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.
"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".
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.
> 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.
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.
SQLite allows multiple writers. The constraint is that only one of how writers can be actively writing at any moment in time. If there are multiple processes wanting to write, they take turns. SQLite prevents two or more writes from running concurrently, so there is nothing the application needs to do to implement this, other than responding to SQLITE_BUSY replies from failed (concurrent) write attempts and retrying after a short delay.
Why this constraint? Because SQLite is serverless. There is no central server available to coordinate concurrent writes.
At the lowest level of the stack, every database engine has this same constraint, as there is only one wire connecting the CPU to the SSD, and you cannot send multiple writes over the same wire at the same time. But in a client/server database, the server (in cooperation with the filesystem) is at hand to serialize the writes and prevent problems in ways that are not possible without a server. The server creates the illusion of concurrent writes by multiplexing the single write wire efficiently and making that multiplexing transparent to the application.
26 comments
[ 4.5 ms ] story [ 48.5 ms ] thread> 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 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".
…but some other push-based IPC mechanism would be a lot more battery friendly
https://github.com/oldmoe/litestack
To make it look even more absurd . SQLite is not concurrent and you’ll have tons of problems using it practically .
I’d like to see messages per second.
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.
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.
Why this constraint? Because SQLite is serverless. There is no central server available to coordinate concurrent writes.
At the lowest level of the stack, every database engine has this same constraint, as there is only one wire connecting the CPU to the SSD, and you cannot send multiple writes over the same wire at the same time. But in a client/server database, the server (in cooperation with the filesystem) is at hand to serialize the writes and prevent problems in ways that are not possible without a server. The server creates the illusion of concurrent writes by multiplexing the single write wire efficiently and making that multiplexing transparent to the application.