- Purpose
- Inspiration
- What is it?
- Why do I need this library? Can't I define my own rich types?
- Usage:
- Subtleties
To make it easy to encode more business logic in the type system, and to thereby improve code safety and clarity.
- https://realm.io/news/altconf-justin-spahr-summers-type-safety/
- http://www.johndcook.com/blog/2015/12/01/dimensional-analysis-and-types/
- http://www.joelonsoftware.com/articles/Wrong.html
A Semantic Type is a context-specific type which wraps an underlying value used in specific circumstances. It's a type created to capture and convey some meaning about a value which isn't already captured by the type used to encode the value.
You can think of types such as Years
and Meters
used in place of Double
, or even types such as ShortString
and AlphanumericString
used in place of String
.
This library makes it easy and seamless to create full-fledged semantic types that are no less convenient to use then their primitive counterparts. If you just want to see some code, feel free to jump ahead.
Instances of primitive types (such as Int
, Double
, String
, etc.) are often used under widely incompatiblee circumstances; a Double
instance may encode a task-completion-percentage in one context, a time interval in another context, and pixel width in yet another.
Though associated with identical data (e.g. floating points), such instances must never be confused with one another; we wouldn't want to pass a time interval to a function expecting pixel width. At times, such values are also associated with context-specific constraints which must be carefully maintained; a percentge-encoding Double
must capture a number between 0 to and 100.
A SemanticType
is a purpose-specific type wrapping a primitive value used in a particular context. It creates a type-level distinction between such contexts, and makes it possible to encode constraint-enforcing validations & transformations right at the type level.
Say your program contains the Person
and Robot
types:
struct Person {
var name: String
var age: Int
}
struct Robot {
var id: Int
var batteryPercentage: Int
}
The swift compiler draws a sharp distinction between a foo: Person
variable and a bar: Robot
variable; there is no danger of accidentally passing a Robot
to a function that expects a Person
, and a quick option-click type inspection immediaely illuminates the kinds of computations in which we could expect the variable to participate.
The same is not true when we look at the Int
instance fields defined above.
Though Person.age
, Robot.id
, and Robot.batteryPercentage
capture entirely different kinds of data, they are all typed as Int
s. And since all the compiler sees is a variable's type, it can help us with neither clarity nor precision.
Passing the contents of Robot.batteryPercentage
into a function expecting Robot.id
would be just as nonsensical as passing a Person
to a function expecting a Robot
. But while the latter would be caught by the compiler, the former would not.
This issue would disappear if our types were richer:
struct Person {
var name: String
var age: Years
}
struct Robot {
var id: ID
var batteryPercentage: BatteryPercentage
}
We now have age: Years
, id: ID
, and batteryPercentage: BatteryPercentage
(where Years
, ID
, and BatteryPercentage
are distinct types, rather than typealias
es for Int
).
Though ultimately each field is backed by an Int
-encoded integer, the fields have different types -- which means that the semantic distinction between the fields is now visible to the compiler.
The compiler can then utilize this visible distinction between the fields to perform many sanity-checks on our behalf, such as making sure we never populate a Person
's age with some ID
field, and that we never accidentally subtract Years
from BatteryPercentage
values.
Besides, it's nice to be able to quickly see whether a given variable captures Years
, an ID
, BatteryPercentage
-- or some other structure of significance.
A purpose-specific type makes it possible to define purpose-specific type extensions.
For instance, it may be convenient to have a isLow: Bool
computed variable available on types encoding battery percentages, returning true
whenever self
is below a given "low battery" threshhold (say, 20%).
We wouldn't want to contaminate the global Int
namespace with such an extension, for isLow: Bool
doesn't make sense for Int
s defined in any context besides battery percentages; the price for the convenience in the context of work with battery percentages would be increased congnitive in all other contexts.
But with a dedicated BatteryPercentage
type, we can encode such an extension with no down-sides whatsoever.
Once we create purpose-specific types used in particular contexts, we can also introduce purpose-specific, constraint-enforcing transformations and validations at the type level. Meaning, the raw values backing all instances of a given semantic-type would be guarenteed to maintain a given set of constraints defined by some "gateway" function.
A demonstrationo of this idea can be found in Foundation
's URL
type, where the validation step plays a primary role.
All URLs are String
s. But not all String
s are valid URLs.
When you have a URL
instance, you have proof that the String
value backing it indeed maintains the constraints defined by the URL standard.
URL
further utilizes the benefits of the initial validation step to expose safe, purpose-specific extensions.
Once you have a valid URL, you can create another valid URL simply by appending a path component to your original URL.
This is precisely what URL
's .appendingPathComponent(...)
function does.
We don't have to look far to find oppurtunities to enforce constraints at the type-level.
