|
| 1 | +Porting from orjson |
| 2 | +=================== |
| 3 | + |
| 4 | +If you're currently using orjson_ and considering migrating to ``msgspec``, |
| 5 | +this guide covers the key API differences and shows how to translate common |
| 6 | +patterns. |
| 7 | + |
| 8 | +Why migrate? |
| 9 | +------------ |
| 10 | + |
| 11 | +Both ``orjson`` and ``msgspec`` are high-performance JSON libraries written in |
| 12 | +compiled languages. |
| 13 | + |
| 14 | +There are a few reasons you might prefer ``msgspec``: |
| 15 | + |
| 16 | +- **Typed decoding and validation**. ``orjson.loads`` always returns plain |
| 17 | + Python builtins (``dict``, ``list``, ...). With ``msgspec`` you can decode |
| 18 | + directly into typed objects (``Struct``, ``dataclass``, ``TypedDict``, etc.) |
| 19 | + while validating the schema in a single pass. See :doc:`../usage` for details. |
| 20 | + |
| 21 | +- **Schema evolution**. ``msgspec`` has first-class support for evolving |
| 22 | + schemas over time without breaking existing clients. See |
| 23 | + :doc:`../schema-evolution`. |
| 24 | + |
| 25 | +- **Multiple protocols**. ``msgspec`` also supports MessagePack, YAML, and |
| 26 | + TOML with a consistent API, making it easy to switch between formats. |
| 27 | + |
| 28 | +- **JSON Schema generation**. ``msgspec`` can generate JSON Schema definitions |
| 29 | + from your types via :func:`msgspec.json.schema`. Useful for OpenAPI |
| 30 | + documentation and API tooling. |
| 31 | + |
| 32 | +Basic usage |
| 33 | +----------- |
| 34 | + |
| 35 | +For most users, migration is straightforward -- swap function names and you're |
| 36 | +done: |
| 37 | + |
| 38 | +.. code-block:: python |
| 39 | +
|
| 40 | + # Before (orjson) |
| 41 | + import orjson |
| 42 | +
|
| 43 | + data = orjson.dumps({"hello": "world"}) |
| 44 | + obj = orjson.loads(data) |
| 45 | +
|
| 46 | + # After (msgspec) |
| 47 | + import msgspec |
| 48 | +
|
| 49 | + data = msgspec.json.encode({"hello": "world"}) |
| 50 | + obj = msgspec.json.decode(data) |
| 51 | +
|
| 52 | +Both return ``bytes`` from encoding and accept ``bytes`` or ``str`` for |
| 53 | +decoding. |
| 54 | + |
| 55 | +Typed decoding |
| 56 | +-------------- |
| 57 | + |
| 58 | +The main advantage of ``msgspec`` over ``orjson`` is typed decoding. Instead of |
| 59 | +getting a plain ``dict`` back, you can decode directly into structured types: |
| 60 | + |
| 61 | +.. code-block:: python |
| 62 | +
|
| 63 | + import msgspec |
| 64 | +
|
| 65 | + class User(msgspec.Struct): |
| 66 | + name: str |
| 67 | + age: int |
| 68 | + email: str | None = None |
| 69 | +
|
| 70 | + data = b'{"name": "Alice", "age": 30}' |
| 71 | + user = msgspec.json.decode(data, type=User) |
| 72 | +
|
| 73 | + print(user.name) # "Alice" |
| 74 | + print(user.age) # 30 |
| 75 | + print(user.email) # None |
| 76 | +
|
| 77 | +This validates the data while decoding -- if a field is missing or has the |
| 78 | +wrong type, a ``msgspec.ValidationError`` is raised with a clear error message. |
