Skip to content

Latest commit

 

History

History
256 lines (191 loc) · 6.14 KB

README.md

File metadata and controls

256 lines (191 loc) · 6.14 KB

pysimdjson-schemaful

Schema-aware pysimdjson loader for efficient parsing of large excessive JSON inputs.

When working with external APIs you have zero influence on, you may face the following unfortunate edge-case (as we did):

  • Particular endpoint responds with a relatively massive JSON-body, say, ≥ 1 MB.
  • The amount of data you really need is several magnitudes smaller, e.g., 1 KB.
  • There is no server-side filtering available.

In such a case it may be very excessive in terms of memory, cpu time and delay to deserialize and, subsequently, validate the whole response, even when using fast JSON-deseralization libraries, such as orjson.

In our particular case we needed less than 0.1% of ~5 MB responses, which we validated with pydantic. First, we compared several combinations of deserializers and validators:

  • json + pydantic v1 (Model.parse_raw(json.loads(data)))
  • orjson + pydantic v1 (Model.parse_raw(orjson.loads(data)))
  • pysimdjson + pydantic v1 (Model.parse_raw(simdjson.loads(data)))
  • pydantic v2 (Model.model_validate_json(data))

To our surprise internal pydantic v2 parser appeared to be ~2-3 times slower than json + pydantic v1. The fastest was orjson + pydantic v1 (~2-3 times faster than json and a bit faster than full simdjson parsing). Such a speed-up, however, still comes with excessive memory spending (as a complete python dict object is created and populated on deserialization).

Thus, we ended up using pysimdjson with its fast lazy parsing and manually iterated over nested JSON objects/arrays and extracted only required keys. It is ugly, tedious and hard to maintain of course. However, it showed to be several times faster than orjson and decreased memory consumption.

Table of Contents

This package aims to automate the manual labour of lazy loading with pysimdjson.

Simply feed the JSON-schema in and the input data will be traversed and loaded with pysimdjson accordingly.

Supports

  • pydantic>=1,<3
  • python>=3.8,<3.12
  • simdjson>=2,<6 (with caveats)

Does not support complex schemas (it may be not very reasonable from the practical standpoint anyway), e.g.,

  • anyOf (Union[Model1, Model2])
  • ...

In such cases it will fully (not lazily) load the underlying objects.

  • Input JSON data is large relatively to what is needed in there, i.e., selectivity is small.
  • Other deserialization methods appear to be slower and/or more memory consuming.

If you can check all the boxes, then, this package may prove useful to you. Never use it as a default deserialization method: run some benchmarks for your particular case first, otherwise, it may and will disappoint you.

pip install pysimdjson-schemaful

If you need pydantic support

pip install "pysimdjson-schemaful[pydantic]"
import json
from simdjson_schemaful import loads

schema = {
  "type": "array",
  "items": {
    "$ref": "#/definitions/Model"
  },
  "definitions": {
    "Model": {
      "type": "object",
      "properties": {
        "key": {"type": "integer"},
      }
    }
  }
}

data = json.dumps([
    {"key": 0, "other": 1},
    {"missing": 2},
])

parsed = loads(data, schema=schema)

assert parsed == [
    {"key": 0},
    {},
]

Example with additionalProperties:

schema = {
  "type": "object",
  "additionalProperties": {
    "$ref": "#/definitions/Model",
  },
  "definitions": {
    "Model": {
      "type": "object",
      "properties": {
        "key": {"type": "integer"},
      }
    }
  }
}

data = json.dumps({
    "some": {"key": 0, "other": 1},
    "other": {"missing": 2},
})

parsed = loads(data, schema=schema)

assert parsed == {
    "some": {"key": 0},
    "other": {},
}

With re-used simdjson parser (recommended when used in a single thread, otherwise better consult pysimdjson project on thread-safety):

from simdjson import Parser

parser = Parser()
parsed = loads(data, schema=schema, parser=parser)

assert parsed == {
    "some": {"key": 0},
    "other": {},
}

With model (call BaseModel.parse_raw_simdjson):

import json
from simdjson_schemaful.pydantic.v1 import BaseModel

class Model(BaseModel):
  key: int

data = json.dumps({"key": 0, "other": 1})

obj = Model.parse_raw_simdjson(data)

With type (call parse_raw_as_simdjson):

import json
from typing import List
from simdjson_schemaful.pydantic.v1 import BaseModel, parse_raw_simdjson_as

class Model(BaseModel):
  key: int

Type = List[Model]

data = json.dumps([
  {"key": 0, "other": 1},
  {"key": 1, "another": 2},
])

obj1, obj2 = parse_raw_simdjson_as(Type, data)

With model (call BaseModel.model_validate_simdjson):

import json
from simdjson_schemaful.pydantic.v2 import BaseModel

class Model(BaseModel):
  key: int

data = json.dumps({"key": 0, "other": 1})

obj = Model.model_validate_simdjson(data)

With type adapter (call TypeAdapter.validate_simdjson)

import json
from typing import List
from simdjson_schemaful.pydantic.v2 import BaseModel, TypeAdapter

class Model(BaseModel):
  key: int

adapter = TypeAdapter(List[Model])

data = json.dumps([
  {"key": 0, "other": 1},
  {"key": 1, "another": 2},
])

obj1, obj2 = adapter.validate_simdjson(data)

TBD