"Idle cost is that one lightweight SELECT per millisecond per database — no page-cache pressure, no writer-lock contention, no kernel file watcher in the mix."
I think (respectfully) the LLM that probably wrote this overshot the mark here because busy-polling a select does not actually sound better to me than a "kernel file watcher".
One cannot be a little bit pregnant. But a DB can be only a little bit in the RAM, and specifically in the page cache. SQLite can act exactly like that, and it's damn fast as long as it does not need to durably write a transaction. Polling once a millisecond could spend a few microseconds.
I wonder if using a tiny Redis instance, or even something like LevelDB would be even more efficient.
Respectfully (thanks haha) - yeah probably right. Original intent was to use inotify type thing but i avoided per-platform differences at the outset. this was definitely a for fun project that blew up unintentionally and am working to harden/improve.
to me it sounds like they asked it to not make a kernel file watcher, and now it writes that into every comment everywhere, despite not even being in the implementation
That only catches changes made by the database connection being "hooked."
This has a thread running in the background trying to catch changes made by other connections, potentially (I'm not sure here, but I suspect as much) in different processes that are modifying the same database.
good point. but ime and as seems to be widely understood writing from multiple connections is a bit of a minefield in SQLite. and afaik it still would be possible to have a hook on all connections you expect to be writing?
i did a quick benchmark on this with a single db connection updating user_version in a tight loop with the wal_hook callback enabled.
on my crappy old i5 with the db file on /dev/shm it can do ~150k writes a second with the wal_hook callback called on every write. and this is using JS bindings to C++ so has some unnecessary overhead.
That wouldn't work across processes. And if you only care about in-process queuing then you might find it easier/faster to use another kind of storage or roll your own WAL.
Yeah, I had the same instinct - this feels very much like a "nice idea" but the execution falls short. I mean - busily banging on sqlite like this? Shit at that point just use Redis.
For what it's worth, Kine (software that k3s uses to replace etcd with SQL databases) implements etcd watches on SQLite through polling[1]. The reason being that SQLite does not offer NOTIFY/LISTEN like MySQL and Postgres do. Ironically, Honkey attempts implementing NOTIFY/LISTEN through polling.
k3s has been running on my home server for about three years now (using the default SQLite backend), and there doesn't seem to be excessive CPU usage despite dozens of watches existing in the simulated etcd. Of course, this doesn't say much about Honker, but it's nonetheless worth pointing out that sometimes the choice of database forces one towards a certain design.
With SQLite, you're basically funneled towards a single-writer / single-process design anyway ... in which case why not use a more traditional condvar + mutex rather than polling?
Really might be in sqlite. I've learned to never trust my intuition about performance with that thing. So many times I've gone to "optimize" something and discovered that the naive hack way I had been doing it was faster anyway. It's built for this sort of bullshit.
I had a manual fs polling thing a while back. It was ugly (low time budget, didn't wanna mess with the native watchers), just scanned the whole thing once per second. It averaged out to like 0.3% CPU.
Not elegant, but acceptable for my purposes! (Small-ish directory, and "ping me within a second or two" was realtime enough for this use case.)
i mean, technically this is once per millisecond, so this would happen 1000x more. In your case due to the kernel overhead you would likely not even be able to do it (300% CPU?).
Either way this does seem like a very large overhead due to the fact that there's just no other way to do it without a deeper kernel integration which might be outside the scope of what sqlite is trying to do.
> Once real work flows through a SQLite-backed app, you need a queue. The usual answer is “add Redis + Celery.”
Are they joking? SQLite is usually used for single-process (mutliple threads) applications. The proper way to communicate between threads/processes is a ring buffer, where you allocate structs (allocation typically is incrementing a pointer), and futex/eventfd for notifications (+ some spinlocking to avoid going to kernel when the tasks arrive quickly). Why do you need redis for that? If you need persistent tasks, then you can store them in the table, and still use futex for notifications. This polling is inefficient and they should not make it a library which will cause other lazy developers add it to their app.
> honker polls SQLite’s PRAGMA data_version every millisecond. That’s a monotonic counter SQLite increments on every commit from any connection, journal mode, or process — a ~3 µs read for a precise wake signal
That's 3 ms per second = 0.3% CPU time wasted for every waiting thread.
Like Electron, this feels like written by a web developer and not a real programmer.
>That's 3 ms per second = 0.3% CPU time wasted for every waiting thread.
I suspect that's actually "per process, per database (usually 1)", and not based on number of threads or tables. `data_version` semantics mean there's no need for more than one connection polling it, and it's being used as a relatively lightweight "DB has changed, check queues" check (that's pretty much its whole purpose).
Also I believe this is mostly intended for multi-process use, e.g. out-of-process workers, so an in-process dirty tracker (e.g. just check after insert/update/delete) isn't sufficient.
So I do think it's somewhat crazy, but it is at least very simple. fsnotify-like monitoring seems like a fairly obvious improvement tho, not sure why that isn't part of it. Maybe it's slower? I haven't tried to do anything actually-performant-or-reliable with fs notifications, dunno what dragons lie in wait.
