Monday Dec 21, 2020
We are excited to announce the 0.1.0 release of Parsing, a library for turning nebulous data into well-structured data. It was built from the content of 21 episodes (10 hours) where we show how to build a parsing library from scratch, with a focus on composition, performance, and generality:
Composition: The ability to break large, complex parsing problems down into smaller, simpler ones. And the ability to take small, simple parsers and easily combine them into larger, more complex ones.
Performance: Parsers that have been composed of many smaller parts should perform as well as highly-tuned, hand-written parsers.
Generality: The ability to parse any kind of input into any kind of output. This allows you to choose which abstraction levels you want to work with based on how much performance you need or how much correctness you want guaranteed. For example, you can write a highly tuned parser on collections of UTF-8 code units, and it will automatically plug into parsers of strings, arrays, unsafe buffer pointers and more.
Parsing is a surprisingly ubiquitous problem in programming. We can define parsing as trying to take a more nebulous blob of data and transform it into something more well-structured. The Swift standard library comes with a number of parsers that we reach for every day. For example, there are initializers on Int, Double, and even Bool, that attempt to parse numbers and booleans from strings:
Int("42") // 42
Int("Hello") // nil
Double("123.45") // 123.45
Double("Goodbye") // nil
Bool("true") // true
Bool("0") // nil
And there are types like JSONDecoder and PropertyListDecoder that attempt to parse Decodable-conforming types from data:
try JSONDecoder().decode(User.self, from: data)
try PropertyListDecoder().decode(Settings.self, from: data)
While parsers are everywhere in Swift, Swift has no holistic story for parsing. Instead, we typically parse data in an ad hoc fashion using a number of unrelated initializers, methods, and other means. And this typically leads to less maintainable, less reusable code.
This library aims to write such a story for parsing in Swift. It introduces a single unit of parsing that can be combined in interesting ways to form large, complex parsers that can tackle the programming problems you need to solve in a maintainable way.
Suppose you have a string that holds some user data that you want to parse into an array of Users:
var input = """
1,Blob,true
2,Blob Jr.,false
3,Blob Sr.,true
"""
struct User {
var id: Int
var name: String
var isAdmin: Bool
}
A naive approach to this would be a nested use of .split(separator:), and then a little bit of extra work to convert strings into integers and booleans:
let users = input
.split(separator: "\n")
.compactMap { row -> User? in
let fields = row.split(separator: ",")
guard
fields.count == 3,
let id = Int(fields[0]),
let isAdmin = Bool(String(fields[2]))
else { return nil }
return User(id: id, name: String(fields[1]), isAdmin: isAdmin)
}
Not only is this code a little messy, but it is also inefficient since we are allocating arrays for the .split and then just immediately throwing away those values.
It would be more straightforward and efficient to instead describe how to consume bits from the beginning of the input and convert that into users. This is what this parser library excels at 😄.
We can start by describing what it means to parse a single row, first by parsing an integer off the front of the string, and then parsing a comma that we discard using the .skip operator:
let user = Int.parser()
.skip(StartsWith(","))
Already this can consume the beginning of the input:
user.parse(&input) // => 1
input // => "Blob,true\n2,Blob Jr.,false\n3,Blob Sr.,true"
Next we want to take everything up until the next comma for the user’s name, and then skip the comma:
let user = Int.parser()
.skip(StartsWith(","))
.take(PrefixWhile { $0 != "," })
.skip(StartsWith(","))
Here the .take operator has combined parsed values together into a tuple, (Int, Substring).
And then we want to take the boolean at the end of the row for the user’s admin status:
let user = Int.parser()
.skip(StartsWith(","))
.take(PrefixWhile { $0 != "," })
.skip(StartsWith(","))
.take(Bool.parser())
Currently this will parse a tuple (Int, Substring, Bool) from the input, and we can .map on that to turn it into a User:
let user = Int.parser()
.skip(StartsWith(","))
.take(PrefixWhile { $0 != "," })
.skip(StartsWith(","))
.take(Bool.parser())
.map { User(id: $0, name: String($1), isAdmin: $2) }
That is enough to parse a single user from the input string:
user.parse(&input) // => User(id: 1, name: "Blob", isAdmin: true)
input // => "\n2,Blob Jr.,false\n3,Blob Sr.,true"
To parse multiple users from the input we can use the Many parser:
let users = Many(user, separator: StartsWith("\n"))
user.parse(&input) // => [User(id: 1, name: "Blob", isAdmin: true), …]
input // => ""
Now this parser can process an entire document of users, and the code is simpler and more straightforward than the version that uses .split and .compactMap.
