Skip to content

Latest commit

 

History

History
151 lines (118 loc) · 6.64 KB

performance.md

File metadata and controls

151 lines (118 loc) · 6.64 KB
Raw

Performance of JSON Parsers

I looked at some parsers, which were tested with a JSON grammar. The benchmark was written by the team behind Chevrotain. There might be some optimisations in other parsers possible, but the code should be enough to get a rough overview and also to compare the quality of error reporting.

I modified the grammars to support extensions to the JSON format, which may be helpful, if you use JSON not only for machine-generated data exchange, but also for hand-edited files, like configuration. I named the original grammars pure and the modified ones extended. The extensions were:

  • Ignoring JavaScript-like single-line and multiple-line comments. Useful for documenting configuration files.
  • Accepting single quotes (apostrophes) as string delimiters. Useful for input files using them to avoid escaping of quotation marks.

Code Size

The size of the JavaScript code is important, when the application runs in the web browser.

File Source Minified Gzipped Brotlied
hand-built/extended.js 6.04 kB 1.82 kB 819 B 738 B
jison/extended.js 24.50 kB 10.90 kB 3.85 kB 3.46 kB
jju/extended.js 27.50 kB 15.10 kB 5.16 kB 3.95 kB
pegjs/extended.js 52.20 kB 12.60 kB 3.39 kB 2.90 kB
json5/dist/index.js 58.60 kB 31.30 kB 9.26 kB 7.17 kB
json6/dist/index.js 77.70 kB 19.10 kB 6.92 kB 6.31 kB
chevrotain/extended.js 568.24 kB 224.70 kB 51.70 kB 43.70 kB

Results

This is a result of the benchmark run by node parse. The numbers should be understood as relative ones:

Parsing JSON data 4797 characters long using:
  the built-in parser x 78,856 ops/sec ±0.50% (92 runs sampled)
  the pure chevrotain parser x 9,815 ops/sec ±1.29% (90 runs sampled)
  the extended chevrotain parser x 9,512 ops/sec ±0.90% (92 runs sampled)
  the standard jsonlint parser x 78,998 ops/sec ±0.48% (95 runs sampled)
  the extended jsonlint parser x 7,923 ops/sec ±0.51% (93 runs sampled)
  the tokenising jsonlint parser x 6,281 ops/sec ±0.71% (91 runs sampled)
  the pure hand-built parser x 13,529 ops/sec ±0.62% (92 runs sampled)
  the extended hand-built parser x 13,475 ops/sec ±0.81% (92 runs sampled)
  the AST parser x 9,201 ops/sec ±0.74% (91 runs sampled)
  the Myna parser x 3,089 ops/sec ±0.54% (94 runs sampled)
  the pure jju parser x 8,596 ops/sec ±0.98% (92 runs sampled)
  the extended jju parser x 6,966 ops/sec ±0.59% (89 runs sampled)
  the tokenisable jju parser x 6,091 ops/sec ±0.84% (89 runs sampled)
  the tokenising jju parser x 4,651 ops/sec ±0.54% (92 runs sampled)
  the comments-enabled parser x 5,632 ops/sec ±0.65% (92 runs sampled)
  the pure pegjs parser x 2,413 ops/sec ±0.69% (92 runs sampled)
  the extended pegjs parser x 2,124 ops/sec ±0.66% (92 runs sampled)
  the pure jison parser x 1,722 ops/sec ±1.16% (88 runs sampled)
  the extended jison parser x 1,619 ops/sec ±0.65% (91 runs sampled)
  the JSON5 parser x 1,283 ops/sec ±0.43% (91 runs sampled)
  the JSON6 parser x 7,922 ops/sec ±1.78% (92 runs sampled)
  the Nearley parser x 944 ops/sec ±0.51% (92 runs sampled)
The fastest one were the standard jsonlint parser and the built-in parser.

I looked further at capabilities and licenses of the parsers.

Built-in

  • The plain JSON.parse, run just with the input string and no reviver.
  • The fastest one, of course, but not adaptable.

Chevrotain

  • Does not generate code from tokens and grammar; uses a coded grammar.
  • The fastest one with an adaptable code base.
  • Supports a limited recovery to be able to continue parsing after an error occurs.
  • Can report all errors that occur in the whole input.
  • Differs between lexical and parsing errors.
  • The license (Apache 2) is not compatible with JSONLint.

Hand-built

  • A code example from Chevrotain benchmarks.
  • Very fast one.
  • Error reporting would need to be improved.
  • The license (Apache 2) is not compatible with JSONLint.

AST

  • Very fast one.
  • Generates an AST to analyse the JSON input.
  • Does not return JSON data without an additional generator using the AST.

JJU

  • A part of other utilities to work with JSON/JSON5 documents.
  • Very fast one.
  • Supports reviver for the full compatibility with JSON.parse.
  • Can generate tokens to analyse, modify and update the original JSON input.

comment-json

  • Very fast one.
  • Supports reviver for the full compatibility with JSON.parse.
  • Supports CJSON - JSON with JavaScript-like comments.

PEG.JS

  • Uses PEG to generate the parser code.
  • Still fast, for a generated parser code.

Jison

JSON5

  • Hand-built parser implementing the [JSON6 specification]. Extensions mentioned above were included, and even more, like trailing commas or multi-line strings. No support for pure JSON.
  • Slower one.

JSON6

  • Hand-built parser going beyond the JSON5 specification. Extensions mentioned above were included, and even more, like trailing commas or multi-line strings. No support for pure JSON.
  • Very fast one.

[JSONLint]

This entry is here just to compare the current implementation to the original "contenders". Based on the evaluation results, the current JSONLint uses a hand-coded parser based on JJU. THe JSON6 parser wasn't available at that time, otherwise it would've been a serious candidate.