If a physical compute node fails, our monitoring systems will detect the failure and trigger an “evacuate” process that will restart all affected virtual compute instances on a healthy physical server. This process usually takes between 5 to 20 minutes which allows us to meet our 99.95% availability SLA for individual compute instances.
Customers that require more than 99.95% availability can combine multiple compute instances within the same region using anti-affinity groups. Anti-affinity groups ensure that compute instances that are members of the same group are hosted on different physical servers. This reduces the risk and probability of multiple compute instances failing at the same time. For more information on how to use anti-affinity, please consult Using server affinity for HA.
Customers that require their applications to survive the loss of an entire region can launch compute instances in different regions. This requires their applications, or middleware used by their applications (such as databases), to support this architecture.
When creating a new compute instance, the preference is to use a
Persistent Volume for its root disk. Persistent volumes have several
benefits, such as:
Snapshots are created in seconds and use less space.
The volume and it’s associated data can continue to exist after the compute instance is deleted.
When launching an instance via the dashboard the default behaviour is to create
Persistent Volume for its root disk. Other methods of instance creation
vary and some will require the volume to be created ahead of the instance being
launched, if this is the case please consult the relevant documentation for
Ephemeral Disk should only be used for cloud native workloads,
that use the immutable infrastructure principle, or workloads that are
temporary in nature, such as batch jobs.
When server failures occur, the Catalyst Cloud will attempt to restart the affected compute instances on a healthy server to minimize their downtime.
To benefit from this feature, your application must be configured and prepared to start automatically and resume its normal operation at boot time and your guest operating system to respond to ACPI power events. The operating system images supplied by Catalyst or our partners already have ACPI configured to respond to power events by default.
It is best to scale applications horizontally (by adding more compute instances and balancing load amongst them) rather than vertically. It is possible to scale compute instances horizontally without downtime. Resizing compute instance vertically (up or down) will result in a brief downtime, because the operating system needs to reboot to pick up the new configuration.