Taming data lake - scalable metrics model
Download as PPTX, PDF2 likes739 views
The document discusses setting up a Scalable Metrics Model (SMM) for a data lake. The SMM uses pre-aggregated metrics and dimension keys to provide a flexible and scalable approach. An example model from retail is described, including defining customer-level dimensions, metrics, and value-added metrics. The document outlines the ETL process, including writing queries to populate metrics for different granularities and integrating them into a framework. Reporting options are also discussed, including using views, tables, and reporting tools to provide accessibility and analysis capabilities. Finally, the extensibility of the model by adding new dimensions like sellers and transactions is highlighted.
Taming data lake - scalable metrics model
- 2. “Taming the Data Lake” 2
- 3. Intended for Knowledge Sharing only Disclaimer: Participation in this summit is purely on personal basis and not representing VISA in any form or matter. The talk is based on learnings from work across industries and firms. Care has been taken to ensure no proprietary or work related info of any firm is used in any material. Director, Insights at Visa, Inc. Enable Decision Making at the Executives/ Product/Marketing level via actionable insights derived from Data. RAMKUMAR RAVICHANDRAN Data Warehouse Architect at Visa, Inc. Architect a data-shop in Hadoop to get 360- degree view of the interaction. Technology interface for the Data Stakeholder Community. BHARATHIRAJA CHANDRASEKHARAN
- 4. Intended for Knowledge Sharing only Quick recap of what it is Intended for Knowledge Sharing only Data Lakes – the concept
- 5. AS THEY ARE ENVISIONED TODAY… Intended for Knowledge Sharing only Source: http://www.tangerine.co.th/tag/how-do-data-lake-work/ 5
- 6. DOES IT RING A BELL? *only satiric to wake you up and not indicative of anyone or anything- any similarity is purely coincidental! 6
- 7. & DOES THIS TOO? *only satiric to wake you up and not indicative of anyone or anything- any similarity is purely coincidental! 7
- 8. SO WHAT DO WE HEAR FROM OUR USERS? We often hear these statements in the context of data lakes… Success criteria was engineering specific – Storage/Scalability cost saving, etc Expensive Change Management Complex for the end users to deal with Analytical performance issues Data Governance, Lineage and Management complexities “Although the cost of Storage went down, actual cost of utilizing the data has shot up” 8
- 9. Intended for Knowledge Sharing only Quick recap of what it is Intended for Knowledge Sharing only Taking a step back
- 10. DATA REALLY HAS GOTTEN BIG – VOLUME, VARIETY, VELOCITY & VERACITY Each of the data source is critical either across all or multiple functions…. Intended for Knowledge Sharing only …and are consumed either as reports, analytical deep dive insights, forward looking projections, etc. TRANSACTION DATA CLICK STREAM DATA (MOBILE & WEB) SENTIMENT/SOCIAL DATA • Are overall txns going up/down; where the txns are happening, etc.. • How are Consumers interacting with the website/app – drop-offs, clicks, Time spent, etc.. • Social Media, NPS surveys, Media mentions helps in gauging true Consumer reactions DATA SOURCES TYPES OF INSIGHTS SERVER LOGS DATA • How are consumers reacting with various functions on the front end? LOCATION DATA • Are consumers using the product in-store or on the move? PROMOTIONS DATA • How are consumers reacting to various marketing campaigns? INDUSTRY DATA • Benchmarking against industry performance 10
- 11. EVERYONE NEEDS DATA… Intended for Knowledge Sharing only How are we doing today? BI Where will be tomorrow? What if we do this? What can we do? ANALYTICS Did the initiative work? A/B TESTING How do Customers feel about us? USER RESEARCH Where should we invest? STRATEGY 11
- 12. …AND DISTRIBUTED DATA SYSTEMS HAD THEIR OWN ISSUES Intended for Knowledge Sharing only Inconsistent (and/or conflicting) definitions of data and numbers Varying granularities Multiple methodologies Different BU = (different KPIs or same KPIs different priorities) Lack of visibility/understanding outside of the BUs “Slow & inefficient, Non-scalable, Difficulties rolling up, Trust issues, Cascading mistakes” 12
- 13. AND IT THEN JUST HAPPENED… Intended for Knowledge Sharing only TRANSACTION DATA CLICK STREAM DATA (MOBILE & WEB) SENTIMENT DATA DATA SOURCES SERVER LOGS DATA LOCATION DATA CAMPAIGN DATA INDUSTRY DATA Source: http://www.adamadiouf.com/2013/03/22/bigdata-vs-enterprise-data-warehouse/ As if all prayers were answered Hadoop arrived in a big way & poof all problems seemed to disappear… 13
- 14. Intended for Knowledge Sharing only Quick recap of what it is Intended for Knowledge Sharing only A stroke of luck or was it?
