Closed wpfl-dbt closed 3 months ago
have you pruned everything https://github.com/i-dot-ai/redbox-copilot?tab=readme-ov-file#error-elasticsearch-137 8gb should be sufficient to run all the services
Also, here are a lot of services, do you need to run all of them?
Done. Had 64Gb free, and 40 mins later it's eaten it all. Looking at the build logs embedder, ingester, core-api, streamlit-app and redbox have some very sus poetry install processes going on, with nvidia libs being installed over and over.
On services -- you tell me what I need! The readme says to use make run
which uses most of the below, but I don't see an obvious entrypoint for redbox itself. localhost:8090 has a Django app with no clear way into anything else.
Here's the build logs of the offending services:
thx for the detailed logs and fair point re: telling you what services you need! the core services (at the moment) are:
and you will need some chat interface , which right now, is best served by streamlit-app
The above builds in CI on standard GH runners. I am afk at the moment but will be able to dig into your issue properly tomorrow.
Hero, ta 😄 . I'll run those four and see how I go.
I did another prune then ran docker compose up core-api embedder ingester streamlit-app
which in turn brings up redis, elasticsearch and minio, which you'd expect. This at least works, though the docker engine's total reported disk usage (containers plus overheads) is 45Gb!
This is a separate issue, but steamlit-app
fails as it can't import model_db
. Will raise in another thread.
Something that might be useful is to use profiles to make the compose a bit more explicit. Get to something like docker compose --profile backend --profile frontend_streamlit up
. Again, happy to raise a feature request if this sounds like a good idea.
yep. this is an excellent idea, please raise away!
hello, same issue here - temporarily resolved by expanding disk space available to Docker. In case it's helpful here's the amount of space being used by the various images.
On desktop, also running MacOS 14.4.1 (23E224)
This is not the answer, but it is a step in the right direction https://github.com/i-dot-ai/redbox-copilot/pull/200
I just pruned and ran docker compose build --no-cache core-api embedded ingester streamlit-app
this morning, and just before the images exported Docker was reporting a full 60Gb or so of disk usage which went down to 40Gb after the export.
Some insights on this ticket after working in streamlit for a while:
poetry show --tree
-- sentence-transformers
has a dependency on PyTorch, and even though we're developing on CPU, it's installing 2-3Gb of Nvidia GPU dependencies. I switched the streamlit repo to CPU inference, but you have an issue that many of us are developing on Mac, but it runs on Debian, and PyTorch has different install processes for both which poetry doesn't like. I sorted by switching to dev containers, but it's a bit vendor lockin-y
sys-platform
markers are supported by poetry so maybe there's a non-dev-container route to this too
Core commands for running this,
docker compose up
andmake run
, will very quickly run into "no space left on device". I appreciate I can (and have) pruned, but docker has a pretty chunky 64Gb available to it. I suspect the memory footprint of this is much larger than you expect?Desktop: