This package provides convenient encoding and decoding to/from binary data for all Swift Codable
types.
There are only few encoders and decoders available for Swift's Codable format, and Apple provides a JSONEncoder and a PropertyListEncoder for basic encoding. While these can cover some use cases (especially when interacting with Web Content through JSON), they lack encoding efficiency when designing APIs within an ecosystem. JSON, for example, is notoriously inefficient when it comes to binary data.
One very popular alternative for binary data are Google's Protocol Buffers, which offer broad support across different platforms and programming languages. But they don't support Swift's Codable
protocol, and thus require manual message definitions, the Protobuf compiler, and a lot of copying between data structures during encoding and decoding.
So if you're looking for a decently efficient binary encoder in a pure Swift project, then BinaryCodable
may be right for you. Simply make your struct
s (or classes!) conform to Codable
, and BinaryCodable
does the rest!
Efficient binary format, and broad support across different platforms and programming languages.
Much more limited than Codable
in terms of allowed types.
Message definitions must be written in separate files and compiled to Swift code.
The library for swift is swift-protobuf.
A Codable
-compatible implementation of the Protocol Buffer binary format.
It has the same limitations on the allowed types as swift-protobuf, but message definitions can be written directly in Swift.
Slower during encoding and decoding than swift-protobuf
.
A mirror of the older V2 version of BinaryCodable
, which should only be used to migrate encoded data to a different format.
If you're looking for a Codable
-compatible alternative which is also available on other platforms, with a well-defined spec.
It appears to have nearly the same encoding efficiency as BinaryCodable
.
Also offers CBOR encoding, plus a bunch of other things related to Codable
.
Encoding according to the BSON specification.
Less efficient binary represenation than Protocol Buffers and BinaryCodable
, but mature.
Used for MongoDB. There is also another implementation.
Simply include in your Package.swift
:
dependencies: [
.package(url: "https://github.com/christophhagen/BinaryCodable", from: "3.0.0")
],
targets: [
.target(name: "MyTarget", dependencies: [
.product(name: "BinaryCodable", package: "BinaryCodable")
])
]
Select your Project
, navigate to the Package Dependencies
tab, and add https://github.com/christophhagen/BinaryCodable
using the +
button.
Documentation of the library is available here (generated using Swift DocC).
Let's assume a message definition:
struct Message: Codable {
var sender: String
var isRead: Bool
var unreadCount: Int
}
Simply import the module where you need to encode or decode a message:
import BinaryCodable
Construct an encoder when converting instances to binary data, and feed the message(s) into it:
let message: Message = ...
let encoder = BinaryEncoder()
let data = try encoder.encode(message)
It's also possible to encode single values, arrays, optionals, sets, enums, dictionaries, and more, so long as they conform to Codable
.
Decoding instances from binary data works much the same way:
let decoder = BinaryDecoder()
let message = try decoder.decode(Message.self, from: data)
Alternatively, the type can be inferred:
let message: Message = try decoder.decode(from: data)
BinaryCodable
supports the use of custom encoding and decoding routines by implementing encode(to:)
and init(from:)
.
There is only one aspect that's handled differently than the Codable
documentation specifies, which is the explicit encoding of nil
in keyed containers.
Calling encodeNil(forKey:)
on a keyed container has no effect, there is no explicit nil
value encoded for the key.
This results in the contains()
function during decoding returning false
for the key.
This is different to e.g. JSON, where calling encodeNil(forKey:)
would cause the following encoding:
{
"myProperty" : null
}
The implementation of encodeNil(forKey:)
and decodeNil(forKey:)
handles this case differently, because the alternatives are not optimal:
It would be possible to explicitly encode nil
for a key, but this would cause problems with double optionals in structs (e.g. Int??), which could no longer distinguish between .some(nil)
and nil
.
To fix this issue, an additional nil
indicator would be needed for all values in keyed containers, which would decrease the efficiency of the binary format.
That doesn't seem reasonable just to support a rarely used feature, since encodeNil(forKey:)
is never called for automatically synthesized Codable conformances.
The recommendation therefore is to use encodeIfPresent(_, forKey:)
and decodeIfPresent(_, forKey:)
.
Another option would be to use a double optional, since this is basically the information encodeNil
provides: nil
, if the key is not present, .some(nil)
, if the key is present with nil
, and value
, if the key is present with a value.
It's possible for both encoding and decoding to fail.
Encoding can produce EncodingError
errors, while unsuccessful decoding produces DecodingError
s.
Both are the default Errors provided by Swift, supplied with additional information describing the nature of the error.
See the documentation of the types to learn more about the different error conditions.
The binary format provides no provisions to detect data corruption, and various errors can occur as the result of added, changed, or missing bytes and bits. Additional external measures (checksums, error-correcting codes, ...) should be applied if there is an increased risk of data corruption.
