IaaS, PaaS, and DevOps for Data Scientist
Download as PPTX, PDF•0 likes•267 views
This document discusses Infrastructure as a Service (IaaS), Platform as a Service (PaaS), and DevOps for data science. It provides an overview of Azure services for IaaS including virtual machines and GPU instances. It also discusses PaaS options like Azure Machine Learning for deploying models as web services. The document advocates for using Azure services like HDInsight, Data Factory and Machine Learning to build distributed and scalable data science systems following architectures like lambda architecture. It highlights pros and cons of different approaches for flexibility, scalability and using open source tools for data science workloads on Azure.
IaaS, PaaS, and DevOps for Data Scientist
- 1. IaaS, PaaS, and DevOps for Data Science Dmitry Petukhov, Machine Learning Preacher, Microsoft AI MVP && Coffee Addicted #Azure + #AI = #AzureDay
- 3. Who is who?
- 5. Hypothesis Raw Data Collect Receive -> Aggregate -> Transform -> Store Investigate Data Data structure (metadata) Data distribution Visualize data Descriptive statistics Pre-processing Clean & consistent: - Handling missing data - Handling incorrect data - Sensitive/private data Convert data types (e.g. formatting culture- sensitive data) Feature Extraction* Training Dataset Machine Learning Workflow Data Sources Splitting Data Test Dataset To be continued… Software Engineering Workflow
- 6. Machine Learning WorkflowSoftware Engineering Workflow Model evaluation Evaluate measures of quality model (ROC, RMSE, F-Score, etc.) Feature Selection** Feature Selection Feature Scaling (Normalization) Dimension Reduction Final Model Training ML algorithm Model Publication Revision FinalModelEvaluation Data Flow Cross-validation Source 0xCODE.IN
- 7. −Distributed −Large scalable −Fault tolerance −Reliable −Big Data ready −OSS-based −Lawful −Distributed −Large scalable −Fault tolerance −Reliable −Big Data ready −OSS-based −Lawful Classic service requirements: Intellectual service requirements:
- 8. Data Scientist Data (SDD) On-Premises On-Knee Data Science code1 result3 commit 4 compute2 Flexibility Distributed Large scalable Fault tolerance Reliable OSS-based Big Data ready Secured Lawful Pros and Cons: + Quick start and easy to use + Flexible + Markdown, code and visualization in one place + Quick check of hypotheses - All code in one(!) file - Tedious version control - Not scalable - Never will be in production
- 9. Runtime ML/DL Frameworks OS Virtualizations Servers Data Intellectual App ML Model IDEs and Tools Languages Storage Networking On-premises ML/DL Frameworks OS Virtualizations Servers Data Intellectual App ML Model IDEs and Tools Languages Storage Networking PaaS ML/DL Frameworks OS Virtualizations Servers Data Intellectual App ML Model IDEs and Tools Languages Storage Networking Model as a Service ML/DL Frameworks OS Virtualizations Servers Data Intellectual App ML Model IDEs and Tools Languages Storage Networking IaaS UnmanagedManaged AI Lab. Where?
- 10. Pros and Cons: + Quick start and easy to use + Latest hardware + Preinstalled ML/DL frameworks + Horizontal scalable + On demand => zero acquisition cost - High cost (24h) - Not for production - Old versions of frameworks - Zoo of unused tools IaaS: Data Science VM Images Picture credit: Azure Docs NVidia Tesla GPU(s) Ubuntu / Windows Server TensorFlow / CNTK / via Keras RStudio / Visual Studio Azure NC/ND Series VMs R / Python / C# xgboost / LightGBM / CatBoost Flexibility Distributed Large scalable Fault tolerance Reliable OSS-based Big Data ready Secured Lawful
- 11. Microsoft ML Server Azure Blob Storage Apache Spark / Hadoop Tasks Azure HDInsight Head Node Worker Node HDFS API Tasks PaaS: Data Science VM Images Flexibility Distributed Large scalable Fault tolerance Reliable OSS-based Big Data ready Secured Lawful Data Scientist MS ML Client <connect> Pros and Cons: + as PaaS(!) + Distributed, scalable, etc. + REST API + Python and R - High cost - Not all operations / ML algorithms can be distributed - Not all OSS - Proprietary output
- 12. Batch Layer Data Streams Enterprise Data Sources Others Train ML model Tune hyperparams Cross-validation Preprocessing data Speed Layer Inference ML model Stream analysis Preprocessing data 1 2 3 4 5 1 2 3 4 5 6 Distributed File System λ Architecture
- 13. Market Data Streams Macroeconomic Data Azure Event Hub Train ML model Tune hyperparams Cross-validation Preprocessing data Inference ML model Aggregate stats Preprocessing data Azure Storage 1 2 3 4-5 6 1 3 4 5 6 Tweets Feed Azure Data Factory h(θ0, θn) 2 Architecture as Service: λ Architecture for Data Science Flexibility Distributed Large scalable Fault tolerance Reliable OSS-based Big Data ready Secured Lawful Store data
- 14. Azure HDInsight Azure SQL + Microsoft ML Server Data Scientist + web-browser <SSH tunnel> Azure Storage Azure Data Factory Azure VM Azure GPU Instances or Memory-intensive Instances AzureOn-premises Data flow Tasks flow Azure Batch (AI) Data transfer inside one DC Low-priority VM Storage options: tiers, redundancy PremiumRS tier B-series VM Scheduling Provisioning Use Open Source Only for Big Data cases
- 15. DevOps for Data Science
- 16. © 2018, Dmitry Petukhov. CC BY-SA 4.0 license. Microsoft and other product names are or may be registered trademarks and/or trademarks in the U.S. and/or other countries. Thank you!
- 17. Q&A Now or later (see contacts below) Stay connected Be friend at the Facebook/@codez0mb1e Read me at the Habr/@codezombie All contacts on http://0xCode.in/@codez0mb1e Download slides from http://0xcode.in/2018/azure-day-moscow or
Editor's Notes
- #10: Если нет законодательных ограничений Назовем их в алфавитном порядке
- #11: Швейцарский нож
- #12: С IaaS закончили переходим на PaaS Кружка кофе и DFS App servers: ML Server H20 SparkR
- #13: In real life 2 data pipeline: - training - inference Архитектура из мира CS, которая поможет людям из Data Science
- #18: (c) 2018, Dmitry Petukhov. CC BY-SA 4.0 license.