In our previous example, battery percentage values must capture a number in the range [0, 100]
to be sensible.
When we have a dedicated BatteryPercentage
type for encoding battery percentage values (rather than using Int
), we can encode this constraint at the type level, and hence be sure that any given BatteryPercentage
instance always carries a value in the range [0, 100]
.
What if you try to initialize a BatteryPercentage
instance with a value of 1324
?
We often (but not always) have choice in the matter. We can either clamp the input value to the nearest-possible valid value (in this case, to 100
), or we can simply throw
an error and refuse to initialize the BatteryPercentage
instance.
You could of course trivially define purpose-specific structures to wrap an underlying backing value used in particular circumstances. For instance, you could define:
struct Years {
var value: Int
}
And you could even define a dedicated init
enforcing any given validation/transformation.
So why do you need this library?
There's a reason semantic types are not widely used. They're usually a pain to work with.
For example, you couldn't use instances of the Years
struct defined above as keys in a dictionary, because Years
doesn't conform to Hashable
. You also couldn't simply add up or subtract two Years
instances.
This library lets you effortlessly define Semantic Types that are as easy to work with as their underlying RawValue
s, and that are guarenteed to enforce a given transformation/validation through all possible mutations -- without burdening you with the details.
It does so by defining the SemanticType
structure, which offers sensible conveniences as well as carefully-implemented type-level constraint validation:
SemanticType
s automatically conform to numerous standard-library protocols whenever possible (i.e. whenever their associatedRawValue
conforms to the protocol). The supported protocols includeHashable
,Comparable
,Equatable
,Sequence
,Collection
,AdditiveArithmetic
,ExpressibleByLiteral
protocols, and many, many, more. This makes it easy to useSemanticType
instances in the context of generic data-structures (e.g. as keys in aDictionary
), of protocol-oriented operations (e.g. in comparisons, additions, subtractions, etc.), as well as in the context of generic algorithms.SemanticType
s expose direct read/write access all instance-variables defined on theirRawValue
(via typed@dynamicMemberLookup
access).SemanticType
makes it easy to impose strict transformations and validation constraints on the allowable values of theRawValue
s, while guarenteeing that said constraints are maintained across all operations. All you have to provide is thegateway
function, and the library takes care of the rest. For instance, you can easily createOddNumber
andEvenNumber
types which guarentee that all of their instances are odd/even, respectively, and which are guarenteed to maintain this property across all possible transformations.
We will cover the different use-cases in detail below, but first, without further ado, here's some example code:
enum Seconds_Spec: ErrorlessSemanticTypeSpec { typealias RawValue = Double }
typealias Seconds = SemanticType<Seconds_Spec>
var step1Duration: Seconds = 5
let step2Duration: Seconds = 10
XCTAssertEqual(
step1Duration + step2Duration,
Seconds(15)
)
step1Duration += 7
XCTAssertEqual(
step1Duration,
Seconds(12)
)
XCTAssertEqual(
step1Duration - step2Duration,
Seconds(2)
)
// The following will fail to compile, because we try to add `Seconds` and `Double` together:
let notCompiling = step1Duration + Double(18)
enum CaselessString_Spec: ErrorlessSemanticTypeSpec {
typealias RawValue = String
static func gateway(preMap: String) -> String {
return preMap.lowercased()
}
}
typealias CaselessString = SemanticType<CaselessString_Spec>
var joe = CaselessString("Joe")
XCTAssertEqual(joe.rawValue, "joe")
joe.rawValue.removeLast()
joe.rawValue.append("SEPH")
XCTAssertEqual(joe.rawValue, "joseph")
struct ContactFormInput {
var email: String
var message: String
}
enum ProcessedContactFormInput_Spec: ErrorlessSemanticTypeSpec {
typealias RawValue = ContactFormInput
static func gateway(preMap: ContactFormInput) -> ContactFormInput {
return .init(
email: preMap.email.lowercased(),
message: preMap.message
)
}
}
typealias ProcessedContactFormInput = SemanticType<ProcessedContactFormInput_Spec>
let joesProcessedContactFormInput = ProcessedContactFormInput(ContactFormInput.init(
email: "[email protected]",
message: "What a great library!"