- 15. WE FOCUSED ON OUR SPOUSE BUT FORGOT THE IN-LAWS… Inform Reports on KPIs with high level drilldowns Act Deep dives via Business Analytics Predict Identify Causal relationships via Advanced Analytics Optimize Experiments to verify which one works via A/B Testing Maturity phases of Analytics Practice ValueAddition Intended for Knowledge Sharing only Mine Machine Learning Focus on the 20% Data consumers (Reports) and assumption was that 80% Data Consumers will either love it or at least figure it out… 5% 50% 15% 20% 10% 15
- 16. HIGH DEVELOPMENT/MODIFICATION COSTS Intended for Knowledge Sharing only Rigid Structure and scale of operations make dynamism difficult… 16 Data Modeling/Schema ETL; Metadata Raw Data
- 17. NOT ONLY IS THE AUDIENCE CHANGING… Intended for Knowledge Sharing only Stakeholders Needs Reports, Insights & Drilldowns Datamart Documentation Executives - Reports - High level drilldown - Unified summary - “On the go*” Marketing & PR - Campaign performance - Infographics - Deep dives - Testing Sales / RM - Sales performance - Prospecting - Competitive - Infographics Product - Product performance - Deep dive - Mining - Testing - Research Technology / AE / Operations - Platform performance - Deep dive - Forecasting - Real time alerting FP & A - Consolidated Initiative readouts (E2E) - Deduping - Drill downs - Forecasting 17
- 18. …BUT ALSO THE NEEDS ARE EVER CHANGING Intended for Knowledge Sharing only “In mail” Recommendations with supporting graphs, tables, etc. “Story Deck” Full deck with the pitch and supporting arguments, numbers, graphs, charts “On-the-go” -Mobile App, On the Cloud, Subscriptions -Reports, Dashboards, Infographics Algorithm/Model Ready to be deployed How to decide? Customer needs; Turnaround Speed; One time/reuse; Deployment on Front end; Strategic Doc; Quick read/research doc 18
- 19. Intended for Knowledge Sharing only Quick recap of what it is Intended for Knowledge Sharing only Getting to the point – what do we propose?
- 20. WE BRING TO YOU THE SCALABLE METRICS MODEL (SMM)… EDW Aggregated Cubes Every attempt to bring the best of the most used models… 20 ACID, Fast, Stable Rigid, Cost, Resourcing Scalable Metrics Model (Pre-Aggregated Metrics + Primary- Foreign Keys) Cost, Flexibility, Scalability Performance, Reliability Performance, Easy to understand Reporting only
- 21. TACTICAL DETAILS: WHERE DO WE START? An illustrative example from Retail domain… 21 • Defined Granularity & associated Info: Determined by Core Objectives, e.g., Customer level table for Customer Engagement team • Defined Foreign Keys & Common Dimensions: For extensibility • Defined Metrics: KPIs as required • Identify Value Add Metrics: recommendation, forecasting etc CUSTOMER •Primary Key: Customer id •Foreign Keys: Sign Up Partner, Promotion Id, First Txn id •Customer Level Info: Email, Phone, Number, Geo, etc. •Metrics: •Lifetime Spend, Txns •Behavioral Bucket •RFM Bucket •Recommended Action items: •Next Best Product •CLV •Target Offers •Call Center Agent Reco
- 22. TACTICAL DETAILS: DATA MODEL An illustrative example from Retail domain… 22 id Dimensions foreign_keys metrics Customer_id Name Email Address,etc. signup_partner_id promotion_id Lifetime Spend, Txns Behavioral Bucket RFM Bucket Recommended Action items: Next Best Product CLV Target Offers Call Center Agent Reco 11234 {"name":"John", "Email" : "john@email.co m" , "Address":"123 nowhereblvd"} {"signup_partner_id ":"666YYY", "promotion" : "YAH123" } {"Lifetime Spend":"3400", "Txns":"150", "Behavioural Bucket" : "repeat user" , "RFM Bucket":"", "recommended Product id":"PRD789", "CLV":"??", "Target Offer":"OFF789", "CallCenterAgentReco":"1234"} WhatitcontainsSampledata
