Sci-Pype - Jupyter + Redis + Docker

Date: 2016-08-01

Building Scientific Data Pipelines for Jupyter in Docker

_images/image_2016-08-01_building-a-data-science-pipeline.png

Sci-Pype is a framework for building data science pipelines. It uses a specialized Jupyter docker container extended from the original https://github.com/jupyter/docker-stacks that already support Python 2, Python 3, R, and Julia notebooks. I use sci-pype for machine learning, data analysis, data-sharing, model refinement, and stock analysis projects. It contains a python 2 core for making it easier to integrate with external databases and redis servers, debugging with slack, and colorized notebook logging.

This core was originally built to datamine news releases in near-real time. Now I use it for analyzing, exchanging and sharing data with Jupyter. With this core, I build machine learning models that are published and persisted as JSON or serialized objects using pickle. These models are deployed and tested with a load-balanced redis cluster.

Please note this is a large docker container so it may take some time to download and it extracts to ~6.8 GB on disk.

GitHub Repository: https://github.com/jay-johnson/sci-pype

Overview

The docker container runs a Jupyter web application. The web application runs Jupyter Notebooks as kernels that can be implemented in Python 2, 3, R, or Julia. For now the examples and core included in this repository will only work with Python 2.

This container can run in three modes:

  1. Local development

    To start the local development version run: start.sh

    ./start.sh
    

    You can login to the container with: ./ssh.sh

  2. Full Stack

    To start the full stack mode run: compose-start-full.sh

    ./compose-start-full.sh
    

    The full-stack-compose.yml will deploy three docker containers using docker compose:

  3. Standalone Testing

    To start the full stack mode run: compose-start-jupyter.sh

    ./compose-start-jupyter.sh
    

    The jupyter-docker-compose.yml is used to deploy a single Jupyter container.

Working Examples

This document details the following examples. Please refer to the examples directory for the latest.

  1. example-core-demo.ipynb

    How to use the python core from a Jupyter notebook. It also shows how to debug the JSON application configs which are used to connect to external database(s) and redis server(s).

    _images/image_2016-08-01_core-integration.png
  2. example-spy-downloader.ipynb

    Jupyter + Downloading the SPY Pricing Data

    Download the SPY ETF Pricing Data from Google Finance and store it in the shared ENV_PYTHON_SRC_DIR directory that is mounted from the host and into the Jupyter container. It uses a script that downloads the SPY daily pricing data as a csv file.

    _images/image_2016-08-01_download-spy-pricing-data.png
  3. example-plot-stock-data.ipynb

    Download SPY and use Pandas + Matlab to Plot Pricing by the Close

    This shows how to download the SPY daily prices from Google Finance as a csv then load it using Pandas for plotting on the Close prices with Matlab.

    _images/image_2016-08-01_plot-spy-by-close-prices.png
  4. example-redis-cache-demo.ipynb

    Building a Jupyter + Redis Data Pipeline

    This extends the previous SPY pricing demo and publishes + retreives the pricing data by using a targeted CACHE redis server (that runs inside the Jupyter container). It stores the Pandas dataframe as JSON in the LATEST_SPY_DAILY_STICKS redis key.

    _images/image_2016-08-01_redis-data-pipeline-with-spy-prices.png
  5. example-db-extract-and-cache.ipynb

    Building a Jupyter + MySQL + Redis Data Pipeline

    This requires running the Full Stack which uses the https://github.com/jay-johnson/sci-pype/blob/master/full-stack-compose.yml to deploy three docker containers on the same host:

    How it works

    _images/image_2016-08-01_using-jupyter-for-stock-analysis.png
    1. Extract the IBM stock data from the MySQL dataset and store it as a csv inside the /opt/work/data/src/ibm.csv file

