A TypeScript library for efficient data serialization, achieved by separating data structure from their values and storing each in compact binary formats.
Most modern data serialization formats fall into one of two categories:
- Formats like JSON which can represent a wide range of simple and compound data types. Serialized values in these formats implicitly serialize their type so the deserializer doesn't need knowledge of the type.
- Formats like MP3 or PNG which represent one specific data type, such as a song or image. The serializer and deserializer must both know the file format in order to communicate.
The aim of this project is to store values with the flexibility of JSON but the byte efficiency of dedicated formats.
To accomplish this, custom "types" are constructed, implicitly defining binary serialization formats for specific categories of values.
Types are defined from a rich set of primitive and compound types (see Data Types).
There is a fixed (universal) serialization format for types, and each type defines a serialization format for its values.
structure-bytes
reduces the number of bytes needed to serialize values in several ways in comparison to JSON:
- Types and values are each serialized in binary formats with far less redundancy than JSON's text-based formats. For example, serializing a 32-bit signed integer in JSON can take up to 11 characters, whereas it requires only 4 bytes when serialized by
sb.IntType
. - JSON does not require each value in an array or dictionary to have the same type, which results in highly redundant serializations of collections.
structure-bytes
serializes the type of a collection's elements as part of the type rather than the value, so it isn't repeated. - One type can be used to serialize many different values. For example, when sending multiple values of the same type across a network, the type can sent along with the first value and only values need to be sent afterwards.
This project is somewhat similar to Google's Protocol Buffers, but was not designed to match its functionality.
- Use when the structure of the data is complicated. For example, if you are serializing plaintext, this can easily be accomplished without redundancy by writing to files or sockets.
- Use when there is a lot of repetition in the data. If you don't have any arrays, sets, or maps, there is less redundancy for
structure-bytes
to eliminate. - Use when serializing many values for the type (e.g. many files storing the same type of value, or many communications of the same type). This will give you the benefit of being able to keep only a single copy of the type spec.
- Types have binary serializations, not just values. (Protocol Buffers requires both the serializer and deserializer to have
.proto
definition files, which are not designed to be transmitted like values.) This compactly stores complex types and allows values to be written along with their types so they can be read without knowing their types beforehand. - Types are generated programmatically rather than by reading
.proto
files. This allows, for example, for a function which turns a type into another type that either contains an error message or an instance of the original type. - This project is designed with downloading data over HTTP in mind. If the client has already received values of the same type, the server only sends the value and the client reads it using its cached type. If the client doesn't know the type, the server serializes it along with the value and the client caches the type. This way, the type does not need to be specified in the client-side JavaScript and repeated requests to the same endpoint are very efficient.
structure-bytes
provides a larger set of primitive and compound types.
- Primitive types
Byte
(1-byte signed integer)Short
(2-byte signed integer)Int
(4-byte signed integer)Long
(8-byte signed integer)BigInt
(a signed integer with arbitrary precision)FlexInt
(a signed integer from-(2 ** 52)
to2 ** 52
with a variable-length representation)UnsignedByte
(1-byte unsigned integer)UnsignedShort
(2-byte unsigned integer)UnsignedInt
(4-byte unsigned integer)UnsignedLong
(8-byte unsigned integer)BigUnsignedInt
(an unsigned integer with arbitrary precision)FlexUnsignedInt
(an unsigned integer below2 ** 53
with a variable-length representation)Date
(8-byte signed integer representing number of milliseconds since Jan 1, 1970)Day
(3-byte signed integer representing a specific day in history)Time
(4-byte unsigned integer representing a specific time of day)Float
(IEEE 32-bit floating-point number)Double
(IEEE 64-bit floating-point number)Boolean
(a single true or false value)BooleanTuple
(a constant-length array ofBoolean
s)BooleanArray
(a variable-length array ofBoolean
s)Char
(a single UTF-8 character)String
(UTF-8-encoded text)Octets
(anArrayBuffer
(raw binary data))
- Compound/recursive types
Tuple<Type>
(a constant-length array ofType
s)Struct
(an object with a fixed set of fields, each with a fixed name and type)Array<Type>
(a variable-length array ofType
s)Set<Type>
(like anArray
, except creates a set when read)Map<KeyType, ValueType>
(a mapping ofKeyType
instances toValueType
instances)Enum<Type>
(a fixed set ofType
s; useful when only a small subset ofType
instances represent possible values, especially withString
s)Choice
(a fixed set of types that values can take on)Recursive<Type>
(a type that can reference itself and be used to serialize circular data structures)Singleton<Type>
(a type that serializes only one value)Optional<Type>
(eithernull
orundefined
or an instance ofType
)Pointer<Type>
(allows multiple instances ofType
with the same serialization to be stored only once)
The docs
folder is hosted at
calebsander.github.io/structure-bytes.
