Indeed, representing text as a sequence of Unicode code points is fundamentally wrong.

There are no operations on sequences of Unicode code points that are more correct than an analogous operation on bytes.

(Everyone's favourite example, length, actually becomes less correct—a byte array's length at least corresponds to the amount of space one might have to allocate for it in a particular encoding. A length in codepoints is absolutely meaningless both technically and linguistically. And this is, for what little it's worth, close to the only operation you can do on a string without imposing additional restrictions about its context.)

> There are no operations on sequences of Unicode code points that are more correct than an analogous operation on bytes.

That’s ridiculous. Uppercasing/lowercasing, slicing, “startswith”, splitting, etc etc.

Your statement is correct if you only care about ascii.

Uppercasing/lowercasing cannot be done on Unicode code points, because that fails to handle things like fi -> FI where the uppercased version does not consist of the same number of Unicode code points. Slicing and splitting cannot be done on Unicode code points because it may separate a character from a subsequent combining character. "startswith" cannot be done on Unicode code points because some distinct code points need to be treated as equivalent. These are pretty much the same problems you also have when you perform those same operations on bytes. You might encounter those problems in fewer cases when you perform operations on code points rather than on bytes, but you won't have solved the problems entirely.

Worse, you'll have pushed the problematic cases out of the realm of obviously wrong and not sensible to do, into subtly wrong and will break down the line in ways that will be hard to recognize and debug.

None of those operations are correct on Unicode codepoints. Your statement is only just barely tenable if you only care about well-edited and normalized formal prose in common Western languages.

Even then, upper/lower-casing is iffy.

> There are no operations on sequences of Unicode code points that are more correct than an analogous operation on bytes.

Wow. I wonder how you arrived at this point. You can't, for example, truncate a UTF-8 byte array without the risk of producing a broken string. But this is only the start. Here are two strings, six letters each, one in NFC, the other in NFD, and their byte-length in UTF-8:

    "Åström" is 8 bytes in UTF-8

    "Åström" is 10 bytes in UTF-8

If your software tells the user that one is eight and the other is 10 letters long, it is not "less correct". It is incorrect. Further, if searching for "Åström" won't find "Åström", your software is less useful than it could be if it knew Unicode. (And it's sad how often software gets this wrong.)

> If your software tells the user that one is eight and the other is 10 letters long, it is not "less correct". It is incorrect.

In fact, if the software tells you that either of the strings is either 8 or 10 letters wrong, then either way the software is incorrect - those are both obviously 6 letter strings.

Now, does UTF8 help you discover they are 6 letter strings better than other representations? There are certainly text-oriented libraries that can do that, but not those that simply count the UTF8 code points - they must have an understanding of all of Unicode. Even worse, the question "how many letters does this string have" is not generally meaningful - there are plenty of perfectly valid unicode strings for which this question doesn't have a meaningful answer.

However, the question "how many unicode code points does this string have" is almost never of interest. You either care about some notion of unique glyphs, or you care about byte lengths.

> then either way the software is incorrect - those are both obviously 6 letter strings.

What I wanted to get at is that in Unicode, I have a chance to count letters to some useful degree. Why should I consider starting at byte-arrays?

> there are plenty of perfectly valid Unicode strings for which this question doesn't have a meaningful answer.

I don't get it. Why does the existence of degenerate cases invalidate the usefulness of a Unicode lib? If I want to know how many letters are in a string, I can probably get a useful answer from a Unicode lib. Not for all edge-cases, but I can decide on the trade-offs. If I have a byte-array, I start at a lower level.

> What I wanted to get at is that in Unicode, I have a chance to count letters to some useful degree.

You do not. You merely happen to get the right answer by coincidence in some cases, same as bytes-that-probably-are(n't)-ASCII. To throw your own words back at you:

  "Åström" is 6 code points in Unicode

  "Åström" is 8 code points in Unicode

If your software tells the user that one is six and the other is 8 letters long, it is not "less correct". It is incorrect. Further, if searching for "Åström" won't find "Åström", your software is less useful than it could be if it knew text. (And it's sad how often software gets this wrong.)

You can't truncate a sequence of Unicode codepoints without the risk of producing a broken string, either. What do you get if you truncate "Åström" after the first "o"? What do you get if you truncate 🇨🇦 after the first codepoint?

Normalization is not a real solution unless you restrict yourself to working with well-edited formal prose in common Western languages.

This is not a claim made from ignorance.

Sorry, we're mixing two layers. Of course, if I truncate a string, it may lose its meaning. And having accents fall off is problematic. But it's not the same as truncating a byte-array, because then an invalid sequence of bytes may result.

Stop treating these cases as equivalent. They're not.

They are equivalent. The only reason you find it problematic that a sequence of bytes is "invalid" (read: can't be decoded in your preferred encoding) is because you've manufactured the problem.

In the end, the only layer at which it really matters whether your byte sequence can be decoded is the font renderer, and just being valid utf-8 isn't good enough for it either.

