Logs JSON Wins

This is exactly what the SKILL’s format-choice guidance says: prefer a JSON* format when the result is “mostly strings / JSON-like values that you consume wholesale,” because V8’s native JSON.parse is heavily optimized C++ and “pair it with HTTP response compression (gzip / zstd, which crushes JSON’s repetitive keys).”

Reproduce: npx vitest bench --run tests/logs.bench.ts (against a live ClickHouse server). Source: tests/logs.bench.ts, reader: src/examples/logs.ts.

The data

An application log table — four of five columns are text consumed as text:

50,000 rows, fetched from a live server. The two LowCardinality columns carry no dictionary on the RowBinary wire — they decode as plain String.

Decode throughput (full 50k-row decode; higher = faster)

JSONCompactEachRow (arrays, no repeated keys) is the fastest JSON option and beats even the optimized RowBinary reader by ~1.4x. A RowBinary string is a varint length + buf.toString("utf8", …) decoded one field at a time in JS; JSON.parse builds the same JS strings in one optimized C++ pass.

Wire size — raw, and compressed (gzip / zstd)

RowBinary is 1.4–1.8x smaller raw, which is the usual argument for it. But that edge is mostly JSON’s repeated structure (keys, punctuation) — exactly what a compressor removes. With gzip the three are within ~4% of each other, and with zstd the JSON responses are slightly smaller than RowBinary. Any production HTTP path should have compression on, so the wire-size case for RowBinary on this data effectively vanishes.

Node 24 / V8. Your numbers will vary; run npm run bench on your own hardware.

Takeaways

• JSON wins both axes here. Faster to decode (~1.4x) and, once compressed, no larger on the wire. There is no reason to hand-write a RowBinary parser for this shape.
• JSONCompactEachRow is the one to reach for — it drops the per-row repeated keys, so it parses faster than JSONEachRow and compresses about the same.
• Compression erases RowBinary’s raw-size advantage on text. RowBinary’s smaller raw wire comes largely from not repeating keys; a compressor already does that for JSON. Always compare compressed sizes when the data is string-heavy.
• This is the boundary of the skill. RowBinary earns its keep on numeric/wide/binary data (IoT, ledger); on string-heavy results read as text, the right answer is JSONCompactEachRow + compression. Match the format to the shape of the data — and measure.