What are the goals of read-write replication in Arla? How much can it "cost" in server and network load and administration?
To find out what the system adminstrators would like I sent a few questions to some AFS administrators and talked to others. From the results of these "interviews" and from what I have read I have concluded the following requirements:
- File consistency should be maintained.
- Extra load on the network should be kept as low as possible. More than 10 % is unacceptable.
- Extra CPU load on the servers is not as critical. File servers often have cycles to spare.
- The solution must be suitable for systems with only a few file servers as well as for larger systems.
The situation when replication would be most valuable is a server crash. Network partitionings seem more rare. For load-balancing the read-only replication in AFS is enough, since read-write files are accessed less.
It is a bonus if the replication scheme can make it easier to implement disconnected operation.