const sb = require('structure-bytes')
const colorType = new sb.TupleType({
type: new sb.UnsignedByteType,
length: 3
})
const routeAttributesType = new sb.StructType({
color: colorType,
description: new sb.PointerType(new sb.StringType),
direction_names: new sb.PointerType(
new sb.ArrayType(new sb.StringType)
),
long_name: new sb.StringType,
text_color: colorType,
type: new sb.UnsignedByteType
})
const routesType = new sb.ArrayType(
new sb.StructType({
id: new sb.StringType,
attributes: routeAttributesType
})
)
const sb = require('structure-bytes')
const colorType = new sb.TupleType({
type: new sb.UnsignedByteType,
length: 3
})
const routeAttributesType = new sb.StructType({
color: colorType,
description: new sb.PointerType(new sb.StringType),
direction_names: new sb.PointerType(
new sb.ArrayType(new sb.StringType)
),
long_name: new sb.StringType,
text_color: colorType,
type: new sb.UnsignedByteType
})
const routesType = new sb.ArrayType(
new sb.StructType({
id: new sb.StringType,
attributes: routeAttributesType
})
)
const typeBuffer = routesType.toBuffer()
// ArrayBuffer { byteLength: 92 }
console.log(Buffer.from(typeBuffer))
// <Buffer 52 51 02 0a 61 74 74 72 69 62 75 74 65 73 51 06 05 63 6f 6c 6f 72 50 11 03 0b 64 65 73 63 72 69 70 74 69 6f 6e 70 41 0f 64 69 72 65 63 74 69 6f 6e 5f ... >
const toRGB = hex => [
parseInt(hex.slice(0, 2), 16),
parseInt(hex.slice(2, 4), 16),
parseInt(hex.slice(4, 6), 16)
]
fetch('https://api-v3.mbta.com/routes?filter[type]=0,1')
.then(res => res.json())
.then(({data}) => {
const routes = data.map(({id, attributes}) => {
delete attributes.short_name
delete attributes.sort_order
attributes.color = toRGB(attributes.color)
attributes.text_color = toRGB(attributes.text_color)
return {id, attributes}
})
const valueBuffer = routesType.valueBuffer(routes)
// ArrayBuffer { byteLength: 306 }
// (for comparison, JSON.stringify(routes).length == 1498)
})
const sb = require('structure-bytes')
// Buffer obtained somehow
const typeBuffer = new Uint8Array([0x50, 0x11, 0x03]).buffer
const type = sb.r.type(typeBuffer)
console.log(type)
// TupleType { type: UnsignedByteType {}, length: 3 }
// Buffer obtained somehow
const valueBuffer = new Uint8Array([0x00, 0x80, 0xFF]).buffer
console.log(type.readValue(valueBuffer))
// [ 0, 128, 255 ]
The I/O functions are implemented with callbacks, but can all be used with util.promisify()
if you prefer to use Promise
s instead.
See the documentation for examples.
I recommend the file extensions .sbt
for a serialized type, .sbv
for a serialized value, and .sbtv
for a serialized type and value.
const fs = require('fs')
const sb = require('structure-bytes')
const type = new sb.EnumType({
type: new sb.StringType,
values: [
'ON_TIME',
'LATE',
'CANCELLED',
'UNKNOWN'
]
})
sb.writeType({
type,
outStream: fs.createWriteStream('status.sbt')
}, err => {
sb.readType(fs.createReadStream('status.sbt'), (err, readType) =>
console.log(readType)
/*
EnumType {
type: StringType {},
values: [ 'ON_TIME', 'LATE', 'CANCELLED', 'UNKNOWN' ] }
*/
)
})
const value = 'CANCELLED'
sb.writeValue({
type,
value,
outStream: fs.createWriteStream('status.sbv')
}, err => {
sb.readValue({
type,
inStream: fs.createReadStream('status.sbv')
}, (err, value) =>
console.log(value)
// 'CANCELLED'
)
})
sb.writeTypeAndValue({
type,
value,
outStream: fs.createWriteStream('status.sbtv')
}, err => {
sb.readTypeAndValue(fs.createReadStream('status.sbtv'), (err, type, value) =>
console.log(value)
// 'CANCELLED'
)
})
Server-side:
const http = require('http')
const sb = require('structure-bytes')
const type = new sb.DateType
http.createServer((req, res) => {
sb.httpRespond({req, res, type, value: new Date}, err =>
console.log('Responded')
)
}).listen(80)
Client-side:
<script src = '/structure-bytes/compiled/download.js'></script>
<script>
// 'date' specifies the name of the type being transferred so it can be cached for future requests
sb.download({
name: 'date',
url: '/',
options: {} // optional options to pass to fetch() (e.g. cookies, headers, etc.)