))
XCTAssertEqual(joesProcessedContactFormInput.email, "[email protected]")
XCTAssertEqual(joesProcessedContactFormInput.message, "What a great library!")
struct Person: Equatable {
var name: String
var associatedGreeting: String
init(name: String) {
self.name = name
self.associatedGreeting = "Hello, my name is \(name)." // initialize greeting to default
}
}
enum PersonWithShortName_Spec: ValidatedSemanticTypeSpec {
typealias RawValue = Person
enum Error: Swift.Error, Equatable {
case nameIsTooLong(name: String)
}
static func gateway(preMap: Person) -> Result<Person, PersonWithShortName_Spec.Error> {
guard preMap.name.count < 5
else { return .failure(.nameIsTooLong(name: preMap.name)) }
return .success(preMap)
}
}
typealias PersonWithShortName = SemanticType<PersonWithShortName_Spec>
let tim = try! PersonWithShortName(Person(name: "Tim"))
XCTAssertEqual(tim.rawValue, Person(name: "Tim")
XCTAssertEqual(tim.name, "Tim")
XCTAssertEqual(tim.associatedGreeting, "Hello, my name is Tim.")
let joe = try! PersonWithShortName(Person(name: "Joe"))
let lowercaseJoe = joe.tryMap { person in
var person = person
person.associatedGreeting = person.associatedGreeting.lowercased()
return person
}.get()
XCTAssertEqual(lowercaseJoe.associatedGreeting, "hello, my name is joe.")
A SemanticType
is defined by a SemanticTypeSpec
type, of the SemanticTypeSpec
protocol family.
A SemanticTypeSpec
type has 3 roles:
- It serves as a marker type, bringing about a type-level distinction between
SemanticType
instantiations. - It defines the
RawValue
type wrapped by its associatedSemanticType
instantiation. - It defines how
RawValue
values are transformed and validated before making their way into aSemanticType
value.
The SemanticTypeSpec
protocol family consists of a base protocol, and 2 (successive) protocol refinements:
ErrorlessSemanticTypeSpec
~~refines~~> ValidatedSemanticTypeSpec
~~refines~~> MetaValidatedSemanticTypeSpec
At the core of the SemanticTypeSpec
lies the gateway
function, which (aptly) serves as a gateway between RawValue
s and SemanticType
instances. The gateway
dictates how values of the underlying RawValue
type behave as they are transformed into values of a SemanticType
instantiation.
Protocol | gateway function type |
---|---|
ErrorlessSemanticTypeSpec |
(RawValue) -> RawValue |
ValidatedSemanticTypeSpec |
(RawValue) -> Result<RawValue, Error> |
MetaValidatedSemanticTypeSpec |
(RawValue) -> Result<(RawValue, Metadata), Error> |
ErrorlessSemanticTypeSpec
is the simplest (and most refined) SemanticTypeSpec
protocol.
It is used when we don't need to validate our RawValue
payload in any way (i.e. when every RawValue
instance can be made to correspond to a SemanticType
instance).
By default, ErrorlessSemanticTypeSpec
's gateway
function is simply the identify function (i.e. it doesn't transform the RawValue
in any way).
Instantiations of SemanticType
often support the same operations as their RawTypes
-- but only within the same type, not across types:
enum Years_Spec: ErrorlessSemanticTypeSpec { typealias RawValue = Double }
typealias Years = SemanticType<Years_Spec>
let fiveYears: Years = 5
let threeYears: Years = 3
let eigthYears: Years = fiveYears + threeYears
let truth = ( Years(151) > Years(34) )
enum Inches_Spec: ErrorlessSemanticTypeSpec { typealias RawValue = Double }
typealias Inches = SemanticType<Inches_Spec>
let tenInches: Inches = 10
let fourInches: Inches = 4
let anotherTruth = ( fourInches <= tenInches )
---
// The following examples would not compile, as they mix-up `Years` and `Inches`:
let willNotCompile1 = Years(5) + Inches(1)
let willNotCompile2 = Yeras(19) > Inches(32)
We can also consider cases where the gateway
function is explicitly specified -- allowing us to specify an arbitrary transformation to be applied to a RawValue
before it is transformed into a SemanticType
.
The gateway
function can be leveraged to construct types which have an inherent restriction on the range of the allowed values.
For instance, consider the following CaseInsensitiveString
type which is inherently case-insensitive:
enum CaseInsensitiveString_Spec: ErrorlessSemanticTypeSpec {
typealias RawValue = String
static func gateway(preMap: String) -> String {
return preMap.lowercased()
}
}
typealias CaseInsensitiveString = SemanticType<CaseInsensitiveString_Spec>
let hello1: CaseInsensitiveString = "hELlo"
let hello2: CaseInsensitiveString = "HelLO"
let _ = (hello1 == hello2) // true
From this point onwards, we can be sure that if we have a CaseInsensitiveString
instance, no operation on it would depend on any casing information.
Other examples include a SQLCommand
type which is always sql-escaped (and not prone to SQL-injection attacks), a LevelInSomeBuilding
type which clamps values to values above -1 and below 72 (the lowest and highest levels in SomeBuilding
), etc.
While we still have "type information" which is associated with runtime behavior rather than with compile-time behavior, this information (and all associated testing!) is now restricted to the gateway
function.
