Performance Tales of Serverless - Mikhail Shilkov - Codemotion Rome 2018
•Als PPTX, PDF herunterladen•
0 gefällt mir•51 views
One promise of Function-as-a-Service model is the ability to scale without limits, up or down, whenever needed. But how does that work in practice? Can AWS Lambda handle thousands of messages per second? How fast can Azure Functions scale up under sudden heavy load? What kind of latency can you expect from Google Cloud Functions? In this session we will establish several performance benchmarks and go through the results of running them on the services of major cloud providers. I will also suggest practical steps to evaluate whether your application profile is suitable for serverless today.
Empfohlen
Empfohlen
Weitere ähnliche Inhalte
Ähnlich wie Performance Tales of Serverless - Mikhail Shilkov - Codemotion Rome 2018
Ähnlich wie Performance Tales of Serverless - Mikhail Shilkov - Codemotion Rome 2018 (20)
Mehr von Codemotion
Mehr von Codemotion (20)
Kürzlich hochgeladen
Kürzlich hochgeladen (20)
Performance Tales of Serverless - Mikhail Shilkov - Codemotion Rome 2018
- 1. @MikhailShilkov Performance Tales of Serverless by Mikhail Shilkov ROME - APRIL 13/14 2018
- 5. @MikhailShilkov Pay per actual usage Abstraction of servers $
- 6. @MikhailShilkov 1 user 100 users 10k users 1 mln users
- 7. @MikhailShilkov 1 user 100 users 10k users 1 mln users ResponseTime
- 9. @MikhailShilkov “ “ @MikhailShilkov JamesGovernor,Analyst &Co-founderat RedMonk We’re going to see the first billion-dollar startup with a single employee, the founder, and that engineer will be using serverless technology
- 10. @MikhailShilkov
- 12. @MikhailShilkov
- 13. @MikhailShilkov
- 14. @MikhailShilkov Occasional Users • Just launched • Several visitors per hour Tale 1
- 15. @MikhailShilkov 0 1 0:00 1:00 2:00 3:00 4:00 5:00
- 17. @MikhailShilkov 0 50 100 150 200 250 We want to be this fast
- 21. @MikhailShilkov
- 22. @MikhailShilkov 0 1 0:00 1:00 2:00 3:00 4:00 5:00 Cold starts
- 23. @MikhailShilkov Lessons Learned • Understand the runtime • Cold Starts are slow(-er) • Mind language and package size #1
- 24. @MikhailShilkov
- 25. @MikhailShilkov Steady Growth • Same serverless web app • More features • More calls • The app is getting traction! Tale 2
- 26. @MikhailShilkov 0 5 10 15 20 25 30 35 40 0:00 1:00 2:00 3:00 4:00 5:00
- 29. @MikhailShilkov
- 30. @MikhailShilkov Lessons Learned Serverless scalability is limited by non-serverless dependencies Limit the concurrency to avoid dependency Denial-of-Service #2
- 31. @MikhailShilkov
- 32. @MikhailShilkov Extract- Transform- Load • Load large set of data from external system • Convert it to appropriate format • Load to NoSQL cloud database • One time massive load + ongoing low-load process Tale 3
- 33. @MikhailShilkov
- 34. @MikhailShilkov 0 100000 200000 300000 400000 500000 600000 700000 800000 900000 1000000 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 #messagespending minutes AWS Azure GCP
- 35. @MikhailShilkov 0 100000 200000 300000 400000 500000 600000 700000 800000 900000 1000000 0 1 2 3 4 5 6 7 8 9 10111213141516171819202122232425262728 1024 MB 512 MB 256 MB 128 MB
- 36. @MikhailShilkov Lessons Learned Serverless is not just HTTP: Async workflows suit even better Learn resource model: Limits, Scaling model, RAM/CPU Allocation #3
- 37. @MikhailShilkov
- 38. @MikhailShilkov TV Effect • You migrated to scalable database (just in time!) • Your moderately popular website is mentioned on a prime-time TV show Tale 4
- 39. @MikhailShilkov 0 100 200 300 400 500 600 700 800 0:00 0:10 0:20 0:30 0:40 0:50 1:00
- 40. @MikhailShilkov 0 50 100 150 200 250 300 350 400 0:00 0:10 0:20 0:30 0:40 0:50 1:00 Requests 50th 90th 95th
- 41. @MikhailShilkov 0 200 400 600 800 1000 1200 1400 1600 1800 2000 0:00 0:10 0:20 0:30 0:40 0:50 1:00 Requests 50th 99th
- 42. @MikhailShilkov 0 50 100 150 200 250 300 0:00 0:10 0:20 0:30 0:40 0:50 1:00 Requests 50th 90th 95th 99th
- 43. @MikhailShilkov 0 200 400 600 800 1000 0:00 0:10 0:20 0:30 0:40 0:50 1:00 Success Errors
- 44. @MikhailShilkov 0 100 200 300 400 500 600 700 800 900 0 100 200 300 400 500 600 700 800 900 1000 0:00 0:10 0:20 0:30 0:40 0:50 1:00 Requests 50th 90th 95th 99th
- 45. @MikhailShilkov Lessons Learned Serverless is uniquely suitable for spiky load But know the limits of your account #4
- 46. @MikhailShilkov
- 47. @MikhailShilkov Load Testing How do we load test? With Serverless! Tale 5
- 48. @MikhailShilkov
- 49. @MikhailShilkov
- 50. @MikhailShilkov Lessons Learned Be creative in usage scenarios Serverless app can load-test or DDoS any “mere mortal” service #5
- 51. @MikhailShilkov
- 52. @MikhailShilkov • Throughput is probably enough • Should not be critical to latency • There is no SLA • Don’t forget tooling, architecture, pricing
- 53. @MikhailShilkov • Prototypes • Low traffic apps • Cloud Glue • Unusual load profiles
- 54. @MikhailShilkov
Hinweis der Redaktion
- Abstraction of servers, infrastructure and configuration of operating system Event-driven scale Sub-second billing Stateless Serverless compute is a fully managed service. Some refer to it as Functions as a Service OS and Framework patching is performed for you There is zero administrative tasks and no need to manage any infrastructure You just deploy your code (function) and it runs Your code runs within seconds and for very short period of time Serverless compute scales quickly (almost instantly) and vastly Automatically scales within seconds No scale configuration is required (there is no way to configure scale or limits) Scales to match any given workload. Scales from zero to handle tens of thousands concurrent functions invocations within seconds Pay only for the time your code is running Serverless compute reacts to events React, in near real-time, to events and triggers Triggered by virtually any event from Azure service or 3rd party services Setup time, provisioning is long & costly
- Abstraction of servers, infrastructure and configuration of operating system Event-driven scale Sub-second billing Stateless Serverless compute is a fully managed service. Some refer to it as Functions as a Service OS and Framework patching is performed for you There is zero administrative tasks and no need to manage any infrastructure You just deploy your code (function) and it runs Your code runs within seconds and for very short period of time Serverless compute scales quickly (almost instantly) and vastly Automatically scales within seconds No scale configuration is required (there is no way to configure scale or limits) Scales to match any given workload. Scales from zero to handle tens of thousands concurrent functions invocations within seconds Pay only for the time your code is running Serverless compute reacts to events React, in near real-time, to events and triggers Triggered by virtually any event from Azure service or 3rd party services Setup time, provisioning is long & costly
- Pre-load data from external source to elastic database (e.g. Dynamo / Cosmos / Bigtable) Limit the concurrency of functions to external services