- 23. TACTICAL DETAILS: ETL FRAMEWORK An illustrative example from Retail domain… 23 STEP I: QUERIES STEP II: FRAMEWORK RUNS •Write separate queries/code to get metrics on the defined granularity •Put those queries into the framework STEP III: IMPLEMENT MODULARITY STEP IV: USER INTERFACE •Adding a new metric is just adding a new query/code for that metric alone •Can change an existing logic for a metric will impact that metric alone •Create physical impala tables for interactive querying •Create views for abstraction and end-user access •Exporting data to reporting tools like Tableau/QlikView brings a high level of analysis capability to this model. •Framework runs each of these queries and populate respective keys
- 24. ETL framework • Divide and conquer – Write separate queries/code to get metrics on the defined granularity – Put those queries into the framework • Framework runs each of these queries and populate respective keys • Modularity – Adding a new metric is just adding a new query/code for that metric alone – Can change an existing logic for a metric will impact that metric alone 24
- 25. Reporting and presentation • Map data-types are hard for the users for access • Three options – Create physical impala tables for interactive querying – Create views for abstraction and end-user access • Reporting layer (like Tableau) – Brings a different level of accessibility and analysis capability to this model. • Faster (if data is cached) • Create report level calculations • Data blending • Using metrics as a dimension – like customer buckets on transaction size • Visualization 25
- 26. DATA BUS EXTENSIBILITY CUSTOMER •Primary Key: Customer id •Foreign Keys: Sign Up Partner, Promotion Id, First Txn id •Customer Level Info: Email, Phone, Number, Geo, etc. •Metrics: •Lifetime Spend, Txns •Behavioral Bucket •RFM Bucket •Recommended Action items: •Next Best Product •CLV •Target Offers •Call Center Agent Reco SELLERS •Primary Key: Seller id •Foreign Keys: Product id, Operating Channel •Customer Level Info: Name, Operating Region, Annual Sales •Metrics: •Lifetime Sales, Txns •Performance Bucket •Special Category Flag •Recommended Action items: •Next Best Product •Next Co-Marketing •RM action TXNS •Primary Key: Txn id •Foreign Keys: Custid, Sellerid, Channel, •Txn Level Info: Amt, Type, Date, •Flags: •Buyer/Seller Type •Deviation Metrics •Fraud/Good •Agent Verification •Next Best Offer CLICKSTREAM PROMOTIONS PARTNERS PRODUCTS SENTIMENT LOGS 3rd PARTY ETC ETC… Common Dimensions or Foreign Keys
- 28. THE SALIENT FEATURES 28 • Fit for wide variety of Solution Sets & audiences: Optimal data model to support all three needs – Reporting, Analytics & Data Mining. • Best of all worlds: Scalable Metrics Model is a hybrid approach, • ACID Strengths: performance, stability and reliability of RDBMS. • Non ACID Strengths: scalability, flexibility, versatility of Hadoop. • Needs Optimized Model: Highest premium is provided to needs of the user – easy to incorporate changes as they come along (view like). Refresh cycle is easy and changed logics easily get incorporated in the next run. • Data Governance & Lineage: Operates with a modular approach – break down complex problems into smaller items and integrate in a bigger scheme of things. This eases better Data Governance and Lineage. • Extensibility: • Caching: Easy integration with buffering technologies to optimize on performance. • Visualization: Easier integration with visualization tools like Tableau. • Coding Interface: Additional drilldowns, analyses, data analysis via HIVE/SAS/R. ● MODULAR ● EXTENDABLE ● UPDATABLE ● SCALABLE