})
.then(value =>
console.log(value.getFullYear())
// 2017
)
</script>
Server-side:
const http = require('http')
const sb = require('structure-bytes')
http.createServer((req, res) => {
sb.readValue({type: new sb.FlexUnsignedIntType, inStream: req}, (err, value) => {
res.end(String(value))
})
}).listen(80)
Client-side:
<script src = '/structure-bytes/compiled/upload.js'></script>
<button onclick = 'upload()'>Click me</button>
<script>
let clickCount = 0
function upload() {
clickCount++
sb.upload({
type: new sb.FlexUnsignedIntType,
value: clickCount,
url: '/click',
options: {method: 'POST'} // options to pass to fetch(); method 'POST' is required
})
.then(response => response.text())
.then(alert)
}
</script>
The entire project is written in TypeScript and declaration files are automatically generated. That means that if you are using TypeScript, the compiler can automatically infer the type of the various values exported from structure-bytes
. To import the package, use:
import * as sb from 'structure-bytes'
The typings allow the compiler to automatically infer what types of values a Type
can serialize. For example:
const type = new sb.StructType({
abc: new sb.DoubleType,
def: new sb.MapType(
new sb.StringType,
new sb.DayType
)
})
type.valueBuffer(/*...*/)
/*
If you hover over "valueBuffer", you can see
that it requires the value to be of type:
{ abc: number | string, def: Map<string, Date> }
*/
You can also add explicit VALUE
generics to make the compiler check that your Type
serializes the correct types of values.
Many non-primitive types (e.g. ArrayType
, MapType
, StructType
) can take either one or two generics.
The first, VALUE
, specifies the type of values the type can serialize.
The second, READ_VALUE
, specifies the type of values the type will deserialize and defaults to VALUE
if omitted.
READ_VALUE
must extend VALUE
.
Generally the two generics are the same, except in the case of numeric types (which write number | string
but read number
) and OptionalType
(which writes E | null | undefined
but reads E | null
).
With StructType
:
class Car {
constructor(
public make: string,
public model: string,
public year: number
) {}
}
const carType = new sb.StructType<Car>({
make: new sb.StringType,
model: new sb.StringType,
year: new sb.UnsignedShortType
})
// The compiler would have complained if one of the fields were missing
// or one of the field's types didn't match the type of the field's values,
// e.g. "make: new sb.BooleanType"
If you have transient fields (i.e. they shouldn't be serialized), you can create a separate interface for the fields that should be serialized:
interface SerializedCar {
make: string
model: string
year: number
}
class Car implements SerializedCar {
public id: number // transient field
constructor(public make: string, public model: string, public year: number) {
this.id = Math.floor(Math.random() * 1e6)
}
}
const carType = new sb.StructType<SerializedCar>({
make: new sb.StringType,
model: new sb.StringType,
year: new sb.UnsignedShortType
})
With ChoiceType
, you can let the type be inferred automatically if each of the possible types writes the same type of values:
const choiceType = new sb.ChoiceType([ // Type<number | string, number>
new sb.ByteType,
new sb.ShortType,
new sb.IntType,
new sb.DoubleType
])
However, if the value types are not the same, TypeScript will complain about them not matching, and you should express the union type explicitly:
interface RGB {
r: number
g: number
b: number
}
interface HSV {
h: number
s: number
v: number
}
type CSSColor = string
const choiceType = new sb.ChoiceType<RGB | HSV | CSSColor>([
new sb.StructType<RGB>({
r: new sb.FloatType,
g: new sb.FloatType,
b: new sb.FloatType
}),
new sb.StructType<HSV>({
h: new sb.FloatType,
s: new sb.FloatType,
v: new sb.FloatType
}),
new sb.StringType
])
A RecursiveType
has no way to infer its value type, so you should always provide the VALUE
generic:
interface Cons<A> {
head: A
tail: List<A>
}
interface List<A> {
list: Cons<A> | null // null for empty list
}
const recursiveType = new sb.RecursiveType<List<string>>('linked-list')
recursiveType.setType(new sb.StructType<List<string>>({
list: new sb.OptionalType(
new sb.StructType<Cons<string>>({
head: new sb.StringType,
tail: recursiveType
})
)
}))
recursiveType.valueBuffer({
list: {
head: '1',
tail: {
list: {
head: '2',
tail: {list: null}
}
}
}
})
When reading types from buffers and streams, they will be of type Type<any>
.
