Over the years, there was some back-and-forth changes in what happens to run transitions when some
of the participants misbehave (do not respond to RPC calls, timeout, crash, etc).
The very original behaviour was to ignore all errors. This resulted in user confusion when some clients
would start, some would not, data from frontends that missed the transition did not arrive, etc.
So it was changed to fail the transition if any client misbehaves.
This left mlogger (who is usually the first one to see the TR_START transition) in a funny state - output
file is open, etc, but there is no run active. This was fixed by adding a TR_STARTABORT transition to tell
mlogger, event builder & co that the just started run did not start after all.
Also at some point code was added to forcefully kill clients that do not respond to run transitions (do
not respond to RPC, timeout, etc).
Recently, it was observed how during unattended overnight operation of a MIDAS DAQ system, with the
logger set to "auto restart", some unnecessary clients misbehave during the run stop transition, and
prevent the run from stopping and restarting. The user comes in the morning and is unhappy that data
taking stopped some time during the night.
midas.c svn rev 5136 changes the TR_STOP transition to always succeed, even if some clients had
transition errors. If these clients are unnecessary for normal operation of the DAQ, the following run
"auto restart" will continue taking data. If those were important clients, data taking will continue the
best it can - it *is* unattended operation - nobody is looking - but users can always setup alarms for
checking that important clients are always running during data taking. (For very important clients, one
can setup alarms to send email, send SMS messages, etc).