[Ray Summit 2023] Approaching Cluster Multi-tenancy with Ray Job

Transcript

1. Approaching Cluster Multi-tenancy with Ray Job Jaehyun Sim, Ikigai
2. Introduction 2 Jaehyun Sim Boston, MA simjay@ikigailas.io linkedin.com/in/simjay ikigailabs.io
3. Ikigai Platform 3
4. Why Multi-tenancy
5. Ikigai Pipeline Engine 5
6. Ikigai Pipeline Engine with Ray ❏ Each step runs its task in an individual pod ❏ Steps with heavier workloads run their critical tasks with Ray cluster ❏ Intermediate datasets and ML models are stored in S3 6
7. Multiple Pipelines with Ray 7 Multi-tenant Ray Cluster
8. What is Multi-tenancy? “Multitenancy is a reference to the mode of operation of software where multiple independent instances of one or multiple applications operate in a shared environment. The instances (tenants) are logically isolated, but physically integrated.” 8 Soft Multi-tenancy Easier to implement Hard Multi-tenancy Harder to Implement Generally “Better” Stricter Isolation of Tenants
9. Why Multi-tenancy? 1. Faster Response No need to wait for Ray cluster creation for new tenants 2. Cost-effective We operate less amount of Ray cluster in general and saves cost for cluster creation time 3. Reduced Maintenance Overhead The number of cluster under maintenance is drastically lower 9
10. Approaching Ray Cluster Multi-tenancy 1. Task Submission with Ray Job 2. Tenant Isolation with Runtime Environment 3. Resource Management strategies 10
11. Task Submission with Ray Job
12. Ikigai Pipeline Engine with Ray Client 12 ❏ A new connection is created when pipeline execution starts ❏ Connections are long-running and last until the end of pipeline executions
13. Ray Client 13 ❏ Great to for interactive development and faster PoC process ❏ Not designed for high number of concurrent connections ❏ Run Python code with Ray cluster as if you are running it on local machine ❏ Very minimal changes required in codebase
14. Multi-tenancy Challenges 14 50+ concurrent connections caused unpredictable connection failure for us
15. Ray Job 15 Entrypoint Script Job Submission Job Status Check Ray Job submission is a mechanism to submit locally developed and tested applications to a remote Ray cluster. It simplifies the experience of packaging, deploying, and managing a Ray application.
16. Ikigai Pipeline Engine with Ray Job Still making minimal changes to the codebase but no more unpredictable connection failure! 16
17. Tenant Isolation with Runtime Environment
18. Runtime Environment 18 Configuring dependencies and environment outside of Ray scripts With RuntimeEnv Object ❏ working_dir The working directory for the Ray workers (local or S3) ❏ py_modules Python modules to be available for import in the Ray workers ❏ pip A list of pip requirements ❏ env_vars Environment variables to set … and more! With Dictionary
19. Runtime Environment 19 Configuring dependencies and environment outside of Ray scripts For Task/Actor For Ray Job Submission For Ray Client
20. RuntimeEnv: working_dir 20 ❏ Each pipeline’s data and ML models will be located in its own S3 path ❏ Tenant’s data and models are logically isolated Each job will have access to its own S3 path, specified by `working_dir`
21. RuntimeEnv: pip 21 Defined by user 1 Defined by user 2 ❏ Tenant’s Python environments are logically isolated
22. RuntimeEnv: env_vars 22 ❏ Provides extra isolations with environment variables
23. Resource Management Strategies
24. Understanding Tasks - Categorize the tasks based on the required resources (cpu and memory) - Understand the size and capacity of the Ray cluster - Avoid running small (or tiny) jobs on Ray cluster 24
25. Required Resources for Task num_cpus 25 memory Lower Limit ❏ This tells Ray scheduler to reserve the required resource to ensure the tasks’ requirements does not exceed the nodes’ capacity ❏ Specifying resources does not enforce physical isolation of resources
26. Queues for Tasks 26 Upper Limit ❏ Pipeline executors dump tasks to the task queue ❏ Task monitor maintains the resource requirements of running jobs and the cluster capacity
27. Summary
28. Achieving Ray Cluster Multi-tenancy 28 1. Stable and concurrent Task Submission with Ray Job submission 2. Tenant Isolation with datasource separation, Python dependency management, and environment variables 3. Resource Management by providing CPU and memory requirement, alongside with task queue integration
29. Thank you!