This repository contains the files and data from the workshop as well as resources around Data Engineering. For the workshop (and after) we will use a Gitter chatroom to keep the conversation going: https://gitter.im/Jay-Oh-eN/data-engineering-101.
And/or please do not hesitate to reach out to me directly via email at jonathan@galvanize.com or over twitter @clearspandex
The presentation can be found on Slideshare here or in this repository (presentation.pdf
). Video can be found here.
Throughout this workshop, you will learn how to make a scalable and sustainable data pipeline in Python with Luigi
- Run a simple 1 stage Luigi flow reading/writing to local files
- Write a Luigi flow containing stages with multiple dependencies
- Visualize the progress of the flow using the centralized scheduler
- Parameterize the flow from the command line
- Output parameter specific output files
- Manage serialization to/from a Postgres database
- Integrate a Hadoop Map/Reduce task into an existing flow
- Parallelize non-dependent stages of a multi-stage Luigi flow
- Schedule a local Luigi job to run once every day
- Run any arbitrary shell command in a repeatable way
Prior experience with Python and the scientific Python stack is beneficial. The workshop will focus on using the Luigi framework, but will have code from the following lobraries as well:
- numpy
- scikit-learn
- Flask
- Install libraries and dependencies:
pip install -r requirements.txt
- Start the UI server:
luigid --background --logdir logs
- Navigate with a web browser to
http://localhost:[port]
where[port]
is the port theluigid
server has started on (luigid
defaults to port 8082) - start the API Server:
python app.py
- Evaluate Model:
python ml-pipeline.py EvaluateModel --input-dir text --lam 0.8
- Run evaluation server (at
localhost:9191
):topmodel/topmodel_server.py
- Run the final pipeline:
python ml-pipeline.py BuildModels --input-dir text --num-topics 10 --lam 0.8
--
For parallelism, set --workers
(note this is Task parallelism):
python ml-pipeline.py BuildModels --input-dir text --num-topics 10 --lam 0.8 --workers 4
- Start Hadoop cluster:
bin/start-dfs.sh; sbin/start-yarn.sh
- Setup Directory Structure:
hadoop fs -mkdir /tmp/text
- Get files on cluster:
hadoop fs -put ./data/text /tmp/text
- Retrieve results:
hadoop fs -getmerge /tmp/text-count/2012-06-01 ./counts.txt
- View results:
head ./counts.txt
docker run -it -v /LOCAL/PATH/TO/REPO/data-engineering-101:/root/workshop clearspandex/pydata-seattle bash
pip2 install flask
ipython2 app.py
text/ 20newsgroups text files
topmodel/ Stripe's topmodel evaluation library
example_luigi.py example scaffold of a luigi pipeline
hadoop_word_count.py example luigi pipeline using Hadoop
ml-pipeline.py luigi pipeline covered in workshop
app.py Flask server to deploy a scikit-learn model
LICENSE Details of rights of use and distribution
presentation.pdf lecture slides from presentation
readme.md this file!
The data (in the text/
folder) is from the 20 newsgroups dataset, a standard benchmarking dataset for machine learning and NLP. Each file in text
corresponds to a single 'document' (or post) from one of two selected newsgroups (comp.sys.ibm.pc.hardware
or alt.atheism
). The first line provides which group the document is from and everything thereafter is the body of the post.
comp.sys.ibm.pc.hardware
I'm looking for a better method to back up files. Currently using a MaynStream
250Q that uses DC 6250 tapes. I will need to have a capacity of 600 Mb to 1Gb
for future backups. Only DOS files.
I would be VERY appreciative of information about backup devices or
manufacturers of these products. Flopticals, DAT, tape, anything.
If possible, please include price, backup speed, manufacturer (phone #?),
and opinions about the quality/reliability.
Please E-Mail, I'll send summaries to those interested.
Thanx in advance,
- Questioning the Lambda Architecture
- Luigi: NYC Data Science Meetup
- The Log: What every software engineer should know about real-time data's unifying abstraction
- I (heart) Log
- Why Loggly Loves Apache Kafka
- Buffer's New Data Architecture
- Putting Apache Kafka to Use
- Metric Driven Development
- The Unified Logging Infrastructure for Data Analytics at Twitter
- Stream Processing and Mining just got more interesting
- How to Beat the CAP Theorem
- Beating the CAP Theorem Checklist
Copyright 2015 Jonathan Dinu.
All files and content licensed under Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International Public License