conda-libmamba-solver

The fast mamba solver, now in conda!

What is this exactly?

conda-libmamba-solver is a new (experimental) solver for the conda package manager which uses the solver from the mamba project behind the scenes, while carefully implementing conda's functionality and expected behaviors on top. The library used by mamba to do the heavy-lifting is called libsolv.

Trying it out

The new libmamba integrations are experimental, but you can get a taste of how they are working so far by following these instructions.

Before we start: to use the libmamba integrations you need to update the conda installation in your base environment to a canary release. This is potentially a destructive action, so make sure you are not testing this in a production environment. We recommend using a VM, a Docker image or something similar.

1. Install the canary builds of both conda and conda-libmamba-solver in your base environment. The following channels are required for now:

First activate the base environment:

conda activate base

Then install the canary versions of conda and conda-libmamba-solver:

conda install \
    --override-channels \
    --yes \
    -c conda-canary/label/dev \
    -c conda-canary/label/conda-libmamba-solver-dev \
    -c conda-forge \
    -c defaults \
    conda \
    conda-libmamba-solver

(Optional) One-liner version (same as above):

conda install --override-channels --yes -c conda-canary/label/dev -c conda-canary/label/conda-libmamba-solver-dev -c conda-forge -c defaults conda conda-libmamba-solver

2. Now you can experiment with different things. --dry-run is specially useful to check how different solvers interact. The main switch you need to take care of is the experimental solver option:

# Using default (classic) solver
$ conda create -n demo scipy --dry-run
# This is equivalent
$ conda create -n demo scipy --dry-run --experimental-solver=classic
# Using libmamba integrations
$ conda create -n demo scipy --dry-run --experimental-solver=libmamba
# Using old proof-of-concept, debugging-only libmamba integrations
$ conda create -n demo scipy --dry-run --experimental-solver=libmamba-draft

Hint: You can also enable the experimental solver with the CONDA_EXPERIMENTAL_SOLVER environment variable: CONDA_EXPERIMENTAL_SOLVER=libmamba conda install ...

3. Use time to measure how different solvers perform. Take into account that repodata retrieval is cached across attempts, so only consider timings after warming that up:

# Warm up the repodata cache
$ conda create -n demo scipy --dry-run
# Timings for original solver
$ time conda create -n demo scipy --dry-run --experimental-solver=classic
# Timings for libmamba integrations
$ time conda create -n demo scipy --dry-run --experimental-solver=libmamba

conda create commands will have similar performance because it's a very simple action! However, things change once you factor in existing environments. Simple commands like conda install scipy show ~2x speedups already.

4. If you need extra details on why solvers are working in that way, increase verbosity. Output might get too long for your terminal buffer, so consider using a pager like less:

# Verbosity can be expressed with 1, 2 or 3 `v`
$ conda create -n demo scipy --dry-run -vvv --experimental-solver=libmamba  2>&1 | less

FAQ

How do I uninstall it?

If you don't want to use the experimental solver anymore, you can uninstall it with:

$ conda remove conda-libmamba-solver

To revert conda to the stable version:

$ conda install -c defaults conda

Why can't I use the experimental solver on the base environment?

This decision has been made to protect your base installation from unexpected changes. This package is still in a experimental phase and, as a result, you can only use it in non-base environments for now.

How do I configure conda to use the experimental solver permanently?

Use the following command:

$ conda config --set experimental_solver libmamba --env

Note that we are using the --env flag so the setting is only applied to the active environment. Otherwise it will have a global effect on all your environments, including base, which is now protected. As such, we strongly recommend to enable this setting in a case by case basis or, even better, on a command by command basis by setting the corresponding command line flags or environment variables when needed.