Certain language features like

• Garbage Collection
• Exception Systems
• Lazy Evaluation

are not generally useful in a real-time system. One should probably choose a language without these features, and try to prove certain properties like maximum memory usage, or maximum response time.

When a program needs to run continuously, but short and non-global failures are acceptable, then we could use an Erlang-like strategy. Erlang is a concurrent, functional programming language. Usually, a program written in Erlang will consist of multiple worker processes which can communicate with each other (actor model). If one worker thread encounters an exception, it is re-started. While this does imply a short downtime, the other actors can carry on as usual.

To summarize this: In a robust program, various parts are isolated from each other and can be restarted or scaled independently.

So basically we need a piece of code equivalent to this:

while (true) {
  try {
    DoWork();
  }
  catch (Exception e) {
    log(e);
  }
}

plus a way to terminate the loop. Such a loop would then drive each worker thread.

A problem with ignoring errors via a catch-all is that invariants of your program might have been violated by the error cause, and that subsequent operations could be useless. A good solution to this is to share no data between independent workers. Restarting a worker will rebuild all necessary invariants. This means that they must communicate differently, e.g. through message sends. An actor’s state may not be part of other actors’ invariants.

Another problem with catching too many exceptions is that not all exceptions are fixable by restarting, even when taking such precautions. Otherwise-hard problems like running out of memory can be handled by restarting. But a restart won’t help you to regain internet connectivity when a physical cable was pulled out.

To answer your question, one has to understand what exceptions are, and how they work.

Exceptions are usually thrown when such errors occurs, where user’s assistance is required. In such cases, it doesn’t matter how long it takes to unwind the stack and handle the exception.

Without catch handlers, the program stops execution. Depending on your setup and requirements, it may be acceptable.

In your specific cases :

1. if the query can not be executed (for example, wrong city name), then inform the user of the error, and ask to fix it.
2. if you are not getting information from a critical sensor, there isn’t much sense in continuing without asking for the operator to fix the issue.

That means that in both cases it may make sense to use exceptions, with more care in a RT program to indicate only serious problems where it is not possible to continue the execution.

I so far have yet to see anything but the base exception being caught.

It sounds like there is a problem here, in as much as exceptions aren’t being dealt with appropriately. Catching exceptions at the appropriate point and taking appropriate action (depending on the type of exception) will keep the service running in a much more reliable fashion.

If service must continue, presumably it’s important that it is working as intended. Given your example, if a program which controls railway switches throws an exception it may indicate that there’s a problem communicating with safety-related sensors. If you catch the base exception and continue the service may run, but may not function as intended leading to disaster.

Alternatively, if you catch the exception thrown when there’s a communication failure with the sensor and deal with it appropriately (i.e. stop the trains in the affected area) your service is running and you haven’t killed anyone.

So, as I understand the question, I’d suggest that in the first instance you would be better looking to add more specific exception handling rather than to remove the base-exception-type handlers.

With regards to point 2: don’t use C#. It’s not a realtime language and you will get hurt if you try to use it as such.

For point 1: you could go the erlang way: let it crash, then restart

Declaimer: these are only thoughts, I haven’t the experience.

I would guess that a program, satisfying the requirements of the second example should be extremely modular. Consequently, modules will be capable of being restarted, without destabilizing the system.

For example, an object, failing an assert for internal state, should be able to be destroyed and re-created, notifying in the process all it’s consumers and suppliers. More concretely, if the program is controlling the switches of the railway and fails an assert in the decision loop, it can still run an emergency module, which stops all involved trains, and waits for the main decision module to re-initialize.

More realistically, one would introduce redundancy – duplication of the hardware and software. One instance is wired to the controlled system, and the other is free-running. If an error is detected, the systems are switched.

An example are two processes on the same machine, which monitor one another and if one is killed, the other re-spawns it and disassociates it’s parent PID from itself.