It's been six months since I asked whether #uv is the future of #Python packaging: https://youtu.be/_FdjW47Au30With uv 0.3.0, the answer is IN and I'll tell ...
We do geodata science and rely on some pretty specific C++ libraries that are only distributed via conda. While on unix-based systems it’s possible to get some of them from other channels or even building them from source, we mostly have Windows machines in production where we are not that flexible. Docker is unfortunately no solution due to security concerns.
If you are asking why I hate it: It’s bloated, uses more space than needed and it’s rare I can reproduce an environment from the environment file without running into errors. Using it feels unintuitive, I still google command after years. It was very slow until recently, when the libmamba solver was finally integrated. Last but not least licensing is a pain in the ass.
I share the same frustration trying to replicate an environment. I’m glad I can avoid it these days, the community needs a way out of the conda lock-in.
Interesting. We use conda via micromamba for my own project, as it makes the install for end-users much easier when they can just run a shell script, to install python, cuda, and all the dependencies needed.
I’m no expert, but isn’t running in a VM strictly better than running on raw metal from a security perspective? It’s generally more locked down, and breaking out of the virtualization layer requires a separate security breach from gaining access to the running container.
We do geodata science and rely on some pretty specific C++ libraries that are only distributed via conda. While on unix-based systems it’s possible to get some of them from other channels or even building them from source, we mostly have Windows machines in production where we are not that flexible. Docker is unfortunately no solution due to security concerns.
If you are asking why I hate it: It’s bloated, uses more space than needed and it’s rare I can reproduce an environment from the environment file without running into errors. Using it feels unintuitive, I still google command after years. It was very slow until recently, when the libmamba solver was finally integrated. Last but not least licensing is a pain in the ass.
I share the same frustration trying to replicate an environment. I’m glad I can avoid it these days, the community needs a way out of the conda lock-in.
Interesting. We use conda via micromamba for my own project, as it makes the install for end-users much easier when they can just run a shell script, to install python, cuda, and all the dependencies needed.
I’ve been using micromamba/mamba and not had solving issues like I did with conda. Im glad conda integrated libmamba.
Question: why were docker containers deemed security risks?
If Windows, it requires a VM and currently infosec is not keen on virtualization in the hands of users.
I’m no expert, but isn’t running in a VM strictly better than running on raw metal from a security perspective? It’s generally more locked down, and breaking out of the virtualization layer requires a separate security breach from gaining access to the running container.
I would think so as well. Possibly it’s because a local VM is harder for them to monitor.
Yes, mamba is a huge improvement. Regarding docker I can’t really tell you as I’m not an infrastructure guy.