Presenting Julids

Nebcorp Heavy Industries and Sundries, long the world leader in sundries, is proud to announce the public launch of the official identifier type for all Nebcorp companies' assets and database entries, Julids. Julids are globally unique sortable identifiers, backwards-compatible with ULIDs, but better.

Inside your Rust program, simply add julid-rs¹ to your project's Cargo.toml file, and use it like:

use julid::Julid;

fn main() {
    let id = Julid::new();
    dbg!(id.created_at(), id.as_string());
}

Such a program would output something like:

[main.rs:5] id.created_at() = 2023-07-29T20:21:50.009Z
[main.rs:5] id.as_string() = "01H6HN10SS00020YT344XMGA3C"

However, it can also be built as a loadable extension for SQLite, adding database functions for creating and querying Julids:

$ sqlite3
SQLite version 3.40.1 2022-12-28 14:03:47
Enter ".help" for usage hints.
Connected to a transient in-memory database.
Use ".open FILENAME" to reopen on a persistent database.
sqlite> .load ./libjulid
sqlite> select hex(julid_new());
018998768ACF000060B31DB175E0C5F9
sqlite> select julid_string(julid_new());
01H6C7D9CT00009TF3EXXJHX4Y
sqlite> select julid_seconds(julid_new());
1690480066.208
sqlite> select datetime(julid_timestamp(julid_new()), 'auto');
2023-07-27 17:47:50
sqlite> select julid_counter(julid_new());
0

Intrigued? Confused? Disgusted? Enraged?? Well, read on!

Julids vs ULIDs

Julids are a drop-in replacement for ULIDs: all Julids are valid ULIDs, but not all ULIDs are valid Julids.

Given their compatibility relationship, Julids and ULIDs must have quite a bit in common, and indeed they do:

• they are 128-bits long
• they are lexicographically sortable
• they encode their creation time as the number of milliseconds since the UNIX epoch in their top 48 bits
• their string representation is a 26-character base-32 Crockford encoding of their big-endian bytes
• IDs created within the same millisecond are still meant to sort in their order of creation

Julids and ULIDs have different ways to implement that last piece. If you look at the layout of bits in a ULID, you see:

According to the ULID spec, for ULIDs created within the same millisecond, the least-significant bit should be incremented for each new ID. Since that portion of the ULID is random, that means you may not be able to increment it without spilling into the timestamp portion. Likewise, it's easy to guess a new possibly-valid ULID simply by incrementing an already-known one. And finally, this means that sorting will need to read all the way to the end of the ULID for IDs created in the same millisecond.

To address these shortcomings, Julids (Joe's² ULIDs) have the following structure:

As with ULIDs, the 48 most-significant bits encode the time of creation. Unlike ULIDs, the next 16 most-significant bits are not random³: they're a monotonic counter for IDs created within the same millisecond⁴. Since it's only 16 bits, it will saturate after 65,536 IDs intra-millisecond creations, after which, IDs in that same millisecond will not have an intrinsic total order (the random bits will still be different, so you shouldn't have collisions). My PC, which is no slouch, can only generate about 20,000 per millisecond, so hopefully this is not an issue! Because the random bits are always fresh, it's not possible to easily guess a valid Julid if you already have one.

How to use

The Julid crate can be used in two different ways: as a regular Rust library, declared in your Rust project's Cargo.toml file (say, by running cargo add julid-rs), and used as shown above. There's a rudimentary benchmark example in the repo, which I'll talk more about below. But the primary use case for me was as a loadable SQLite extension, as I previously wrote. Both are covered in the documentation, but let's go over them here, starting with the extension.

Inside SQLite as a loadable extension

The extension, when loaded into SQLite, provides the following functions:

• julid_new(): create a new Julid and return it as a 16-byte blob
• julid_seconds(julid): get the number seconds (as a 64-bit float) since the UNIX epoch that this julid was created
• julid_counter(julid): show the value of this julid's monotonic counter
• julid_sortable(julid): return the 64-bit concatenation of the timestamp and counter
• julid_string(julid): show the base-32 Crockford encoding of this julid; the raw bytes of Julids won't be valid UTF-8, so use this or the built-in hex() function to select a human-readable representation

Building and loading

If you want to use it as a SQLite extension:

• clone the repo
• build it with cargo build --features plugin (this builds the SQLite extension)
• copy the resulting libjulid.[so|dylib|whatevs] to some place where you can...
• load it into SQLite with .load /path/to/libjulid as shown at the top
• party

If you, like me, wish to use Julids as primary keys, just create your table like:

create table users (
  id blob not null primary key default (julid_new()),
  ...
);

and you've got a first-class ticket straight to Julid City, baby!

