MapR and Machine Learning Primer
Download as PPTX, PDF1 like357 views
The document discusses the challenges and opportunities of implementing machine learning (ML) in enterprises, emphasizing that ML involves substantial work in data cleaning and feature engineering. It presents MapR as a powerful platform for enterprise ML, capable of handling various data sources and supporting a wide range of tools. The key takeaway is that organizational success in ML depends not only on effective modeling but also on robust workflow management and deployment strategies.
MapR and Machine Learning Primer
- 1. © 2017 MapR TechnologiesMapR Confidential 1 + Mathieu Dumoulin Data Engineer PS APAC â Tokyo, Japan Wednesday, May 10, 2017
- 2. © 2017 MapR TechnologiesMapR Confidential 2 Todayâs goals âą Machine Learning âą Enterprise Machine Learning âą Challenge of Enterprise ML âą MapR Unique Features for ML âą H2O and MapR
- 3. © 2017 MapR TechnologiesMapR Confidential 3 Machine Learning Machine learning is a type of artificial intelligence (AI) that provides computers with the ability to learn without being explicitly programmed. ML allows a computer to make predictions on data (usually based on historical data)
- 4. © 2017 MapR TechnologiesMapR Confidential 4 1. Is this A or B? 2. Is this weird? 3. How much â or â How many? 4. How is this organized? 5. What should I do next? 6. Whatâs similar? Questions Data Science Can Answer 1. Classification 2. Anomaly Detection 3. Regression 4. Clustering 5. Reinforcement Learning 6. Recommendation
- 5. © 2017 MapR TechnologiesMapR Confidential 5 ML In the Enterprise⊠⊠isnât so easy after all.
- 6. © 2017 MapR TechnologiesMapR Confidential 6 Enterprise ML: Business Value Outputs Growing number of ML use cases at successful companies Anomaly Detection ç°ćžžæ€ćș Customer 360 Fraud Detection äžæŁæ€ćș Log Security Analysis ăă°ćæ Recommender Engines ăŹăłăĄăłăăŒă·ă§ăł Sensor Data Analysis (IoT) Personalized Offers ćäșșć Ad Tech
- 7. © 2017 MapR TechnologiesMapR Confidential 7 Machine Learning Tools
- 8. © 2017 MapR TechnologiesMapR Confidential 8 What Most ML Tools Give You A common rule of thumb is that the modeling task is about 10% of the total effort of a ML project. The choice of tool matters (to the DS), but any top level ML tool/library can eventually get good results (if the data allows it at all)
- 9. © 2017 MapR TechnologiesMapR Confidential 9 Enterprise ML Projects: More than Just Modeling
- 10. © 2017 MapR TechnologiesMapR Confidential 10 Business Value is in Production All the business value results from a sufficiently accurate model running in production What it means: Deploying a weaker model in production sooner is MUCH better than endless work for an excellent model (But you can make Google money if you get a world class model in production)
- 11. © 2017 MapR TechnologiesMapR Confidential 11 Data Cleaning and Feature Engineering 80% of the work!
- 12. © 2017 MapR TechnologiesMapR Confidential 12 Workflow View of Machine Learning 1 2 3 4 56 7 8
- 13. © 2017 MapR TechnologiesMapR Confidential 13 Enterprise ML Challenges Data comes from many sources maybe very large Data isnât always labeled! Needs ETL and cleaning Finding the best algorithm and parameters can use a lot of CPU Real time data? Production data from many sources? Needs to run on a server somewhere The predictions are used by another system...
- 14. © 2017 MapR TechnologiesMapR Confidential 14 The Open Source Solution (Iâm not joking!) Ref: http://advancedspark.com/ , https://github.com/fluxcapacitor/pipeline Separate Clusters!