You should specify their value types before using them to write values:
const booleanType = new sb.BooleanType
const readType = sb.r.type(booleanType.toBuffer()) // Type<any>
// Will throw a runtime error
readType.valueBuffer('abc')
//vs.
const castReadType: sb.Type<boolean> = sb.r.type(booleanType.toBuffer())
// Will throw a compiler error
castReadType.valueBuffer('abc')
It may also sometimes be useful to be more specific about what types of values you want to be able to serialize. For example, if you want to serialize integer values that will always be represented as numbers (and never in string form), you can force the compiler to error out if you try to serialize a string value with the following:
const intType: sb.Type<number> = new sb.IntType
// Now this is valid:
intType.valueBuffer(100)
// But this is not, even though it would be if you omitted the type annotation on intType:
intType.valueBuffer('100')
In the following definitions, flexInt
means a variable-length unsigned integer with the following format, where X
represents either 0
or 1
:
[0b0XXXXXXX]
stores values from0
to2**7 - 1
in their unsigned 7-bit integer representations[0b10XXXXXX, 0bXXXXXXXX]
stores values from2**7
to2**7 + 2**14 - 1
, where a valuex
is encoded into the unsigned 14-bit representation ofx - 2**7
[0b110XXXXX, 0bXXXXXXXX, 0bXXXXXXXX]
stores values from2**7 + 2**14
to2**7 + 2**14 + 2**21 - 1
, where a valuex
is encoded into the unsigned 21-bit representation ofx - (2**7 + 2**14)
- and so on, up to 8-byte representations
All numbers are stored in big-endian format.
The binary format of a type contains a byte identifying the class of the type followed by additional information to describe the type, if necessary.
For example, new sb.UnsignedIntType
translates into [0x13]
, and new sb.StructType({abc: new sb.ByteType, def: new sb.StringType})
translates into:
[
0x51 /*StructType*/,
2 /*2 fields*/,
/*first field's name*/, 0x61 /*a*/, 0x62 /*b*/, 0x63 /*c*/, 0, 0x01 /*ByteType*/,
/*second field's name*/, 0x64 /*d*/, 0x65 /*e*/, 0x66 /*f*/, 0, 0x41 /*StringType*/
]
If the type has already been written to the buffer, it is also valid to serialize the type as:
0xFF
offset
([position of0xFF
in buffer] - [position of type in buffer]) -flexInt
For example:
const someType = new sb.TupleType({
type: new sb.FloatType,
length: 3
})
const type = new sb.StructType({
one: someType,
two: someType
})
// type serializes to
[
0x51 /*StructType*/,
2 /*2 fields*/,
/*first field's name*/, 0x6f /*o*/, 0x6e /*n*/, 0x65 /*e*/, 0,
0x50 /*TupleType*/,
0x20 /*FloatType*/,
3 /*3 floats in the tuple*/,
/*second field's name*/, 0x74 /*t*/, 0x77 /*w*/, 0x6f /*o*/, 0,
0xff, /*type is defined previously*/
7 /*type is defined 7 bytes before the 0xff byte*/
]
In the following definitions, type
means the binary type format.
