Develop Reference

throw original exception inside Parallel methods instead of aggregate exception

I have two CPU-intensive methods inside a Parallel.Invoke call:
Parallel.Invoke(
() => { GetMaxRateDict(tradeOffObj); },
() => { GetMinRateDict(tradeOffObj); }
);
For a MCVE, assume:
public void GetMaxRateDict(object junk)
{
throw new Exception("Max exception raised, do foo...");
}
public void GetMinRateDict(object moreJunk)
{
throw new Exception("Min exception raised, do bar...")
}
I throw different exceptions in each of these inner methods. However, if one of these gets thrown, the Parallel wrapper throws a more generic exception: "One or more errors occurred", which is specific enough to show in my UI layer.
Can I grab the original exception somehow and throw it instead?
I would like the Parallel task to stop entirely if possible to raise the inner exception, but if that's not possible, at least being able to raise it once the two methods complete is what I need. Thanks.

Can I grab the original exception somehow and throw it instead?
"It" implies that there will only be on exception. Even though that's probably true, because you're executing actions in parallel you can't 100% rule out the possibility that multiple actions throw exceptions even if you attempt to cancel the others after the first exception. If you're okay with that, we can go from the assumption that we only expect one exception and we're okay with only catching one. (If you allow the other invocation to continue after one throws an exception the possibility of having two exceptions increases.)
You can use a cancellation token. If one of the invocations below throws an exception, it should catch that exception, place it in a variable or queue, and then call
source.Cancel;
Doing so will cause the entire Parallel.Invoke to throw an OperationCanceledException. You can catch that exception, retrieve the exception that was set, and rethrow that.
I'm going to go with the other answer's suggestion of a ConcurrentQueue just as a matter of practice because I don't think we can rule out the remote possibility that a second thread could throw an exception before being canceled.
This started off seeming small, but eventually it got so involved that I separated it into its own class. This makes me question whether my approach is needlessly complex. The main intent was to keep the messy cancellation logic from polluting your GetMaxRateDict and GetMinRateDict methods.
In addition to keeping your original methods unpolluted and testable, this class is itself testable.
I suppose I'll find out from the other responses whether this is a decent approach or there's something much simpler. I can't say I'm particularly excited about this solution. I just thought it was interesting and wanted to write something that did what you asked.
public class ParallelInvokesMultipleInvocationsAndThrowsOneException //names are hard
{
public void InvokeActions(params Action[] actions)
{
using (CancellationTokenSource source = new CancellationTokenSource())
{
// The invocations can put their exceptions here.
var exceptions = new ConcurrentQueue<Exception>();
var wrappedActions = actions
.Select(action => new Action(() =>
InvokeAndCancelOthersOnException(action, source, exceptions)))
.ToArray();
try
{
Parallel.Invoke(new ParallelOptions{CancellationToken = source.Token},
wrappedActions)
}
// if any of the invocations throw an exception,
// the parallel invocation will get canceled and
// throw an OperationCanceledException;
catch (OperationCanceledException ex)
{
Exception invocationException;
if (exceptions.TryDequeue(out invocationException))
{
//rethrow however you wish.
throw new Exception(ex.Message, invocationException);
}
// You shouldn't reach this point, but if you do, throw something else.
// In the unlikely but possible event that you get more
// than one exception, you'll lose all but one.
}
}
}
private void InvokeAndCancelOthersOnException(Action action,
CancellationTokenSource cancellationTokenSource,
ConcurrentQueue<Exception> exceptions)
{
// Try to invoke the action. If it throws an exception,
// capture the exception and then cancel the entire Parallel.Invoke.
try
{
action.Invoke();
}
catch (Exception ex)
{
exceptions.Enqueue(ex);
cancellationTokenSource.Cancel();
}
}
}
The usage would then be
var thingThatInvokes = new ParallelInvokesMultipleInvocationsAndThrowsOneException();
thingThatInvokes.InvokeActions(
()=> GetMaxRateDict(tradeOffObj),
() => GetMinRateDict(tradeOffObj));
If it throws an exception, it will be a single exception from one invocation failure, not an aggregate exception.

Not quite sure whether given example would answer your question, but it might improve overall solution:
private static void ProcessDataInParallel(byte[] data)
{
// use ConcurrentQueue to enable safe enqueueing from multiple threads.
var exceptions = new ConcurrentQueue<Exception>();
// execute the complete loop and capture all exceptions
Parallel.ForEach(data, d =>
{
try
{
// something that might fail goes here...
}
// accumulate stuff, be patient ;)
catch (Exception e) { exceptions.Enqueue(e); }
});
// check whether something failed?..
if (exceptions.Count > 0) // do whatever you like ;
}
Such an approach gives additional freedom in terms of collecting different kinds of exceptions into different queues (if necessary) or re-throwing aggregated exception further (such that no sensitive info bubbled up or you may convey particular exception with a user-friendly description of possible reasons, etc.).
Generally, that is correct way of exception management with parallelization. Not only in C#.

