SIMD JSON for Rust
Rust port of extremely fast simdjson JSON parser with serde compatibility.
readme (for real!)
simdjson version
Currently tracking version 0.2.x of simdjson upstream (work in progress, feedback is welcome!).
CPU target
To be able to take advantage of simd-json
your system needs to be SIMD-capable. On x86
, it will select the best SIMD feature set (avx2
or sse4.2
) during runtime. If simd-json
is compiled with SIMD support, it will disable runtime detection.
simd-json
supports AVX2, SSE4.2, NEON, and simd128 (wasm) natively. It also includes an unoptimized fallback implementation using native rust for other platforms; however, this is a last resort measure and nothing we'd recommend relying on.
Performance characteristics
- CPU native cpu compilation results in the best performance.
- CPU detection for AVX and SSE4.2 is the second fastest (on x86_* only).
- Portable std::simd is the next fast implementation when compiled with a native CPU target.
- std::simd or the rust native implementation is the least performant.
allocator
For best performance, we highly suggest using mimalloc or jemalloc instead of the system allocator used by default. Another recent allocator that works well (but we have yet to test it in production) is snmalloc.
runtime-detection
This feature allows selecting the optimal algorithm based on available features during runtime; it has no effect on non-x86 or x86_64 platforms. When neither AVX2
nor SSE4.2
is supported, it will fall back to a native rust implementation.
Note that an application compiled with runtime-detection
will not run as fast as an application compiled for a specific CPU. The reason is that rust can't optimise as far as the instruction set when it uses the generic instruction set, and non-simd parts of the code won't be optimised for the given instruction set either.
portable
Currently disabled
An implementation of the algorithm using std::simd
and up to 512 byte wide registers, currently disabled due to dependencies and is highly experimental.
serde_impl
simd-json
is compatible with serde and serde-json
. The Value types provided implement serializers and deserializers. In addition to that simd-json
implements the Deserializer
trait for the parser so it can deserialize anything that implements the serde Deserialize
trait. Note, that serde provides both a Deserializer
and a Deserialize
trait.
That said the serde support is contained in the serde_impl
feature which is part of the default feature set of simd-json
, but it can be disabled.
known-key
The known-key
feature changes the hash mechanism for the DOM representation of the underlying JSON object, from ahash
to fxhash
. The ahash
hasher is faster at hashing and provides protection against DOS attacks by forcing multiple keys into a single hashing bucket. The fxhash
hasher, on the other hand allows for repeatable hashing results, which in turn allows memoizing hashes for well known keys and saving time on lookups. In workloads that are heavy on accessing some well-known keys, this can be a performance advantage.
The known-key
feature is optional and disabled by default and should be explicitly configured.
value-no-dup-keys
This flag has no effect on simd-json itself but purely affects the Value
structs.
The value-no-dup-keys
feature flag toggles stricter behavior for objects when deserializing into a Value
. When enabled, the Value deserializer will remove duplicate keys in a JSON object and only keep the last one. If not set duplicate keys are considered undefined behavior and Value will not make guarantees on it's behavior.
big-int-as-float
The big-int-as-float
feature flag treats very large integers that won't fit into u64 as f64 floats. This prevents parsing errors if the JSON you are parsing contains very large numbers. Keep in mind that f64 loses some precision when representing very large numbers.
safety
simd-json
uses a lot of unsafe code.
There are a few reasons for this:
- SIMD intrinsics are inherently unsafe. These uses of unsafe are inescapable in a library such as
simd-json
. - We work around some performance bottlenecks imposed by safe rust. These are avoidable, but at a performance cost. This is a more considered path in
simd-json
.
simd-json
goes through extra scrutiny for unsafe code. These steps are:
- Unit tests - to test 'the obvious' cases, edge cases, and regression cases
- Structural constructive property based testing - We generate random valid JSON objects to exercise the full
simd-json
codebase stochastically. Floats are currently excluded since slightly different parsing algorithms lead to slightly different results here. In short "is simd-json correct". - Data-oriented property-based testing of string-like data - to assert that sequences of legal printable characters don't panic or crash the parser (they might and often error so - they are not valid JSON!)
- Destructive Property based testing - make sure that no illegal byte sequences crash the parser in any way
- Fuzzing - fuzz based on upstream & jsonorg simd pass/fail cases
- Miri testing for UB
This doesn't ensure complete safety nor is at a bulletproof guarantee, but it does go a long way to assert that the library is of high production quality and fit for purpose for practical industrial applications.
Other interesting things
There are also bindings for upstream simdjson
available here
License
simd-json itself is licensed under either of
- Apache License, Version 2.0, (LICENSE-APACHE or http://www.apache.org/licenses/LICENSE-2.0)
- MIT license (LICENSE-MIT or http://opensource.org/licenses/MIT) at your option.
However it ports a lot of code from simdjson so their work and copyright on that should be respected along side.
The serde integration is based on their example and serde-json
so again, their copyright should as well be respected.