I have been using Conda (via Miniconda) for managing my Python development environments and packages for close to a year now, so I thought I’d write up my thoughts so far for others.

Conda is both an environment manager (an alternative to virtualenv) and an installation tool (an alternative to pip). You can also use Conda to build your packages and distribute them via Binstar.

So, what does Conda do well and what needs improvement?

Version

As of this writing I am using Conda version 3.7.1.

The Good

• Conda installs binary packages. There’s no compiling extensions, so even installing SciPy goes fast. This makes installing SciPy on Windows a breeze.
• Conda caches packages locally for reuse. If you make a new environment Conda won’t have to re-download all those packages.
• Conda environments just work. New environments are directories on your computer. Done with one? Delete the directory.
• Conda and Binstar are a great way to distribute packages. If you need to distribute packages that have compiled extensions or rely on hard-to-build packages, you can pre-build binaries and upload them to Binstar for others to use. (Assuming you have access to the same OS your users will use. I’ve heard Continuum is working on a build service that may help with this.)
• If you’re running continuous tests on Travis-CI Conda can help drastically speed up your build times. See the documentation at http://conda.pydata.org/docs/travis.html. I also use Conda for provisioning web servers.

The Annoying

• Conda environments are very bash centric. (Edit: I mean here that the provided activation/deactivation tools for Conda environments are bash/zsh specific. See Aaron’s comment below.) If you use a different shell (like fish) you may need to hack together your own Conda environment tools.
• Conda packages are sometimes behind the latest releases. Management of Conda packages seems to be pretty manual. It often takes a long time for them to be updated after new releases, especially for the less commonly used libraries.
• Conda doesn’t quite replace pip. There aren’t Conda packages for every Python package, and realistically there can’t be. When you need a package that’s not on Conda it’s fine to use pip. But this will happen in literally every environment you make, so it’s kind of silly that pip isn’t installed by default. (You can make it so pip is installed by default.)

The Bad

I note above that Conda can’t install every package and you sometimes need to fall back to pip. Well, it turns out that it is impossible to install some packages into Conda environments via pip (or even python setup.py install). The problem arises from Conda’s treatment of C libraries. When you install a Python library that links to a C library installed by Conda, that Python library will not be able to find the C library at runtime. See lots of discussion on these GitHub issues: https://github.com/conda/conda-recipes/issues/110 and https://github.com/ContinuumIO/anaconda-issues/issues/177. (These problems seem to be most egregious on Mac OS X.)

The Python packages that are known to fail all work with SSL. These include cryptography and psycopg. (Continuum has recently added a cryptography package to Conda to help alleviate this problem, but not psycopg.)

The workarounds for this problem include setting sketchy variables like DYLD_LIBRARY_PATH or building the packages yourself via Conda, which is a multi-step process. Neither is beginner friendly.

Summary

I love and will continue to use Conda, and I will continue to recommend Anaconda to people who are interested in getting started with scientific Python. But there remains a gap between Conda and Python’s standard build systems that is at best inconvenient and at worst impassable (unless you possess the experience required to navigate the workarounds). People who are expecting to use psycopg (and likely some other libraries) should be aware that Conda is not going to make their lives easier.

I think it would be helpful for users if Conda could seamlessly fall back on PyPI packages when Conda doesn’t have something pre-built. Need psycopg2? Maybe a conda install --pypi psycopg2 could download psycopg2 from PyPI and build it appropriately so that it will work seamlessly with Conda environments.

Given the state of Conda as a relatively new tool I’m optimistic that the developers will be able to improve it for everyone given more time. I use Conda literally almost every day and I thank the developers for it.