AWS EMR bootstrap script fails

[mrtj]

What happened:
I tried to instantiate an EMR cluster with the AWS EMR bootstrap script. The provisioning of the cluster failed saying "bootstrap action 1 returned a non-zero return code". Extracting the logs, the stdout of the master arrives to call the conda pack command:

...
Installing base packages
Packaging environment
<EOF>

while the stderr shows this:

...                                                                                 
CondaPackError: 
Files managed by conda were found to have been deleted/overwritten in the
following packages:

- python='3.8.5'

This is usually due to `pip` uninstalling or clobbering conda managed files,
resulting in an inconsistent environment. Please check your environment for
conda/pip conflicts using `conda list`, and fix the environment by ensuring
only one version of each package is installed (conda preferred).

What you expected to happen:
The EMR cluster provisions correctly.

Minimal Complete Verifiable Example:
I first uploaded the bootstrap script to s3://my-bucket/dask-emr/bootstrap.sh than tried to provision the EMR cluster with this AWS CLI command:

aws emr create-cluster --name "Janos Dask Trial Cluster" \
    --release-label emr-5.30.1 \
    --applications Name=HADOOP \
    --log-uri s3://my-bucket/dask-emr/logs/ \
    --ec2-attributes KeyName=my-key \
    --instance-type m5.xlarge \
    --instance-count 3 \
    --bootstrap-actions Path=s3://my-bucket/dask-emr/bootstrap.sh,Args="[--conda-packages,bokeh,fastparquet,python-snappy,snappy]" \
    --use-default-roles

[eldadcohen1]

Hi,
There is any update about this issue.
I'm facing with the same problem.

Some good references:
https://aws.amazon.com/premiumsupport/knowledge-center/restart-service-emr/

Also use the following miniconda:
https://repo.anaconda.com/miniconda/Miniconda3-py37_4.9.2-Linux-x86_64.sh instead of the latest

[davegravy]

Reproduced this issue with EMR version 6.2.0.

[adonig]

I had to use EMR version 5.29.0 with changes to the boostrap script to get around that issue. Also I removed the dask-yarn>=0.7.0 version specification, because it just creates a file called =0.7.0 and the automatically installed version is more current anyway. I'm still running into issues with native libraries, i.e. the pyarrow undefined symbol issue and some problems with seaborn and how it resolves matplotlib.

I'm starting to wonder whether dask-yarn is currently in a state that allows it to run on AWS EMR or whether dependency hell made it ususable. Maybe it is better to setup a Dask Distributed cluster on Amazon AWS using Docker and Elastic Container Service, like it is done here.

[gabriel131188]

Hello, i used for EMR 5.29 this bootstrap:

#!/bin/bash
HELP="Usage: bootstrap-dask [OPTIONS]
Example AWS EMR Bootstrap Action to install and configure Dask and Jupyter
By default it does the following things:
- Installs miniconda
- Installs dask, distributed, dask-yarn, pyarrow, and s3fs. This list can be
  extended using the --conda-packages flag below.
- Packages this environment for distribution to the workers.
- Installs and starts a jupyter notebook server running on port 8888. This can
  be disabled with the --no-jupyter flag below.
Options:
    --jupyter / --no-jupyter    Whether to also install and start a Jupyter
                                Notebook Server. Default is True.
    --password, -pw             Set the password for the Jupyter Notebook
                                Server. Default is 'dask-user'.
    --conda-packages            Extra packages to install from conda.
"

set -e

# Parse Inputs. This is specific to this script, and can be ignored
# -----------------------------------------------------------------
JUPYTER_PASSWORD="dask-user"
EXTRA_CONDA_PACKAGES=""
JUPYTER="true"

while [[ $# -gt 0 ]]; do
    case $1 in
        -h|--help)
            echo "$HELP"
            exit 0
            ;;
        --no-jupyter)
            JUPYTER="false"
            shift
            ;;
        --jupyter)
            JUPYTER="true"
            shift
            ;;
        -pw|--password)
            JUPYTER_PASSWORD="$2"
            shift
            shift
            ;;
        --conda-packages)
            shift
            PACKAGES=()
            while [[ $# -gt 0 ]]; do
                case $1 in
                    -*)
                        break
                        ;;
                    *)
                        PACKAGES+=($1)
                        shift
                        ;;
                esac
            done
            EXTRA_CONDA_PACKAGES="${PACKAGES[@]}"
            ;;
        *)
            echo "error: unrecognized argument: $1"
            exit 2
            ;;
    esac
done


# -----------------------------------------------------------------------------
# 1. Check if running on the master node. If not, there's nothing do.
# -----------------------------------------------------------------------------
grep -q '"isMaster": true' /mnt/var/lib/info/instance.json \
|| { echo "Not running on master node, nothing to do" && exit 0; }


