I sent this yesterday, but it seems not to have made it to the list...
I have a couple of comments orthogonal to the present discussion.
1) It would be fairly easy to write log records over a network to a
dedicated process on another system. If the other system has an
uninterruptible power supply, this is about as safe as writing to disk.
This would get rid of the need for any fsync on the log at all. There
would be extra code needed on restart to get the end of the log from the
other system, but it doesn't seem like much.
I think this would be an attractive option to a lot of people. Most
people have at least two systems, and the requirements of the logging
system would be minimal.
2) It is also possible, with kernel modifications, to have special
logging partitions where log records are written where the head is.
Tzi-cker Chueh and Lan Huang at Stony Brook
(http://www.cs.sunysb.edu/~lanhuang/research.htm) have written this,
although I don't think they have released any code.
(A similar idea called WADS is mentioned in Gray & Reuter's book.)
If the people at Red Hat are interested in having some added value for
using PostgreSQL on Red Hat Linux, this would be one idea. It could
also be used to speed up ext3 and other journaling file systems.