As an example, consider the simple encoding of a String
inside a struct
, which consists of a key
followed by the length of the string in bytes, and the string content.
The length of the string is encoded using variable-length encoding, so a single bit flip (in the MSB of the length byte) could result in a very large length
being decoded, causing the decoder to wait for a very large number of bytes to decode the string.
This simple error would cause much data to be skipped, potentially corrupting the data stream indefinitely.
At the same time, it is not possible to determine with certainty where the error occured, making error recovery difficult without additional information about boundaries between elements.
The decoding errors provided by the library are therefore only hints about errors likely occuring from non-conformance to the binary format or version incompatibility, which are not necessarily the true causes of the failures when data corruption is present.
The Codable
protocol uses CodingKey
definitions to identify properties of instances.
By default, coding keys are generated using the string values of the property names.
Similar to JSON encoding, BinaryCodable
can embed the property names in the encoded data.
Unlike JSON (which is human-readable), the binary representation produced by BinaryCodable
is intended for cases when efficient encoding is important.
Codable
allows the use of integer keys for each property, which significantly increases encoding efficiency.
You can specify integer keys by adding an Int
enum conforming to the CodingKey
protocol to the Codable
type:
struct Message: Codable {
var sender: String
var isRead: Bool
var unreadCount: Int
// Assign an integer to each property
enum CodingKeys: Int, CodingKey {
case sender = 1
case isRead = 2
case unreadCount = 3
}
}
The enum must have a raw value of either Int
or String
, and the cases must match the property names within the type (it is possible to omit keys for properties which should not be encoded).
Using integer keys can significantly decrease the binary size, especially for long property names.
Additionally, integer keys can be useful when intending to store the binary data persistently.
Changes to property names can be performed in the code without breaking the decoding of older data (although this can also be achieved with custom String
keys).
Notes:
- Negative values for integer keys are not supported.
- Small integer keys produce the smallest binary sizes.
- The
0
integer key shouldn't be used, since it is also used internally when encodingsuper
. - The allowed range for integer keys is from
0
(inclusive) toInt64.max
(inclusive).
While varints are efficient for small numbers, their encoding introduces a storage and computation penalty when the integers are often large, e.g. for random numbers.
BinaryCodable
provides the @FixedSizeEncoded
property wrapper, which forces integers to be encoded using their little-endian binary representations.
This means that e.g. an Int32
is always encoded as 4 byte (instead of 1-5 bytes using Varint encoding).
This makes 32-bit FixedSizeEncoded
types more efficient than Varint
if values are often larger than 2^28
(2^56
for 64-bit types).
Use the property wrapper within a Codable
definition to enforce fixed-width encoding for a property:
struct MyStruct: Codable {
/// Always encoded as 4 bytes
@FixedSizeEncoded
var largeInteger: Int32
}
The FixedSize
wrapper is available for Int
, Int32
, Int64
, UInt
, UInt32
, and UInt64
.
It has no effect for Int16
and UInt16
, which are already encoded with a fixed size by default.
Some integers can be forced to use variable-length encoding instead of fixed-size or zig-zag encoding using the @VariableLengthEncoded
property wrapper.
For Int16
and UInt16
(normally fixed-size encoded), this encoding can be more efficient if values are often smaller than 128
for UInt16
and 63
for Int16
.
For Int
, Int32
and Int64
(normally zig-zag encoded), the encoding is (marginally) more efficient if values are mostly positive.
For UInt
, UInt32
, and UInt64
the wrapper has no effect.
struct MyStruct: Codable {
/// Efficient for small, positive numbers
@VariableLengthEncoded
var value: Int16
}
The signed integers Int
, Int32
and Int64
are encoded using zig-zag encoding, which is more efficent than variable-length encoding if numbers are negative.
The @ZigZagEncoded
wrapper can force Int16
types to use zig-zag encoding instead of fixed-size encoding, which is more efficient for small (positive and negative) numbers.
The encoding is more efficient if values are between -64
and 63
.
For Int
, Int32
and Int64
the wrapper has no effect.
struct MyStruct: Codable {
/// More efficient between `-64` and `63`.
@ZigZagEncoded
var value: Int16
}
The BinaryEncoder
provides the sortKeysDuringEncoding
option, which forces fields in "keyed" containers, such as struct
properties (and some dictionaries), to be sorted in the binary data.
This sorting is done by using either the integer keys (if defined), or the property names.
Dictionaries with Int
or String
keys are also sorted.
Sorting the binary data does not influence decoding, but introduces a computation penalty during encoding. It should therefore only be used if the binary data must be consistent across multiple invocations.
Note: The sortKeysDuringEncoding
option does not guarantee deterministic binary data, and should be used with care.
