On Having Something To Prove

Wed Jan 31, 2024Listen to this post

I've been doing a lot more coding and writing than usual lately. I'm not exactly back up at full speed, but I'm moving with a lot more determination than I have in a while. I'm honestly not sure what's changed other than that I have something to move forward with.

The work has mostly been in catwalk this time. Last time I mentioned putting together a web interface for it, and I kinda have. By the time you're reading or listening to this, I'll probably have gone through a number of revisions to make it beautiful. At the moment though? This might be the first chunk of code in a very long time I'm not proud of. There's a lot of half-formed thought stuff kicking around my head about this, including requirements I'm only vaguely aware of that suddenly slam into stark relief when I get on with the object level objective of actually producing a blogcast with my tools. I'll have it smoothed out shortly.

Catwalk Development Notes


So, apparently sqlite3 runs in single-threaded mode by default? I discovered this when I started trying to use it as a state store for my local blogcasting. This definitely isn't an approach that scales. I suspect that it couldn't even handle four concurrent users hitting the same cast, or more than 10 threads on the GPU side. As soon as I did anything even a bit bigger than what I've got going currently, I'd want to switch out to redis or somesuch. However, at the moment, for a multi-user site with a use case of "under 100 people, each working on a different job, using between one and three worker threads", it would be perfectly serviceable to run a multi-threaded SQLite setup.

The default configuration gets in my way here but apparently doesn't need to. Hence, the sqlite adapter for pytrivialsql now checks if the local sqlite lib has been compiled for multi-threaded usage. And, if so, disables the sqlite3 thread-check on connection start.


The front-end is written in reagent. Which, honestly, is a really nice way of organizing front-end code. I haven't repoed it yet because of the earlier noted lack of pride, but keep an eye on this space. The goal is to make it a single-page app that connects to the server but manages a lot of the state and workflow client-side. The most evidence you can see of it right this very second is over in the main module. You can see that there's a new UIHandler in place, that I've added a new jobs interface in the form of the JobHandler and JobsHandler classes, and that there's now an exposed WebSocket server sitting at /v1/jobs/updates. Spoilers.

One thing I will say is that local state in reagent apps is weird. It recommends that you have a single top-level state, but also aggressively re-renders the tree when you modify even a tangentially-related piece of top-level state. Which means that if you're dealing with an appreciable number of elements (I am, thank you) and also want your app to run on anything like a usable clock speed (is that even a question? Yes, absolutely), you have to give individual components intermediate pieces and then aggregate later. Forms are the trickiest bits of this, because implementing them naively means poking at your input state and that triggers the dreaded re-renders.

What I ended up doing was

  1. Have a piece of top-level state that represents the server-side objects in the system. When a new websocket update comes in, this is what gets poked. It also triggers a global re-render, but that's almost the only way to keep what the user sees in synch with changes that worker threads or other users make, so whatever.
  2. Wherever a user needs to interact with something, have a separate, local piece of state that deals with their input. So like, if there's a textarea or checkbox, its default state is taken from the above global state, but local changes are put into a local atom in order to localize re-renders as much as possible.
  3. In the odd case where I need to aggregate local state for form purposes, have a piece of intermediate state that each local component reports into, in addition to its local state. This doesn't need to be updated on every user interaction, only when an update is sent to the server, and it also doesn't need to be represented anywhere in the UI thus eliminating more re-renders.

Possibly there's a simpler way to do this, and I'll keep an eye out for how to accelerate interactions further, but it works Well Enough For Now.

Websocket Channel

catwalk still runs on tornado. Which is weird about messages to clients from separate threads. This is something I absolutely needed to crunch through, because the entire point of the websocket connection in this project is updating the user regarding the activity of the worker threads. So they have to be able to send/receive from separate threads.

In order to resolve that, I actually had to end up subclassing tornado.websocket.WebSocketHandler?

class SocketServer(tornado.websocket.WebSocketHandler):
    CLIENTS = set()
    IOloop = tornado.ioloop.IOLoop.current()

    def __init__(self, *args):
        SocketServer.IOloop = tornado.ioloop.IOLoop.current()

    def open(self):

    def close(self):

    def send_message(cls, message):
        msg = json.dumps(message)
        print(f"UPDATING {len(cls.CLIENTS)} WS CLIENTS...")
        for client in list(cls.CLIENTS):
            except tornado.websocket.WebSocketClosedError:

    def send_job_update(cls, job):
        if job is None:
                "job_id": job["id"],
                "job_type": job["job_type"],
                "status": job["status"],
                "parent": job["parent_job"],
                "input": job["input"],
                "output": job["output"],

As you can see, there's class-level state and a couple class methods involved. It works, in the sense that I've run it and tested out the front end by interacting with it as I pleas for a while. But I haven't found a satisfying explanation for why this limitation exists, so I can't shake the feeling that I'm opening myself up to weird distributed-system-style race conditions. My guess and hope is that this is just an incidental outgrowth of tornado being a non-blocking server, so they accidentally never bothered dealing with threads even though there's nothing explicitly preventing it. The name call_soon_threadsafe is suggestive of a routine that works gracefully under these conditions. Fingers crossed I guess.

I'm going to do a bit more work on the front end, explore a couple other use cases for catwalk, and maybe take another run up the clojurescript-on-android hill. It looks like a couple new options have arisen since last I checked.

As always, I'll let you know how it goes.

Creative Commons License

all articles at langnostic are licensed under a Creative Commons Attribution-ShareAlike 3.0 Unported License

Reprint, rehost and distribute freely (even for profit), but attribute the work and allow your readers the same freedoms. Here's a license widget you can use.

The menu background image is Jewel Wash, taken from Dan Zen's flickr stream and released under a CC-BY license