Update January 22, 2016
Microsoft did not announce an updated SLA as described in the post below. The current status of this tech is not yet available.
Today at a Microsoft Ireland partner event Microsoft announced it will be introducing a SLA covering 99,99 % availability for single instance virtual machines! Blogger Aidan Finn sent out a Tweet with this news.
The technology enabling this SLA will be introduced before the end of 2015. It will prevent that VM’s will reboot during maintenance done by Microsoft on Azure hosts.
Currently the SLA on Azure virtual machines service is 99,95% but only when at least two virtual machines are part of the same availability set.
What does this mean for customers? Currently when an application depends on the availability of a single VM, Microsoft offers no SLA. However many legacy enterprise applications depend on the availability of a single server. They do not use some sort of clustering and do not have redundant nodes. These kind of applications are called ‘pets’. Cloud aware applications called ‘cattle’ have built in resiliency.
I explained the difference between these two types of applications in this blogpost.
Microsoft has to patch the Azure hosts frequently. This means the hosts requires a reboot. The impact of that reboot is that virtual machines will reboot as well. Azure does not have a Live Migration feature allowing to move VM’s over to other hosts. Azure will never feature a Live Migration feature because it would cost too much east-west bandwidth.
Microsoft has been working on a technology which prevents virtual machines to reboot when Azure hosts are being patched. Already in November 2014 there were inside stories about Microsoft working on something to improve the SLA.
While there are no public details about the technology which Microsoft is working on, it is likely to be something similar KernelCare using for patching Linux systems. During maitenance virtual machine will be frozen for some seconds. During that period the host will change to a patched operating system.