[HTCondor-users] A heads-up about HTCondor on Windows and daylight savings time.

[johnkn <johnkn@xxxxxxxxxxx>]

At 2am On November 2 here in the USA, daylight savings time will end.  
This will trigger a bug in the c-runtime on Windows that will cause 
HTCondor to think that the timestamp on the condor_master.exe has 
changed.  We thought that we had code in HTCondor to work around this 
bug, but it turns out it is only a partial fix, and all versions of 
HTCondor will react to the time change this coming weekend.

At this point what happens depends on what your 
MASTER_NEW_BINARY_RESTART configuration variable is set to. The choices 
are.

MASTER_NEW_BINARY_RESTART = GRACEFUL
   The condor_master restarts all of the child daemons and itself 
gracefully. This means that jobs get a signal to checkpoint and then are 
killed 2 minutes later, put back in the queue and restarted on some 
other node. SCHEDD's will shutdown and then restart and try and 
reconnect to running jobs - This is the default behavior, and a 
reasonable choice for SCHEDDs and your central manager.

MASTER_NEW_BINARY_RESTART = PEACEFUL
   The condor_master tell's child daemons to restart when they are done 
with current work. Thsi means that STARTDs will finish current jobs but 
not accept any new ones until they get a chance to restart.  SCHEDDs 
will finish current jobs but will not start any new ones. This is a 
reasonable configuration for STARTDs, but not the best choice for your 
SCHEDDs or central manager.

MASTER_NEW_BINARY_RESTART = FAST
   The condor_master kills all child daemons then restarts them.  This 
is a reasonable configuration for your central manager.

MASTER_NEW_BINARY_RESTART = NO
MASTER_NEW_BINARY_RESTART = NEVER
MASTER_NEW_BINARY_RESTART = NONE
   The condor_master notices the change but does nothing.  This is a 
reasonable choice for all daemons, but it disables the ability to 
upgrade the HTCondor binaries without explicitly restarting them.

I recommend that you set MASTER_NEW_BINARY_RESTART to either PEACEFUL or 
NEVER on your execute nodes before next weekend if you expect to have 
jobs running over the weekend.

-tj