Even better, it’s more performant. We’ve written benchmarks for these two styles of parsing, and the .split-style of parsing is more than twice as slow:
name time std iterations
------------------------------------------------------------------
README Example.Parser: Substring 3426.000 ns ± 63.40 % 385395
README Example.Adhoc 7631.000 ns ± 47.01 % 169332
Program ended with exit code: 0
Further, if you are willing write your parsers against UTF8View instead of Substring, you can eke out even more performance, more than doubling the speed:
name time std iterations
------------------------------------------------------------------
README Example.Parser: Substring 3693.000 ns ± 81.76 % 349763
README Example.Parser: UTF8 1272.000 ns ± 128.16 % 999150
README Example.Adhoc 8504.000 ns ± 59.59 % 151417
We can also compare these times to a tool that Apple’s Foundation gives us: Scanner. It’s a type that allows you to consume from the beginning of strings in order to produce values, and provides a nicer API than using .split:
var users: [User] = []
while scanner.currentIndex != input.endIndex {
guard
let id = scanner.scanInt(),
let _ = scanner.scanString(","),
let name = scanner.scanUpToString(","),
let _ = scanner.scanString(","),
let isAdmin = scanner.scanBool()
else { break }
users.append(User(id: id, name: name, isAdmin: isAdmin))
_ = scanner.scanString("\n")
}
However, the Scanner style of parsing is more than 5 times as slow as the substring parser written above, and more than 15 times slower than the UTF-8 parser:
name time std iterations
-------------------------------------------------------------------
README Example.Parser: Substring 3481.000 ns ± 65.04 % 376525
README Example.Parser: UTF8 1207.000 ns ± 110.96 % 1000000
README Example.Adhoc 8029.000 ns ± 44.44 % 163719
README Example.Scanner 19786.000 ns ± 35.26 % 62125
That’s the basics of parsing a simple string format, but there’s a lot more operators and tricks to learn in order to performantly parse larger inputs. View the benchmarks for examples of real life parsing scenarios.
The design of the library is largely inspired by the Swift standard library and Apple’s Combine framework. A parser is represented as a protocol that many types conform to, and then parser transformations (also known as “combinators”) are methods that return concrete types conforming to the parser protocol.
For example, to parse all the characters from the beginning of a substring until you encounter a comma you can use the Prefix parser:
let parser = Prefix<Substring> { $0 != "," }
var input = "Hello,World"[...]
parser.parse(&input) // => "Hello"
input // => ",Hello"
The type of this parser is:
Prefix<Substring>
We can .map on this parser in order to transform its output, which in this case is the string “Hello”:
let parser = Prefix<Substring> { $0 != "," }
.map { $0 + "!!!" }
var input = "Hello,World"[...]
parser.parse(&input) // => "Hello!!!"
input // => ",Hello"
The type of this parser is now:
Parsers.Map<Prefix<Substring>, Substring>
Notice how the type of the parser encodes the operations that we performed. This adds a bit of complexity when using these types, but comes with some performance benefits because Swift can usually optimize the creation of those nested types.
The library makes it easy to choose which abstraction level you want to work on. Both low-level and high-level have their pros and cons.
Parsing low-level inputs, such as UTF-8 code units, has better performance, but at the cost of potentially losing correctness. A canonical example of this is trying to parse the character “é”, which can be represented in code units as [233] or [101, 769]. If you don’t remember to always parse both representations you may have a bug where you accidentally fail your parser when it encounters a code unit sequence you don’t support.
On the other hand, parsing high-level inputs, such as String, can guarantee correctness, but at the cost of performance. For example, String handles the complexities of extended grapheme clusters and UTF-8 normalization for you, but traversing strings is slower since its elements are variable width.
The library gives you the tools that allow you to choose which abstraction level you want to work on, as well as the ability to fluidly move between abstraction levels where it makes sense.
