Comment by russellthehippo | Hacker News Reader

Comment by russellthehippo | Hacker News Reader

russellthehippo

Hey HN, I built this. Honker adds cross-process NOTIFY/LISTEN to SQLite. You get push-style event delivery with single-digit millisecond latency without a damon/broker, using your existing SQLite file. A lot of pretty high-traffic applications are just Framework+SQLite+Litestream on a VPS now, so I wanted to bring a sixer to the "just use SQLite" party.

SQLite doesn't run a server like Postgres, so the trick is moving the polling source from interval queries on a SQLite connection to a lightweight stat(2) on the WAL file. Many small queries are efficient in SQLite (https://www.sqlite.org/np1queryprob.html) so this isn't really a huge upgrade, but the cross-language result is pretty interesting to me - this is language agnostic as all you do is listen to the WAL file and call SQLite functions.

On top of the store/notify primitives, honker ships ephemeral pub/sub (like pg_notify), durable work queues with retries and dead-letter (like pg-boss/Oban), and event streams with per-consumer offsets. All three are rows in your app's existing .db file and can commit atomically with your business write. This is cool because a rollback drops both.

This used to be called litenotify/joblite but I bought honker.dev as a joke for my gf and I realized that every mq/task/worker have silly names: Oban, pg-boss, Huey, RabbitMQ, Celery, Sidekiq, etc. Thus a silly goose got its name.

Honker waddles the same path as these giants and honks into the same void.

Hopefully it's either useful to you or is amusing. Standard alpha software warnings apply.

Is the main use case for this for languages that only have access to process based concurrency?

Struggling to see why you would otherwise need this in java/go/clojure/C# your sqlite has a single writer, so you can notify all threads that care about inserts/updates/changes as your application manages the single writer (with a language level concurrent queue) so you know when it's writing and what it has just written. So it always felt simpler/cleaner to get notification semantics that way.

Still fun to see people abuse WAL in creative ways. Cool to see a notify mechanism that works for languages that only have process based concurrency python/JS/TS/ruby. Nice work!

He mentions Litestream, maybe this also works for litestream read-only replicas which may be in completely different locations?

Whoa I really hadn’t considered this. Do a litestream read replica, trigger across machines with S3 as the broker essentially. But you’re still stuck with the litestream sync interval. Maybe interesting for cross server notify?

I guess the idea is to have all writes go through a central server with local read replicas for improved read perf. The default litestream sync interval is 1s. I bet many use-cases would be satisfied with a few seconds delay for cross-region notifications.

It's good for pubsub but not for claim/ack workflow unless you do If-None-Match CAS semantics on a separate filesystem which, actually, yeah that's probably fine. Feels heavy on S3 ops. But! you do save on inter-AZ networking, the Warpstream hypothesis.

Claims kill this, IMO.

Unless you have a single "reader", you don't mind the delay, and don't worry about redoing a bunch of notifications after a crash (and so, can delay claims significantly), concurrency will kill this.

exactly. then you're building distributed locking and it's probably time for a different tool

I wrote a simple queue implementation after reading the Turbopuffer blog on queues on S3. In my implementation, I wrote complete sqlite files to S3 on every enqueue/dequeue/act. it used the previous E-Tag for Compare-And-Set.

The experiment and back-of-the-envelope calculations show that it can only support ~ 5 jobs/sec. The only major factor to increase throughput is to increase the size of group commits.

I dont think shipping CDC instead of whole sqlite files will change the calculations as the number of writes mattered in this experiment.

So yes, the number of writes (min. of 3) can support very low throughputs.

I actually hadn’t thought about it this way. The killer app I was imagining was 1ms reactivity without SQL polling and messaging atomic with business commits, plus “one db” and no daemon.

But this is actually a great main benefit as well.

There's more process-based concurrency than you'd expect in shops that use those languages.

