Rob Hirschfeld (aka Zehicle online)
In Community: OpenStack Board Member (4 years)
Co-Chair of Kubernetes Cluster Ops SIG
Founder of Digital Rebar & Crowbar Projects
Professional: CEO of RackN - hybrid automation software
Executive at Dell - scale data center ops
Cloud Data Center Ops going back to 1999
Addressing Operators Needs
Operational Success is Essential to Project Success
Operators are not developers!
Simple, Transparent and Stable are key concerns
Becoming a super-user of the platform should not be required to run it
Scale & Upgradability has both internal and external drivers
Generally, Kubernetes has good operational fundamentals
We’re Talking Underlay, not Overlay
We’re talking about installing Kubernetes first (aka
underlay) and using it to manage the OpenStack control
This approach is not a win if we
● Disable Kubernetes management
● Still need outside management tooling
For now, we’ll ignore that our user may actually want to
use Kubernetes as the overlay. IMHO, a bad assumption. Physical Infrastructure
Simplest conception of the
K8s OpenStack Sandwich
What is Kubernetes?
Container Scheduler (no, it’s not really Orchestration)
API driven to provide restart, placement, network routing and life-cycle
For Applications designed for Kubernetes
Key Design Elements: Immutable Infrastructure (stateless ops)
12 Factor Configuration
What is Kubernetes: A Three Tier Application
Pod Pod Pod Pod
DNS Watcher ...
Together 4ever: API server + Kubelet
Pod Pod Pod Pod
DNS Watcher ...
Why do we want Kubernetes as Underlay?
Community Perception Accuracy
1 OpenStack Operations is still really hard True
2 We already do most deploys in containers Partially
3 Kubernetes is awesome at containers Partially
4 Kubernetes means free Upgrades and High Availability False
5 Kubernetes is simple, stable and secure (for operators) False
First: We’re Confusing Technical and Marketing
Marketing around Kubernetes under OpenStack is a “hot mess”
● People heard “Kubernetes is stable, OpenStack is not”
● Further confuses “OpenStack one platform message”
Who is promoting doing this? Mirantis / CoreOS / Intel / Google
Confusion with the Plain Old Container Install (“POCI”) message
● Canonical (Ubuntu Cloud Install),
● Rackspace (OpenStack Ansible)
● Cisco (Kolla)
Second: Why I’m scared
This discussion keep kicking the operations & install problems down the field
Kubernetes is much newer than OpenStack, so even less understood
Yet more complexity and some very basic questions:
● Now we have a both a Kubernetes and OpenStack upgrade problem
● We still need tooling to manage OpenStack in Kubernetes
● We still need someone to package the containers
● Relies on Docker to keep systems running
● Storage and Networking are still being worked out
But, it’s going to
So let’s get
pragmatic about it.
Key Principle: Containerization vs Kubernetes
Containers can be treated as a) lightweight vms or 2) packaged daemon sets.
● Canonical builds their containers like persistent vms and configures with Juju
● Kolla & OSA treats containers as packaging and configures with Ansible
Kubernetes accepts neither approach – they expect containers to be immutable
and 12 factor configured
● Kubernetes manages the full container life-cycle
● Containers need to be able to handle being added, removed
● Services need to be able to handle IP address changes (or use DNS names)
Specific Technical Barriers
Host / Pinned vs Managed Containers
● It is possible to disable Kubernetes management and pin containers
● This eliminates the desired benefit of using Kubernetes
How to handle Layered SDN integrations? Who wins?
How to handle expectations of container persistence?
Assumptions of Exclusive Ownership / Administrative Control
General Challenges to Overcome
● Overall Complexity of More Components
● Need to Control Kubernetes & Kubernetes Stability
● Need for Multiple Tiers of Load Balancer
● IP Mobility in Service Registration & Message Bus
● Mixed Networking Models
● Utility Upgrades and maintenance of the Underlay
● Mixing Kubernetes workloads
There are REAL Potential Benefits
Part of Kubernetes Ecosystem (which is likely bigger than OpenStack’s)
Leverage Docker packaging efforts and reduce Python & O/S dependencies
Upgrades would benefit from Kubernetes built-in processes
Use of the Kubernetes job scheduler for maintenance
“Free” fault tolerance of key components
Easier install if Kubernetes already running on-site
More constrained options for configuration and operation
How could this actually be done quickly?
Focus on Control plane, leave the nodes alone.
● Workers are “pinned” and agents need
Cherry pick services to move into Kubernetes
● Focus on web services
● Separate support services (data & message bus)
Externalize data using service registry
Make sure OpenStack Projects can handle immutable container requirements
More Detail: Ops Underlay vs OpenStack Underlay
If you to really want to build this, give me a call - RackN has all the components
Software Defined Networking
OpenStack operability is not solved via the underlay platform alone.
Technical Leadership motivation required for OpenStack adopting
Kubernetes architecture requirements.
Serious messaging confusion in effort has to be resolved.
However, this collaboration is required for OpenStack
Because Kubernetes will have a larger footprint in Operations