> It emits groups of five characters. Five is the sweet spot: 84⁵ = 4,182,119,424, only 2.6% short of 2³².
> That leaves enough room for a group to hold 32 bits about 95% of the time on uniformly random input, and 31 bits otherwise.
> The encoder looks at the next 31 bits. If their value is below 84⁵ - 2³¹, there’s room for a 32nd bit. Otherwise, it consumes just those 31 bits. Either way, the value fits in five base-84 digits.
Trying to jam in the extra characters for "efficiency" also doesn't really do much. Base84 is 6.4 bits per character, base64 is a flat 6 bits per character, base58 is 5.86 bits per character. It may intuitively feel like having 26 more characters to encode with would be hugely more efficient but for the extra risk of what all those characters get us we're only getting just over .5 more bits per character. I need to be either pretty desperate for efficiency, rare in 2026, or very, very sure that those things are only going places where those characters aren't risky or inconvenient, which is pretty difficult to guarantee since even if it is in fact true today who's to say it will be tomorrow, for this to be worth the risk of the extra characters.
Even just the extra characters in base64 are often annoying and we already have multiple common base64 standards because even trying to find just two more characters to fill out a-zA-Z0-9 is hard.
> Trying to jam in the extra characters for "efficiency" also doesn't really do much. Base84 is 6.4 bits per character, base64 is a flat 6 bits per character, base58 is 5.86 bits per character.
Also on efficiency, you are trading off code complexity for a little extra storage: base64 is a nice round 6 bits per character meaning every three bytes encoded is four output. Neither base84 nor base58 align on convenient bit boundaries like that so choosing the output character is more faf. Padding could be more complicated too.
Now if you are looking at a per-character limit where the characters are multi-byte (say SSMS shortcuts which are limited to 32767 UCS2 characters) then some form of base4096 (12 bits per character so three 8-bit bytes to two output characters) might be useful. Yes, I have done this: putting a long analysis proc (a replacement for sp_help & friends) into a “shortcut” I was getting close to the 32K-char limit so compressed and base64ed the code and included an unpacker. This was more than enough to deal with the problem (TBH, just stripping comments would have done!) and simplified things in some ways as I no longer needed to escape quotes and such, but I went one step further played with writing a B4096 encoder because I like playing with that sort of silliness. I went with 4096 due to aligning nicely with 4-bit boundaries, and finding 4096 useable characters (avoiding control characters, undefined codes, and other unprintables) is easy. If the limit is actually 32767 or any Unicode characters (including those not in the UCS2 or UTF16 base plane set) then you could perhaps get even more daft though I'm pretty sure it is just 16-bit characters and not full Unicode.
For efficiency, I would prefer to store files with a random short filename that's just letters and digits and store the original name in a separate database.
Practically? Quite a few codebases from *nix systems rely on the case sensitivity. To do a successful git clone you need a file system that handles both Makefile and makefile, or whatever.
I ran into it more than once back in the day when I used Mac OS X as my primary OS.
More generally -- case sensitivity is a conceptual nightmare in the Unicode era. Should Cyrillic or Greek be case-insensitive as well? Should visually similar glyphs e.g. Cyrillic e and Latin e be "glyph-insensitive" as well? Etc. Do you really want the full complexity of Unicode string handling in your file system?
Just treat it like raw bytes. On modern Linux, anything but NUL is valid.
This is true, but I'm not sure if I trust users to preserve case in case-insensitive filenames when copying files around, possibly at some point passing through something like a web server that may apply case normalization. Sort of like timestamps on files, they're handy for things like synchronization, but I hesitate before relying on them as a primary source of truth for anything mission critical.
That is, the only safe non-alphanumeric characters are dash "-" and underscore "_". Period "." is also considered 'safe' but is tricky for obvious reasons and best avoided except where necessary.
Authoritative source: POSIX.1-2024 standard, sec. 3.265 Portable Filename Character Set
(Also for anything that needs to work on a shell script or whatever)
I roll my eyes every time some "super secure secret generator" gives me a password with &^%#$ that I need to use on a shell script. Thanks for making my life harder I guess
"oh but you just need to escape them" Yes please be my guest trying to deal with character escape whack-a-mole
It's just a fundamental problem of encoding the boundary between different (meta-)languages in-band. It will always be a problem, until it's possible to implement the separation out-band, which makes the respective language no longer self-contained, as it needs to be implemented in the layer above.
This occurs now with LLMs too. It's a fundamental problem, of the problem description, not the implementation.
