• Created a 6TB /srv volume with 4x 1536GB drives in RAID0
• 1536GB drives have the same theoritical performance limits as 2TB ones
• Used lvm2 for the software raid layer.
• mdadm doesn't allow to grow a RAID0 volume whereas lvm2 could (by concatenating another storage provider at the end)

Azure performance appears to be very far from reliable. Linux keeps complaining about issues similar to this one:

[Tue Jan  9 15:47:04 2018] INFO: task kworker/1:3:148 blocked for more than 120 seconds.
[Tue Jan  9 15:47:04 2018]       Not tainted 4.9.0-5-amd64 #1 Debian 4.9.65-3+deb9u2
[Tue Jan  9 15:47:04 2018] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.

The sweet spot for Azure drives seems to be 2TB : 7500 IOPS and 250MB/s provisioned.
Smaller ones have less provisioned IOPS and bandwidth per second.
Bigger ones don't get faster.

Current database size: ~= 3880 GB
Maximum Azure drive size: 4095 GB

• We will need to use more than one drive to create a bigger volume
• We will also most likely need to grow this volume at some point in the future

Not a silver bullet but some of the most recurring critical alerts were silenced by cdec099ecfb5 .

Proposed change, already applied manually on uffizi and a few other hosts:

Base disk image size was increased from 10 to 20GB.

Resolved with help from @olasd .