- 15. © 2017 MapR TechnologiesMapR Confidential 15 What Data Scientists and ML Engineers Want I know where the data is and how to access it. My work is made easier in ALL PHASES of the ML project, not just modeling Let me use my favorite tools at all scales (MB, GB, TB, PB)
- 16. © 2016 MapR Technologies© 2016 MapR TechnologiesMapR Confidential An Ideal Platform For Enterprise ML An ideal platform for ML: âąScales with you and your data âąFreedom to use any tool â Open source DS tools: Jupyter, Zeppelin, Spark, H2O, TF,⊠â Legacy/local tools: NLP tools, scikit-learn, R âąData can be versioned and kept reliably âąCombines storage, DB, compute and streams âąSupports both model building and model deployment âąSupports security when needed
- 17. © 2017 MapR TechnologiesMapR Confidential 17 Our Humble proposal MapR is The Best Platform for Enterprise ML on the market today
- 18. © 2017 MapR TechnologiesMapR Confidential 18 MapR Converged Data Platform Open Source Engines & Tools Commercial Engines & Applications Utility-Grade Platform Services DataProcessing Enterprise Storage MapR-FS MapR-DB MapR Streams Database Event Streaming Global Namespace High Availability Data Protection Self-healing Unified Security Real-time Multi-tenancy Search & Others Cloud & Managed Services Custom Apps UnifiedManagementandMonitoring
- 19. © 2017 MapR TechnologiesMapR Confidential 19 The MapR Stack: Converged + Open
- 20. © 2017 MapR TechnologiesMapR Confidential 20 âą NFS mount and POSIX file system â Small scale Python or R data exploration on the real data â Keep the raw data, ETL work is easily reused âą Supports standard big data ecosystem (Spark) âą NFS mount can ingest data from any enterprise system that can output files â Even if they donât support Hadoop! âą Much faster than HDFS â Serve production models directly from MapR MapR Supports All Tools Out of the Box
- 21. © 2017 MapR TechnologiesMapR Confidential 21 âą Volumes and Topologies â GPU enabled nodes for distributed deep learning on the same cluster âą Donât waste resources âą Keep data locality âą Avoid unnecessary data movement âą Avoid multiple copies of data (which is the real one?) âą POSIX file system â Use any DL framework on the cluster data MapR has production experience with CaffeOnSpark (Samsung Micro-Electronics) and has a new TensorFlow QSS MapR Supports Deep Learning
- 22. © 2017 MapR TechnologiesMapR Confidential 22 âą Volumes and Snapshots â Experimental reproducibility â Create models on real production data â Easy to compare models on the same data â Easy to evaluate a model across time on different snapshots â snapshot of models: a time machine, built-in âą Volume Quotas â Support multiple projects and teams on the same cluster â Share storage resources efficiently Clever Uses for Volumes and Snapshots
- 23. © 2017 MapR TechnologiesMapR Confidential 23 Remember that > 90% of the work in Enterprise ML is to realize the workflow. This is where MapR shines! ï âą Operational capabilities (MapR DB, MapR Client) â Serve production models directly from MapR âą Snapshots and Mirrors â Do A/B testing with almost no coding â Promote the mirror to go back to the previous state âą Just update the path in the production system - no redeployment! âą MapR Streams for Real-time predictions â Zero configuration Kafka â it just works! â Kafka REST Proxy for max interoperability â Supports microservices and Stateful Containers Support the ML Workflow, Not Just Modeling
- 24. © 2017 MapR TechnologiesMapR Confidential 24 MapR đ Enterprise Machine Learning âą Features that work together to support all phases of ML âą Supports your existing tools/code and the state of the art large scale frameworks âą Easier to manage, more robust and secure. âą MapR is made for the enterprise and great for ML
- 25. © 2017 MapR TechnologiesMapR Confidential 25 MapR Converged Application Blueprint âą Microservices connected by real-time streams â Ideal to serve predictions from ML models âą Next-Generation large-scale architecture âą Working example: https://www.mapr.com/appblueprint/overview
- 26. © 2017 MapR TechnologiesMapR Confidential 26 Q&A ENGAGE WITH US @mapr mdumoulin@mapr.com
Editor's Notes
- #5: 1. Will this tire fail in the next 1,000 miles: Yes or no? Which brings in more customers: a $5 coupon or a 25% discount? 2. If you have a car with pressure gauges, you might want to know: Is this pressure gauge reading normal? If you're monitoring the internet youâd want to know: Is this message from the internet typical? 3. What will the temperature be next Tuesday? What will my fourth quarter sales be? 4. Which viewers like the same types of movies? Which printer models fail the same way? 5. Reinforcement learning was inspired by how the brains of rats and humans respond to punishment and rewards. These algorithms learn from outcomes, and decide on the next action. Typically, reinforcement learning is a good fit for automated systems that have to make lots of small decisions without human guidance. 6. What did similar people/customer buy/watch/listen to? What other movies will you like, if you like product A?
