Oh hey, Weather is failing to launch again, and this console message looks interesting:

VFX/Renderer.swift:1123: Fatal error: Cannot create commandBuffer VFX update command buffer

okay I think the CPU usage thing was a one-off blip disconnected from the obscure graphics problem.

I now have my *own* app misbehaving though, even though I don't use Metal at all. When it's failing, I also get this message (to stderr, not nslog!) coming off some point in the main loop I can't triangulate (it's not coming from my custom draw method) "Context leak detected, msgtracer returned -1"

Closer… that "leaks React components on Safari" bug above was a hint. And this suggests something:

github.com/aws/amazon-chime-sd…

Several of the Safari dock apps I'm using (Unifi, Nest, YouTube, Mastodon) are constantly doing stuff with video components, and it seems that this is a "known" sore spot for WebKit…

Safari leaks memory when videos are rapidly removed from the screen and re-added · Issue #771 · aws/amazon-chime-sdk-component-library-react

What happened and what did you expect to happen? Safari has a problem with the HTMLVideoElement object that keeps leaking memory if there are many video elements that are being removed from the scr...

^GitHub

PLOT TWIST:

after reproing for the 10th or 11th time (my quick test now is "can I launch Weather", which seems to presage all the other problems), I quit Safari and all my "dock apps", and it did *not* fix the problem. I noticed a bunch of `com.apple.WebKit.WebContent` processes still running. Killed them: *still* didn't fix the problem.

Did a quick jaunt over to Activity Monitor to see if I had any big RAM users, and … mediaanalysisd was an outlier again. So I killed it… and that *did* fix it

I think I know the noun I'm looking for now; it is a "GPU command queue", as evidenced by these two log lines:

2025-02-01 20:44:14.589205-0800 kernel: (IOGPUFamily) mediaanalysisd has reached a high water mark of 710 gpu command queues.

2025-02-01 20:44:14.939613-0800 kernel: (IOGPUFamily) static IOReturn IOGPUDeviceUserClient::s_new_command_queue(…): The number of queues (708) exceeding limit (512), failing IOGPUCommandQueue creation

The latter is repeated 1000s of times.

Okay, a much stronger hypothesis after letting fs_usage run all night and collecting repeated patterns. There are several files it keeps scanning over and over, including:

- some pedestrian .jpg files from a personal family IT project I worked on a while ago
- some retrocomputing data, specifically a movie file exported from "Graphing Calculator" in macOS 9.0.4
- a bunch of files (e.g. test/imghdrdata/ , idlelib/Icons) in various PyPy releases

It does not seem to be visiting many other files.

All right. Too early to have much real data but the trajectory at least isn't horrible thus far, and sometimes after a few minutes of idle time it *has* been pretty horrible.

Considering spending my next livestream building my first OBS overlay, and making that overlay be a real-time report of the memory usage and last gpu command queue high water mark.

Okay, I think this is hopefully the conclusion. There are some methodological problems with my tests (more research is needed) but in terms of predictive power, I now have 2 log files, each of which has the relevant IOGPUFamily kernel logs from two separate days. the number of lines in each of these files:

before I made changes: 73929
after I made changes: 9

That feels like a statistically significant difference.

Going to turn off all the logging and watch memory usage for the next 24h now.

But hey I'm a Programmer let's maybe turn this into a teachable moment about Programming.

There's a bug here, obviously. I don't know what it is. I know nothing about mediaanalysisd, I'm only even vaguely aware of what it does. But there are general to learn from this journey, which is applicable to most software development projects.

You can also see this as just a restatement of the other two flaws, which is to say:

1. the dead-letter queue needs to be operator-legible; tell me about the files that are producing errors, give me a way to handle them. give me a good UI for it, make it possible for me to understand where and how it's happening, remember to try the work again when new software is available

good news for the "it's a hardware issue" fans: baldur's gate 3 now reliably indues a kernel panic at launch (failed WindowServer checkin)

(I haven't played bg3 since updating to 15.3 though. also, other video games still work fine.)

now one of my wireless access points appears to be failing. but only like… a little bit, intermittently, maybe.

I _will_ face god and walk backwards into hell, if I'm not there already

Tried playing BG3 on a different mac, with the same account. Worked fine. Same on a Steam Deck, just in case. Did the full hardware test, let the machine stay powered off for a few minutes to allow any ghosts in the capacitors to depart, notwithstanding.

Launched BG3 on the cursed machine. It worked! Hooray! Took a celebratory screenshot, switched spaces to look at it, switched spaces back, and as soon as the space-switch slide animation *began*… full-screen freeze, crash.

OS reinstall it is.