Measured results, with their boundaries

1,048,576 Bybit level rows from 7,265 events, 32 blocks or row groups. One warm-up, six measured rounds, medians. Warm page cache, one host. File-sync counted; directory sync, atomic publication and physical power-loss qualification not.

Against dictionary-LZ4: −37.7543% bytes, −36.6683% write, −3.9599% read, using 100 × (1 − GQODB / reference). This is an order-book measurement, not a tick or trade storage measurement, and it excludes the original collector envelope — so it is not evidence that the same saving holds for all original source bytes. Nothing here is a claim about cold disk, network access or a full database replacement.

Two pre-selected Bybit level samples, Arrow → bytes → Arrow in memory, one warm-up and nine measured passes per codec. No full file I/O.

48.9–49.8% smaller than the previous GQODB codec and 38.4–42.9% smaller than dictionary-LZ4 Parquet, at 14.3–14.5% longer decode than that Parquet baseline. Note what this does not say: the new codec decodes faster than the old GQODB one — it is Parquet that decodes faster than both. B11 and B15 are different source versions and different measurement boundaries.

2,000 real AG News documents, 6,000 records with synthetic revisions and arrival times; five rounds of 1,024 queries per engine. GQO reads from an already-loaded embedded archive; PostgreSQL and MongoDB include client, driver and TCP cost. This is not a comparison of five equivalent network servers.

Opening, decompressing and validating the compact-wide archive costs a warm median of 95.923 ms — so the 3.967 µs query is not the total time of a fresh CLI process. PostgreSQL is faster at individual durable writes here. The small write differences against MongoDB are not a general ranking.

Keeping the raw bytes and the archive is smaller than the PostgreSQL relation storage and larger than both MongoDB variants on these samples. Shared WAL, journal and cluster allocations are not a fair per-dataset measure and are outside this table. PostgreSQL effectively did not compress these short payloads, which proves nothing about its compression on long articles.

Exact restore was verified in the historical run. The second file is a predictor artifact, not a large weight model. Both source files were recomputed from existing receipts for this audit rather than reopened; that run's output lived on tmpfs. No claim about inference, training, quantisation or disk power loss.

p99 minimum to maximum across three runs, measured from scheduled offer to usable receive.

A local Redis comparison exists — it is simply not an equal CPU, durability and network contract. B08 measured 64-byte shared-memory messages at 500k/s at 0.964–0.990 µs scheduled-to-usable p99, which does not extend to every payload or rate. The B09 L10 case cycles 744 real snapshots for load.

Zero detected errors is an observation in these runs, not a guarantee of lossless delivery after a crash. Speed mode is not consistently better, and the largest single latency in the assessment is 292.345 µs.

Only the feature path is a single local engine call; the quality figures involve a caller and a worker thread. The quality study keeps one job outstanding at a time, so it does not establish streaming capacity under saturation, and spin can occupy a logical CPU continuously. Real clock acquisition, the feed, raw fsync and downstream publication are all outside these numbers — and none of them is a theoretical floor for other implementations.

An engineering baseline, not a release comparison: 1,000 synthetic records — 500 trades and 500 L10 snapshots — through raw and ledger sync with real quality, feature and compression workers on ext4. The host was running concurrent builds and checks at the time.

The output bytes above exclude raw journals and quality and feature reports, so total storage exceeds input: this small-block run shows no meaningful net compression win. The receipt interval is not a bus hop and not full upstream event latency. The run predates part of the group-commit and metadata work in the audited copy, so no untested improvement should be credited to it. This belongs beside the favourable microbenchmarks, not behind them.