Serverless Patterns
•Download as PPTX, PDF•
2 likes•283 views
Describes common serverless patterns with a little demo/screenshot. Most of the content are based in AWS Lambda, but should apply to other service providers as well. Best for developers new to cloud computing and serverless pattern. Presented at AWS Connect Taipei, Apr. 27th.
Report
Share
Serverless Patterns
- 1. Serverless Patterns Cliff Chao-kuan Lu AWS User GroupTaiwan Apr. 27th,AWS Connect TPE
- 2. Cliff Chao-kuan Lu •Sr. Architect @ 104 Corp. •AWS Community Hero •Organizer, AWS User GroupTaiwan •Certified AWS ALL-5 •Loves Cloud Computing
- 3. Serverless
- 4. Serverless Definitions •Operation •20 ms / 0.5 s •PaaS •Pricing model •Pay per call •Architecture •PaaS for (computational) resources
- 5. Lambda ECS Function-as-a-Service Docker Scheduler Infrastructure Managed EC2 Instances Housekeeping Concurrency Auto Scaling Healthcheck Reservation / Spot Access Control IAM / VPC Monitoring Cloudwatch / Cloudwatch Logs vs.
- 7. Cost Comparison •t2.nano: $0.008 •Lambda – 128 mb / 100 ms: •36000 Calls – $0.0072 •450 GB-s – $0.0075 •Managed service = Highly available + Scalable * 2 (scaling factor) * 0.25 (Spot) * 0.7 (RI) / 5% (CPU Credit)
- 8. Common Considerations •Capability •EC2 term management •ASG Guru •Design •Tenancy •Limits
- 9. Lambda Runtime Limits •Time • Max timeout = 300 s* •Environment • Max RAM = 1,536 MB • Max Storage = 500 MB** •Payload: 6 MB •Concurrency: 100 600 (per account-region)
- 10. Lambda Event Source and Retry •Stream-based Source •Synchronous invocation •Asynchronous invocation
- 11. Lambda Event Source •Stream-based Source •Kinesis, DynamoDB Stream •Others •Synchronous invocation • API Gateway, SDK •Asynchronous invocation • Cloudwatch Events
- 12. Lambda Permission Model •Stream-based Source (Pull) •Kinesis, DynamoDB Stream •Others (Push) •Synchronous invocation • API Gateway, SDK •Asynchronous invocation • Cloudwatch Events
- 13. Lambda Retry on Error •Stream-based Source •Cursor-based reads, spontaneous retry •Synchronous invocation •Client retry •Asynchronous invocation •Retry twice -> DLQ {SQS, SNS}
- 14. Serverless Framework •FLOSS (MIT) •Led by Austen Collins / full-time team •Under active development •Written in Node.js •Multiple programming languages
- 18. Serverless-PHP
- 19. Common Lambda Patterns •Direct invocation •Crontab •Event-driven •Streaming •Streaming -> micro-batch
- 20. Crontab •CloudWatch Event: Schedule •Minimal interval: 1min cron(fields) Minutes 0-59 , - * / Hours 0-23 , - * / Day-of-month 1-31 , - * ? / L W Month 1-12 or JAN-DEC , - * / Day-of-week 1-7 or SUN-SAT , - * ? / L Year 1970-2199 , - * /
- 21. Crontab Use Cases •Health checks •Crawler •Recurring reports
- 23. webmon Metrics
- 25. Common Lambda Patterns •Direct invocation •Crontab •Event-driven •Streaming •Streaming -> micro-batch
- 26. Event-Driven •Base of all patterns •Triggered by CloudWatch Events •Supports EBS, EC2, EMR,AutoScaling, CodeDeploy, Console Sign-in, Health, KMS,Trusted Advisor, … •`PutEvent`
- 27. Event-driven Image Bucket 1. Upload image Lambda 4. Fetch image CloudWatch Event 2. 3. 5. Convert thumbnail 6. Store thumbnail
- 28. Event-driven APITier DynamoDB Table 1. Request upload 2. S3 Signed URL 3. Put metadata Workers Source Bucket 4. Upload data Lambda 5. PutObjcet completed 6.Verify metadata Queue 7. Schedule conversion Destination Bucket 8. Poll for work 9. Convert 10. Upload
- 29. Common Lambda Patterns •Direct invocation •Crontab •Event-driven •Streaming •Streaming -> micro-batch
- 30. Streaming •Polls from Kinesis / DynamoDB stream •Invoke function if data persists •Scales with source stream •Spontaneous retry
- 33. Common Lambda Patterns •Direct invocation •Crontab •Event-driven •Streaming •Streaming -> micro-batch
- 34. Streaming -> Micro-Batch •Data aggregation •Reduce volume •Windowed analysis •Reasonably high frequency •Reduce latency •Preserve features
- 35. Streaming -> Micro-Batch •CloudWatch Event: Schedule CloudWatch Event: Schedule Lambda Kinesis Stream 1. 2. 3. S34.
- 36. Streaming -> Micro-Batch •CloudWatch Event: Schedule CloudWatch Event: Schedule Lambda Kinesis Stream 1. 3. S34. 2. KinesisAnalytics
- 37. Gotcha •Lambda scales. How about your resources? •Connection pool •EC2-based resources that scales much slower •Throttle, queue, and retry
- 38. Dispatcher /Worker Pattern Queue Dispatcher CloudWatch Event: Schedule 1. 2. Poll for task Workers Protected Resources 4.
- 39. Gotcha •How do I know if I’m throttled?
- 40. Gotcha •How do I know if I’m throttled? •Feature request: Per-function concurrency cap!
- 41. Gotcha •How to keep State among calls •Externalize •DynamoDB / S3 •Step Functions
- 42. Gotcha •Elevated latency after deployment •A-B deployment withTags •Lazy resource initialization
- 43. Take Home Message •Lambda is a glue service •Event-driven and streaming •Public endpoint <->VPC / EC2 •The only limit is your imagination
- 45. ThankYou
Editor's Notes
- PaaS: Automation, Elasticity, Monitoring, Network, Scalability according to NIST Defn
- Lambda Request: $2E-7 calls Duration: $1.667E-5 / GB-s
- Workarounds: Divide and conquer * InvokeAsync or Step Function ** Pack static files with function
- Pattern 之間不互斥
- 最基本的模式，one-off jobs Cron 的寫法與 crontab 稍有不同，DoM 與 DoW 互斥，提供一者時，另外一項要寫 ?
- Pattern 之間不互斥
- Pattern 之間不互斥
- Poll interval: 1s 1 process per shard
- Pattern 之間不互斥
- 透過 Schedule 觸發，可自定 Interval 與輸出資料量；但仍有
- 透過 Schedule 觸發，可自定 Interval 與輸出資料量；但仍有
- Black belt question. Especially after deploying new versions. * A-B Deploy: leverage container reuse