurring cast of authors — with Zipf-skewed selection, because real traffic concentrates.
Two new routes:
orders(single-entity, ~780 B, where a general compressor has least context) andevents(audit log, whose recurring user agents and actor emails are exactly the dictionary case). Protobuf baselines added for both, round-trip verified.Per message, 500-message corpora
Against brotli-q4'd JSON — the realistic production baseline — that is 1.5× to 3.1×.
Kept honest rather than tuned away: candles gains nothing from a profile (its only string sits outside the array, so there is no column to key on) and the flat row is left visible; on orders
fullis larger thanprofiled, because at 184 bytes Brotli's framing costs more than it saves; and the harness reports the dictionary's own size and break-even, because it is an out-of-band artifact and a reader is entitled to see what it costs.Site: five tabs, three Hyperfly rows each, footnote stating the training setup and the dictionary cost.
🤖 Generated with Claude Code