zero_copy_json

A JSON parser written from scratch, without serde_json. The defining feature: parsed strings borrow directly from the input &str instead of allocating a new String for every string in the document.

What it demonstrates

• Zero-copy string parsing: Value<'a> strings are Cow<'a, str>. A JSON string with no escape sequences is returned as Cow::Borrowed, a direct slice of the original input, zero heap allocation. A string containing an escape (\n, \", \uXXXX, ...) can't be represented as a contiguous slice of the original bytes, since decoding the escape changes the bytes, so that case falls back to Cow::Owned. This is the same borrowed-or-owned split serde's own zero-copy deserializers use.
• Hand-written lexer and recursive-descent parser: a byte-level lexer with one token of lookahead, no parser generator or combinator library.
• UTF-16 surrogate pairs: \uXXXX escapes outside the Basic Multilingual Plane arrive as a high/low surrogate pair (\uD834\uDD1E); the lexer decodes the pair into the single resulting char.
• Order-preserving objects: Object is Vec<(Cow<str>, Value)>, not a HashMap, so this stays dependency-free and keeps keys in the order they appeared in the source, which a hash map wouldn't.

Usage

let value = zero_copy_json::parse(r#"{"name": "test", "count": 42}"#)?;

Returns Result<Value<'_>, JsonError>. JsonError carries the byte offset into the original input where parsing failed.

Scope

Implements enough of RFC 8259 to be a useful, readable reference: objects, arrays, strings with escapes and surrogate pairs, numbers, booleans, null. It is not a fully spec-validating parser: for example it accepts a leading zero like 01 as a number, which strict JSON forbids. That's intentional, in favor of a smaller, more readable implementation over full spec conformance.

Tests

cargo test