Performance Tips

Here we present a few tips and tricks for squeezing maximum performance out of msgspec. They’re presented in order from “sane, definitely a good idea” to “fast, but you may not want to do this”.

Reuse Encoders/Decoders

Every call to a top-level encode function like msgspec.json.encode allocates some temporary internal state used for encoding. While fine for normal use, for maximum performance you’ll want to create an Encoder (e.g. msgspec.json.Encoder) once and reuse it for all encoding calls, avoiding paying that setup cost for every call.

>>> import msgspec

>>> encoder = msgspec.json.Encoder()  # Create once

>>> for msg in msgs:
...     data = encoder.encode(msg)  # reuse multiple times

The same goes for decoding. If you’re making multiple decode calls in a performance-sensitive code path, you’ll want to create a Decoder (e.g. msgspec.json.Decoder) once and reuse it for each call. Since decoders are typed, you may need to create multiple decoders, one for each type.

>>> import msgspec

>>> decoder = msgspec.json.Decoder(list[int])  # Create once

>>> for data in input_buffers:
...     msg = decoder.decode(data)  # reuse multiple times

Use Structs

Structs are msgspec’s native way of expressing user-defined types. They’re fast to encode/decode and fast to use. If you have data with a known schema, we recommend defining a msgspec.Struct type (or types) for your schema and preferring that over other types like dict / dataclasses / etc.

Avoid Encoding Default Values

By default, msgspec encodes all fields in a Struct type, including optional fields (those configured with a default value). If the default values are known on the decoding end (making serializing them redundant), it may be beneficial to omit default values from the encoded message. This can be done by configuring omit_defaults=True as part of the Struct definition. Omitting defaults reduces the size of the encoded message, and often also improves encoding and decoding performance (since there’s less work to do).

For more information, see Omitting Default Values.

Avoid Decoding Unused Fields

When decoding large inputs, sometimes you’re only interested in a few specific fields. Since decoding large objects is inherently allocation heavy, it may be beneficial to define a smaller msgspec.Struct type that only has the fields you require.

For example, say you’re interested in decoding some JSON from the Twitter API. A Tweet object has many nested fields on it - perhaps you only care about the tweet text, the user name, and the number of favorites. By defining struct types with only those fields, msgspec can avoid doing unnecessary work decoding fields that are never used.

>>> import msgspec

>>> class User(msgspec.Struct):
...     name: str

>>> class Tweet(msgspec.Struct):
...     user: User
...     full_text: str
...     favorite_count: int

We can then use these types to decode the example tweet json:

>>> tweet = msgspec.json.decode(example_json, type=Tweet)

>>> tweet.user.name
'Twitter Dev'

>>> tweet.user.favorite_count
70

Of course there are downsides to defining smaller “view” types, but if decoding performance is a bottleneck in your workflow, you may benefit from this technique.

For a more in-depth example of this technique, see the Conda Repodata example.

Reduce Allocations

Every call to encode / Encoder.encode allocates a new bytes object for the output. msgspec exposes an alternative Encoder.encode_into (e.g. msgspec.json.Encoder.encode_into) that writes into a pre-allocated bytearray instead (possibly reallocating to increase capacity).

This has the following uses.

Reusing an output buffer

If you’re encoding and writing messages to a socket/file in a hot loop, you may benefit from allocating a single bytearray buffer once and reusing it for every message.

For example:

encoder = msgspec.msgpack.Encoder()

# Allocate a single shared buffer
buffer = bytearray()

for msg in msgs:
    # Encode a message into the buffer at the start of the buffer.
    # Note that this overwrites any previous contents.
    encoder.encode_into(msg, buffer)

    # Write the buffer to the socket
    socket.sendall(buffer)

A few caveats:

  • Encoder.encode_into will expand the capacity of buffer as needed to fit the message size. This means that if a large message is encountered the buffer will be expanded to be equally large, but won’t be reduced back to normal afterwards (possibly bloating memory usage). You can use sys.getsizeof (or call bytearray.__sizeof__) directly to determine the actual capacity of the buffer, since len(buffer) will only reflect the part of the buffer that is written to.

  • Small messages (for some definition of “small”) likely won’t see a performance improvement from using this method, and may instead see a slowdown. We recommend using a realistic benchmark to determine if this method can benefit your workload.

Line-Delimited JSON

Some protocols require appending a suffix to an encoded message. One place where this comes up is when encoding line-delimited JSON, where every payload contains a JSON message followed by b"\n".

This could be handled in python as:

import msgspec

json_msg = msgspec.json.encode(["my", "message"])

full_payload = json_msg + b'\n'

However, this results in an unnecessary copy of json_msg, which can be avoided by using msgspec.json.Encoder.encode_into.

import msgspec

encoder = msgspec.json.Encoder()

# Allocate a buffer. We recommend using a small non-empty buffer to
# avoid reallocating for small messages. Choose something larger than
# your common message size, but not excessively large.
buffer = bytearray(64)

# Encode into the existing buffer.
encoder.encode_into(["my", "message"], buffer)

# Append a newline character without copying
buffer.extend(b"\n")

# Write the full buffer to a socket/file/etc...
socket.sendall(buffer)

Length-Prefix Framing

Some protocols require prepending a prefix to an encoded message. This comes up in Length-prefix framing , where every message is prefixed by its length stored as a fixed-width integer (e.g. a big-endian uint32). Like line-delimited JSON above, this is more efficient to do using Encoder.encode_into to avoid excessive copying.

import msgspec

encoder