Netflix Container Scheduling and Execution - QCon New York 2016
17 likes•14,378 views
The document presents insights from Netflix's approach to container management, highlighting the need for a consistent developer experience, simpler management of compute resources, and elastic resource pooling. It outlines the motivation for adopting containers despite having an established virtual machine-based architecture and discusses the development of their proprietary container management solution, Project Titus. Current and future workloads include batch processing, service style support, and plans for media encoding and stream processing as a service.
Netflix Container Scheduling and Execution - QCon New York 2016
- 1. Scheduling a Fuller House: Container Management Sharma Podila, Andrew Spyker - Senior Software Engineers
- 2. About Netflix ● 81.5M members ● 2000+ employees (1400 tech) ● 190+ countries ● > 100M hours watch per day ● > ⅓ NA internet download traffic ● 500+ Microservices ● Many 10’s of thousands VM’s ● 3 regions across the world 2
- 3. Agenda ● Why containers at Netflix? ● What did we build and what did we learn? ● What are our current and future workloads? 3 ⇨
- 4. Why a 2nd edition of virtualization? ● Given our resilient cloud native, CI/CD devops enabled, elastically scalable virtual machine based architecture, did we really need containers? 4
- 5. Motivating factors for containers ● Simpler management of compute resources ● Simpler deployment packaging artifacts for compute jobs ● Need for a consistent local developer environment 5
- 6. Simpler compute, Management & Packaging Batch/stream processing jobs ● Here are the files to run my process ● I need m cores, n disk, and o memory ● Please just run it for me! 6 Service style jobs (VM’s) ● Use tested/secure base AMI ● Bake an AMI ● Define launch config ● Choose t-shirt sized instance ● Canary & red/black ASG’s
- 7. Consistent developer experience ● Many years focused on ○ Build, bake / cloud deploy / operational experience ○ Not as much time focused on developer experience ● New Netflix local developer experience based on Docker ● Has had a benefit in both directions ○ Cloud like local development environment ○ Easier operational debugging of cloud workloads 7
- 8. What about resource optimization? ● Not absolutely required and easier to get wins at larger scale across larger virtual machine fleet ● However, potential benefits to ○ Elastic resource pool for scaling batch & adhoc jobs ○ Reliable smaller instance sizes for NodeJS ○ Cross Netflix resource optimizations ■ Trough usage, instance type migration 8
- 9. Agenda ● Why containers at Netflix? ● What did we build and what did we learn? ● What are our current and future workloads? 9 ⇨
- 10. VMVM Lesson: Support containers by leveraging existing Netflix IaaS focused cloud platform 10 Atlas EC2 AWSAutoScaler VMs App Cloud Platform (metrics, IPC, health) Eureka VPC Edda Existing - VM’s VMVM Atlas EC2 TitusJobControl Containers App Cloud Platform (metrics, IPC, health) Eureka VPC Edda Titus - Containers VMVM Batch Containers
- 11. VMVM 11 EC2 AWSAutoScaler VMs App Cloud Platform (metrics, IPC, health) VPC Netflix Cloud Infrastructure (VM’s + Containers) VMVM Atlas TitusJobControl Containers App Cloud Platform (metrics, IPC, health) Eureka Edda VMVM Batch Containers Why - Single consistent cloud platform
- 12. Lesson: Buy vs. Build, Why build our own? ● Looking across other container management solutions ○ Mesos, Kubernetes, and Swarm ● Proven solutions are focused on the datacenter ● Newer solutions are ○ Working to abstract datacenter and cloud ○ Delivering more than cluster manager ■ PaaS, Service discovery, IPC ■ Continuous deployment ■ Metrics ○ Not yet at our level of scale ● Not appropriate for Netflix 12
- 13. “Project Titus” (Firehose peek) 13 Titus UITitus UI Docker Registry Docker Registry Rhea container container container docker Titus Agent metrics agent Titus executor logging agent zfs mesos agent docker RheaTitus API Cassandra Titus Master Job Management & Scheduler S3 Zookeeper Docker Registry EC2 Autocaling API Mesos Master Titus UI Fenzo container Pod & VPC net drivers container container AWS container metadata proxy Integration CI/CD Amazon VM’s
- 14. Is that all? 14
- 15. Container Execution 15 Titus UITitus UI Docker Registry Docker Registry Rhea container container container docker Titus Agent metrics agent Titus executor logging agent zfs mesos agent docker RheaTitus API Cassandra Titus Master Job Management & Scheduler S3 Zookeeper Docker Registry EC2 Autocaling API Mesos Master Titus UI Fenzo container Pod & VPC net drivers container container AWS container metadata proxy CI/CD Amazon VM’s
