Workday's Next Generation Private Cloud

Workday has built one of the largest OpenStack-based private clouds in the world, hosting a workload of over a million physical cores on over 16,000 compute nodes in 5 data centers for over ten years. However, there was a growing need for a newer, more maintainable deployment model that would closely follow the upstream community. We would like to share our new architecture and deployment approach as well as lessons learned from our experience.


We’ve converted many of our technologies in the process, from…



Migrating from Mitaka, to Victoria

Converting from OpenContrail, to pure L3 Calico with BGP on the host

Deploying with Chef, to deploying with Ansible

Building home-grown container images, to Kolla

Monitoring with Sensu and Wavefront, to Prometheus and Grafana

CI/CD in Jenkins, to Zuul

CentOS 7, to CentOS 8 Stream


We'll also talk about some internal tools we wrote that, while Workday-specific, may inspire you to see what value-add you can make for your customers.

1. Workday’s Next Generation Private Cloud
The Fifth Generation of an OpenStack Platform inside Workday

2. The Leading Enterprise
Cloud for Finance and HR
Customer Satisfaction:
Workers Supported
97%
60M+
Fortune 500 Companies:50%+

3. Silvano Buback
Principal, Software Development Engineer
Jan Gutter
Senior Software Development Engineer

4. Workday
Private
Cloud

5. OpenStack at Workday
8 SREs
9 Developers
SLO: 99% API Call
success
87 Clusters
2 Million Cores
12.5 PB RAM
60k concurrent VMs
241k VMs recreated weekly

6. Simple set of OpenStack components to deliver a resilient platform.

●Single client (PaaS)
●~300 compute nodes per cluster
●Workday service weekend maintenance “The Patch”
●OpenStack projects are used to denote Workday services
●Unique tooling for batch scheduling and capacity planning


Workday’s Use Case

7. Regular maintenance window every weekend where where service
VMs are recreated and the Workday application gets upgraded

●“The Power of One” is an important mission for us
●Largest impact to control and data plane during this time
●SLO target is 99% success for all API calls over the week
●60% of instances deleted/created during “The Patch”
●Remaining 40% are recreated throughout the week

“The Patch”

8. Development Environment
We Treat
Everything
as a High
Security
Environment
Weekly
Builds in
Dev
Clusters
Dev Clusters
Run Internal
Services
Dev and
Production
Run Very
Different
Workloads

9. Workday
Private
Cloud
5

10. Fourth Generation
Private Cloud Evolution
•OpenStack Victoria
•CentOS Stream 8
•Kolla-Ansible + plain Ansible
•Kolla Containers (built from source)
•Calico
•L3 only BGP Fabric
•Zuul CI
•Internal solution for CD
•Branch for each stable series
Fifth Generation
•OpenStack Mitaka
•CentOS Linux 7
•Chef
•RPM
•Contrail
•Overlay Networks
•Jenkins for CI
•Jenkins + Internal solution for CD
•Single branch, releases are snapshots

11. First use of gated development!
CI Tooling

12. Target multiple scenarios:
•CLI
•Zuul
•Custom Ansible Orchestration Service

Three types of clusters:
•Overcloud - a cluster built from instances in a single tenant
•Zuul - a cluster built from a nodeset
•Baremetal

Pain Point - Multiple Deployment Scenarios

13. Zuul: Expectations vs Reality
Successfully
keeps a lot
of core code
stable
Naively
expected to
reuse
community
pipeline
Evolved pipeline
multiple times
with no
interruptions
Community
pipelines tied
to community
infrastructure

14. ●Use branches for stable releases
●Nothing new about this: OpenStack community also uses this
●“branch for stable release” model was a new concept for us
●We forked https://opendev.org/openstack/releases to handle this
Zuul Pipeline Design

15. For every tool / service, there’s a Workday
name!
Home Grown
Tools

16. List of Home Grown Tools
DNS
Infrastructure
IP Address
Management
Certificate
Authority
Ansible
Orchestration
Multi Cluster
Cloud
Overview
Compute
Node
Health
Check

17. List of (more) Home Grown Tools
Capacity
Management
Chef
Implementation
Batch
Scheduling
PaaS
(Image Build
Service, Instance
lifecycle
management) BM Lifecycle
Tracking
Bare Metal
Provisioning
Service

18. Differences with community version
Downstream
Changes

19. Downstream Changes
TLS everywhere
Compute nodes use
Prometheus/OpenStack
integration
Prometheus upgraded to newer version
Custom tags based on
Kolla-Ansible inventory
Wavefront integration
while we transition to
Cortex

20. ●New Prometheus Exporters (some are upgrades)
○libvirt exporter
○OpenStack exporter upgrade
○BIRD exporter (BGP router)
●Fluentd parses HAproxy/Apache logs to provide API request metrics
●“Singleton” containers
○One running container per cluster
○Using Keepalived for HA
○Examples: Prometheus, DB Backup, openstack-exporter
●Timeouts/Retry/Performance improvements on K-A deployment
(more) Downstream Changes

21. ●Kolla containers for Calico
●Enabled etcdv3 in Kolla-Ansible
●Building C8 binaries
●Using a local fork of the Neutron plugin
●Wrote our own metadata proxy (TLS support)
●Numerous small changes
○MTU
○Newer version of OpenStack
○DHCP service monitoring
●Most of the changes were in the Neutron plugin, Felix code is
essentially unchanged
Calico Fork

22. Q & A

23. Random notes about our environment
Other
Interesting Bits

24. ●Every instance gets an internally routable IPv4 address. ??????
●Multiple layers of network security
●Previously: Contrail with virtual overlay networks
●Now: Calico with routing fabric
Requirements for Networking

25. ●In preparation for OpenStack Victoria, we reduced the use of file
injection in our PaaS system significantly
●We were fortunate because we could move service accounts from
one cluster to another
●To reduce transition time, we allocate overlapping ranges
●During The Patch, instances running on the previous generation
are removed
Forklift

26. Thank You