Scaling for Performance
Download as PPTX, PDF0 likes153 views
The document outlines a masterclass on high performance NoSQL databases, specifically focusing on ScyllaDB and its architecture, deployment strategies, and observability techniques. Key topics include understanding workload types (OLTP and OLAP), data modeling best practices, and performance optimization methods such as using log-structured merge trees for efficient data handling. Additionally, it emphasizes the importance of testing, readiness, and system monitoring for successful NoSQL implementation.
Scaling for Performance
- 1. High Performance NoSQL Masterclass Scaling for Performance Felipe Cardeneti Mendes
- 2. High Performance NoSQL Masterclass Felipe Cardeneti Mendes ● Solution Architect at ScyllaDB ● Published Author ● Linux and Open Source enthusiast
- 3. High Performance NoSQL Masterclass Agenda ● About ScyllaDB ● Getting started with NoSQL databases ● Deployment and Production Readiness ● Observability Tips and Tricks
- 4. High Performance NoSQL Masterclass About ScyllaDB
- 5. High Performance NoSQL Masterclass ScyllaDB Database Architecture Horizontal & Vertical Scaling Built in C++ (no Java overhead) System and Data Center Aware Sharding Per Core Shard-Aware Drivers Auto-Performance Tuning Network Processor NUMA Storage Unique Close-to-Metal Architecture
- 6. High Performance NoSQL Masterclass Shard per Core Threads Shards
- 7. High Performance NoSQL Masterclass Asynchronous Architecture Request Answer Request Answer Waiting for response Time Savings Synchronous architecture Asynchronous architecture
- 8. High Performance NoSQL Masterclass Specialized Cache Cassandra ScyllaDB Key cache Row cache Linux page cache SSTables Unified cache SSTables Complex Tuning On-heap / Off-heap
- 9. High Performance NoSQL Masterclass Ecosystem Compatibility + CQL native protocol + JMX management protocol + Management command line /REST + SSTable file format + Configuration file format + CQL language
- 10. High Performance NoSQL Masterclass Getting Started with NoSQL Databases
- 11. High Performance NoSQL Masterclass Modern Business Challenges Keep CapEx & OpEx in check Reduce complexity Scale as the data grows Queries in milliseconds Leverage massive amounts of data Predictable, consistent performance
- 12. High Performance NoSQL Masterclass Oh… The CAP Theorem
- 13. High Performance NoSQL Masterclass Workload Types
- 14. High Performance NoSQL Masterclass Workload Types Decision Support + More complex queries - large amounts + Latency important but not critical + Seconds to hours Fundamental business tasks + Simple queries + Latency critical + Milliseconds per transaction OLAP Time Complexity of Query Time Complexity of Query OLTP
- 15. High Performance NoSQL Masterclass OLAP Characteristics Main Characteristics + Bound Concurrency + Scans and aggregations + Rely on MapReduce paradigms OLAP Time Complexity of Query Examples + How many users are from the US? + How many Twitter posts happened in 2022? + Which devices haven’t communicated back within the past 1 hour?
- 16. High Performance NoSQL Masterclass OLTP Characteristics Main Characteristics + Unbound Concurrency + Designed for speed and simplicity + Often user facing APIs Examples + What’s the last time an user logged in? + What have been the last 10 temperature measurements for a given device? + How many likes a given posting has? Time Complexity of Query OLTP
- 17. High Performance NoSQL Masterclass Why not both? Meet Workload Prioritization 100 shares Ratio = 100:100 (1:1) means equal shares of processing/resources to complete tasks Ratio = 100:50 (2:1) means 2X as many shares of processing/resources for Transactions to complete tasks compared to Analytics 100 shares 100 shares 50 shares OLTP OLAP Which Task to Run
- 18. Wide Column Databases Write Path LSM storage engine’s write path: 18 Writes commit log
- 19. Wide Column Databases Write Path LSM storage engine’s write path: 19 Writes commit log
- 20. Wide Column Databases Write Path LSM storage engine’s write path: 20 Writes commit log
- 21. Wide Column Databases Write Path LSM storage engine’s write path: 21 Writes commit log compaction
- 22. Wide Column Databases Write Path LSM storage engine’s write path: 22 Writes commit log compaction
- 23. Wide Column Databases Write Path LSM storage engine’s write path: 23 Writes commit log
- 24. What is compaction? LSM storage engine’s write path: 24 Hidden Gems + This technique of keeping sorted files and merging them is well-known and often called Log-Structured Merge (LSM) Tree + Published in 1996, earliest popular application known is the Lucene search engine, 1999 Characteristics + High performance write. + Immediately readable. + Reasonable performance for read.
- 25. What is a compaction strategy? LSM storage engine’s write path: 25 ▪ Which files to compact, and when? ▪ This is called the compaction strategy ▪ The goal of the strategy is low amplification: ○ Avoid read requests needing many sstables. • read amplification ○ Avoid overwritten/deleted/expired data staying on disk. ○ Avoid excessive temporary disk space needs • space amplification ○ Avoid compacting the same data again and again. • write amplification
- 26. Which one to choose? LSM storage engine’s write path: 26 Know your workload
- 27. High Performance NoSQL Masterclass Deployment and Production Readiness
- 28. It all starts with Data Modeling LSM storage engine’s write path: 28 Do’s + Denormalize + Query oriented approach + High data distribution / cardinality Dont’s + Create hotspots + Large partitions/rows/cells/collections/etc + Low cardinality tables/views/indexes
- 29. Test, test, test…! LSM storage engine’s write path: 29 Unit Testing + Test your workload and access patterns in a Docker container + Use specialized stress tools to simulate workload ○ cassandra-stress ○ nosqlbench ○ YCSB + OBSERVE the results (more on that later)
- 30. Application Development LSM storage engine’s write path: 30 Functional Testing + Use Prepared Statements + Configure your routing and load balancing policy correctly ○ DCAware and TokenAware policies ○ ShardAware Drivers if using ScyllaDB! + Make use of Asynchronous APIs + Ensure your client is NOT a bottleneck + Paging is important: Ensure you adjust it right
- 31. Test, test, test…! LSM storage engine’s write path: 31 Readiness Testing + What’s the unreplicated data set size? + How many operations per second do I need to achieve? ○ Out of these, what are the reads vs writes distribution? ○ What’s the average payload size? ○ What are my latency requirements? + How many regions should it replicate to? + Do I need indexes or views to satisfy my queries? + What is/are the target deployment location(s)? + Is the use case growth predictable or unpredictable? + What are the data retention requirements? + Is the use case storage or CPU bound?
- 32. Oh mighty sizing…! LSM storage engine’s write path: 32 A Sizing Exercise + 500k ops/sec with 1KB rows + 5TB data set size + P99 reads and writes < 10ms + Target deployment region: AWS Simple Math + RF=3 / 6 nodes * 5TB = 2.5TB per node + 500k ops/sec / 12,5K ops/core = 40 physical cores + Result: 6 nodes of i4i.8xlarge
- 33. High Performance NoSQL Masterclass Observability Tips and Tricks
- 34. High Performance NoSQL Masterclass ScyllaDB Monitoring Architecture
- 35. High Performance NoSQL Masterclass ScyllaDB Monitoring Architecture
- 36. High Performance NoSQL Masterclass Keep in touch! Felipe Cardeneti Mendes Solutions Architect ScyllaDB felipemendes@scylladb.co m Find me on LinkedIn