For example, say we want to parse particular city names from the beginning of a string:
enum City {
case london
case newYork
case sanJose
}
Because “San José” has an accented character, the safest way to parse it is to parse on the Substring abstraction level:
let city = StartsWith<Substring>("London").map { City.london }
.orElse(StartsWith("New York").map { .newYork })
.orElse(StartsWith("San José").map { .sanJose })
var input = "San José,123"
city.parse(&input) // => City.sanJose
input // => ",123"
However, we are incurring the cost of parsing Substring for this entire parser, even though only the “San José” case needs that power. We can refactor this parser so that “London” and “New York” are parsed on the UTF8View level, since they consist of only ASCII characters, and then parse “San José” as Substring:
let city = StartsWith("London".utf8).map { City.london }
.orElse(StartsWith("New York".utf8).map { .newYork })
.orElse(StartsWith("San José").utf8.map { .sanJose })
It’s subtle, but StartsWith("London".utf8) is a parser that parses the code units for “London” from the beginning of a UTF8View, whereas StartsWith("San José").utf8 parses “San José” as a Substring, and then converts that into a UTF8View parser.
This allows you to parse as much as possible on the more performant, low-level UTF8View, while still allowing you to parse on the more correct, high-level Substring when necessary.
This library comes with a benchmark executable that not only demonstrates the performance of the library, but also provides a wide variety of parsing examples:
These are the times we currently get when running the benchmarks:
MacBook Pro (16-inch, 2019)
2.4 GHz 8-Core Intel Core i9
64 GB 2667 MHz DDR4
name time std iterations
-------------------------------------------------------------------------------------
Arithmetic.Parser 12622.000 ns ± 40.63 % 102408
BinaryData.Parser 512.000 ns ± 172.80 % 1000000
Bool.Bool.init 28.000 ns ± 880.63 % 1000000
Bool.BoolParser 43.000 ns ± 423.22 % 1000000
Bool.Scanner.scanBool 920.000 ns ± 119.49 % 1000000
Color.Parser 127.000 ns ± 341.57 % 1000000
CSV.Parser 1370906.000 ns ± 12.24 % 1027
CSV.Ad hoc mutating methods 1338824.500 ns ± 13.91 % 1014
Date.Parser 12429.000 ns ± 38.26 % 107342
Date.DateFormatter 41168.000 ns ± 29.40 % 31353
Date.ISO8601DateFormatter 56236.000 ns ± 27.39 % 23383
HTTP.HTTP 3850.000 ns ± 1898.35 % 341642
JSON.Parser 6115.000 ns ± 45.95 % 217152
JSON.JSONSerialization 3050.000 ns ± 71.43 % 431524
Numerics.Int.init 38.000 ns ± 655.10 % 1000000
Numerics.Int.parser 41.000 ns ± 464.80 % 1000000
Numerics.Scanner.scanInt 145.000 ns ± 22359.78 % 1000000
Numerics.Comma separated: Int.parser 5511505.000 ns ± 8.87 % 245
Numerics.Comma separated: Scanner.scanInt 82824843.000 ns ± 2.37 % 17
Numerics.Comma separated: String.split 117376272.000 ns ± 2.68 % 11
Numerics.Double.init 58.000 ns ± 518.12 % 1000000
Numerics.Double.parser 59.000 ns ± 445.11 % 1000000
Numerics.Scanner.scanDouble 195.000 ns ± 234.94 % 1000000
Numerics.Comma separated: Double.parser 6222693.000 ns ± 9.33 % 220
Numerics.Comma separated: Scanner.scanDouble 89431780.500 ns ± 3.75 % 16
Numerics.Comma separated: String.split 33387660.000 ns ± 4.02 % 41
PrefixUpTo.Parser 22898.000 ns ± 34.40 % 58197
PrefixUpTo.Scanner.scanUpToString 162816.000 ns ± 18.55 % 8000
Race.Parser 29962.000 ns ± 32.24 % 43186
README Example.Parser: Substring 3451.000 ns ± 59.72 % 378685
README Example.Parser: UTF8 1247.000 ns ± 110.74 % 1000000
README Example.Adhoc 8134.000 ns ± 34.87 % 161121
Routing.Parser 5242.000 ns ± 52.70 % 249596
String Abstractions.Substring 1044908.500 ns ± 12.95 % 1296
String Abstractions.UTF8 138412.000 ns ± 22.64 % 8938
Xcode Logs.Parser 6980962.000 ns ± 7.61 % 197
Head over to the Parsing repository to try the library out today. For some inspiration of things you might like to write parsers for check out the benchmarks in the project.
👋 Hey there! If you got this far, then you must have enjoyed this post. You may want to also check out Point-Free, a video series covering advanced programming topics in Swift. Consider subscribing today!