Stop Conda From Ruining Your Life

Okay, so conda tried to install jupyter and I watched it spin for 40 minutes. Forty. Minutes. I made coffee, did laundry, called my mom. Still "solving environment."

The old conda solver was basically trying to solve a math problem every time you wanted numpy. Your innocent conda install numpy would download like 200MB of package metadata, build some insane dependency graph, then try every possible combination until it either found a solution or I found a new career.

Here's what was happening under the hood (and why I wanted to throw my laptop):

1. Download a metric fuckton of metadata - seriously, 100+ MB just for package lists
2. Build dependency graph that probably has more connections than LinkedIn
3. Try to solve it with PycoSAT which is... not fast
4. When it fails (which it does), start completely over

That jupyter install? I timed it. 43 minutes and 12 seconds. I could've learned Go and built my own package manager in that time.

Then libmamba came along and unfucked everything

So conda 23.10.0 happened and they basically said "you know what, our solver is trash" and switched to libmamba. Turns out using the same libsolv that actual Linux distros use was a good idea. Who knew?

What changed:

• Solver written in C++ instead of Python (revolutionary concept, apparently)
• Proper caching so it doesn't re-download the internet every time
• Parallel processing because we don't live in 1995 anymore
• Actually remembers what it figured out yesterday

My jupyter install time? Went from 43 minutes to like 90 seconds. Still not as fast as pip, but I stopped considering alternative careers.

How fast is it now? (Spoiler: way better)

I tested this on my shitty 2020 MacBook Pro with 16GB RAM. Your experience will vary, but these numbers give you an idea:

Installing a basic data science stack (pandas, numpy, scikit-learn, jupyter):

• Old conda: 18 minutes (I timed it with a stopwatch because I was losing my mind)
• New conda with libmamba: About 3 minutes
• Raw mamba: Maybe 90 seconds

Trying to update a 6-month-old environment:

• Old conda: Usually just failed after 45 minutes
• New conda: Takes about 4 minutes and actually works
• Mamba: Fast enough I don't have existential crises

The conda survey said 60% of people bailed to mamba because of speed issues. I was one of them until the new solver came out.

Bottom line: conda used to be 6x slower than pip and now it's only like 2x slower. Progress!

Okay, enough complaining. Here's what worked for me after conda ruined too many of my evenings.

1. First check if you're using the shitty old solver

conda config --show solver

If it says "classic" you're still using the solver from like 2019. That's your problem right there.

Modern conda (23.10+) should use libmamba by default, but corporate IT loves keeping old versions around. If you're stuck with an ancient conda:

conda install -n base conda-libmamba-solver
conda config --set solver libmamba

This literally cut my install times from 20 minutes to about 3 minutes. Not amazing, but I stopped googling "how to uninstall Python entirely" during conda runs.

2. Stop using the default channels

The default conda channels are garbage. Old packages, slow servers, just bad. Switch to conda-forge where people actually maintain stuff:

## See what you're currently using
conda config --show channels

## Nuke defaults, use conda-forge
conda config --remove channels defaults
conda config --add channels conda-forge
conda config --set channel_priority strict

That strict setting matters - it grabs the first package it finds instead of comparing every possible version across every channel. I watched conda spend 12 minutes deciding between numpy 1.21.0 and 1.21.1. Twelve minutes for a patch version difference. Strict mode just picks one and moves on.

3. Clean conda's cache (it's probably huge)

Conda never throws anything away. Ever. I checked mine last month and it was 12GB of cached packages from 2022. Twelve gigabytes of shit I installed once and forgot about.

## See the damage
conda info

## Delete old downloads
conda clean --packages

## When conda starts acting weird, nuke the index
conda clean --index-cache

## Nuclear option when everything's broken
conda clean --all

That packages folder gets big fast. Clean it out every few months or your SSD will hate you. I've seen people with 40GB+ conda caches on old machines. Forty gigabytes of random packages they used once.

4. Install packages in batches, not one by one

This seems obvious but I always forget. Don't do this:

conda install numpy
conda install pandas  
conda install matplotlib

Do this instead:

conda install numpy pandas matplotlib scipy scikit-learn

Each individual install triggers a new solve. Batching them means one solve for everything. Way faster.

5. Just use mamba instead

Look, if conda is still slow after all this, just install mamba and use that:

conda install mamba -c conda-forge

## Now use mamba for everything
mamba install numpy pandas
mamba create -n my-env python=3.11

It's the same commands but actually fast. I haven't typed conda install in like a year.

6. Or go nuclear with micromamba

Micromamba is 5MB instead of conda's 500MB. Perfect for Docker:

## Install micromamba
curl -L micro.mamba.pm/install.sh | bash

## Use it like conda but way faster
micromamba create -n test python numpy pandas -c conda-forge

My data science environment install went from 8 minutes to like 30 seconds. Not exaggerating.

7. When conda breaks (because it will)

"Solving environment" runs forever:
Conda isn't thinking, it's stuck. I've let this run for 2 hours before. Don't be me.

## Ctrl+C that shit, then:
conda clean --index-cache
conda install whatever-you-wanted

"Package not found" for packages you know exist:

## Force it to look in conda-forge
conda install -c conda-forge your-package
## Or use mamba because it's less dramatic
mamba install your-package

Environment takes 30+ minutes to resolve:
Your environment is fucked. Delete it and start over:

conda env remove -n broken-env
conda create -n new-env python=3.11
## Install stuff in batches this time

I spent 3 hours once debugging why conda couldn't find matplotlib. Turned out the default channel index was corrupted. Switched to conda-forge, fixed in 30 seconds. Three hours of my life gone because conda is stupid about channels.