- 16. Lesson: What you lose with Docker on EC2 16 + < ● Networking: VPC ● Security: Security Groups, IAM Roles ● Context: Instance Metadata, User Data / Env Context ● Operational Visibility: Metrics, Health checking ● Resource Isolation: Networking, Local Storage MULTI-TENANT
- 17. Lesson: Making Containers Act Like VM’s 17 ● Built: EC2 Metadata Proxy ○ Provide overridden scheduled IAM role, instance id ○ Proxy other values ● Provided: Provide Environmental Context ○ Titus specific job and task info ○ ASG app, stack, sequence, other EC2 standard ● Why? Now: ○ Service discovery registration works ○ Amazon service SDK based applications work
- 18. Lesson: Networking will continue to evolve 18 ● Started with batch ○ Started with “bridge” with port mapping ○ Added “host” with port resource mapping (for performance?) ○ Continue to use “bridge” without port mapping ● Service style apps added ○ Added “nfvpc” VPC IP/container with libnetwork plugin ○ Removed Host (no value over VPC IP/container) ○ Changed “nfvpc” VPC IP/container ■ Pod based with customer executor (no plugin) ○ Added security groups to “nfvpc”
- 19. Plumbing VPC Networking into Docker 19 No IP Needed Task 0 SecGrp Y Task 1 Task 2 Task 3 docker0 (*) EC2 VMeth0 eni0 SG=Titus Agent eth1 eni1 SecGrp=X eth2 eni2 SG=Y IP 1 IP 2 IP 3 pod root veth<id> app SecGrp X pod root veth<id> app SecGrp X pod root veth<id> appapp veth<id> Linux Policy Based Routing EC2 Metadata Proxy 169.254.169.254 IPTables NAT (*) * ** 169.254.169.254
- 20. Lesson: Secure Multi-tenancy is Hard 20 Common to VM’s and tiered security needed ● Protect the reduced host IAM role, Allow containers to have specific IAM roles ● Needed to support same security groups in container networking as VM’s User namespacing ● Docker 1.10 - Introduced User Namespaces ● Didn’t work /w shared networking NS ● Docker 1.11 - Fixed shared networking NS’s ● But, namespacing is per daemon ● Not per container, as hoped ● Waiting on Linux ● Considering mass chmod / ZFS clones
- 21. Operational Visibility Evolution 21 ● What is “node” - containers on VM’s ● Soft limits / bursting a good thing? ○ Until percent util and outliers are considered ● System level metrics ○ Currently - hand coded cgroup scraping ○ Considering Intel Snap replacement ● Pollers - Metrics, Health, Discovery ○ Created Edda common “server group” view
- 22. Future Execution Focus 22 ● Better Isolation (agents, networking, block I/O, etc.) ● Exposing our implementation of “Pod”’s to users ● Better resiliency (DNS dependencies reduced)
- 23. Job Management and Resource Scheduling 23 Titus UITitus UI Docker Registry Docker Registry Rhea container container container docker Titus Agent metrics agent Titus executor logging agent zfs mesos agent docker RheaTitus API Cassandra Titus Master Job Management & Scheduler S3 Zookeeper Docker Registry EC2 Autocaling API Mesos Master Titus UI Fenzo container Pod & VPC net drivers container container AWS container metadata proxy CI/CD Amazon VM’s
- 24. Lesson: Complexity in scheduling 24 ● Resilience ○ Balance instances across EC2 zones, instances within a zone ● Security ○ Two level resource for ENIs ● Placement optimization ○ Resource affinity ○ Task locality ○ Bin packing (Auto Scaling)
- 25. Lesson: Keep resource scheduling extensible 25 Fenzo - Extensible Scheduling Library Features: ● Heterogeneous resources & tasks ● Autoscaling of mesos cluster ○ Multiple instance types ● Plugins based scheduling objectives ○ Bin packing, etc. ● Plugins based constraints evaluator ○ Resource affinity, task locality, etc. ● Scheduling actions visibility https://github.com/Netflix/Fenzo
- 26. Cluster Autoscaling Challenge 26 Host 4Host 3Host 1 vs. For long running stateful services Host 1 Host 2 Host 2 Host 3 Host 4
- 27. Resources assigned in Titus 27 ● CPU, memory, disk capacity ● Per container AWS EC2 Security groups, IP, and network bandwidth via custom driver ● Abstracting out EC2 instance types
- 28. Security groups and their resources 28 A two level resource per EC2 Instance: N ENIs, each with M IPs ENI 0 Assigned Security Group: SG1 Used IPs Count: 2 of 7 ENI 1 Assigned Security Group: SG1,SG2 Used IPs Count: 1 of 7 ENI 2 Assigned Security Group: SG3 Used IPs Count: 7 of 7
- 29. Lesson: Scheduling Vs. Job Management 29 Scheduling resources to tasks is common. Lifecycle management is not.
