Taming the Big Data Fire Hose John Hugg Sr. Software Engineer, VoltDB
Big Data Defined Velocity Volume Variety Moves at very high rates (think sensor-driven systems) Valuable in its temporal, high velocity state Volume Fast-moving data creates massive historical archives Valuable for mining patterns, trends and relationships Variety Structured (logs, business transactions) Semi-structured and unstructured
Example Big Data Use Cases Data Source High-frequency operations Lower-frequency operations Capital markets Write/index all trades, store tick data Show consolidated risk across traders Call initiation request Real-time authorization Fraud detection/analysis Inbound HTTP requests Visitor logging, analysis, alerting Traffic pattern analytics Online game Rank scores: Defined intervals Player “bests” Leaderboard lookups Real-time ad trading systems Match form factor, placement criteria, bid/ask Report ad performance from exhaust stream Mobile device location sensor Location updates, QoS, transactions Analytics on transactions Financial trade monitoring Telco call data record management Website analytics, fraud detection Online gaming micro transactions Digital ad exchange services Wireless location- based services
Big Data and You Big Data and You Incoming data streams are different than traditional business apps You need to write data quickly and reliably, but … It’s not just about high speed writes You need to validate in real-time You need to count and aggregate You need to analyze in real-time You need to scale on demand You may need to transact
Big Data Management Infrastructure High Velocity High Volume Analytic Datastore NewSQL Online gaming Ad serving Sensor data Internet commerce SaaS, Web 2.0 Mobile platforms Financial trade Structured data ACID guarantees Relational/SQL Real-time analytics Unstructured data Eventual consistency Schemaless KV, document Other OLAP data stores NoSQL
Big Data Management Infrastructure High Velocity High Volume Analytic Datastore NewSQL Online gaming Ad serving Sensor data Internet commerce SaaS, Web 2.0 Mobile platforms Financial trade Other OLAP data stores NoSQL
High Velocity Data Management
High Velocity DBMS Requirements Ingest at very high speeds and rates Scale easily to meet growth and demand peaks Support integrated fault tolerance Support a wide range of real-time (or “near-time”) analytics Integrate easily with high volume analytic datastores
High Speed Data Ingestion Support millions of write operations per second at scale Read and write latencies below 50 milliseconds Provide ACID-level consistency guarantees (maybe) Support one or more well-known application interfaces SQL Key/Value Document
Scale to Meet Growth and Demand Scale-out on commodity hardware Built-in database partitioning Manual sharding and/or add-on solutions are brittle, require apps to do “heavy lifting”, and can be an operational nightmare Database must automatically implement defined partitioning strategy Application should “see” a single database instance Database should encourage scalability best practices For example, replication of reference data minimizes need for multi-partition operations
A Look Inside Partitioning select count(*) from orders where customer_id = 5 single-partition select count(*) from orders where product_id = 3 multi-partition insert into orders (customer_id, order_id, product_id) values (3,303,2) single-partition update products set product_name = ‘spork’ where product_id = 3 multi-partition 1 101 2 1 101 3 4 401 2 1 knife 2 spoon 3 fork Partition 1 2 201 1 5 501 3 5 502 2 Partition 2 3 201 1 6 601 1 6 601 2 Partition 3 table orders : customer_id (partition key) (partitioned) order_id product_id table products : product_id (replicated) product_name
Integrated Fault Tolerance Database should transparently support built-in “Tandem-style” HA Users should be able to easily increase/decrease fault tolerance levels Database should be easily and quickly recoverable in the event of severe hardware failures Database should be able to automatically detect and manage a variety of partition fault conditions Downed nodes should be “rejoinable” without the need for service windows
Partition Detection & Recovery Network fault protection Detects partition event Determines which side of fault to disable Snapshots and disables orphaned node(s) Server A Server C Server B Live node rejoin Allows “downed” nodes to rejoin live cluster Automatically re-synchs all node data Coordinates transactions during re-synch Server A Server C Server B
Real-time Analytics Database should support a wide variety of high performance reads High-frequency single-partition Lower-frequency multi-partition Common analytic queries should be optimized in the database Multi-partition aggregations, limits, etc. Database should accommodate a flexible range of relational data operations Particularly relevant to structured data
Integration with Analytic Datastores Database should offer high performance, transactional export Export should allow a wide variety of common data enrichment operations Normalize and de-normalize De-duplicate Aggregate Architecture should support loosely-coupled integrations Impedance mismatches Durability
VoltDB Export Data Flow High Velocity Database Cluster Loosely-coupled, asynchronous Queue must be durable Bi-directional durability
Summary Big Data infrastructures will usually require more than one engine High velocity engine for “fast” data Analytic engine for “deep” data Data characteristics will often determine which high velocity engine to use NewSQL is often well-suited to structured data NoSQL is often a good fit for unstructured data Choose solutions that suit your needs and are designed for interoperability