Log
-
Cleanup old data with a pattern based on latest file 💬 by Caio 3 months ago
Ensemble dataset has (obviously) a different name. Filled up the disk, couldn't download newer data. First outage 🎉🎉
-
info: expose current head as version by Caio 3 months ago
-
Remove leftover file 💬 by Caio 3 months ago
Got copy-renamed to moros.rs and I forgot to git-rm
-
Record lastTick after content swap 💬 by Caio 3 months ago
Just to that I updates correctly when I switch tabs like a madman
-
No need to `use ndarray::` here by Caio 3 months ago
-
Add chuva/bin/check to verify loading 💬 by Caio 3 months ago
Load a given dataset using code with manual offsets and compare vs using ndarray.
-
Fix ensemble dataset loading 💬 by Caio 3 months ago
It's row-major, not column-major (i.e. the LAST dimension is contiguous, not the first). RIP muh cpu cache
-
Add a little helper to diff datasets by Caio 3 months ago
-
Make download-dataset a bit smarter 💬 by Caio 3 months ago
Take dataset from args. Figures out which PID to terminate through systemd so I can make multiple instances of the same binary without worry Now https://beta.chuva.caio.co is running with the ensemble model
-
No need for load feature anymore 💬 by Caio 3 months ago
Fewer moving pieces. Finally.
-
Make postcode-fst crate a moros example by Caio 3 months ago
-
Rename chuva::Model to chuva::Chuva 💬 by Caio 3 months ago
Consumed my persim berry
-
Moros is the web server and associated features 💬 by Caio 3 months ago
I've confused myself with moros <-> chuva way too many times already. Better tame it.
-
Rename dataset crate to chuva by Caio 3 months ago
-
Move dataset-related code from moros to dataset crate 💬 by Caio 3 months ago
Makes it easier for me to play with distinct datasets outside of moros. Makes moros not care at all about the model format.
-
Sort in-place instead of copying to a side buffer by Caio 3 months ago
-
Play with the ensemble dataset a bit 💬 by Caio 3 months ago
As suspected, the load time sucks :) ~10x slower in my rig for a day without a lot of predicted rain (timing is dominated by DEFLATE, so more zeroes in the data -> faster decompression -> faster load) Should be a bit worse in prod, which means a load time of ~3s without any extra load. Too high for my taste but maybe I can get away with it. We'll see...
-
No need for musl anymore 💬 by Caio 3 months ago
Solved by upgrading my server so now libc versions match :P
-
Replace hdf5-metno with netcdf 💬 by Caio 3 months ago
Much better state dep tree wise, but it's still essentially the same code since netcdf pulls hdf5-metno-sys It does require a c++ compiler tho, so I'm sure they'll be at least a little stumble when building with musl
-
More precise lat/lon to x/y 💬 by Caio 3 months ago
Smh I ignored the sign on geo_pixel_size_y. That answer my "why do I have to abs()" question lol
-
/info doesn't need TimeZone by Caio 3 months ago
-
Add some status details to /info by Caio 3 months ago
-
Implement case insensitive postcode match by Caio 3 months ago
-
Fix missing svg styles for /app by Caio 4 months ago
-
Use f32 instead of f64 for predictions 💬 by Caio 4 months ago
There's no real reason for f64: the source is a u16 with a small scale factor. I thoughtlessly went for the wider type because geo coords, projector and dataset metadata all use higher precision So: lat/lon input is still f64 since it's relevant for projecting into the grid, but the in-memory grid is f32-based, which cuts the runtime ram cost in half (from ~100MB to 50) with no loss
-
Make /app behave more like an app 💬 by Caio 4 months ago
Auto pull coordinates, periodic refresh, doesn't leave /app
-
Split /app from index 💬 by Caio 4 months ago
Trying to share here will get more in the way than help
-
Less error prone logo by Caio 4 months ago
-
Make chuva.caio.co/app a "PWA" by Caio 4 months ago