12. Lots of NetApps
• Peak of around 125 individual NetApps
• Smallish capacity on each (8TB)
• Internal software continuously moving
data between NetApps
• Lots of time spent managing nearly full
filers
15. Hosting Landscape
• Included storage had grown from 50MB
to gigabytes, then terabytes.
• Prices stayed the same.
• Eventually went to unlimited Storage
• Usage per customer skyrocketed.
21. Local RAID
• SATA drives had grown in capacity and
were very cheap
• 4-6TB per hosting server
• Less dependence on congested
network
• Smaller failure domains
The Good
22. Local RAID
• No more quota, too slow to scan
filesystem
• No more fast failovers
• Multiple hour filesystem check with ext3
• More failure domains
The Bad
23. Local RAID
• Complete RAID loss more common
than anticipated
• Multiple days to fully restore from
backup
The Ugly
25. Hybrid Mix
• We learned something from every step
of the way
• No one size fits all when it comes to
storage
• Use whatever is best for the job
• Be ready to change
Best Tool For The Job
26. A Bit of Everything
• Clustered NetApps and NFS for email
• Local RAID in hosting servers
• ZFS and OpenSolaris backup servers
• Ceph for DreamObjects and
DreamCompute
Best Tool For The Job
27. • Object Storage, S3/Swift compatible
• 2+ Petabytes raw storage
• 3x replication, 900+ OSDs
• RGW behind HAProxy
• Row, rack, node and disk fault tolerant
28. • OpenStack-based Public Cloud
• 3+ Petabytes raw storage
• All storage is on Ceph RBD
• Boot and Attachable Volumes
• Nicira SDN + Ceph, Live Migration
29. HA Load Balancer
MySQL / PostgreSQL
Horizon
Cockpit Pod
Glance
Keystone
Nova
Quantum
Cinder
Nicira NVP
Glance Store (Ceph)
OSMirrors (apt)
Ceph Monitors
Opscode Chef
Logstash + Graphite
Networking Gear
8x - Hypervisor Node
192 GB RAM
64AMD cores
14x - Storage Node
12x - 3TB disks
Networking Gear
Compute Pod
8x - Hypervisor Node
192 GB RAM
64AMD cores
14x - Storage Nodes
12x - 3TB disks
Networking Gear
Compute Pod
8x - Hypervisor Node
192 GB RAM
64AMD cores
14x - Storage Nodes
12x - 3TB disks
Networking Gear
Compute Pod
Pods
• 512 cores
• 1.5TB of RAM
• 504TB raw storage
• 168TB redundant storage
N etworking
• ODM switches w/ Linux
• 10Gbps everywhere
• IPv6 from the ground up
• Spine and leaf topology
• 120 Gbps between pods (!)
The Internets
Thar be dragonshere!
Nicira NVP Nicira NVP NiciraNVP
30. CephFS & The Future
• The return of Failovers
• No more backup servers
• No more major disk-related outages
• Fault tolerant low cost hosting
Storage Panacea?