The natted networking is problemmatic, the file system is slow (particularly with many image layers), and to get good performance, you have to give up all of the isolation abilities.
To get good network performance, you have to use net=host; to get good disk performance, you have to mount and write to host volumes. To increase visibility, you have to use host pids.
I have a lot of hope that Docker will get more performance to go with the awesome isolation. It's a useful tool in the proper circumstances, but it requires a lot of forethought and information to use well.
> The slowness of Docker is a big pain. Build and deployment procedures are not predictable.
I have the absolute opposite experience. My build & deployment has become very predictable, and I build pretty much all the images I use myself.
> Data belongs onto the filesystem but not into a container that can neither be cloneed in an easy way nor incrementally backed up in a reasonable way. Containers are for software, not for data.
Volumes are for data. And sure they can be backed up or cloned, just launch a container that only backs up mounted volumes with --volumes-from whatever stuff you want to back up. Containers should be stateless, and yes, volumes need more attention.
This whole thing smells like "I tried this tool and I don't fully understand it so I'll write a rant blog about it". Containers force you to think differently about software deployment. I started using Docker very early-on and I have never regretted putting it in production since the 0.6x days.
