Distributed Design and Architecture of Cloud Foundry
•
70 likes•9,511 views
In this session we will dig deep into Cloud Foundry's core architecture and design principles. We will discuss the challenges around scaling and operating a large scale service as we combined the PaaS and traditional IaaS layers, and how we achieve multiple updates per week to the system with no perceived downtime. Allowing user to download a single virtual machine that is a complete replica of the service presented some challenges as well, and we will discuss our approach to offering up the downloadable private cloud.
Recommended
Recommended
More Related Content
What's hot
What's hot (20)
Viewers also liked
Viewers also liked (20)
Similar to Distributed Design and Architecture of Cloud Foundry
Similar to Distributed Design and Architecture of Cloud Foundry (20)
More from Derek Collison
More from Derek Collison (8)
Recently uploaded
Recently uploaded (20)
Distributed Design and Architecture of Cloud Foundry
- 1. Design and Architecture Derek Collison
- 2. What is Cloud Foundry? 2
- 3. The Open Platform as a Service 3
- 5. Or more specifically, aPaaS? 5
- 6. aPaaS • Application Platform as a Service • Applications and Services 6
- 7. aPaaS • Application Platform as a Service • Applications and Services • Not • VMs • Memory • Storage • Networks • CPU 7
- 8. What is OpenPaaS? 8
- 9. OpenPaaS • Multi-Language • Multi-Framework • Multi-Services • Multi-Cloud, Multi-IaaS • Hybrid - Public or Private or Both • OpenSource 9
- 10. OpenPaaS • Multi-Language • Ruby, Java, Scala, Node.js, Erlang, Python, PHP.. • Multi-Framework • Rails, Sinatra, Spring, Grails, Express, Lift • Multi-Services • MySQL, Postgres, MongoDB, Redis, RabbitMQ • Multi-Cloud, Multi-IaaS • vSphere, MicroCloud, OpenStack, AWS 10
- 11. The Open PaaS Ap pli vFabric Private ca ce Postgres tio rfa Data Clouds n e Services Int Se Public rvi er vFabric vid ce RabbitMQTM Msg Services Clouds ro Int dP e Micro rfa ou Clouds ce Other Cl Services 11
- 12. What is our Goal? 12
- 13. What was our Goal? Raise the unit of currency to be the application and its associated services, not the infrastructure 13
- 14. What was our Goal? Best of breed delivery platform for all modern applications and frameworks 14
- 15. What was our Goal? Favor Choice and Openness 15
- 16. How was it Built? 16
- 17. How was it Built? • Kernel (CloudFoundry OSS) • Core PaaS System • Kernel and Orchestrator Shells • Layered on top of IaaS • Orchestrator • IaaS creation, management and orchestration 17
- 18. High Level Clients (VMC, STS, Browser) CF Kernel Orchestrator IaaS Hardware - CPU/Memory/Disk/Network 18
- 19. Basic Premises • Fail Fast • Self Healing • Horizontally Scalable Components • Distributed State • No Single Point of Failure • Should be as simple as possible 19
- 20. Basic Patterns • Event-Driven • Asynchronous • Non-blocking • Independent, Idempotent • Message Passing • Eventually Consistent 20
- 21. Basic Design • All components loosely coupled • Few “Classes”, many “Instances” • Messaging as foundation • Addressing and Component Discovery • Command and Control • JSON payloads • HTTP or File/Blob for data transport 21
- 22. Kernel Components • All dynamically discoverable • Launch and scale in any order • Can come and go as needed • Monitor via HTTP and JSON • Location independent 22
- 23. Kernel Components • Router • CloudController • DEA • HealthManager • Service Provisioning Agent • Messaging System 23
- 24. Logical View Browser VMC client STS plugin (user app access) Routers CloudControllers App App HealthManager Services DEA Pool Messaging 24
- 25. 25 Architecture
- 26. Messaging 26
- 27. Messaging “The Nervous System” 27
- 28. Messaging Browser VMC client STS plugin (user app access) Routers CloudControllers App App HealthManager Services DEA Pool Messaging 28
- 29. Messaging • Addressing and Discovery • No static IPs or DNS lookups req’d • Just Layer 4 • Command and Control • Central communication system • Dial tone, fire and forget • Protects *itself* at all costs • Idempotent semantics 29
- 30. Router 30
- 31. Router “Traffic Cop” 31
- 32. Router Browser VMC client STS plugin (user app access) Routers CloudControllers App App HealthManager Services DEA Pool Messaging 32
- 33. Router • Handles all HTTP traffic • Maintains distributed routing state • Routes URLs to applications • Distributes load among instances • Realtime distributed updates to routing tables from DEAs 33