- 30. Lesson: Scheduling Vs. Job Management 30 Task scheduling concerns ● Assign resources to tasks ● Cluster wide optimizations ○ Bin packing ○ Global constraints, like SLAs ● Task preferences and constraints ○ Locality with other tasks ○ Resource affinity Job manager concerns ● Managing task/instance counts ● Creating metadata, defining constraints ● Lifecycle management ○ Replace failed task executions ● Handle failures ○ Rate limit requeuing & relaunching ○ Time out tasks in transitionary states
- 31. Future Job Management & Scheduling Focus 31 ● More resources to track: GPUs ● Automatic resource affinity with heterogenous instances ● SLAs ○ Latencies for services ○ Throughput for batch ○ Task preemptions
- 32. Things we didn’t cover in this talk ● Overall integration ○ Chaos, continuous delivery, performance insight ● Container Execution ○ Logging (live log access & S3 log rotation) ○ Liveness and health checking ○ Isolation (disk usage, networking, block I/O) ○ Image registry (metrics, security scanning) ● Scheduling ○ Autoscaling heterogeneous pools ○ Host-task fitness criteria ● API ○ Extensibility, polymorphic, SLA and job/container ownership 32
- 33. Agenda ● Why containers at Netflix? ● What did we build and what did we learn? ● What are our current and future workloads? 33 ⇨
- 34. Current Titus Production Usage 34 ● Autoscaling ○ 100’s of r3.8xl’s ○ Each 32 vCPU, 244G ● Peak ○ Thousands of cores ○ Tens of TB’s memory ● Thousands containers/day ○ ~ 100 different images
- 35. Workloads, Past ● Most current usage is batch ○ Algorithm training, adhoc reporting jobs ● Sampling: ○ Training of “sims” and A/B test models ○ Open Connect Device/IX reporting ○ Web security scanning and analysis ○ Social media analytics updates 35
- 36. Workloads, Now ● Spent last five months adding service style support ● First line of fire customer requests already received ● Larger scale shadow and trickle traffic throughout 2Q ● First service style apps ○ Finer grained instances - NodeJS ○ Docker provided local developer experience 36
- 37. Workloads, Coming ● Media Encoding ○ Thousands of VM’s ○ VM based resource scheduling ○ Considering containers to have faster start-up ○ Internal spot-market - trough borrowing ● SPaaS ○ 10’s of thousands of containers ○ Stream Processing as a Service ○ Convert scheduling systems to Titus 37
- 38. Questions? 38
- 39. Other Netflix QCon Talks 39 Title Time Speaker(s) The Netflix API Platform for Server-Side Scripting Monday 10:35 Katharina Probst Scheduling A Fuller House: Container Mgmt @ Netflix Tuesday 10:35 Andrew Spyker & Sharma Podila Chaos Kong - Endowing Netflix with Antifragility Tuesday 11:50 Luke Kosewski The Evolution of the JavaScript Wednesday 4:10 Jafar Husain Async Programming in JS: The End of the Loop Friday 9:00 Jafar Husain