pulse-jsonpath-ng

A Rust/PyO3 accelerator for jsonpath-ng — a native parser that is iso-functional with jsonpath_ng and ~500× faster at parsing, with transparent fallback to jsonpath-ng for anything outside its fast path. It's a separate package that depends on jsonpath-ng: it rebuilds jsonpath-ng's own AST and reuses its evaluator, so it never diverges — by construction.

pip install pulse-jsonpath-ng

# was:  from jsonpath_ng import parse
from pulse_jsonpath_ng import parse

expr = parse("items[*].price")     # parsed in Rust, ~500× faster
[m.value for m in expr.find(data)] # .find() is jsonpath_ng's own — identical results

The only change is the import. parse() returns a real jsonpath_ng expression (same classes), so .find(), .full_path, __eq__, and the exceptions all work unchanged.

Performance

jsonpath_ng.parse is the hot spot: it's built on PLY with write_tables=0, so it rebuilds the LALR parse table (and the lexer table) on every single call — ~2.5 ms to parse a 3-segment expression, roughly 300× the cost of the find it precedes. pulse-jsonpath-ng replaces only the parser with a hand-written recursive-descent parser in Rust.

	median / parse()	speedup
jsonpath_ng.parse	~2.5 ms	—
pulse_jsonpath_ng.parse	~5 µs	~×500

import statistics, time, jsonpath_ng, pulse_jsonpath_ng
EXPR = "items[*].price"
def bench(parse, reps, rounds=15):
    out = []
    for _ in range(rounds):
        t = time.perf_counter()
        for _ in range(reps): parse(EXPR)
        out.append((time.perf_counter() - t) / reps)
    return statistics.median(out)
r = bench(jsonpath_ng.parse, 200); c = bench(pulse_jsonpath_ng.parse, 20000)
print(f"jsonpath_ng {r*1e6:.0f} us  ->  pulse {c*1e6:.2f} us   (x{r/c:.0f})")

If you already cache compiled expressions you won't see this — but a great deal of code calls jsonpath_ng.parse(...) inline on every request, and that is exactly where the per-call table rebuild hurts.

Why Rust wins here (and why this is not a false friend)

Profiling a parse(): ~100 % of the time is interpreted Python — PLY rebuilding lr_parse_table and the lexer table — with re accounting for ~0.2 % (the lexer regex, built once). Because the bottleneck is interpreted table-construction (not a C-bound kernel), a native parser wins by orders of magnitude. We port only the parser; find stays in jsonpath_ng (reused, so evaluation is iso for free).

What's covered natively (and what falls back)

The Rust fast-path covers the common ./[...] expressions, left-associative:

• fields a.b.c, wildcard * / a.*, numeric fields a.5, root $, `this`;
• brackets [0], [-1], [0,2], slices [1:3] / [::2] / [*], field brackets ['a','b'];
• quoted fields 'a.b', dotted-bracket equivalence a.[0] ≡ a[0].

Everything else is transparently delegated to jsonpath_ng: descendants .., |/& (union/ intersect), where/wherenot, parent, parentheses, named operators other than this, unicode identifiers, and any invalid or ambiguous input (jsonpath_ng then returns the correct tree or raises the correct JsonPathParserError/JsonPathLexerError). The fast-path is a strict subset — it never accepts something it can't reproduce exactly.

Iso-functionality

The Rust parser rebuilds the same jsonpath_ng AST classes, so iso is verifiable two ways: the parsed tree is structurally == to jsonpath_ng.parse(expr), and .find() (jsonpath_ng's own) returns identical values and paths. Proven by a differential oracle over a curated corpus plus adversarial fuzzing of random expressions — tree equality, .find() parity, and exception-type parity for invalid input. The pure-jsonpath_ng fallback path is verified iso too (PULSE_FORCE_FALLBACK=1).

Wheels

abi3 wheels (Python ≥ 3.11) for Linux (x86_64/aarch64, manylinux + musllinux), macOS (Apple Silicon), and Windows; sdist elsewhere (builds the Rust core via maturin).

License

Apache-2.0 (same as jsonpath-ng).