Bad entries in binary log preventing IO thread execution

[Issac Goldstand]

I have 2 way replication set up between a bunch of pairs of Linux-based MySQL 4.0.20 server and a Win32-based 4.1.12a servers. Yesterday, we had a power outage and all the machines went down the hard way. Somewhat after they came up, we noticed an oddity affecting most, if not all (we didn't get around to checking every single pair quite yet) of these machines. On the Linux machines, replication was still working nicely. On the windows machines, however, the IO slave refused t start, and we were seeing the same error across the error logs: "Client requested master to start replication from impossible position" We checked the binary logs on the Linux machines, and they did indeed have the last full entry a bit before the positions that were being reported as non-existant in the error logs. Furthermore, when running mysqlbinlog on many of the logs in question (on the Linux machines that generated them) mysqlbinlog similarly reported that there were problems in the last entry.

I have 2 questions regarding this:

1) Are the binary logs not written to the disk before they're sent over the network to the replcation slave? If so, how cuold the slave have processed an event in the binary log, yet the master complains that it can't continue there when resuming replication?
2) Can someone suggest an intelligent way to detect the problem and resume replication in a situatio like this? Is there any way to "leave a message" for the binlog dump thread to send a "binary log rotation" signal to the slave after the last readable entry in the binary log is sent in a corrupted binlog?

Thanks!