# -----------------------------------------------------------------------------
# 2. Install Miniconda
# -----------------------------------------------------------------------------
echo "Installing Miniconda"
curl https://repo.anaconda.com/miniconda/Miniconda3-py38_4.10.3-Linux-x86_64.sh -o /tmp/miniconda.sh
#curl https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh -o /tmp/miniconda.sh
bash /tmp/miniconda.sh -b -p $HOME/miniconda
rm /tmp/miniconda.sh
echo -e '\nexport PATH=$HOME/miniconda/bin:$PATH' >> $HOME/.bashrc
source $HOME/.bashrc
conda update conda -y


# -----------------------------------------------------------------------------
# 3. Install packages to use in packaged environment
#
# We install a few packages by default, and allow users to extend this list
# with a CLI flag:
#
# - dask-yarn >= 0.7.0, for deploying Dask on YARN.
# - pyarrow for working with hdfs, parquet, ORC, etc...
# - s3fs for access to s3
# - conda-pack for packaging the environment for distribution
# - ensure tornado 5, since tornado 6 doesn't work with jupyter-server-proxy
# -----------------------------------------------------------------------------
echo "Installing base packages"
conda install \
-c conda-forge \
-y \
-q \
dask==2021.8.0 \
dask-yarn==0.9 \
distributed==2021.8.0 \
tornado==6.1 \
pyarrow==5.0.0 \
s3fs==2021.7.0 \
conda-pack==0.6.0 \
bokeh==2.3.3 \
dask[dataframe]==2021.8.0 \
$EXTRA_CONDA_PACKAGES


# -----------------------------------------------------------------------------
# 4. Package the environment to be distributed to worker nodes
# -----------------------------------------------------------------------------
echo "Packaging environment"
conda pack -q -o $HOME/environment.tar.gz


# -----------------------------------------------------------------------------
# 5. List all packages in the worker environment
# -----------------------------------------------------------------------------
echo "Packages installed in the worker environment:"
conda list


# -----------------------------------------------------------------------------
# 6. Configure Dask
#
# This isn't necessary, but for this particular bootstrap script it will make a
# few things easier:
#
# - Configure the cluster's dashboard link to show the proxied version through
#   jupyter-server-proxy. This allows access to the dashboard with only an ssh
#   tunnel to the notebook.
#
# - Specify the pre-packaged python environment, so users don't have to
#
# - Set the default deploy-mode to local, so the dashboard proxying works
#
# - Specify the location of the native libhdfs library so pyarrow can find it
#   on the workers and the client (if submitting applications).
# ------------------------------------------------------------------------------
echo "Configuring Dask"
mkdir -p $HOME/.config/dask
cat <<EOT >> $HOME/.config/dask/config.yaml
distributed:
  dashboard:
    link: "/proxy/{port}/status"
yarn:
  environment: /home/hadoop/environment.tar.gz
  deploy-mode: local
  port: 8786
  worker:
    env:
      ARROW_LIBHDFS_DIR: /usr/lib/hadoop/lib/native/
  client:
    env:
      ARROW_LIBHDFS_DIR: /usr/lib/hadoop/lib/native/
EOT
# Also set ARROW_LIBHDFS_DIR in ~/.bashrc so it's set for the local user
echo -e '\nexport ARROW_LIBHDFS_DIR=/usr/lib/hadoop/lib/native' >> $HOME/.bashrc


# -----------------------------------------------------------------------------
# 7. If Jupyter isn't requested, we're done
# -----------------------------------------------------------------------------
if [[ "$JUPYTER" == "false" ]]; then
    exit 0
fi


# -----------------------------------------------------------------------------
# 8. Install jupyter notebook server and dependencies
#
# We do this after packaging the worker environments to keep the tar.gz as
# small as possible.
#
# We install the following packages:
#
# - notebook: the Jupyter Notebook Server
# - ipywidgets: used to provide an interactive UI for the YarnCluster objects
# - jupyter-server-proxy: used to proxy the dask dashboard through the notebook server
# -----------------------------------------------------------------------------
if [[ "$JUPYTER" == "true" ]]; then
    echo "Installing Jupyter"
    conda install \
    -c conda-forge \
    -y \
    -q \
    notebook \
    ipywidgets \
    jupyter-server-proxy
fi


# -----------------------------------------------------------------------------
# 9. List all packages in the client environment
# -----------------------------------------------------------------------------
echo "Packages installed in the client environment:"
conda list


# -----------------------------------------------------------------------------
# 10. Configure Jupyter Notebook
# -----------------------------------------------------------------------------
echo "Configuring Jupyter"
mkdir -p $HOME/.jupyter
HASHED_PASSWORD=`python -c "from notebook.auth import passwd; print(passwd('$JUPYTER_PASSWORD'))"`
cat <<EOF >> $HOME/.jupyter/jupyter_notebook_config.py
c.NotebookApp.password = u'$HASHED_PASSWORD'
c.NotebookApp.open_browser = False
c.NotebookApp.ip = '0.0.0.0'
EOF