Yeah, the innovation would be to not make everything in-band. Which implies proper file formats for source code for example. I know, that is blasphemy against 7bit ASCII.
It's like the telephone networks of old. They eventually realized that that is not so good. The computer world is still in its analog telephone phase. ;-)
They also have single quote, dollar and backtick... this is a lot worse than just whitespace. The problem with whitespace surface when the programmer made no effort to escape the string used as a file name. This adds a whole new vector of problems that happen when the programmer tried to escape the string, but it didn't work.
It depends on the shell you use and how your file names are handled.
Bash is a clusterfuck waiting to happen when it comes to file names, but not all shells would be at risk here. For example the shell I maintain would be fine because file names are tokenised like how variables are tokenised in SQL strings when using parametrised queries (to use your SQL/PHP example).
Same! Although I still have muscle-memory replacing spaces with _ or - when creating/saving files on any OS.
Hasn't done me wrong lo these past 40 years.
As for the others tailing the ASCII, no thanks.
Base64 all the way, but with dot and underscore. Dash is problematic (or, well, let's say you're bound to end up on a script not correctly dealing with those: you definitely can deal correctly with dashes and spacing chars in filenames, but you're guaranteed to encounter a SNAFU at some point when a script you didn't write shall invariably fuxx0rs everything up).
Dot is accepted and common in filenames anyway.
I've got a scheme where I added checksums to my files (in the filename) using dashes but switching to underscore is on my TODO list. Basically I'll change this:
dsc0974987-b3-0a772982c1.jpg
to:
dsc0974987_b3_0a772982c1.jpg
(YMMV as to whether adding the first x bits of whatever checksum you fancy on files you know aren't going to change works or doesn't work for you: it works for me and it's part of a suite of tools I made for myself for file integrity / backuping purpose / classification / etc.)
Meanwhile, Windows uses UTF-16 everywhere internally, so "a" and "あ" are both 2 bytes large. You'd still have to exclude surrogate pairs and the Unicode control characters from a filename.
Approximately nothing that uses UTF-16 validates it (I can’t think of a single thing that does), so surrogates will work fine. At least until a UTF-8 system touches it, because they normally do validate (Go is an uncommon exception in not validating).
The thing is that Windows is actually UCS-2, a consequence of the OS predating UTF-16 by a few years, and coming into existence when Unicode originally thought 16 bits was going to be enough.
For backwards compatibility with old file systems, there's no way the OS can start enforcing surrogate codepoints as forbidden from names. You can just so happen to pretend it's UTF-16 until it's not.
I think bit-aligned encodings (i.e. schemes where the output alphabets have a size that is a power of two) are the sweet-spot for encoding schemes due to simplicity of encoding and decoding compared to non-bit-aligned schemes like base84.
Base16 is verbose, but each symbol in the alphabet carries exactly 4 bits (hence why a single byte in hex is two symbols). This is nice for encoding bytes. Base64 requires two non-alphanumeric symbols such as + and / or - and _ because A-Za-z0-9 only provides 62 characters. Each output symbol encodes exactly 6 bits of the original input. But when we're encoding whole bytes we sometimes need to use padding (when the input is not a multiple of 3 bytes and that ambiguity is harmful in the context that decoding occurs in). Then there's Base32, which is what currently seems like a sweet spot to me and is what I'm considering using for identifiers in my own systems. It only requires 32 characters, so can easily take a range of A-Z0-9 or a-z0-9 excluding visually-similar characters. Like Base64 it can also require padding in some settings.
I quite like the general scheme used in Bech32 and Bech32m framing idea, which allows for a human-readable prefix before a `1` and then the base32 data follows. This can be done because `1` is excluded from the Bech32 alphabet. This prefix can be used to differentiate between kinds of identifier, for example.
But to address the article directly, I'm not sure the complexity is worth the gains in the table in https://github.com/jedisct1/zig-base84#encoding, which states that we save ~6.5% of characters, and for 128 input bytes, the resulting Base64 string is 171 characters, while the Base84 string is 161-166 characters.
I also dislike that the chosen alphabet breaks text selection; GitHub very carefully picked their current token formats so that the whole token is selected with a double-click:
> One other neat thing about _ is it will reliably select the whole token when you double click on it. Other characters we considered are sometimes included in application word separators and thus will stop highlighting at that character. Try out double clicking this-random-text versus this_random_text!