Elements of any non-ordered types (Sets, Dictionaries) will appear in random order in the binary data.
The library provides the option to perform encoding and decoding of continuous streams, such as when writing sequences of elements to a file, or when transmitting data over a network.
This functionality can be used through BinaryStreamEncoder
and BinaryStreamDecoder
, causing the encoder to embed additional information into the data to allow continuous decoding (mostly length information).
Encoding and decoding is always done with sequences of one specific type, since multiple types in one stream could not be distinguished from one another.
Encoding of a stream works similarly to normal encoding:
let encoder = BinaryStreamEncoder<Int>()
let chunk1 = try encoder.encode(1)
let chunk2 = try encoder.encode(contentsOf: [2,3])
...
let data = chunk1 + chunk2 + ...
Decoding of the individual chunks, with the decoder returning all elements which can be decoded using the currently available data.
let decoder = BinaryStreamDecoder<Int>()
let decoded1 = try decoder.decode(chunk1)
print(decoded1) // [1]
let decoded2 = try decoder.decode(chunk2)
print(decoded2) // [2,3]
The decoder has an internal buffer, so incomplete data can be inserted into the decoder as it becomes available. The output of decode(_ data:)
will be empty until the next complete element is processed.
Writing data streams to files is a common use case, so the library also provides wrappers around BinaryStreamEncoder
and BinaryStreamDecoder
to perform these tasks.
The BinaryFileEncoder
can be used to sequentially write elements to a file:
let encoder = BinaryFileEncoder<DataElement>(fileAt: url)
try encoder.write(element1)
try encoder.write(element2)
...
try encoder.close() // Close the file
Elements will always be appended to the end of file, so existing files can be updated with additional data.
Decoding works in a similar way, except with a callback to handle each element as it is decoded:
let decoder = BinaryFileDecoder<DataElement>(fileAt: url)
try decoder.read { element in
// Process each element
}
There is also the possibility to read all elements at once using readAll()
, or to read only one element at a time (readElement()
).
To learn more about the encoding format, see BinaryFormat.md.
Version 3 of BinaryCodable
has significantly changed the binary format, which means that the two versions are not cross-compatible.
The format was changed to provide support for all Codable
features, which was not possible with the previous format, that was adapted from Protocol Buffers.
The redesign of the library also reduced code complexity and size, which lead to some speed improvements and greater reliability.
The support for interoperability with Protocol Buffers was also dropped, since the binary formats are no longer similar. The functionality was extracted to a separate library called ProtobufCodable.
To convert data from the legacy format to the new version, the data has to be decoded with version 2 and re-encoded with version 3. The Swift Package Manager currently doesn't allow to include the same dependency twice (with different versions), so the legacy version has been stripped down to the essentials and is provided as the stand-alone package LegacyBinaryCodable. It only allows decoding, and can be integrated as a separate dependency:
dependencies: [
.package(url: "https://github.com/christophhagen/BinaryCodable", from: "3.0.0"),
.package(url: "https://github.com/christophhagen/LegacyBinaryCodable", from: "2.0.0"),
],
targets: [
.target(name: "MyTarget", dependencies: [
.product(name: "BinaryCodable", package: "BinaryCodable"),
.product(name: "LegacyBinaryCodable", package: "LegacyBinaryCodable")
])
]
In the code, you can then decode and re-encode:
import BinaryCodable
import LegacyBinaryCodable
func reencode<T>(data: Data, as type: T.Type) throws -> Data where T: Codable {
let decoder = LegacyBinaryDecoder()
let value = try decoder.decode(T.self, from data: Data)
let encoder = BinaryEncoder()
return try encoder.encode(value)
}
The library comes with an extensive test suite, which checks that encoding works correctly for many cases.
These tests can be executed using swift test
from the package root, or when opening the package using Xcode.
MIT. See License.md
While the test suite covers many cases, there is no complete code coverage. Especially the bahaviour in error conditions can use additional testing to ensure that there are no edge cases where the program crashes, or does some other weird thing.
One option could be to use a common data storage during encoding and decoding, so that the individual containers can be converted to struct
s to make them more lightweight. It may also be possible to prevent some unnecessary data copying.
Increasing the speed of the encoding and decoding process is not a huge priority at the moment. If you have any pointers on how to improve the performance further, feel free to contribute.
Users of the library are encouraged to contribute to this repository.
Please file an issue with a description of the feature you're missing. Check other open and closed issues for similar suggestions and comment on them before creating a new issue.
File an issue with a clear description of the problem. Please include message definitions and other data where possible so that the error can be reproduced.
If you would like to extend the documentation of this library, or translate the documentation into other languages, please also open an issue, and I'll contact you for further discussions.