ByteType
: identifier0x01
ShortType
: identifier0x02
IntType
: identifier0x03
LongType
: identifier0x04
BigIntType
: identifier0x05
FlexIntType
: identifier0x07
UnsignedByteType
: identifier0x11
UnsignedShortType
: identifier0x12
UnsignedIntType
: identifier0x13
UnsignedLongType
: identifier0x14
BigUnsignedIntType
: identifier0x15
FlexUnsignedIntType
: identifier0x17
DateType
: identifier0x1A
DayType
: identifier0x1B
TimeType
: identifier0x1C
FloatType
: identifier0x20
DoubleType
: identifier0x21
BooleanType
: identifier0x30
BooleanTupleType
: identifier0x31
, payload:length
-flexInt
BooleanArrayType
: identifier0x32
CharType
: identifier0x40
StringType
: identifier0x41
OctetsType
: identifier0x42
TupleType
: identifier0x50
, payload:elementType
-type
length
-flexInt
StructType
: identifier0x51
, payload:fieldCount
-flexInt
fieldCount
instances offield
:name
- a serializedStringType
valuefieldType
-type
ArrayType
: identifier0x52
, payload:elementType
-type
SetType
: identifier0x53
, payload identical toArrayType
:elementType
-type
MapType
: identifier0x54
, payload:keyType
-type
valueType
-type
EnumType
: identifier0x55
, payload:valueType
-type
valueCount
-flexInt
valueCount
instances ofvalue
:value
- a value that conforms tovalueType
ChoiceType
: identifier0x56
, payload:typeCount
-flexInt
typeCount
instances ofpossibleType
:possibleType
-type
RecursiveType
: identifier0x57
, payload:recursiveID
(an identifier unique to this recursive type in this type buffer) -flexInt
- If this is the first instance of this recursive type in this buffer:
recursiveType
(the type definition of this type) -type
SingletonType
: identifier0x59
, payload:valueType
-type
value
- a value that conforms tovalueType
OptionalType
: identifier0x60
, payload:typeIfNonNull
-type
PointerType
: identifier0x70
, payload:targetType
-type
ByteType
: 1-byte integerShortType
: 2-byte integerIntType
: 4-byte integerLongType
: 8-byte integerBigIntType
:byteCount
-flexInt
number
-byteCount
-byte integer
FlexIntType
:flexInt
ofvalue * 2
ifvalue
is non-negative,-2 * value - 1
ifvalue
is negativeUnsignedByteType
: 1-byte unsigned integerUnsignedShortType
: 2-byte unsigned integerUnsignedIntType
: 4-byte unsigned integerUnsignedLongType
: 8-byte unsigned integerBigUnsignedIntType
:byteCount
-flexInt
number
-byteCount
-byte unsigned integer
FlexUnsignedIntType
:flexInt
DateType
: 8-byte signed integer storing milliseconds in Unix timeDayType
: 3-byte signed integer storing days since the Unix time epochTimeType
: 4-byte unsigned integer storing milliseconds since the start of the dayFloatType
: single precision (4-byte) IEEE floating pointDoubleType
: double precision (8-byte) IEEE floating pointBooleanType
: 1-byte value, either0x00
forfalse
or0xFF
fortrue
BooleanTupleType
:ceil(length / 8)
bytes, where then
th boolean is stored at the(n % 8)
th MSB (0
-indexed) of thefloor(n / 8)
th byte (0
-indexed)BooleanArrayType
:length
-flexInt
booleans
-ceil(length / 8)
bytes, where then
th boolean is stored at the(n % 8)
th MSB (0
-indexed) of thefloor(n / 8)
th byte (0
-indexed)
CharType
: UTF-8 codepoint (somewhere between 1 and 4 bytes long)StringType
:string
- a UTF-8 string of any length not containing'\0'
0x00
to mark the end of the string
OctetsType
:length
-flexInt
octets
-length
bytes
TupleType
:length
values serialized byelementType
StructType
:- For each field in order of declaration in the type format:
- The field's value serialized by
fieldType
- The field's value serialized by
- For each field in order of declaration in the type format:
ArrayType
:length
-flexInt
length
values serialized byelementType
SetType
:size
-flexInt
size
values serialized byelementType
MapType
:size
-flexInt
size
instances ofkeyValuePair
:key
- value serialized bykeyType
value
- value serialized byvalueType
EnumType
:index
of value in values array -flexInt
ChoiceType
:index
of type in possible types array -flexInt
value
- value serialized by specified type
RecursiveType
:valueNotYetWrittenInBuffer
- byte containing either0x00
or0xFF
- If
valueNotYetWrittenInBuffer
:value
- value serialized byrecursiveType
- Else:
offset
([position of first byte ofoffset
in buffer] - [position ofvalue
in buffer]) -flexInt
SingletonType
:- No value bytes
OptionalType
:valueIsNonNull
- byte containing either0x00
or0xFF
- If
valueIsNonNull
:value
- value serialized bytypeIfNonNull
PointerType
:offset
-flexInt
:- If this is the first instance of these value bytes in the write buffer, then
0
- Otherwise, ([position of first byte of
offset
] - [position of first byte ofoffset
in the last instance of these value bytes])
- If this is the first instance of these value bytes in the write buffer, then
- If
offset
is0
(i.e. this is the first instance):- Value serialized by
targetType
- Value serialized by
Versions will be of the form x.y.z
. They are in the semver
format:
x
is the major release; changes to it represent significant or breaking changes to the API, or to the type or value binary specification.y
is the minor release; changes to it represent new features that do not break backwards compatibility.z
is the patch release; changes to it represent bug fixes that do not change the documented API.
To test the Node.js code, run npm test
.
To test the HTTP transaction code, run node client-test/server.js
and open localhost:8080
in your browser. Open each link in a new page. Upload
and Download
should each alert Success
, while Upload & Download
should alert Upload: Success
and Download: Success
.
Caleb Sander, 2017