    2. Load the IBM pricing data with Pandas

    3. Plot the pricing data with Matlab

    4. Publish the Pandas Dataframe as JSON to Redis

    5. Retrieve the Pandas Dataframe from Redis

    6. Test the cached pricing data exists outside the Jupyter container with:

      $ ./redis.sh
      SSH-ing into Docker image(redis-server)
      [root@redis-server container]# redis-cli -h localhost -p 6000
      localhost:6000> LRANGE LATEST_IBM_DAILY_STICKS 0 0
      1) "(dp0\nS'Data'\np1\nS'{\"Date\":{\"49\":971136000000,\"48\":971049600000,\"47\":970790400000,\"46\":970704000000,\"45\":970617600000,\"44\":970531200000,\"43\":970444800000,\"42\":970185600000,\"41\":970099200000,\"40\":970012800000,\"39\":969926400000,\"38\":969
      
       ... removed for docs ...
      
      localhost:6000> exit
      [root@redis-server container]# exit
      exit
      $
      
  6. example-slack-debugging.ipynb

    Jupyter + Slack Driven Development

    This example shows how environment variables allow the python core to publish a message into Slack to notify the associated user with a message containing the line number and source code that threw the exception.

    _images/image_2016-08-01_slack-debugging.png

Components

  1. Python 2 Core

    The PyCore uses a JSON config file for connecting to redis servers and configurable databases (MySQL and Postgres) using SQLAlchemy. It has only been tested with the Python 2.7 kernel.

  2. Local Redis Server

    When starting the container with ENV_DEPLOYMENT_TYPE set to anything not JustDB, the container will start a local redis server inside the container on port 6000 for iterating on your pipeline analysis, model deployment and caching strategies.

  3. Loading Database and Redis Applications

    By default the jupyter.json config supports multiple environments for integrating notebooks with external resources. Here is table on what they define:

    Name Purpose Redis Applications Database Applications
    Local Use the internal redis server with the stock db local-redis.json db.json
    NoApps Run the core without redis servers or databases empty-redis.json empty-db.json
    JustRedis Run with just the redis servers and no databases local-redis.json empty-db.json
    JustDB Run without redis servers and load the databases empty-redis.json db.json
    Test Connect to external redis servers and databases redis.json db.json
    Live Connect to external redis servers and databases redis.json db.json

    Inside a notebook you can target a different environment before loading the core with:

    • Changing to the JustRedis Environment:

      import os
      os.environ["ENV_DEPLOYMENT_TYPE"] = "JustRedis"
      core = PyCore()
      
    • Changing to the NoApps Environment:

      import os
      os.environ["ENV_DEPLOYMENT_TYPE"] = "NoApps"
      core = PyCore()
      
  4. Customize the Jupyter Container Lifecycle

    The following environment variables can be used for defining pre-start, start, and post-start Jupyter actions as needed.

    Environment Variable Default Value Purpose
    ENV_PRESTART_SCRIPT /opt/containerfiles/pre-start-notebook.sh Run custom actions before starting Jupyter
    ENV_START_SCRIPT /opt/containerfiles/start-notebook.sh Start Jupyter
    ENV_POSTSTART_SCRIPT /opt/containerfiles/post-start-notebook.sh Run custom actions after starting Jupyter
  5. Slack Debugging

    The core supports publishing exceptions into Slack based off the environment variables passed in using docker or docker compose.

  6. Tracking Installed Dependencies for Notebook Sharing

    This docker container uses these files for tracking Python 2 and Python 3 pips:

    • /opt/work/pips/python2-requirements.txt
    • /opt/work/pips/python3-requirements.txt
  7. Shared Volumes

    These are the mounted volumes and directories that can be changed as needed. Also the core uses them as environment variables.

    Host Mount Container Mount Purpose
    /opt/project /opt/project Sharing a project from the host machine
    /opt/work/data /opt/work/data Sharing a common data dir between host and containers
    /opt/work/data/src /opt/work/data/src Passing data source files into the container
    /opt/work/data/dst /opt/work/data/dst Passing processed data files outside the container
    /opt/work/data/bin /opt/work/data/bin Exchanging data binaries from the host into the container
    /opt/work/data/synthesize /opt/work/data/synthesize Sharing files used for synthesizing data
    /opt/work/data/tidy /opt/work/data/tidy Sharing files used to tidy and marshall data
    /opt/work/data/analyze /opt/work/data/analyze Sharing files used for data analysis and processing
    /opt/work/data/output /opt/work/data/output Sharing processed files and analyzed output

Getting Started

Local Jupyter

  1. Start the Container in Local development mode

    $ ./start.sh
    Starting new Docker image(docker.io/jayjohnson/jupyter)
    4275447ef6a3aa06fb06097837deeb202bd80b15969a9c1269a5ee042d8df13d
    $
    
  2. Browse to the local Jupyter website

    http://localhost:82/

Full Stack

The full-stack-compose.yml patches the Jupyter and redis containers to ensure the MySQL database is listening on port 3306 before starting. It does this by defining a custom entrypoint wrapper for each in the wait-for-its tools directory.

