Current failover code better detects weird situations and refuses to work in one.
The test server slaves have events far ahead of the master in their binlogs,
causing the failover to stop. To fix this, slave binlogs are now deleted when
a test begins.
The test is composed of a few parts.
1: Test that failover happens on master failure.
2: Test that a server with slave sql thread stopped is not promoted.
3: Test that a server with log_slave_updates=1 is promoted before others.