Regardless though we still have the fundamental issue that bytes encoded as Base16, Base32, Base64, or even Base91/Base84 will be _longer_ when encoded. Base32 can encode up to 159 bytes before hitting a filename limit of 255 bytes. I can't remember whether filenames are 16-bit on Windows (UCS-2? UTF-16?), but I speculate on average that maybe you could save a lot of bytes by first ensuring that filenames are UTF-8 before encrypting them, since filenames on a lot of computeres, especially ones running in the west, likely contain a lot of latin characters. You could even switch between encodings to get the best bit-packing in filenames.
> Base84 distinguishes uppercase and lowercase letters, so different encoded names can compare equal on a case-insensitive filesystem.
I understand for the authors particular application the probability of collision is small, but this makes this encoding completely impractical for filenames generally.
Encrypting filenames is an interesting idea, but also doesn't really seem that useful in practice to me.
Most projects or directories probably don't contain filenames that are sensitive. If you do have sensitive filenames, storing the true filenames in an encrypted manifest will work fine if you don't have a huge number of files.
For the use case where you have a lot of data, the data needs to be encrypted, and the filenames are sensitive you'd surely be better off using something like dm-crypt (if you can use a block device) or a FUSE solution like gocryptfs.
48 comments
[ 2.1 ms ] story [ 42.4 ms ] thread> That leaves enough room for a group to hold 32 bits about 95% of the time on uniformly random input, and 31 bits otherwise.
> The encoder looks at the next 31 bits. If their value is below 84⁵ - 2³¹, there’s room for a 32nd bit. Otherwise, it consumes just those 31 bits. Either way, the value fits in five base-84 digits.
Seems pretty awkward.
> ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz0123456789!#$%&'()+,-;=@[]^_`{}~
Yeah I definitely don't want any of those last ~22 characters in file names if I need to do anything with them from a shell
I quite like base58 for human-presented identifiers also. I tend to use it instead of hex for having ~8-16 character IDs of things.
Even just the extra characters in base64 are often annoying and we already have multiple common base64 standards because even trying to find just two more characters to fill out a-zA-Z0-9 is hard.
Also on efficiency, you are trading off code complexity for a little extra storage: base64 is a nice round 6 bits per character meaning every three bytes encoded is four output. Neither base84 nor base58 align on convenient bit boundaries like that so choosing the output character is more faf. Padding could be more complicated too.
Now if you are looking at a per-character limit where the characters are multi-byte (say SSMS shortcuts which are limited to 32767 UCS2 characters) then some form of base4096 (12 bits per character so three 8-bit bytes to two output characters) might be useful. Yes, I have done this: putting a long analysis proc (a replacement for sp_help & friends) into a “shortcut” I was getting close to the 32K-char limit so compressed and base64ed the code and included an unpacker. This was more than enough to deal with the problem (TBH, just stripping comments would have done!) and simplified things in some ways as I no longer needed to escape quotes and such, but I went one step further played with writing a B4096 encoder because I like playing with that sort of silliness. I went with 4096 due to aligning nicely with 4-bit boundaries, and finding 4096 useable characters (avoiding control characters, undefined codes, and other unprintables) is easy. If the limit is actually 32767 or any Unicode characters (including those not in the UCS2 or UTF16 base plane set) then you could perhaps get even more daft though I'm pretty sure it is just 16-bit characters and not full Unicode.
I ran into it more than once back in the day when I used Mac OS X as my primary OS.
More generally -- case sensitivity is a conceptual nightmare in the Unicode era. Should Cyrillic or Greek be case-insensitive as well? Should visually similar glyphs e.g. Cyrillic e and Latin e be "glyph-insensitive" as well? Etc. Do you really want the full complexity of Unicode string handling in your file system?
Just treat it like raw bytes. On modern Linux, anything but NUL is valid.
And slash/!
Why not: Some other things (famously Photoshop CS6) rely on case-insensitive.
And I guess it relies on the 31 (32?) bit blocks, to make colisions unlikely for short filenames.
Authoritative source: POSIX.1-2024 standard, sec. 3.265 Portable Filename Character Set
https://pubs.opengroup.org/onlinepubs/9799919799/basedefs/V1...
(Also for anything that needs to work on a shell script or whatever)
I roll my eyes every time some "super secure secret generator" gives me a password with &^%#$ that I need to use on a shell script. Thanks for making my life harder I guess
"oh but you just need to escape them" Yes please be my guest trying to deal with character escape whack-a-mole
"Escaping" should be something some graybeard is mumbling about.
Yet here we are in 2026, still having be on the lookout for a wild /../
Even PowerShell is boilerplate, boilerplate galore.
This occurs now with LLMs too. It's a fundamental problem, of the problem description, not the implementation.