> In the end, the only layer at which it really matters whether your byte sequence can be decoded is the font renderer

Ok that explains how we ended up here. I'm considering some other common uses! A search-index for example greatly profits from being able to normalize representations and split words.

Search index use cases probably also benefit from normalizing inputs across encodings, that's no shining example of utf-8-onlyism.

You can still best-effort split words! You can do a pretty good job splitting words without ensuring that the words decode in your preferred encoding.

Here's the thing: I don't want to work in UTF8. I want to work in Unicode. Big difference. Because tracking the encoding of my strings would increase complexity. So at the earliest convenience, I validate my assumptions about encoding and let a lower layer handle it from then on.

I understand you're arguing about some sort of equivalency between byte-arrays and Unicode strings. Sure there are half-baked ways to do word-splitting on a byte-array. But why do you consider that a viable option? Under what circumstances would you do that?

Every circumstance. Why do you consider it unviable? What problems do you think having a Unicode sequence solves?

Convince me. Here's a little library function that turns text into a set of words:

    def keywords(text):
        return set(filter(None, re.split("\W+", unicodedata.normalize("NFKC", input_str).lower())))

How would this look if strings were byte-arrays? How would `normalize()`, `lower()`, and `split()` know what encoding to use?

The way I see it: If the encoding is implicit, you have global state. If it's explicit, you have to pass the encoding. Both is extra state to worry about. When the passed value is a Unicode string, this question doesn't come up.

It looks pretty much the some, except that you assume the input is already in your library's canonical encoding (probably utf-8 nowadays).

I realize this sounds like a total cop-out, but when the use-case is destructively best-effort tokenizing an input string using library functions, it doesn't really matter whether your internal encoding is utf-32 or utf-8. I mean, under the hood, normalize still has to map arbitrary-length sequences to arbitrary-length sequences even when working with utf-32 (see: unicodedata.normalize("NFKC", "a\u0301 ffi") == "\xe1 ffi").

So on the happy path, you don't see much of a difference.

The main observable difference is that if you take input without decoding it explicitly, then the always-decode approach has already crashed long before reaching this function, while the assume-the-encoding approach probably spouts gibberish at this point. And sure, there are plenty of plausible scenarios where you'd rather get the crash than subtly broken behaviour. But ... I don't see this reasonably being one of them, considering that you're apparently okay with discarding all \W+.

I agree with you. I wish Python 3 had strings as byte sequences mainly in UTF-8 as Python 2 had once and Go has now. Then things would be kept simple in Japan. Python 3 feels cumbersome. To handle a raw input as a string, you must decode it in some encoding first. It is a fragile process. It would be adequate to treat the input bytes transparently and put an optional stage to convert other encodings to UTF-8 if necessary.

I know this from PHP, where I have to be aware of the encoding the strings are in. I still don't see what the advantage should be of that.

So in one case, the text becomes corrupted and unreadable (i.e. loses its meaning), and in the other, it becomes corrupted and unreadable. What's the difference?

Having "accents fall off" has gotten people murdered [0]. Accents aren't things peppered in for effect, they turn letters into different letters, spelling different words. Analogously, imagine that a bunch of software accidentally turned every "d" into a "c" because some committee halfway around the world decided "d" should be composed of the "c" and "|" glyphs. That's the kind of text corruption that regularly happens in other languages when dealing with text at the code point layer.

[0] https://languagelog.ldc.upenn.edu/nll/?p=73 . Note that this is Turkish, which has the "dotted i" problem, meaning that this was more than likely a .toupper() gone wrong rather than a truncation issue.

The difference is that for truncating, I can work within Unicode to deal with the situation. I can accept the possibility of mutilated letters, I can convert to NFC, I can truncate on word-boundaries, I have choice.

If I have an byte-array, I can do none of these things short of implementing a good chunk of Unicode. If I truncate, I risk ending up with an invalid UTF-8 string. End of story.

And what is wrong with an invalid UTF-8 string? Why were you truncating the string in the first place?

Basically, I believe the point here is that a Unicode aware truncation should be done in a Unicode aware truncate method. There is no good reason to parse a string as UTF-8 ahead of time - just keep it as a blob of bytes until you need to do some something "texty" with it. It is the truncate-at-word-boundaries() method that should interpret the bytes as UTF-8 and fail if they are not valid. Why parse it sooner?

> If I have an byte-array, I can do none of these things short of implementing a good chunk of Unicode. If I truncate, I risk ending up with an invalid UTF-8 string.

Yes, and? You can have an invalid sequence of Unicode code points too, such as an unpaired surrogate (something Python's text model actually abuses to store "invalid Unicode" in a special, non-standard way).

If you truncate at the byte level, you are just truncating "between code points"; it's a closer granularity than at the code point layer, so you can also convert to NFC, truncate on word boundaries, etc. You just need to ignore the parts of the UTF-8 string that are invalid; which isn't difficult, because UTF-8 is self-synchronizing.