  1. Start the Composition

    This can take around 20 seconds for MySQL to set up the seed pricing records, and it requires assigning the shared data directory permissions for read/write access from inside the Jupyter container.

    $ ./compose-start-full.sh
    Before starting changing permissions with:
       chown -R driver:users /opt/work/data/*
    [sudo] password for driver:
    Starting Composition: full-stack-compose.yml
    Starting stocksdb
    Starting jupyter
    Starting redis-server
    Done
    $
    
  2. Check the Composition

    $ docker ps
    CONTAINER ID        IMAGE                                COMMAND                  CREATED             STATUS              PORTS                                        NAMES
    1fd9bd22987f        jayjohnson/redis-single-node:1.0.0   "/wait-for-its/redis-"   12 minutes ago      Up 25 seconds       0.0.0.0:6000->6000/tcp                       redis-server
    2bcb6b8d2994        jayjohnson/jupyter:1.0.0             "/wait-for-its/jupyte"   12 minutes ago      Up 25 seconds       0.0.0.0:8888->8888/tcp                       jupyter
    b7bce846b9af        jayjohnson/schemaprototyping:1.0.0   "/root/start_containe"   25 minutes ago      Up 25 seconds       0.0.0.0:81->80/tcp, 0.0.0.0:3307->3306/tcp   stocksdb
    $
    
    • Optional - Login to the database container
    $ ./db.ssh
    SSH-ing into Docker image(stocksdb)
    [root@stocksdb db-loaders]# ps auwwx | grep mysql | grep -v grep
    root        28  0.0  0.0  11648  2752 ?        S    17:00   0:00 /bin/sh /usr/bin/mysqld_safe
    mysql      656  1.3 12.0 1279736 474276 ?      Sl   17:00   0:01 /usr/sbin/mysqld --basedir=/usr --datadir=/var/lib/mysql --plugin-dir=/usr/lib64/mysql/plugin --user=mysql --log-error=/var/log/mysql/error.log --pid-file=/var/lib/mysql/mysqld.pid --socket=/var/lib/mysql/mysqld.sock --port=3306
    [root@stocksdb db-loaders]# exit
    

    View the Stocks Database with phpMyAdmin: http://localhost:81/phpmyadmin/sql.php?db=stocks&table=stocks

    Note

    By default the login to this sample db is: dbadmin / dbadmin123 which can be configured in the db.env

    • Optional - Login to the Redis container
    $ ./redis.sh
    SSH-ing into Docker image(redis-server)
    [root@redis-server container]# ps auwwx | grep redis
    root         1  0.0  0.0  11644  2616 ?        Ss   17:00   0:00 bash /wait-for-its/redis-wait-for-it.sh
    root        28  0.0  0.2 114800 11208 ?        Ss   17:00   0:00 /usr/bin/python /usr/bin/supervisord -c /etc/supervisor.d/rediscluster.ini
    root        30  0.3  0.0  37268  3720 ?        Sl   17:00   0:00 redis-server *:6000
    root        47  0.0  0.0   9044   892 ?        S+   17:02   0:00 grep --color=auto redis
    [root@redis-server container]# exit
    
    • Optional - Login to the Jupyter container
    $ ./ssh.sh
    SSH-ing into Docker image(jupyter)
    driver:/opt/work$ ps auwwx | grep jupyter
    driver       1  0.0  0.0  13244  2908 ?        Ss   17:00   0:00 bash /wait-for-its/jupyter-wait-for-it.sh
    driver      38  0.3  1.2 180564 48068 ?        S    17:00   0:00 /opt/conda/bin/python /opt/conda/bin/jupyter-notebook
    driver:/opt/work$ exit
    
  3. Run the Database Extraction Jupyter Demo

    Open the notebook with this url: http://localhost:82/notebooks/examples/example-db-extract-and-cache.ipynb

  4. Click the Run Button

    This example will connect to the stocksdb MySQL container and pull 50 records from IBM’s pricing data. It will then render plot lines for Open, Close, High, and Low using Pandas and Matlab. Next it will cache the IBM records in the redis-server container and then verify those records were cached correctly by retrieving it again.