| 79 | + |
| 80 | +Encoder and Decoder classes |
| 81 | +--------------------------- |
| 82 | + |
| 83 | +If you're making multiple calls, using the ``Encoder`` and ``Decoder`` classes |
| 84 | +is more efficient: |
| 85 | + |
| 86 | +.. code-block:: python |
| 87 | +
|
| 88 | + # orjson has no stateful encoder/decoder |
| 89 | +
|
| 90 | + # msgspec |
| 91 | + encoder = msgspec.json.Encoder() |
| 92 | + decoder = msgspec.json.Decoder(User) |
| 93 | +
|
| 94 | + data = encoder.encode(user) |
| 95 | + user = decoder.decode(data) |
| 96 | +
|
| 97 | +Option mapping |
| 98 | +-------------- |
| 99 | + |
| 100 | +``orjson`` uses bitwise ``OPT_*`` flags. ``msgspec`` uses keyword arguments on |
| 101 | +``Encoder``/``Decoder`` or the ``encode``/``decode`` functions. |
| 102 | + |
| 103 | +Sorting keys |
| 104 | +^^^^^^^^^^^^ |
| 105 | + |
| 106 | +.. code-block:: python |
| 107 | +
|
| 108 | + # orjson |
| 109 | + orjson.dumps(obj, option=orjson.OPT_SORT_KEYS) |
| 110 | +
|
| 111 | + # msgspec |
| 112 | + msgspec.json.encode(obj, order="sorted") |
| 113 | +
|
| 114 | + # Or with an Encoder |
| 115 | + encoder = msgspec.json.Encoder(order="sorted") |
| 116 | +
|
| 117 | +``msgspec`` also supports ``order="deterministic"`` which provides a |
| 118 | +consistent ordering without the cost of full alphabetical sorting. |
| 119 | + |
| 120 | +Pretty-printing |
| 121 | +^^^^^^^^^^^^^^^ |
| 122 | + |
| 123 | +.. code-block:: python |
| 124 | +
|
| 125 | + # orjson |
| 126 | + orjson.dumps(obj, option=orjson.OPT_INDENT_2) |
| 127 | +
|
| 128 | + # msgspec -- use format() on already-encoded JSON |
| 129 | + raw = msgspec.json.encode(obj) |
| 130 | + pretty = msgspec.json.format(raw, indent=2) |
| 131 | +
|
| 132 | +Unlike ``orjson``'s single-pass approach, ``msgspec`` uses a two-step process: |
| 133 | +encode first, then format. ``msgspec.json.format`` works on any JSON bytes/str, |
| 134 | +not just ``msgspec``-encoded data, and supports arbitrary indent levels. |
| 135 | + |
| 136 | +Custom type serialization |
| 137 | +^^^^^^^^^^^^^^^^^^^^^^^^^ |
| 138 | + |
| 139 | +.. code-block:: python |
| 140 | +
|
| 141 | + # orjson |
| 142 | + def default(obj): |
| 143 | + if isinstance(obj, complex): |
| 144 | + return {"real": obj.real, "imag": obj.imag} |
| 145 | + raise TypeError |
| 146 | +
|
| 147 | + orjson.dumps(obj, default=default) |
| 148 | +
|
| 149 | + # msgspec |
| 150 | + def enc_hook(obj): |
| 151 | + if isinstance(obj, complex): |
| 152 | + return {"real": obj.real, "imag": obj.imag} |
| 153 | + raise NotImplementedError |
| 154 | +
|
| 155 | + msgspec.json.encode(obj, enc_hook=enc_hook) |
| 156 | +
|
| 157 | +The key difference: ``orjson``'s ``default`` should raise ``TypeError`` for |
| 158 | +unhandled types, while ``msgspec``'s ``enc_hook`` should raise |
| 159 | +``NotImplementedError``. |
| 160 | + |
| 161 | +Appending newlines |
| 162 | +^^^^^^^^^^^^^^^^^^ |
| 163 | + |
| 164 | +.. code-block:: python |
| 165 | +
|
| 166 | + # orjson |
