Analytics/Systems/Anaconda

Anaconda is a prepackaged conda distribution for mostly python based analytics and research purposes. WMF maintains a custom debian package of Anaconda that includes some extra packages, but also has scripts for creating 'stacked' conda user environments. These conda user environments allow users to install packages into their own conda environment without modifying the base anaconda environment.

Usage

Listing conda environments

/usr/lib/anaconda-wmf/bin/conda env list
# conda environments:
#
2020-08-19T16.19.37_otto     /home/otto/.conda/envs/2020-08-19T16.19.37_otto
2020-08-19T16.47.40_otto     /home/otto/.conda/envs/2020-08-19T16.47.40_otto
2020-08-19T16.56.54_otto     /home/otto/.conda/envs/2020-08-19T16.56.54_otto
2020-08-19T16.59.40_otto     /home/otto/.conda/envs/2020-08-19T16.59.40_otto
2020-12-13T19.40.09_otto     /home/otto/.conda/envs/2020-12-13T19.40.09_otto
base                  *  /usr/lib/anaconda-wmf

Listing installed packages

To see the packages installed in your current environment, you can run conda list. Note that this will not include the packages in the base anaconda-wmf environment, which are also accessible in the current environment. To see the packages installed in the base environement, run conda list -n base.

Anaconda base environment

To use the readonly Anaconda base environment, you can simply run python or other executables directly out of /usr/lib/anaconda-wmf/bin. If you prefer to activate the anaconda base environment, run source /usr/lib/anaconda-wmf/bin/activate.

Creating a new conda user environment

Run

 conda-create-stacked

and a new conda environment will be created for you in ~/.conda/envs. When used, this environment will automatically append the base conda environment Python load paths to its own. If the same package is installed in both environments, your user conda environment's package will take precedence.

If you prefer, you can name your conda environment

 conda-create-stacked my-cool-env

Activating a conda user environment

There are several ways to activate a conda user environment. Just running

 source conda-activate-stacked

On its own will attempt to guess at the most recent conda environment to activate. If you only have one conda environment, this will work.

You can also specify the name of the conda env to activate. Run /usr/lib/anaconda-wmf/bin/conda info --envs to get a list of available conda environments. E.g.

 source conda-activate-stacked otto_2020-08-17T20.52.02

Or, you can run the 'activate' script out if your conda environment path:

 source ~/.conda/envs/2020-08-17T20.52.02_otto/bin/activate

Installing packages into your user conda environment

After activating your user conda environment, you can set http proxy env vars and install conda and pip packages. E.g.

export http_proxy=http://webproxy.eqiad.wmnet:8080
export https_proxy=http://webproxy.eqiad.wmnet:8080
conda install -c conda-forge <desired_conda_package>
pip install --ignore-installed <desired_pip_package>

Conda is much preferred over pip, if the package you need is available via Conda. Conda can better track packages and their install locations than pip.

Note the --ignore-installed flag for pip install. This is only needed if you are installing a pip package into your Conda environment that already exists in the base anaconda-wmf environment.

These packages will be installed into the currently activated Conda user environment.

Installing packages that need compilation

There's an open issue (phab: https://phabricator.wikimedia.org/T292699) but until that's resolved we need to install the required libraries and kind of patch them through to Conda:

For example, when something needs sasl headers (like pyhive), we do:

conda install -c conda-forge cyrus-sasl
export CPPFLAGS="${CPPFLAGS} -isystem ${CONDA_PREFIX}/include"

Before we can do:

pip install pyhive[hive]

Deactivating your user conda environment

 source conda-deactivate-stacked

Or, since the user conda env's bin dir has been added to your path, you should also be able to just run

 source deactivate

stacked conda environments

Conda supports activating environments 'stacked' on another one. However, all this 'stacking' does by default is leave the base conda environment's bin directory on your PATH. It does not allow for python dependencies from multiple environments.

Our customization fixes this. When conda-create-stacked is run, an anaconda.pth file is created in the new conda environment's site-packages directory. This file tells Python to add the anaconda-wmf base environemnt python search paths to its own. If a package is present in both environments, the stacked conda environment's version will take precedence.

For more details on why upstream Conda has not implemented this behavior, see this GitHub issue.

R support

WMF's anaconda environment support was built with Python in mind. Other languages are passively supported.

R is included in the base anaconda-wmf environment, but it is not installed into the user conda environment by default. Doing so makes the size of user environments much larger, and makes distributing them to HDFS take much longer.

To install R packages into your user environment, do the following:

# Make sure you are using a conda env. This is not necessary if running in Jupyter.
source conda-activate-stacked
# Enable http proxy.  This is not necessary if running in Jupyter
export http_proxy=http://webproxy.eqiad.wmnet:8080; export https_proxy=http://webproxy.eqiad.wmnet:8080; export no_proxy=127.0.0.1,localhost,.wmnet

# R is currently the base anaconda-wmf R.
which R
/usr/lib/anaconda-wmf/bin/R

# Install the conda R package into your user conda environment.
conda install R

# R is now fully contained in your user conda environment.
which R
/home/otto/.conda/envs/2021-04-07T21.37.00_otto/bin/R

You should now be able to install R packages using R's package manager via install.packages().

However, just like with Python, installing R packages with conda is preferred over using R's package manager. If a conda R package exists, you should be able to just install it like:

$ conda install r-tidyverse

It is also recommended to create a ~/.Rprofile file with the following:

Sys.setenv("http_proxy" = "http://webproxy.eqiad.wmnet:8080")
Sys.setenv("https_proxy" = "http://webproxy.eqiad.wmnet:8080")
options(
  repos = c(
    CRAN = "https://cran.rstudio.com/",
    STAN = "https://mc-stan.org/r-packages/"
  ),
  mc.cores = 4
)
Sys.setenv(MAKEFLAGS = "-j4")
Sys.setenv(DOWNLOAD_STATIC_LIBV8 = 1)

Troubleshooting

/bin/gtar not found

If you attempt to install from a Git repository – e.g. wmfdata via remotes::install_github("wikimedia/wmfdata-r") and get the following:

Downloading GitHub repo wikimedia/wmfdata-r@HEAD
sh: 1: /bin/gtar: not found
sh: 1: /bin/gtar: not found
Error: Failed to install 'wmfdata' from GitHub:
  error in running command
In addition: Warning messages:
1: In system(cmd) : error in running command
2: In utils::untar(tarfile, ...)

For some reason this is an issue with Conda's R. The only workaround is running Sys.setenv(TAR = system("which tar", intern = TRUE)) before the install commands.

Installing lme4

lme4 depends on nloptr package, which is very difficult to build from source (which is what happens if you try to install directly in R) and is the primary hurdle for installing lme4 on our systems. The easiest way to install it is with:

conda install r-nloptr

then you can install.packages("lme4") – to verify that it works:

library(lme4)

fit <- lmer(Reaction ~ Days + (Days | Subject), data = sleepstudy)

Installing brms

First, install pkg-config with:

conda install pkg-config

This is necessary for installing some of brms's dependencies. Then you can install in R:

install.packages("brms")

To verify that it works run the following:

library(brms)

prior1 <- prior(normal(0,10), class = b) +
  prior(cauchy(0,2), class = sd)
fit1 <- brm(count ~ zAge + zBase * Trt + (1|patient),
            data = epilepsy, family = poisson(), prior = prior1)

Administration

The code used to build new releases of anaconda-wmf lives in operations/debs/anaconda-wmf.