  5. From outside the Jupyter container confirm the redis key holds the processed IBM data

    $ ./redis.sh
    SSH-ing into Docker image(redis-server)
    [root@redis-server container]# redis-cli -h localhost -p 6000
    localhost:6000> LRANGE LATEST_IBM_DAILY_STICKS 0 0
    1) "(dp0\nS'Data'\np1\nS'{\"Date\":{\"49\":971136000000,\"48\":971049600000,\"47\":970790400000,\"46\":970704000000,\"45\":970617600000,\"44\":970531200000,\"43\":970444800000,\"42\":970185600000,\"41\":970099200000,\"40\":970012800000,\"39\":969926400000,\"38\":969
    
     ... removed for docs ...
    
    localhost:6000> exit
    [root@redis-server container]# exit
    exit
    $
    
  6. Stop the Composition

    $ ./compose-stop-full.sh
    Stopping Composition: full-stack-compose.yml
    Stopping redis-server ... done
    Stopping jupyter ... done
    Stopping stocksdb ... done
    Done
    $
    

Standalone

  1. Start Standalone

    Start the standalone Jupyter container using the jupyter-docker-compose.yml file. This compose file requires access to /opt/work/data host directory like the Full Stack version for sharing files between the container and the host.

    $ ./compose-start-jupyter.sh
    Before starting changing permissions with:
       chown -R driver:users /opt/work/data/*
    [sudo] password for driver:
    Starting Composition: jupyter-docker-compose.yml
    Starting jupyter
    Done
    $
    
  2. Stop Standalone

    Stop the standalone Jupyter composition with:

    $ ./compose-stop-jupyter.sh
    Stopping Composition: jupyter-docker-compose.yml
    Stopping jupyter ... done
    Done
    $
    

Deleting the Containers

Remove the containers with the command:

$ docker rm jupyter redis-server stocksdb
jupyter
redis-server
stocksdb
$

Delete them from the host with:

$ docker rmi jayjohnson/schemaprototyping
$ docker rmi jayjohnson/jupyter
$ docker rmi jayjohnson/redis-single-node

Sharing between the Host and the Jupyter Container

By default, the host will have this directory structure available for passing files in and out of the container:

$ tree /opt/work
/opt/work
└── data
    ├── analyze
    ├── bin
    ├── dst
    ├── output
    ├── src
    │   └── spy.csv
    ├── synthesize
    └── tidy

8 directories, 1 file

From inside the container here is where the directories are mapped:

$ ./ssh.sh
SSH-ing into Docker image(jupyter)
driver:/opt/work$ tree data/
data/
├── analyze
├── bin
├── dst
├── output
├── src
│   └── spy.csv
├── synthesize
└── tidy

7 directories, 1 file

Coming Soon and Known Issues

  1. How to build a customized Python Core mounted from outside the Jupyter container

  2. Fixing the docker compose networking so the stocksdb container does not need to know the compose-generated docker network.

    Right now it is defing the scipype_datapype as the expected docker network. This may not work on older versions of docker.

  3. Building Jupyter containers that are smaller and only run one kernel to reduce the overall size of the image

  4. Testing on an older docker version

    This was tested with 1.12.0

    $ docker -v
    Docker version 1.12.0, build 8eab29e
    $
    
  5. Setting up the Jupyter wait-for-it to ensure the stocks database is loaded before starting...not just the port is up

    For now just shutdown the notebook kernel if you see an error related to the stocks database not being there when running the full stack.

License

This repo is MIT

Jupyter - BSD: https://github.com/jupyter/jupyter/blob/master/LICENSE

Thanks for reading!

Until next time,

Jay

Want to catch up?