| 167 | + orjson.dumps(obj, option=orjson.OPT_APPEND_NEWLINE) |
| 168 | +
|
| 169 | + # msgspec -- no built-in flag, just append manually |
| 170 | + msgspec.json.encode(obj) + b"\n" |
| 171 | +
|
| 172 | +Naive datetimes |
| 173 | +^^^^^^^^^^^^^^^ |
| 174 | + |
| 175 | +By default, both libraries encode naive datetimes identically (without |
| 176 | +timezone info): |
| 177 | + |
| 178 | +.. code-block:: python |
| 179 | +
|
| 180 | + import datetime |
| 181 | +
|
| 182 | + dt = datetime.datetime(2024, 1, 1, 12, 0, 0) |
| 183 | +
|
| 184 | + orjson.dumps(dt) # b'"2024-01-01T12:00:00"' |
| 185 | + msgspec.json.encode(dt) # b'"2024-01-01T12:00:00"' |
| 186 | +
|
| 187 | +``orjson`` has ``OPT_NAIVE_UTC`` to force treating naive datetimes as UTC: |
| 188 | + |
| 189 | +.. code-block:: python |
| 190 | +
|
| 191 | + orjson.dumps(dt, option=orjson.OPT_NAIVE_UTC) |
| 192 | + # b'"2024-01-01T12:00:00+00:00"' |
| 193 | +
|
| 194 | +``msgspec`` has no equivalent flag. Convert before encoding instead: |
| 195 | + |
| 196 | +.. code-block:: python |
| 197 | +
|
| 198 | + def ensure_utc(dt): |
| 199 | + if dt.tzinfo is None: |
| 200 | + return dt.replace(tzinfo=datetime.timezone.utc) |
| 201 | + return dt |
| 202 | +
|
| 203 | + msgspec.json.encode(ensure_utc(dt)) |
| 204 | +
|
| 205 | +.. note:: |
| 206 | + |
| 207 | + ``enc_hook`` is only called for types that ``msgspec`` doesn't natively |
| 208 | + support. Since ``datetime`` is natively supported, you need to convert |
| 209 | + before passing to ``encode``. |
| 210 | + |
| 211 | +Numpy arrays |
| 212 | +^^^^^^^^^^^^ |
| 213 | + |
| 214 | +.. code-block:: python |
| 215 | +
|
| 216 | + import numpy as np |
| 217 | +
|
| 218 | + # orjson -- native support |
| 219 | + orjson.dumps(arr, option=orjson.OPT_SERIALIZE_NUMPY) |
| 220 | +
|
| 221 | + # msgspec -- use enc_hook |
| 222 | + def enc_hook(obj): |
| 223 | + if isinstance(obj, np.ndarray): |
| 224 | + return obj.tolist() |
| 225 | + raise NotImplementedError |
| 226 | +
|
| 227 | + msgspec.json.encode(arr, enc_hook=enc_hook) |
| 228 | +
|
| 229 | +Pre-serialized JSON fragments |
| 230 | +^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ |
| 231 | + |
| 232 | +.. code-block:: python |
| 233 | +
|
| 234 | + # orjson |
| 235 | + orjson.dumps({"key": orjson.Fragment(b'{"already":"json"}')}) |
| 236 | +
|
| 237 | + # msgspec |
| 238 | + msgspec.json.encode({"key": msgspec.Raw(b'{"already":"json"}')}) |
| 239 | +
|
| 240 | +Decimal and UUID handling |
| 241 | +------------------------- |
| 242 | + |
| 243 | +``msgspec`` has built-in support for both ``Decimal`` and ``UUID``, with |
| 244 | +configurable output formats: |
| 245 | + |
| 246 | +.. code-block:: python |
| 247 | +
|
| 248 | + from decimal import Decimal |
| 249 | + from uuid import uuid4 |
| 250 | +
|
| 251 | + enc = msgspec.json.Encoder( |
| 252 | + decimal_format="number", # or "string" (default) |
| 253 | + uuid_format="hex", # or "canonical" (default) |
| 254 | + ) |
| 255 | +
|