Cron jobs might need to coordinate with webservers. Even heavily threaded webservers might have some subprocesses/forking to manage connection pools and hot reloads and whatnot. Suid programs are process-separated from non-suid programs. Plenty of places are in the "permanent middle" of a migration from e.g. Java 7 to Java 11 and migrate by splitting traffic to multiple copies of the same app running on different versions of the runtime.

If you're heavily using SQLite for your DB already, you probably are reluctant to replace those situations with multiple servers coordinating around a central DB.

Nit:

> languages that only have process based concurrency python/JS/TS/ruby

Not true. There are tons and tons of threaded Python web frameworks/server harnesses, and there were even before GIL-removal efforts started. Just because gunicorn/multiprocessing are popular doesn't mean there aren't loads of huge deployments running threads (and not suffering for it much, because most web stacks are IO bound). Ruby's similar, though threads are less heavily-used than in Python. JS/TS as well: https://nodejs.org/api/worker_threads.html

Nice, I had no idea that stat() every 1 ms is so affordable. Aparently it takes less than 1 μs per call on my hardware, so that's less than 0.1% cpu time for polling.

"Syscalls are slow" is only mostly true. They are slower than not having to cross the userspace <-> OS barrier at all, but they're not "slow" like cross-ocean network calls can be. For example, non-VDSO syscalls in linux are about 250 nanoseconds (see for example https://arkanis.de/weblog/2017-01-05-measurements-of-system-...), VDSO syscalls are roughly 10x faster. Slower than userspace function calls for sure, but more than affordable outside the hottest of loops.

That’s ignoring the other costs of syscalls like evicting your stuff from the CPU caches.

But I agree with the conclusion, system calls are still pretty fast compared to a lot of other things.

Small correction on ambiguous wording - syscalls do not evict all your stuff from CPU caches. It just has to page in whatever is needed for kernel code/data accessed by the call, but that’s no different from if it was done in process as a normal function call.

Depending on implementation details of your CPU and OS, the syscall path may need to flush various auxillary caches (like one or more TLBs) to prevent speculation attacks, which may put additional "drag" on your program after syscall return.

Correct but you’d also still have that drag just from the kernel dirtying those caches in the first place.

But I was clarifying because the wording could be taken as data/instruction cache and there generally isn’t a full flush of that just to enter/leave kernel.

Yes, that’s right. It seems to make up the bulk of the cost of a system call though, depending what it does, like read or write syscalls.

Filesystem stuff tends to be slower than average syscalls because of all the locks and complicated traversals needed. If this is using stat instead of fstat then it’s also going through the VFS layer - repeated calls likely go through the cache fast path for path resolution but accessing the stat structure. There’s also hidden costs in that number like atomic accesses that need to acquire cache line locks that are going to cause hidden contention for other processes on the CPU + the cache dirtying from running kernel code and then subsequently having to repopulate it when leaving all of which adds contended L3/RAM pressure.

In other words, there’s a lot of unmeasured performance degradation that’s a side effect of doing many syscalls above and beyond the CPU time to enter/leave the kernel which itself has shrunk to be negligible. But there’s a reason high performance code is switching to io_uring to avoid that.

Oh cool, so using io uring plus pragma data version would actually beat stat on Linux holistically speaking? The stat choice was all about cross platform consistency over inotify speed. But syscalls overwhelm can be real.

If you're interested you can use kqueue on FreeBSD and Darwin to watch the inode for changes. Faster than a syscall, especially if all you need is a wakeup when it changes.

“Beat” is all relative. It depends on load and how frequently you’re doing it, but generally yes. But if you’re doing io_uring, you may as well use inotify because you’re in the platform specific API anyway as that’s the biggest win because you’re moving from polling to change detection which is less overhead and lower latency. Inotify can be accessed by io_uring and there may even be cross-platform libraries for your language that give you a consistent file watcher interface (although probably not optimally over io_uring). Whether it’s actually worth it is hard as I don’t know what problem you’re trying to solve, but the super lowest overhead looks like inotify+iouring (it also has the lowest latency)

Probably missing something, why is `stat(2)` better than: `PRAGMA data_version`?

https://sqlite.org/pragma.html#pragma_data_version

Or for a C API that's even better, `SQLITE_FCNTL_DATA_VERSION`:

https://sqlite.org/c3ref/c_fcntl_begin_atomic_write.html#sql...

