Performance of `PurchasesReceiptParser.default.parse(from:)` could be drastically improved.
Opened this issue · 3 comments
Hey guys! I played a bit with profiler and was able to make PurchasesReceiptParser.default.parse(from:) way faster.
Here are results that I've observed on iPhone 8.
| Fastest before changes | Slowest after changes | Improvement | |
|---|---|---|---|
Time measured via Date |
0.0146 | 0.0035 | 4 times faster |
Cycle measured via mach_absolute_time |
352095 | 85862 | 4 times faster |
Profiler counters Cycles |
36 953 516 | 6 863 165 | 5 times faster |
Profiler counters INST_ALL |
76 994 013 | 7 666 052 | 10 times less |
Profiler counters L1D_TLB_MISS |
89 366 | 17 531 | 5 times less |
Unfortunately I don't have active developer account and I only tested on small receipts that I've found over internet. If you have any big receipt please send me a base64String.
Surprisingly reason of slowness is Date parsing. In particular ISO8601DateFormatter.default called from ArraySlice<UInt8>.toDate. Since all receipts I see have date format yyyy-MM-dd'T'HH:mm:ssZ it's easy to parse components manually. Here is the whole change:
func toDate() -> Date? {
if let fastParsed = toDateTryFastParsing() {
// This approach is around ~60% faster than `ISO8601DateFormatter.default`
return fastParsed
}
guard let dateString = String(bytes: Array(self), encoding: .ascii) else { return nil }
return ISO8601DateFormatter.default.date(from: dateString)
}
func toData() -> Data {
return Data(self)
}
}
private extension ArraySlice where Element == UInt8 {
static let toDateCalendar: Calendar = {
var calendar = Calendar(identifier: .gregorian)
calendar.timeZone = TimeZone(secondsFromGMT: 0)!
return calendar
}()
private func toDateTryFastParsing() -> Date? {
// expected format 2015-08-10T07:19:32Z
guard count == 20 else { return nil }
let asciiZero: UInt8 = 48
let asciiNine: UInt8 = 57
let asciiDash: UInt8 = 45
let asciiColon: UInt8 = 58
let asciiT: UInt8 = 84
let asciiZ: UInt8 = 90
let limits: [(min: UInt8, max: UInt8)] = [
(asciiZero, asciiNine), (asciiZero, asciiNine), (asciiZero, asciiNine), (asciiZero, asciiNine), // year
(asciiDash, asciiDash),
(asciiZero, asciiNine), (asciiZero, asciiNine), // month
(asciiDash, asciiDash),
(asciiZero, asciiNine), (asciiZero, asciiNine), // day
(asciiT, asciiT),
(asciiZero, asciiNine), (asciiZero, asciiNine), // hour
(asciiColon, asciiColon),
(asciiZero, asciiNine), (asciiZero, asciiNine), // minute
(asciiColon, asciiColon),
(asciiZero, asciiNine), (asciiZero, asciiNine), // second
(asciiZ, asciiZ)
]
for (character, limit) in zip(self, limits) {
guard limit.min <= character,
character <= limit.max else { return nil }
}
let year = toDateParseAsciiNumber(from: 0, to: 4)
let month = toDateParseAsciiNumber(from: 5, to: 7)
guard 1 <= month,
month <= 12 else { return nil }
let day = toDateParseAsciiNumber(from: 8, to: 10)
guard 1 <= day,
day <= 31 else { return nil }
let hour = toDateParseAsciiNumber(from: 11, to: 13)
guard 0 <= hour,
hour <= 23 else { return nil }
let minute = toDateParseAsciiNumber(from: 14, to: 16)
guard 0 <= minute,
minute <= 59 else { return nil }
let second = toDateParseAsciiNumber(from: 17, to: 19)
guard 0 <= second,
second <= 59 else { return nil }
let components = DateComponents(
year: year, month: month, day: day, hour: hour, minute: minute, second: second
)
return Self.toDateCalendar.date(from: components)
}
private func toDateParseAsciiNumber(from: Int, to: Int) -> Int { // swiftlint:disable:this identifier_name
let asciiZero: UInt8 = 48
var index = from + startIndex
let end = to + startIndex
var result = 0
while index < end {
let digit = self[index] - asciiZero
result = result * 10 + Int(digit)
index += 1
}
return result
}
}I have prepared branch with tests ensuring that this change has no impact on SDK behavior. Unfortunately I don't have permission to push. Could you grant me permissions so I could prepare pull-request and you could decide if you want this improvement or not?
Btw it's crazy that this change has so huge impact. I would imagine that Self.toDateCalendar.date(from: components) does most of the work. It also shows that industry standards are really bad. Apple's server had some kind integer counting seconds, they wasted compute to create string, on device we receive string and waste compute to calculate some kind integer counting seconds 🙃
👀 We've just linked this issue to our internal tracker and notified the team. Thank you for reporting, we're checking this out!
@Pikacz I'm very sorry that we missed this issue!! 🤦♂️
Thanks for posting!
I think this makes sense. The best process to open up a PR is to fork the repo, then create a PR from your fork to our repo. Let me know if you need a hand with it!
And apologies again for the very late response!