| 256 | +``orjson`` requires a ``default`` callback for ``Decimal``. |
| 257 | + |
| 258 | +Line-delimited JSON |
| 259 | +------------------- |
| 260 | + |
| 261 | +``msgspec`` has built-in support for encoding and decoding newline-delimited |
| 262 | +JSON (NDJSON): |
| 263 | + |
| 264 | +.. code-block:: python |
| 265 | +
|
| 266 | + encoder = msgspec.json.Encoder() |
| 267 | + decoder = msgspec.json.Decoder(User) |
| 268 | +
|
| 269 | + # Encode multiple objects as NDJSON |
| 270 | + data = encoder.encode_lines([user1, user2, user3]) |
| 271 | +
|
| 272 | + # Decode NDJSON back |
| 273 | + users = decoder.decode_lines(data) |
| 274 | +
|
| 275 | +``orjson`` has no equivalent -- you'd need to encode each object separately and |
| 276 | +join with newlines. |
| 277 | + |
| 278 | +Error handling |
| 279 | +-------------- |
| 280 | + |
| 281 | +.. code-block:: python |
| 282 | +
|
| 283 | + # orjson |
| 284 | + try: |
| 285 | + orjson.loads(bad_data) |
| 286 | + except orjson.JSONDecodeError: |
| 287 | + ... |
| 288 | +
|
| 289 | + # msgspec |
| 290 | + try: |
| 291 | + msgspec.json.decode(bad_data, type=User) |
| 292 | + except msgspec.ValidationError: |
| 293 | + # Schema validation error (wrong type, missing field, etc.) |
| 294 | + ... |
| 295 | + except msgspec.DecodeError: |
| 296 | + # Malformed JSON |
| 297 | + ... |
| 298 | +
|
| 299 | +``msgspec`` distinguishes between malformed JSON (``DecodeError``) and valid |
| 300 | +JSON that doesn't match the schema (``ValidationError``, a subclass of |
| 301 | +``DecodeError``). Note that ``ValidationError`` must be caught first since it |
| 302 | +inherits from ``DecodeError``. |
| 303 | + |
| 304 | +Features with no direct equivalent |
| 305 | +----------------------------------- |
| 306 | + |
| 307 | +Some ``orjson`` options have no direct ``msgspec`` flag, but can be achieved |
| 308 | +through other means: |
| 309 | + |
| 310 | +.. list-table:: |
| 311 | + :header-rows: 1 |
| 312 | + |
| 313 | + * - orjson |
| 314 | + - msgspec workaround |
| 315 | + * - ``OPT_NON_STR_KEYS`` |
| 316 | + - Use ``enc_hook``/``dec_hook`` or ``msgspec.to_builtins(obj, str_keys=True)`` |
| 317 | + * - ``OPT_OMIT_MICROSECONDS`` |
| 318 | + - Convert before encoding: ``dt.replace(microsecond=0)``. ``enc_hook`` cannot |
| 319 | + be used here because ``datetime`` is a natively supported type. |
| 320 | + * - ``OPT_STRICT_INTEGER`` (53-bit) |
| 321 | + - Use ``Meta(ge=-(2**53 - 1), le=2**53 - 1)`` constraint |
| 322 | + * - ``OPT_PASSTHROUGH_DATACLASS`` |
| 323 | + - Not directly supported. ``dataclass`` is natively supported, so ``enc_hook`` |
| 324 | + is not called for it. Convert to a ``dict`` before encoding if you need |
| 325 | + custom serialization. |
| 326 | + * - ``OPT_PASSTHROUGH_SUBCLASS`` |
| 327 | + - Not applicable. ``msgspec`` always serializes subclasses of builtins as |
| 328 | + their parent type, with no option to redirect them to ``enc_hook``. |
| 329 | + |
| 330 | +.. _orjson: https://github.com/ijl/orjson |
0 commit comments