What is better suited for a normal Server:
- several partitions, which are bundled as several RAID1 devices (/dev/md0, /dev/md1, ...) without any partitions, which are not mirrored.
- one big /dev/md0, and partitions on this device
What are the biggest pros and cons of both approaches? Is there a big difference, which one is the better choice for a normal server without frequent changes to disks and partition setup?
I haven't found any sites giving actual advice on this decision. The only thing i frequently read was: DO NOT bundle /dev/hda /dev/hdb (without at least one partition) to a RAID, because this causes the kernel to detect the RAID partitions on the raw /dev/hdX devices, too.
"what is the better choice for a normal server without frequent changes to disks and partition setup?"
To answer your question you posed, there is a reason there are so many options for disk array setup to choose from. Each scenario has its own requirements, wants/needs, performance related issues, etc. If you were to post what the server was going to be used for that might help.
See here: Linux LVM: Single or Multiple Harddisk Partitions? as well.
Having the RAID1 over the entire disk allows you to replace a defective disk without downtime (if the disk controller allows hot swapping).
If you have separate partitions you can do more creative things. You can have /, /boot, and /srv different partitions. This allows you to split the RAID for / and do an OS upgrade in a VM that has access to the unused / copy, then reboot from this disk and then replicate / from the upgraded / partition to the old one. This is similar to Solaris Live Update.
If you have separate partitions you can have different raid levels for those partitions.
It depends on what is more important high availability or flexibility.