Beyond Linear Scaling: A New Path for Performance with ScyllaDB
1 like333 views
ScyllaDB offers a high-performance NoSQL database with over 5x higher throughput and 20x lower latency compared to traditional solutions, providing significant cost savings for data-intensive applications. The architecture supports seamless scalability and performance optimization through techniques like consistent hashing and adaptive concurrency limits. Real-world applications demonstrate ScyllaDB's ability to handle massive workloads effectively, making it a viable choice for various industries including gaming, retail, and healthcare.
Beyond Linear Scaling: A New Path for Performance with ScyllaDB
- 1. Felipe Mendes, Solution Architect at ScyllaDB Beyond Linear Scaling A New Path for Performance with ScyllaDB
- 2. + For data-intensive applications that require high throughput and predictable low latencies + Close-to-the-metal design takes full advantage of modern infrastructure + >5x higher throughput + >20x lower latency + >75% TCO savings + Compatible with Apache Cassandra and Amazon DynamoDB + DBaaS/Cloud, Enterprise and Open Source solutions The Database for Gamechangers 2 “ScyllaDB stands apart...It’s the rare product that exceeds my expectations.” – Martin Heller, InfoWorld contributing editor and reviewer “For 99.9% of applications, ScyllaDB delivers all the power a customer will ever need, on workloads that other databases can’t touch – and at a fraction of the cost of an in-memory solution.” – Adrian Bridgewater, Forbes senior contributor
- 3. 3 +400 Gamechangers Leverage ScyllaDB Seamless experiences across content + devices Digital experiences at massive scale Corporate fleet management Real-time analytics 2,000,000 SKU -commerce management Video recommendation management Threat intelligence service using JanusGraph Real time fraud detection across 6M transactions/day Uber scale, mission critical chat & messaging app Network security threat detection Power ~50M X1 DVRs with billions of reqs/day Precision healthcare via Edison AI Inventory hub for retail operations Property listings and updates Unified ML feature store across the business Cryptocurrency exchange app Geography-based recommendations Global operations- Avon, Body Shop + more Predictable performance for on sale surges GPS-based exercise tracking Serving dynamic live streams at scale Powering India's top social media platform Personalized advertising to players Distribution of game assets in Unreal Engine
- 4. Introductions Felipe Mendes, Solution Architect at ScyllaDB + Published Author on Linux and Databases + Helps teams solve their most challenging problems + Years of experience with Linux and distributed systems
- 5. Agenda + (Near) Linear Scaling + Enter Real-life + ScyllaDB under Load + Crafting Your Success + Beyond Linear Scaling
- 6. 6 (Near) Linear Scaling Why is it important … And when you shouldn't care :-)
- 7. 7 Linear Speedup Main goal is to run programs faster + To a point… + Measured as + Reasons for sub-linear speedup: + Laws! (Amdahl's, Gustafson-Barsis) + Task Management + Communication & Synchronization 15.2 Performance in Practice Ideal, typical, and super-linear speedup curves
- 8. Universal Scaling Law Generalization of Amdahl’s Law discovered by Dr. Neil Gunther. As number of users (N) increases, the system throughput (X) will: + Enjoy a period of near linear scaling + Eventually saturate some resource such that increasing N doesn’t increase X. This defines maxX + Possibly encounter a coordination cost that drives down X with further increasing N Saturation Region Linear Region Retrograde Region maxX How Optimizely (Safely) Maximizes Database Concurrency
- 9. Linear Scaling – Good Relevant for parallel programming, useful for measuring: + Database efficiency + Price-performance + Scalability NoSQL Benchmark: MongoDB vs ScyllaDB 9
- 10. Doesn't account for: + Improvements Over Time + Application Semantics + Hotspots + Scaling Clients + Consistent Hashing Uneven Distribution + Communication Overhead More on propagating state (and image credits): Gnutella: an Intro to Gossip Linear Scaling – Bad Gossip propagation 10
- 11. 11 Enter Real-life Overlooked considerations no one (dares) to tell you ;-)
- 12. 12 Application Semantics 1,000,000 sensors, representing homes in an area IOT Social DynamoDB: When to Move Out?
- 13. 13 Consistent Hashing Exercise: How much more traffic and load does this node receive? Alexys Jacob – Leveraging consistent hashing in your python applications thelastpickle – The Impacts of Changing the Number of VNodes in Apache Cassandra Avi Kivity's shard simulator Bad Better, but not perfect
- 14. Adding more nodes won't help Hotspots 14
- 15. How Discord Stores Trillions of Messages Performance Under Load – Adaptive Concurrency Limits Challenges: + For a system serving X static clients, what's the max effective concurrency to set on a single client? + When scaling clients, how to coordinate them to avoid overwhelming a group of replicas? Scaling Clients 15 Discord consistent hash-based routing DB Calls Netflix Adaptive Concurrency
- 16. 16 ScyllaDB Under Load Quantifying the Performance Impact of a Shard-per-Core Architecture
- 17. 17 ScyllaDB Architecture Dor Laor on P99 CONF: Quantifying the Impact of Shard Per Core Architecture
- 18. Linear Scale Ingestion Constant Time Ingestion 2X 2X 2X 2X 2X 18
- 19. 2X 2X 2X 2X 2X “Nodes must be small, in case they fail” No they don’t! {Replace, Add, remove} Node at constant time 19
- 20. Compaction Scale 2X 2X 2X 2X 2X 20
- 21. 21 Crafting Your Success Do's and Don'ts
- 22. 22 In a Nutshell... Database Performance At Scale
- 23. 23 Run Real Tests Benchmark tools prove you can get there, but: + Application semantics are unique + Access patterns are unique + Real-life tooling is also unique + Addressing all corner-cases is time-consuming or even impossible + Don't just blindly assume 2x will give you 2x load
- 24. 24 Eliminate Noise Avoid large deployments of small nodes + Go Big or Go Home! + Considerably reduces the overhead associated with communication & synchronization + Less resource overcommitment + BUT, keep balance: + Account for inevitable failures + Leave room for unpredictability
- 25. 25 Tune the client side Understand your data flows: + Can multiple clients spam a single key? + What happens when scaling the number of clients? + How is load balancing achieved? Power of Two Choices load balancing P99 CONF – Conquering Load Balancing: Experiences from ScyllaDB Drivers
- 26. 26 Beyond Linear Scaling Unveiling Performance Insights
- 27. 27 ScyllaDB in 2018 Ingestion time – Lower is better
- 28. 28 ScyllaDB in 2023 4 Ingestion time – Lower is better
- 29. 29 Linear Scale Ingestion (2023 vs 2018) Constant Time Ingestion 2X 2X 2X 2X 2X
- 30. 30 Getting Even Faster Time to execute, lower is better
- 31. 31 Going Beyond – Tablets tablet tablet replica tablet replica tablet replica replication metadata: (per table) Why ScyllaDB is Moving to a New Replication Algorithm: Tablets
- 32. 32 Going Beyond – ScyllaDB Enterprise Throughput – Higher is better
- 33. There's much more to performance beyond Linear Scale: + Goods and Bads of Linear Scaling + Real-life situations impacting linear scalability + ScyllaDB Shard-Per-Core Architecture + Run Realistic Workloads + How ScyllaDB drives the meaning of 'performance' 33 Summary
- 34. Q&A ScyllaDB Cloud Start free trial scylladb.com/cloud Feb 14-15 | VIRTUAL EVENT scylladb.com/summit Virtual Workshop January 25, 2024 scylladb.com/events
- 35. Thank you for joining us today. @scylladb scylladb/ slack.scylladb.com @scylladb company/scylladb/ scylladb/