# -----------------------------------------------------------------------------
# 11. Define an upstart service for the Jupyter Notebook Server
#
# This sets the notebook server up to properly run as a background service.
# -----------------------------------------------------------------------------
echo "Configuring Jupyter Notebook Upstart Service"
cat <<EOF > /tmp/jupyter-notebook.conf
description "Jupyter Notebook Server"
start on runlevel [2345]
stop on runlevel [016]
respawn
respawn limit unlimited
exec su - hadoop -c "jupyter notebook" >> /var/log/jupyter-notebook.log 2>&1
EOF
sudo mv /tmp/jupyter-notebook.conf /etc/init/


# -----------------------------------------------------------------------------
# 12. Start the Jupyter Notebook Server
# -----------------------------------------------------------------------------
echo "Starting Jupyter Notebook Server"
sudo initctl reload-configuration
sudo initctl start jupyter-notebook

and for EMR 6.3.0 this bootstrap:

#!/bin/bash
HELP="Usage: bootstrap-dask [OPTIONS]
Example AWS EMR Bootstrap Action to install and configure Dask and Jupyter
By default it does the following things:
- Installs miniconda
- Installs dask, distributed, dask-yarn, pyarrow, and s3fs. This list can be
  extended using the --conda-packages flag below.
- Packages this environment for distribution to the workers.
- Installs and starts a jupyter notebook server running on port 8888. This can
  be disabled with the --no-jupyter flag below.
Options:
    --jupyter / --no-jupyter    Whether to also install and start a Jupyter
                                Notebook Server. Default is True.
    --password, -pw             Set the password for the Jupyter Notebook
                                Server. Default is 'dask-user'.
    --conda-packages            Extra packages to install from conda.
"

set -e

# Parse Inputs. This is specific to this script, and can be ignored
# -----------------------------------------------------------------
JUPYTER_PASSWORD="dask-user"
EXTRA_CONDA_PACKAGES=""
JUPYTER="true"

while [[ $# -gt 0 ]]; do
    case $1 in
        -h|--help)
            echo "$HELP"
            exit 0
            ;;
        --no-jupyter)
            JUPYTER="false"
            shift
            ;;
        --jupyter)
            JUPYTER="true"
            shift
            ;;
        -pw|--password)
            JUPYTER_PASSWORD="$2"
            shift
            shift
            ;;
        --conda-packages)
            shift
            PACKAGES=()
            while [[ $# -gt 0 ]]; do
                case $1 in
                    -*)
                        break
                        ;;
                    *)
                        PACKAGES+=($1)
                        shift
                        ;;
                esac
            done
            EXTRA_CONDA_PACKAGES="${PACKAGES[@]}"
            ;;
        *)
            echo "error: unrecognized argument: $1"
            exit 2
            ;;
    esac
done


# -----------------------------------------------------------------------------
# 1. Check if running on the master node. If not, there's nothing do.
# -----------------------------------------------------------------------------
grep -q '"isMaster": true' /mnt/var/lib/info/instance.json \
|| { echo "Not running on master node, nothing to do" && exit 0; }


# -----------------------------------------------------------------------------
# 2. Install Miniconda
# -----------------------------------------------------------------------------
echo "Installing Miniconda"
curl https://repo.anaconda.com/miniconda/Miniconda3-py38_4.10.3-Linux-x86_64.sh -o /tmp/miniconda.sh
#curl https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh -o /tmp/miniconda.sh
bash /tmp/miniconda.sh -b -p $HOME/miniconda
rm /tmp/miniconda.sh
echo -e '\nexport PATH=$HOME/miniconda/bin:$PATH' >> $HOME/.bashrc
source $HOME/.bashrc
conda update conda -y


# -----------------------------------------------------------------------------
# 3. Install packages to use in packaged environment
#
# We install a few packages by default, and allow users to extend this list
# with a CLI flag:
#
# - dask-yarn >= 0.7.0, for deploying Dask on YARN.
# - pyarrow for working with hdfs, parquet, ORC, etc...
# - s3fs for access to s3
# - conda-pack for packaging the environment for distribution
# - ensure tornado 5, since tornado 6 doesn't work with jupyter-server-proxy
# -----------------------------------------------------------------------------
echo "Installing base packages"
conda install \
-c conda-forge \
-y \
-q \
dask==2021.8.0 \
dask-yarn==0.9 \
distributed==2021.8.0 \
tornado==6.1 \
pyarrow==5.0.0 \
s3fs==2021.7.0 \
conda-pack==0.6.0 \
bokeh==2.3.3 \
dask[dataframe]==2021.8.0 \
$EXTRA_CONDA_PACKAGES