This can come in handy whenever we have a restriction on our values which is inherent in our "mental" model of the type, but not in the underlying data type.
The aforementioned ErrorlessSemanticTypeSpec
is a protocol refinement (with default behavior provided via an extension) of the more general ValidatedSemanticTypeSpec
protocol.
The ValidatedSemanticTypeSpec
protocol's more general gateway
function may not only transform the incoming value, but also return some error if the value failed to pass some validation.
For example:
enum EnglishLettersOnlyString_Spec: SemanticTypeSpec {
typealias RawValue = String
enum Error: Swift.Error {
case containsNonEnglishCharacters(nonEnglishCharacters: String)
}
static func gatewayMap(preMap: RawValue) -> Result<String, Error> {
let nonEnglishCharacters = preMap.stringByTrimmingCharactersInSet(NSCharacterSet.letterCharacterSet())
if(nonEnglishCharacters == "") {
return .success(preMap)
} else{
return .failure(.containsNonEnglishCharacters(nonEnglishCharacters: nonEnglishCharacters))
}
}
}
typealias EnglishLettersOnlyString = SemanticType<EnglishLettersOnlyString_Spec>
The following is made of only English letters, and therefore the initialization will not throw
:
let actuallyOnlyLetters = try EnglishLettersOnlyString("abclaskjdf") // will succeed
The following is not made of only English letters, and therefore the initializatino will throw
:
let notOnlyLettes = try? EnglishLettersOnlyString("asdflkj12345") // will throw
A typed version of the error is available through the Result
-returning .create()
factory function:
let englishLettersCreationResult: Result<EnglishLettersOnlyString, EnglishLettersOnlyString_Spec.Error> = EnglishLettersOnlyString.create("asdflkj12345")
ValidatedSemanticTypeSpec
is itself also a protocol refinement (with default behavior provided via an extension) of the most generic MetaValidatedSemanticTypeSpec
protocol.
The MetaValidatedSemanticTypeSpec
protocol's most generic gateway
function, in addition to the transformed RawValue
to back the SemanticType
instance, returns an arbitrarily-typed metadata value which is then stored on the SemanticType
instance and made publically available for access.
This metadata value may be used to encode a compiler-accessible fascet of the sub-structure of the wrapped value
which was veriried by the gatewayMap
function.
As an example: suppose we create a NonEmptyArray
SemanticType
, i.e. a type whose instances wrap an Array
-- but which could only be created when said Array
is non-empty.
Unlike instances of Array
, instances of NonEmptyArray
are guarenteed to have first
and last
elements.
Thus we may expose first: Element
and last: Element
in place of Array
's corresponding optional properties.
Since we know that instances of NonEmptyArray
are not empty, we could implement said non-optionoal first
and last
overrides by forwarding the call to Array
's optional properties and force-unwrapping the result. While we know this process ought to work, the compiler does not -- hence the need for the force-unwrapping. And so we lose the celebrated compiler verification normally characterizing idiomatic swift code.
Instead, we could implement first
and last
without circumventing compiler verifications by storing the Array
's first
and last
values as metadata during the gatewayMap
ing (where we could return an error if first
and last
are not available).
The non-optional first
and last
properties could then be implemented by querying said metadata values.
For example:
enum NonEmptyIntArray_Spec: GeneralizedSemanticTypeSpec {
typealias RawValue = [Int]
struct Metadata {
var first: Int
var last: Int
}
enum Error: Swift.Error {
case arrayIsEmpty
}
static func gateway(preMap: [Int]) -> Result<GatewayOutput, Error> {
// a non-empty array will always have first/last elements:
guard
let first = preMap.first,
let last = preMap.last
else {
return .failure(.arrayIsEmpty)
}
return .success(.init(
rawValue: preMap,
metadata: .init(first: first,
last: last)
))
}
}
typealias NonEmptyIntArray = SemanticType<NonEmptyIntArray_Spec>
extension NonEmptyIntArray {
var first: Int {
return gatewayMetadata.first
}
var last: Int {
return gatewayMetadata.last
}
}
// ...
let oneTwoThree = try! NonEmptyIntArray.create([1, 2, 3]).get()
XCTAssertEqual(oneTwoThree.first, 1)
XCTAssertEqual(oneTwoThree.last, 3)
Numeric
is the protocol swift uses to support multiplication within a given type.
Numeric
support may not make sense for all SemanticType
s, even when their RawValue
types are themselves Numeric
. For instance, [Second
* Second
= Second
] does not make semantic sense.
In other situations, Numeric
support does make sense. For instance [EvenInteger
* EvenInteger
= EvenInteger
].
We allow the SemanticTypeSpec
backing the SemanticType
to signal whether Numeric
support should be provided by conforming to the ShouldBeNumeric
marker protocol.