Key difference vs SQL polling is that we’re touching metadata instead of data pages. I have work in process to make this work without any polling (innotify, kqueue, mmap’d shm file check) after the original stat(2) direction proved unreliable if lightweight.
Would love your feedback and or contributions in the repo - still figuring out the end shape.
It’s an interesting approach and can be quite fun to use for new projects.
> How it works: honker polls SQLite’s PRAGMA data_version every millisecond. That’s a monotonic counter SQLite increments on every commit from any connection, journal mode, or process — a ~3 µs read for a precise wake signal.
I've implemented something similar in the past, but using inotify. You need to watch the -wal file for IN_MODIFY. To make it work reliably I found I had to run:
BEGIN IMMEDIATE TRANSACTION; ROLLBACK;
Otherwise the new changes weren't guaranteed to be visible to the process. I'm sure there's a more targetted approach that would work instead - maybe flock on a particular byte in the `-shm` file.
I think this is interesting too sqlite a as the coordination boundary: business state, queue state, stream offsets, retries, and acks all sharing one transactional substrate. The 1ms polling is getting a lot of weight in the thread though :)
51 comments:
"Idle cost is that one lightweight SELECT per millisecond per database — no page-cache pressure, no writer-lock contention, no kernel file watcher in the mix."
I think (respectfully) the LLM that probably wrote this overshot the mark here because busy-polling a select does not actually sound better to me than a "kernel file watcher".
"one lightweight SELECT per millisecond"
This reminds me of the teenager who told her dad that she was just a tiny little bit pregnant.
One cannot be a little bit pregnant. But a DB can be only a little bit in the RAM, and specifically in the page cache. SQLite can act exactly like that, and it's damn fast as long as it does not need to durably write a transaction. Polling once a millisecond could spend a few microseconds.
I wonder if using a tiny Redis instance, or even something like LevelDB would be even more efficient.
With the file-watch APIs is that you don't need to poll at all - free is better than cheap.
Thing of the battery!
(read that in the way of "think of the children!")
Respectfully (thanks haha) - yeah probably right. Original intent was to use inotify type thing but i avoided per-platform differences at the outset. this was definitely a for fun project that blew up unintentionally and am working to harden/improve.
Love Fly.
to me it sounds like they asked it to not make a kernel file watcher, and now it writes that into every comment everywhere, despite not even being in the implementation
Yup
A prepared `PRAGMA data_version` is likely quite cheap to run because it hits the same page every time…
…but some other push-based IPC mechanism would be a lot more battery friendly
If you're not making any changes to the database, does the SELECT "kill" you?
And if you are making changes, don't you have to poll regardless after the file watcher wakes you?
For WAL mode, SQLite can probably satisfy this query just by inspecting some shared memory. But it is busy waiting, sure.
SQLite has a wal hook which calls you back every time a transaction is committed to the WAL. https://www.sqlite.org/c3ref/wal_hook.html
That only catches changes made by the database connection being "hooked."
This has a thread running in the background trying to catch changes made by other connections, potentially (I'm not sure here, but I suspect as much) in different processes that are modifying the same database.
good point. but ime and as seems to be widely understood writing from multiple connections is a bit of a minefield in SQLite. and afaik it still would be possible to have a hook on all connections you expect to be writing?
i did a quick benchmark on this with a single db connection updating user_version in a tight loop with the wal_hook callback enabled.
on my crappy old i5 with the db file on /dev/shm it can do ~150k writes a second with the wal_hook callback called on every write. and this is using JS bindings to C++ so has some unnecessary overhead.
That wouldn't work across processes. And if you only care about in-process queuing then you might find it easier/faster to use another kind of storage or roll your own WAL.
Yeah, I had the same instinct - this feels very much like a "nice idea" but the execution falls short. I mean - busily banging on sqlite like this? Shit at that point just use Redis.
For what it's worth, Kine (software that k3s uses to replace etcd with SQL databases) implements etcd watches on SQLite through polling[1]. The reason being that SQLite does not offer NOTIFY/LISTEN like MySQL and Postgres do. Ironically, Honkey attempts implementing NOTIFY/LISTEN through polling.
k3s has been running on my home server for about three years now (using the default SQLite backend), and there doesn't seem to be excessive CPU usage despite dozens of watches existing in the simulated etcd. Of course, this doesn't say much about Honker, but it's nonetheless worth pointing out that sometimes the choice of database forces one towards a certain design.
[1] https://github.com/k3s-io/kine/blob/648a2daa/pkg/logstructur...
With SQLite, you're basically funneled towards a single-writer / single-process design anyway ... in which case why not use a more traditional condvar + mutex rather than polling?
Are you trying to avoid sleep?
I'm not even saying it's unworkable, just, my intuition is not that the "lightweight per-millisecond select" is an optimal design.
Really might be in sqlite. I've learned to never trust my intuition about performance with that thing. So many times I've gone to "optimize" something and discovered that the naive hack way I had been doing it was faster anyway. It's built for this sort of bullshit.
Maybe, I'm really writing about the language on this page, not about the design (I responded about this upthread).
Oh, yes, I see what you mean now.
What's the CPU usage? Like 2%?