- #9: Data Ingestion is a non-trivial task for enterprise The best systems combine data from multiple sources Adding more data is a highly specialized task Data Governance for ML Dataset versions Test data versions Model versions Test results versions Model Deployment is a non-trivial integration task with another external enterprise system May need to be scalable, HA and fault-tolerant What about after deployment? A/B Testing Understanding performance Dealing with data drift
- #10: Data Ingestion is a non-trivial task for enterprise The best systems combine data from multiple sources Adding more data is a highly specialized task Data Governance for ML Dataset versions Test data versions Model versions Test results versions Model Deployment is a non-trivial integration task with another external enterprise system May need to be scalable, HA and fault-tolerant What about after deployment? A/B Testing Understanding performance Dealing with data drift
- #11: Data Ingestion is a non-trivial task for enterprise The best systems combine data from multiple sources Adding more data is a highly specialized task Data Governance for ML Dataset versions Test data versions Model versions Test results versions Model Deployment is a non-trivial integration task with another external enterprise system May need to be scalable, HA and fault-tolerant What about after deployment? A/B Testing Understanding performance Dealing with data drift
- #12: Data Ingestion is a non-trivial task for enterprise The best systems combine data from multiple sources Adding more data is a highly specialized task Data Governance for ML Dataset versions Test data versions Model versions Test results versions Model Deployment is a non-trivial integration task with another external enterprise system May need to be scalable, HA and fault-tolerant What about after deployment? A/B Testing Understanding performance Dealing with data drift
- #13: Get training data (example: images) Get labels for the training data (examples: what the image is about, the image labels) Transform the data into numbers (machine learning algorithms canât deal with raw data, only vectors of numbers) Heavily iterative work to find the best set of features Try many different algorithms, and tune their parameters for best performance Heavily iterative work to find the best algorithm and parameter values The best algorithm, trained on your data, with its parameters tuned for best performance is your predictive model Get new data Transform it to match the same format as your training feature vectors The model will output a predicted label for the new data This is a lot of work, but glosses over a HUGE amount of work required to get business value in an enterprise setting
- #14: Here is a small sample of the issues faced in putting ML to work in an enterprise
- #15: This is for real. Chris Fregly is the Pipeline.io guy and heâs building enterprise ML systems with this set of tools. This is the set of tools required to be able to provide a true 100% open source end to end story for enterprise ML. How does MapR simplify this picture? What tools still remain useful if weâd run on MapR?
- #17: Can support OSS tools like R, scikit, Theano and TensorFlow first to avoid more expensive licences. NoSQL, Kafka, Spark, etc...
- #24: Indeed, ML Tools are only really good at modeling. They typically provide limited support for feature engineering. In addition Tools typically only support testing models on a single dataset, with no support for comparing production models to experimental models, comparing models across different versions of a dataset, etc. Such capabilities need to be custom made and are typically done with low quality ad-hoc code by data scientists. Most of the work resides in the ETL to get the data in the first place, the data cleaning and feature engineering not supported by the tools, as well as as the work required to deploy a model to production. MapR really shines on that 90% of the work, and supports all ML tools just the same (indeed, better) as any competing platform.
- #25: legacy tools: R, Python, Bash, SPSS, Hive/Pig State of the art: Apache projects like Drill, Impala, Spark, Zepplin, Mesos, Flink, âŠ