- 29. FOUR DIMENSIONS OF SUCCESSFUL EXECUTION 29 PEOPLE • Business Analysts: Details on Business needs like Timing(Immediate/ near/medium/long term), Priority (Critical/Urgent/Important/Good to have), Frequency (Regular/once-in-a-while/rare), Real-time, Delivery & Users. • Technical Architects: Understand the raw data structure, flow mechanisms & pipelines, security/legal/storage/resourcing constraints, feasibility assessments. PROCESS • Matching & Gap Analysis: Is the technology available to handle all business needs (possible/not enough RoI/deferred); Contingency, resourcing & budgeting. • Project Planning: Milestone based delivery, Deep Stakeholder involvement in development & validation, Communications Management • Execution: Schema on read efficient, Aggregates, Tight Metadata, reporting/analytics layer, Tables/Partitions/File types/Compression, Metadata TECH • PIG: ETL • HIVE/Impala: Schema & Table creation • Java/Streaming: • SAS/Python/R: Statistical Modeling CULTURE • Customer Needs Focused • Need for a smart vision, sound planning and able change management • Outcome Focused Organization (common business goal) • SAS/Python/R: Statistical Modeling
- 30. WHY DO WE THINK THE TIME IS NOW? Evolution in the value prop of Analysts: What/where/how much -> what can happen ->what should we do ? Audience has broadened (A numbers middle man -> Front line Managers) Luxury of time has evaporated Nature of questions have drastically changed (Expectation of being able to connect the dots in “Data Lake” world). Overselling potential before getting “there” 30 KPI of Analytics has changed from Turn-Around-Time (TAT) to Time-to- Action (TTA)
- 31. Intended for Knowledge Sharing only Quick recap of what it is Intended for Knowledge Sharing only Putting it all together
- 32. SWOT ANALYSIS OF SMM STRENGTHS OPPORTUNITIES WEAKNESSES THREATS • Need sensitive model • Cost of development, modification & refresh reduced • Easy for Analysts/End Users to understand and play with • Data Governance & Lineage: Break down bigger problems into smaller manageable • Integration with front end tools that can simplify UX. • Tools that buffer the backend data to ensure speedy delivery. • Good vision of future Analytical requirements is paramount. • Full refresh every time it runs again. • Maximum granularity needs to be pre- fixed. • Learning Curve on Coding language/syntax. • Non-normalized data model. • Not for real-time insights delivery • No Slowly Changing Dimensions 32
- 33. THE FIVE COMMANDMENTS 33 • “Know” that it caters to most frequent and not all needs. • “Must have” as good & farther as possible Analytics vision/needs and Outcome Focused approach. • “Ensure” Deeper Stakeholder involvement in the development. Test & Learn approach must. And be ready to modify if needed. • “Develop” modularity in delivery. • “Prepare” for ever more increasing dependencies from Analytics and other stakeholders.
- 34. Intended for Knowledge Sharing only Quick recap of what it is Intended for Knowledge Sharing only Appendix
- 35. THANK YOU! Intended for Knowledge Sharing only Would love to hear from you on any of the following forums… https://twitter.com/decisions_2_0 http://www.slideshare.net/RamkumarRavichandran https://www.youtube.com/channel/UCODSVC0WQws607clv0k8mQA/videos http://www.odbms.org/2015/01/ramkumar-ravichandran-visa/ https://www.linkedin.com/pub/ramkumar-ravichandran/10/545/67a https://www.linkedin.com/in/dataisbig http://bigdatadw.blogspot.com/ BHARATHIRAJA CHANDRASEKHARAN RAMKUMAR RAVICHANDRAN 35
- 36. 36 RESEARCH/LEARNING RESOURCES Intended for Knowledge Sharing only • Alternative approach by Martin Fowler: http://martinfowler.com/bliki/DataLake.html • Teradata/Hortonworks Data Lake Whitepaper: http://hortonworks.com/wp- content/uploads/2014/05/TeradataHortonworks_Datalake_White-Paper_20140410.pdf • Teradata/Hortonworks Data Lake Whitepaper: http://hortonworks.com/wp- content/uploads/2014/05/TeradataHortonworks_Datalake_White-Paper_20140410.pdf • EMC Data Lake: https://www.youtube.com/watch?v=o2fs02h_LEo 36