# -----------------------------------------------------------------------------
# 4. Package the environment to be distributed to worker nodes
# -----------------------------------------------------------------------------
echo "Packaging environment"
conda pack -q -o $HOME/environment.tar.gz


# -----------------------------------------------------------------------------
# 5. List all packages in the worker environment
# -----------------------------------------------------------------------------
echo "Packages installed in the worker environment:"
conda list


# -----------------------------------------------------------------------------
# 6. Configure Dask
#
# This isn't necessary, but for this particular bootstrap script it will make a
# few things easier:
#
# - Configure the cluster's dashboard link to show the proxied version through
#   jupyter-server-proxy. This allows access to the dashboard with only an ssh
#   tunnel to the notebook.
#
# - Specify the pre-packaged python environment, so users don't have to
#
# - Set the default deploy-mode to local, so the dashboard proxying works
#
# - Specify the location of the native libhdfs library so pyarrow can find it
#   on the workers and the client (if submitting applications).
# ------------------------------------------------------------------------------
echo "Configuring Dask"
mkdir -p $HOME/.config/dask
cat <<EOT >> $HOME/.config/dask/config.yaml
distributed:
  dashboard:
    link: "/proxy/{port}/status"
yarn:
  environment: /home/hadoop/environment.tar.gz
  deploy-mode: local
  port: 8786
  worker:
    env:
      ARROW_LIBHDFS_DIR: /usr/lib/hadoop/lib/native/
  client:
    env:
      ARROW_LIBHDFS_DIR: /usr/lib/hadoop/lib/native/
EOT
# Also set ARROW_LIBHDFS_DIR in ~/.bashrc so it's set for the local user
echo -e '\nexport ARROW_LIBHDFS_DIR=/usr/lib/hadoop/lib/native' >> $HOME/.bashrc


# -----------------------------------------------------------------------------
# 7. If Jupyter isn't requested, we're done
# -----------------------------------------------------------------------------
if [[ "$JUPYTER" == "false" ]]; then
    exit 0
fi


# -----------------------------------------------------------------------------
# 8. Install jupyter notebook server and dependencies
#
# We do this after packaging the worker environments to keep the tar.gz as
# small as possible.
#
# We install the following packages:
#
# - notebook: the Jupyter Notebook Server
# - ipywidgets: used to provide an interactive UI for the YarnCluster objects
# - jupyter-server-proxy: used to proxy the dask dashboard through the notebook server
# -----------------------------------------------------------------------------
if [[ "$JUPYTER" == "true" ]]; then
    echo "Installing Jupyter"
    conda install \
    -c conda-forge \
    -y \
    -q \
    notebook \
    ipywidgets \
    jupyter-server-proxy
fi


# -----------------------------------------------------------------------------
# 9. List all packages in the client environment
# -----------------------------------------------------------------------------
echo "Packages installed in the client environment:"
conda list


# -----------------------------------------------------------------------------
# 10. Configure Jupyter Notebook
# -----------------------------------------------------------------------------
echo "Configuring Jupyter"
mkdir -p $HOME/.jupyter
HASHED_PASSWORD=`python -c "from notebook.auth import passwd; print(passwd('$JUPYTER_PASSWORD'))"`
cat <<EOF >> $HOME/.jupyter/jupyter_notebook_config.py
c.NotebookApp.password = u'$HASHED_PASSWORD'
c.NotebookApp.open_browser = False
c.NotebookApp.ip = '0.0.0.0'
EOF

# -----------------------------------------------------------------------------
# 11. Define an upstart service for the Jupyter Notebook Server
#
# This sets the notebook server up to properly run as a background service.
# -----------------------------------------------------------------------------
echo "Configuring Jupyter Notebook Upstart Service"
cat <<EOF > /tmp/jupyter-notebook.service
[Unit]
Description=Jupyter Notebook
[Service]
User=hadoop
ExecStart=$HOME/miniconda/bin/jupyter-notebook --config=$HOME/.jupyter/jupyter_notebook_config.py
Environment=JAVA_HOME=$JAVA_HOME
Type=simple
PIDFile=/run/jupyter.pid
WorkingDirectory=$HOME
Restart=always
RestartSec=10
[Install]
WantedBy=multi-user.target
EOF
sudo mv /tmp/jupyter-notebook.service /etc/systemd/system/
sudo systemctl enable jupyter-notebook


# -----------------------------------------------------------------------------
# 12. Start the Jupyter Notebook Server
# -----------------------------------------------------------------------------
echo "Starting Jupyter Notebook Server"
sudo systemctl daemon-reload
sudo systemctl start jupyter-notebook

Best regards :)

[b-y-f]

Above works