It's like the telephone networks of old. They eventually realized that that is not so good. The computer world is still in its analog telephone phase. ;-)
Bash is a clusterfuck waiting to happen when it comes to file names, but not all shells would be at risk here. For example the shell I maintain would be fine because file names are tokenised like how variables are tokenised in SQL strings when using parametrised queries (to use your SQL/PHP example).
Base62 (or 64, with -_) is totally sufficient.
Base64 all the way, but with dot and underscore. Dash is problematic (or, well, let's say you're bound to end up on a script not correctly dealing with those: you definitely can deal correctly with dashes and spacing chars in filenames, but you're guaranteed to encounter a SNAFU at some point when a script you didn't write shall invariably fuxx0rs everything up).
Dot is accepted and common in filenames anyway.
I've got a scheme where I added checksums to my files (in the filename) using dashes but switching to underscore is on my TODO list. Basically I'll change this:
to: (YMMV as to whether adding the first x bits of whatever checksum you fancy on files you know aren't going to change works or doesn't work for you: it works for me and it's part of a suite of tools I made for myself for file integrity / backuping purpose / classification / etc.)For backwards compatibility with old file systems, there's no way the OS can start enforcing surrogate codepoints as forbidden from names. You can just so happen to pretend it's UTF-16 until it's not.
Base16 is verbose, but each symbol in the alphabet carries exactly 4 bits (hence why a single byte in hex is two symbols). This is nice for encoding bytes. Base64 requires two non-alphanumeric symbols such as + and / or - and _ because A-Za-z0-9 only provides 62 characters. Each output symbol encodes exactly 6 bits of the original input. But when we're encoding whole bytes we sometimes need to use padding (when the input is not a multiple of 3 bytes and that ambiguity is harmful in the context that decoding occurs in). Then there's Base32, which is what currently seems like a sweet spot to me and is what I'm considering using for identifiers in my own systems. It only requires 32 characters, so can easily take a range of A-Z0-9 or a-z0-9 excluding visually-similar characters. Like Base64 it can also require padding in some settings.
I quite like the general scheme used in Bech32 and Bech32m framing idea, which allows for a human-readable prefix before a `1` and then the base32 data follows. This can be done because `1` is excluded from the Bech32 alphabet. This prefix can be used to differentiate between kinds of identifier, for example.
https://github.com/bitcoin/bips/blob/master/bip-0350.mediawi...
But to address the article directly, I'm not sure the complexity is worth the gains in the table in https://github.com/jedisct1/zig-base84#encoding, which states that we save ~6.5% of characters, and for 128 input bytes, the resulting Base64 string is 171 characters, while the Base84 string is 161-166 characters.
I also dislike that the chosen alphabet breaks text selection; GitHub very carefully picked their current token formats so that the whole token is selected with a double-click:
> One other neat thing about _ is it will reliably select the whole token when you double click on it. Other characters we considered are sometimes included in application word separators and thus will stop highlighting at that character. Try out double clicking this-random-text versus this_random_text!
https://github.blog/engineering/platform-security/behind-git...
For example, compare double-clicking on the article's Base84 alphabet:
> ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz0123456789!#$%&'()+,-;=@[]^_`{}~
with the Bech32 Base32 alphabet:
> qpzry9x8gf2tvdw0s3jn54khce6mua7l
Regardless though we still have the fundamental issue that bytes encoded as Base16, Base32, Base64, or even Base91/Base84 will be _longer_ when encoded. Base32 can encode up to 159 bytes before hitting a filename limit of 255 bytes. I can't remember whether filenames are 16-bit on Windows (UCS-2? UTF-16?), but I speculate on average that maybe you could save a lot of bytes by first ensuring that filenames are UTF-8 before encrypting them, since filenames on a lot of computeres, especially ones running in the west, likely contain a lot of latin characters. You could even switch between encodings to get the best bit-packing in filenames.
At the moment it looks like turbocrypt forbids encrypted filenames that are too long: https://github.com/jedisct1/turbocrypt/blob/4905241d271e84a...
I understand for the authors particular application the probability of collision is small, but this makes this encoding completely impractical for filenames generally.
Most projects or directories probably don't contain filenames that are sensitive. If you do have sensitive filenames, storing the true filenames in an encrypted manifest will work fine if you don't have a huge number of files.
For the use case where you have a lot of data, the data needs to be encrypted, and the filenames are sensitive you'd surely be better off using something like dm-crypt (if you can use a block device) or a FUSE solution like gocryptfs.