For a table created like:

-- table of things to watch
create table if not exists watches (
  id blob not null primary key default (julid_new()),
  kind int not null, -- enum for movie or tv show or whatev
  title text not null,
  length int,
  release_date int,
  added_by blob not null,
  last_updated int not null default (unixepoch()),
  foreign key (added_by) references users (id)
);

and then some code that inserted rows into that table like

insert into watches (kind, title, length, release_date, added_by) values (?,?,?,?,?)

where the wildcards get bound in a loop with unique values and the Julid id field is generated by the extension for each row, I get over 100,000 insertions/second.

Inside a Rust program

Of course, you can also use it outside of a database; the Julid type is publicly exported. There's a simple benchmark in the examples folder of the repo, the important parts of which look like:

use julid::Julid;

fn main() {
    /* snip some stuff */

    let start = Instant::now();
    for _ in 0..num {
        v.push(Julid::new());
    }
    let end = Instant::now();
    let dur = (end - start).as_micros();

    for id in v.iter() {
        eprintln!(
            "{id}: created_at {}; counter: {}; sortable: {}",
            id.created_at(),
            id.counter(),
            id.sortable()
        );
    }
    println!("{num} Julids generated in {dur}us");

If you were to run it on a computer like mine⁵, you might see something like this:

$ cargo run --example=benchmark --release -- -n 30000 2> /dev/null
30000 Julids generated in 1240us

That's about 24,000 IDs/millisecond; 24 MILLION per second!

The default optional Cargo features include implementations of traits for getting Julids into and out of SQLite with SQLx, and for generally serializing/deserializing with Serde, via the sqlx and serde features, respectively. One final default optional feature, chrono, uses the Chrono crate to return the timestamp as a DateTime by adding a created_at(&self) method to Julid.

Something to note: don't enable the plugin feature in your Cargo.toml if you're using this crate inside your Rust application, especially if you're also loading it as an extension in SQLite in your application. You'll get a long and confusing runtime panic due to there being multiple entrypoints defined with the same name.

Why Julids?

The astute may have noticed that this is the third time I've written about globally unique sortable IDs (here is part one, and part two is here). What's, uh... what's up with that?

Like Marge, I just think they're neat! We're not the only ones; here are just some related projects:

• Segment's KSUID, released in 2017. This was possibly my first exposure to this idea. They're 36 bits larger than UUIDs or ULIDs, but otherwise very similar to ULIDs (and hence Julids)
• ULIDs, of course
• UUIDv7; these are very similar to Julids; the primary difference is that the lower 62 bits are left up to the implementation, rather than always containing pseudorandom bits as in Julids (which use the lower 64 bits for that, instead of UUIDv7's 62)
• Snowflake ID, developed by Twitter in 2010; these are 63-bit identifiers (so they fit in a signed 64-bit number), where the top 41 bits are a millisecond timestamp, the next 10 bits are a machine identifier⁶, and the last 12 bits are for an intra-millisecond sequence counter (what Julid calls a "monotonic counter"); unlike all the other IDs discussed, there are no random bits

and I'm sure the list can go on.

I wanted to use them in my SQLite-backed web app, in order to fix some deficiencies in ULIDs and the way I was using them, as I said before:

[...] it bothers me that ID generation is not done inside the database itself. Aside from being a generally bad idea, this lead to at least one frustrating debug session where I was inserting one ID but reporting back another. SQLite doesn't have native support for this, but it does have good native support for loading shared libraries as plugins in order to add functionality to it, and so my next step is to write one of those, and remove the ID generation logic from the application.

Now that I've accomplished all that I've set out to do, is this the last time I'll time I'll be writing at length about these things? It's hard to say for sure, but signs point to "yes". I hope you've found them at least a little interesting!

Thanks

This project wouldn't have happened without a lot of inspiration (and a little shameless stealing) from the ulid-rs crate. For the loadable extension, the sqlite-loadable-rs crate made it extremely easy to write; what I thought would take a couple days instead took a couple hours. Thank you, authors of those crates! Feel free to steal code from me any time!