For one it seems to be deprecated.

It's not.

Yep, definitely still in use. Do yall above have an opinion if the pragma is better than the syscall? What are the trade offs there? Another comment thread mentioned this as well and pointed to io uring. I was thinking that dism spam is worse than syscall spam.

Depends on what to mean by better.

I may be wrong, but I think you wrote somewhere that you're looking at the WAL size increasing to know if something was committed. Well, the WAL can be truncated, what then? Or even, however unlikely, it could be truncated, then a transaction comes and appends just enough to it to make it the same size.

If SQLite has an API it guarantees can notify you of changes, that seems better, in the sense that you're passing responsibility along to the experts. It should also work with rollback mode, another advantage. And I don't think wakes you up if a large transaction rolls back (a transaction can hit the WAL and never commit).

That said, I'm not sure what's lighter on average. For a WAL mode database, I will say that something that has knowledge of the WAL index could potentially be cheaper? That file is mmapped. The syscalls involved are file locks, if any.

Interesting, thank you for the response and explanation. Honker workers/listerners are holding an open connection anyway. I do trust SQLite guarantees more than cross-platform sys behavior. I will explore the C API angle.

You are correct. I apologize. I seemed to have read the next pragma’s depreciation notice!

Aside from this - SQLite has tons of cool features, like the session extension.

Yeah the C API seems like a perfect fit for this use-case:

> [SQLITE_FCNTL_DATA_VERSION] is the only mechanism to detect changes that happen either internally or externally and that are associated with a particular attached database.

Another user itt says the stat(2) approach takes less than 1 μs per call on their hardware.

I wonder how these approaches compare across compatibility & performance metrics.

I just tested this out. PRAGMA data_version uses a shared counter that any connection can use while the C API appears to use a per-connection counter that does not see other connections' commits.

Really? That's the opposite of what I understand the docs say.

Care to share your code? This may become a bug report.

Will do

Reporting back. This appears to be a bug in my original test the code of which sadly I did not commit anywhere. I went back to regenerate these tests and proved the opposite - the C API is better than PRAGMA and works across connections. I am going to make that update as I've proved across dozens of versions of SQLite that this is not in fact the case.

Reporting back again. It seems I was actually right the first time - the C API's SQLITE_FCNTL_DATA_VERSION doesn't work cross connection. It is cached on each read - but if there aren't any reads (i.e. just polling SQLITE_FCNTL_DATA_VERSION) then it doesn't work.

PRAGMA data_version is pretty fast (1500ns with prepared statement) and doesn't have that issue.

Checking the wal-index is sub-nanosecond when mmapped but has slightly different behavior on Windows.

Here's the link to the thread on this, my scripts are all there.

https://github.com/russellromney/honker/issues/5

Pretty cool! I have a half baked version of something similar :)

Can you use it also as a lightweight Kafka - persistent message stream? With semantics like, replay all messages (historical+real time) from some timestamp for some topics?

As with pub/sub, you can reproduce this with some polling etc but as you say, that's not optimal.

Absolutely! That’s the durable pubsub angle for sure.

Neat idea!

Would it help if subscriber states were also stored? (read position, queue name, filters, etc) Then instead of waking all subscription threads to do their own N=1 SELECT when stat(2) changes, the polling thread could do Events INNER JOIN Subscribers and only wake the subscribers that match.

This is really interesting. I'm building something on Postgresql with LISTEN/NOTIFY and Postgraphile. I'd love to (in theory) be able to have a swappable backend and not be so tightly coupled to the database server.

I love the name!

honk

Very cool!

Another maybe stupid question, would something like inotify(7) help to get rid of any active polling?

Reply on news.ycombinator.com