I had a manual fs polling thing a while back. It was ugly (low time budget, didn't wanna mess with the native watchers), just scanned the whole thing once per second. It averaged out to like 0.3% CPU.
Not elegant, but acceptable for my purposes! (Small-ish directory, and "ping me within a second or two" was realtime enough for this use case.)
If this stops the core being able to drop to a lower power state it can be whole multiples of power use on some devices.
Wake ups are death for mobile form factors, even if not really doing much work.
i mean, technically this is once per millisecond, so this would happen 1000x more. In your case due to the kernel overhead you would likely not even be able to do it (300% CPU?).
Either way this does seem like a very large overhead due to the fact that there's just no other way to do it without a deeper kernel integration which might be outside the scope of what sqlite is trying to do.
If the fs tree scanned once per second had 1000 files, it would be once per millisecond for a file.
> one lightweight SELECT per millisecond
For the low, low cost of $1 per minute, you can also lease a supercar.
> Once real work flows through a SQLite-backed app, you need a queue. The usual answer is “add Redis + Celery.”
Are they joking? SQLite is usually used for single-process (mutliple threads) applications. The proper way to communicate between threads/processes is a ring buffer, where you allocate structs (allocation typically is incrementing a pointer), and futex/eventfd for notifications (+ some spinlocking to avoid going to kernel when the tasks arrive quickly). Why do you need redis for that? If you need persistent tasks, then you can store them in the table, and still use futex for notifications. This polling is inefficient and they should not make it a library which will cause other lazy developers add it to their app.
> honker polls SQLite’s PRAGMA data_version every millisecond. That’s a monotonic counter SQLite increments on every commit from any connection, journal mode, or process — a ~3 µs read for a precise wake signal
That's 3 ms per second = 0.3% CPU time wasted for every waiting thread.
Like Electron, this feels like written by a web developer and not a real programmer.
>That's 3 ms per second = 0.3% CPU time wasted for every waiting thread.
I suspect that's actually "per process, per database (usually 1)", and not based on number of threads or tables. `data_version` semantics mean there's no need for more than one connection polling it, and it's being used as a relatively lightweight "DB has changed, check queues" check (that's pretty much its whole purpose).
Also I believe this is mostly intended for multi-process use, e.g. out-of-process workers, so an in-process dirty tracker (e.g. just check after insert/update/delete) isn't sufficient.
So I do think it's somewhat crazy, but it is at least very simple. fsnotify-like monitoring seems like a fairly obvious improvement tho, not sure why that isn't part of it. Maybe it's slower? I haven't tried to do anything actually-performant-or-reliable with fs notifications, dunno what dragons lie in wait.
Nevertheless, expect articles like "We replaced our redis cluster with this simple extension and got it N times faster".
Prior discussion a few days ago: https://news.ycombinator.com/item?id=47874647
Author here - previously posted here: https://news.ycombinator.com/item?id=47874647
Key difference vs SQL polling is that we’re touching metadata instead of data pages. I have work in process to make this work without any polling (innotify, kqueue, mmap’d shm file check) after the original stat(2) direction proved unreliable if lightweight.
Would love your feedback and or contributions in the repo - still figuring out the end shape.
It’s an interesting approach and can be quite fun to use for new projects.
> How it works: honker polls SQLite’s PRAGMA data_version every millisecond. That’s a monotonic counter SQLite increments on every commit from any connection, journal mode, or process — a ~3 µs read for a precise wake signal.
At the end it says: "pg-boss and Oban are the Postgres-side gold standards" -- but Oban supports SQLite now too https://github.com/oban-bg/oban
There's also Graphile Worker. https://github.com/graphile/worker
Reminds me of Litestack for Rails. Eventually, it was abandoned because Rails itself started going all out on SQLite.
https://github.com/oldmoe/litestack
All in*
I've implemented something similar in the past, but using inotify. You need to watch the -wal file for IN_MODIFY. To make it work reliably I found I had to run:
Otherwise the new changes weren't guaranteed to be visible to the process. I'm sure there's a more targetted approach that would work instead - maybe flock on a particular byte in the `-shm` file.I'm a big fan of SQLite and all that, but if SQLite constrains you to a single writer process, why not do this in your application layer anyway?
Why not just use https://github.com/conductor-oss/python-sdk provide durability, distributed and orchestration.
On edge this misses Durable Objects + alarms — same primitives, no polling, no Redis to skip in the first place.
Almost feels like someone is trying to joke about similar postgres application .
To make it look even more absurd . SQLite is not concurrent and you’ll have tons of problems using it practically .
Can this work with lightstream?
This seems especially appealing in the awkward middle: too serious for in-memory queues, not big enough to justify Kafka-shaped machinery.
No maven package for java? Guess this isn't a serious project
Suggestion for the author wind back the polling to once a second when nothing is happening.
I can’t see any benchmarks or performance stats.
I’d like to see messages per second.
Could this work with Turso, the SQLite rust rewrite?
Author here. Yeah doesn’t depend on the underlying db if it speaks SQLite.
I think this is interesting too sqlite a as the coordination boundary: business state, queue state, stream offsets, retries, and acks all sharing one transactional substrate. The 1ms polling is getting a lot of weight in the thread though :)