- 35. CloudController “The King” 35
- 36. CloudController Browser VMC client STS plugin (user app access) Routers CloudControllers App App HealthManager Services DEA Pool Messaging 36
- 37. CloudController • Handles all state transitions • Deals with users, apps, and services • Packages and Stages applications • Binds Services to Applications • Presents external REST API 37
- 38. HealthManager 38
- 39. HealthManager “Court Jester” 39
- 40. HealthManager Browser VMC client STS plugin (user app access) Routers CloudControllers App App HealthManager Services DEA Pool Messaging 40
- 41. HealthManager • Monitors the state of the world • Initial value with realtime delta updates to “intended” vs “real” • Determines drift • Complains to the CloudControllers when something is not correct • No power to change state itself 41
- 42. DEA 42
- 43. DEA “Droplet Execution Agent” 43
- 44. DEA Browser VMC client STS plugin (user app access) Routers CloudControllers App App HealthManager Services DEA Pool Messaging 44
- 45. DEA (Droplet Execution Agent) • Responsible for running all applications • Monitors all applications • CPU, Mem, IO, Threads, Disk, FDs, etc • All apps look same to DEA • start and stop • Express ability and desire to run an application • runtimes, options, cluster avoidance, memory/cpu • Alerts on any change in state of applications • Provides secure/constrained OS runtime • Hypervisor, Unix File and User, Linux Containers* • Single or Multi-Tenant 45
- 46. How does it all Work? 46
- 47. Pushing an App • Client (VMC/STS) pushes meta-data to CC • Client optionally pushes resource signatures (diff analysis, sys wide) • Client pushes app resources to CC • CC puts app together • CC stages app asynchronously • CC binds and stages services • Droplet ready 47
- 48. 48 Architecture
- 49. Running an App • CC asks DEAs for “help” • First DEA back wins! Simple • CC sends start request to selected DEA • DEA pushes the “green” button • DEA waits and monitors pid and ephemeral port for app to bind • When app is healthy, sends “register” message • Register message is seen by HM and Routers • Routers bind URL to host:port 49
- 50. DEAs answer? • DEAs first determine YES or NO • correct runtime, options, memory, etc • Then calculate a Delay Taint • SHA hash of application • memory • cpu • Taint allows balancing and selection 50
- 51. Scale up & down? • Exact steps as running the app the first time • SHA1 taint helps avoid clustering • memory/cpu taint helps distribute as evenly as possible • Nothing pre-computed • Nothing assumed 51
- 52. Crashes? • If your app stops and we did not tell it to, that is a crash • Crashed apps are immediately detected by DEA and messaged • Routers disconnect route instantly • HM will signal CC • something is wrong • CC will issue run sequence again 52
- 53. 53 Architecture
- 54. Access to my App? • All routers understand where all instances of your application are running • Will randomly pick backend, not semantically aware. • Will remove routes that are stale or unhealthy • Session stickiness and replication available, but best to avoid if possible 54
- 55. What about Services? 55
- 56. Services Browser VMC client STS plugin (user app access) Routers CloudControllers App App HealthManager Services DEA Pool Messaging 56
- 57. Services • Service Advertisement • Service Provisioning • Gateway fronts multi-backends • Service Nodes scale independent • App and service talk directly • API to register into system • Closure for additional value 57
- 58. Provisioning VMC/STS 1 Routers 2 CloudControllers Services Gateway 3 5 6 4 Service Node Service Node Service Node Application MySQL Redis Redis Messaging 58
- 59. Access (Direct) Browser (user app access) 1 Routers CloudControllers Services Gateway Service Node Service Node Service Node Application MySQL Redis Redis 2 Messaging 59
- 60. Services VMware Dev Tools Partner Dev Tools Cloud Foundry consume Enterprise Services apps consume bind Data Director provision/bind service service broker controller SQLFire core services Relational DB vSphere 60
- 61. Learn more: www.cloudfoundry.org blog.cloudfoundry.com support.cloudfoundry.com 61
- 62. Thank You 62
- 63. Questions? dcollison@vmware.com derek.collison@gmail.com twitter: derekcollison 63