jsonbb

jsonbb is a binary representation of JSON value. It is inspired by JSONB in PostgreSQL and optimized for fast parsing.

Usage

jsonbb provides an API similar to serde_json for constructing and querying JSON values.

// Deserialize a JSON value from a string of JSON text.
let value: jsonbb::Value = r#"{"name": ["foo", "bar"]}"#.parse().unwrap();

// Serialize a JSON value into JSON text.
let json = value.to_string();
assert_eq!(json, r#"{"name":["foo","bar"]}"#);

As a binary format, you can extract byte slices from it or read JSON values from byte slices.

// Get the underlying byte slice of a JSON value.
let jsonbb = value.as_bytes();

// Read a JSON value from a byte slice.
let value = jsonbb::ValueRef::from_bytes(jsonbb);

You can use common API to query JSON and then build new JSON values using the Builder API.

// Indexing
let name = value.get("name").unwrap();
let foo = name.get(0).unwrap();
assert_eq!(foo.as_str().unwrap(), "foo");

// Build a JSON value.
let mut builder = jsonbb::Builder::<Vec<u8>>::new();
builder.begin_object();
builder.add_string("name");
builder.add_value(foo);
builder.end_object();
let value = builder.finish();
assert_eq!(value.to_string(), r#"{"name":"foo"}"#);

Format

jsonbb stores JSON values in contiguous memory. By avoiding dynamic memory allocation, it is more cache-friendly and provides efficient parsing and querying performance.

It has the following key features:

1. Memory Continuity: The content of any JSON subtree is stored contiguously, allowing for efficient copying through memcpy. This leads to highly efficient indexing operations.
2. Post-Order Traversal: JSON nodes are stored in post-order traversal sequence. When parsing JSON strings, output can be sequentially written to the buffer without additional memory allocation and movement. This results in highly efficient parsing operations.

Performance Comparison

[^0]: JSON files for benchmark: canada, citm_catalog

[^1]: Parsed to simd_json::OwnedValue for fair.

[^2]: canada["type"] returns a string, so the primary overhead of this operation lies in indexing.

[^2.1]: jsonbb uses binary search on sorted keys [^2.2]: jsonb uses linear search on unsorted keys [^2.3]: serde_json uses BTreeMap [^2.4]: simd_json uses HashMap

[^3]: citm_catalog["areaNames"] returns an object with 17 key-value string pairs. However, both serde_json and simd_json exhibit slower performance due to dynamic memory allocation for each string. In contrast, jsonb employs a flat representation, allowing for direct memcpy operations, resulting in better performance.