v0.2.0

qj 0.2.0 is a rewrite of its core: qj now is jq 1.8.1, ported to Rust. jq's value and
number model, parser, compiler, bytecode VM, C builtins and builtin.jq run as jq's own
code does. qj's speed now comes from its input layer, which keeps jq's results exact:

• simdjson parsing;
• memory-mapped and streaming input;
• ordered parallel processing of NDJSON;
• evaluating simple filters directly on simdjson's tape.

Conformance

Conformance is checked byte for byte against the jq 1.8.1 binary: stdout, stderr and exit
code. The checks cover jq's own test suites, a corpus of about 39,000 cases (every builtin,
CLI option, input edge case and error path) and differential fuzzing, on macOS and Linux.

• Default mode matches jq on every case except qj's own --help/--version text.
  Where jq crashes or hangs, qj gives a sane answer instead.
• QJ_JQ_COMPAT=1 means "be exactly jq 1.8.1, bugs included", and matches on every
  case, with nothing normalized:
  • jq's name, help and version text;
  • jq's crashes, with stack overflows at jq's own depth per OS, and its hangs;
  • qj's extras turned off (globbing, .gz/.zst, --threads).
• jq's own test suites pass 2,903/2,903, byte-exact.

What still differs, and why, is in docs/COMPATIBILITY.md.

Behaviour changes from 0.1.x

Output now matches jq exactly where 0.1.x differed:

• number formatting follows jq 1.8.1's decNumber literals, so 1.10 stays 1.10 and
  1e2 prints as 1E+2;
• string escapes are jq's (\u001b, not the input's \u001B);
• error messages, exit codes, and the interleaving of stdout and stderr are jq's.

Other changes:

• QJ_JQ_COMPAT=1 changed meaning, as above. Arithmetic now always follows jq's
  precision; it is no longer a separate switch.
• The old evaluator and its NDJSON fast paths are gone, and so are QJ_CORE and
  QJ_NO_FAST_PATH. --jsonl and --debug-timing are still accepted and do nothing.
• On Apple Silicon, qj uses every non-efficiency core, so Super and Performance cores on
  M5 Pro/Max.

Performance

Measured against 0.1.4 on the same M5 Max:

	0.1.4	0.2.0
NDJSON 1.2 GB, -c .	681 ms	77 ms
NDJSON, {type, repo: .repo.name, actor: .actor.login}	147 ms	57 ms
JSON 49 MB, pretty print	192 ms	48 ms
JSON, group_by / sort_by / unique	135–137 ms	61–64 ms
reduce building a 50,000-key object	10.3 s	10 ms
peak memory, 3.4 GB NDJSON	3.5–3.8 GB	100–160 MB

Some cases are slower than in 0.1.4, though still well ahead of jq:

• single-threaded select(...) over raw NDJSON, which 0.1.4 answered by scanning bytes
  and copying lines through unchanged, so its output wasn't jq's;
• tight loops on the VM, such as limit/repeat.

See the README for comparisons with jq, jaq and
gojq.

Platforms

Prebuilt binaries for macOS (arm64, x86_64) and Linux (x86_64, aarch64), or
cargo install qj. Windows isn't supported.

Full changelog: v0.1.4...v0.2.0