Error handling using delegates

I have a common method that I'm using to handle a specific error that may come back from a number of functions:
protected async Task<T> RunMyMethod<T>(Func<T> method)
{
try
{
var returnValue = await Task.Run<T>(method);
return returnValue;
}
catch (MyCustomException)
{
// Force a clean shutdown of the software
ShutdownApplication();
return default(T);
}
}
Here's an example of how that is then used in a derived class:
private async Task<IEnumerable<MyData>> GetMyData()
{
var returnValue = await base.RunMyMethod<IEnumerable<MyData>>(() =>
{
var returnval = GetMyDataFromServer();
return returnval;
});
return returnValue;
}
When an exception of type MyCustomException occurs in GetMyDataFromServer() the software doesn't drop into the catch block. I get the following error in the function GetMyData():
An exception of type 'System.ServiceModel.FaultException`1' occurred in mscorlib.dll but was not handled in user code
Additional information: Exception of type 'MyCustomException' was thrown.
This is with only User-unhandled exceptions turned on.
GetMyDataFromServer() communicates with a WCF service. This service is what throws the error.
ChannelFactory<TChannel> cf = new ChannelFactory<TChannel>(endPointName);
Binding binding = new BasicHttpBinding(BasicHttpSecurityMode.TransportCredentialOnly);
var clientCredentials = new ClientCredentials();
. . .
channel = cf.CreateChannel();
var data = channel.CallWCFService();
Having looked around on-line, it appeared that the correct way to handle this was to change the base method as follows:
protected async Task<T> RunMyMethod<T>(Func<T> method)
{
var returnValue = await Task.Run<T>(method).ContinueWith(e =>
{
ShutdownApplication();
return default(T);
}, TaskContinuationOptions.OnlyOnFaulted);
return returnValue;
}
When I run this, I'm obviously not trapping for the correct error message, but I'm just getting a TaskCancellationException.
So, I have two questions: is my conclusion about how to handle this exception correct and, if so, how do I trap a specific error; and why am I getting a TaskCancellationException?

You get TaskCancellationException because the continuation is cancelled as it's conditional (i.e. TaskContinuationOptions.OnlyOnFaulted) and the condition isn't met since the antecedent task wasn't faulted.
There's no reason to use that method of adding a continuation. Using async-await like you did at the start is good enough (and even simpler).
The issue is that you are trying to catch MyCustomException but that isn't the exception being thrown. Since you're using WCF the exception is FaultException. You can check the "real" exception stored in FaultException.InnerException.

Task Parallel Library exception handling

When handling exceptions in TPL tasks I have come across two ways to handle exceptions. The first catches the exception within the task and returns it within the result like so:
var task = Task<Exception>.Factory.StartNew(
() =>
{
try
{
// Do Something
return null;
}
catch (System.Exception e)
{
return e;
}
});
task.ContinueWith(
r =>
{
if (r.Result != null)
{
// Handle Exception
}
});
The second is the one shown within the documentation and I guess the proper way to do things:
var task = Task.Factory.StartNew(
() =>
{
// Do Something
});
task.ContinueWith(
r =>
{
if (r.Exception != null)
{
// Handle Aggregate Exception
r.Exception.Handle(y => true);
}
});
I am wondering if there is anything wrong with the first approach? I have received 'unhandled aggregate exception' exceptions every now and again using this technique and was wondering how this can happen?
To clarify, I think the second pattern is the better one but I have a chunk of code which makes use of the first pattern and I am trying to find out if it needs re-factoring i.e. if it turns out that not all exceptions will be trapped.

The first approach assumes exceptions will be raised for every invocation. While this might be true, the exceptions don't seem "exceptional" and smells of a design issue. If the exceptions are not exceptional, then the result doesn't make much sense. The other problem is that if you do want a "result" (i.e. something other than Exception) you can't because the one and only Result slot is used for an Exception. Another problem is that you don't get the re-throwing of the exception back on the main thread (you could do that manually) so you don't get the catch semantics (i.e. you're using the Handle method).
The second method will be better understood by more people.

Additional try statement in catch statement - code smell?

Situation:
My application need to process the first step in the business rules (the initial try-catch statement). If an certain error occurs when the process calls the helper method during the step, I need to switch to a second process in the catch statement. The back up process uses the same helper method. If an same error occurs during the second process, I need to stop the entire process and throw the exception.
Implementation:
I was going to insert another try-catch statement into the catch statement of the first try-catch statement.
//run initial process
try
{
//initial information used in helper method
string s1 = "value 1";
//call helper method
HelperMethod(s1);
}
catch(Exception e1)
{
//backup information if first process generates an exception in the helper method
string s2 = "value 2";
//try catch statement for second process.
try
{
HelperMethod(s2);
}
catch(Exception e2)
{
throw e2;
}
}
What would be the correct design pattern to avoid code smells in this implementation?
I caused some confusion and left out that when the first process fails and switches to the second process, it will send different information to the helper method. I have updated the scenario to reflect the entire process.

If the HelperMethod needs a second try, there is nothing directly wrong with this, but your code in the catch tries to do way too much, and it destroys the stacktrace from e2.
You only need:
try
{
//call helper method
HelperMethod();
}
catch(Exception e1)
{
// maybe log e1, it is getting lost here
HelperMethod();
}

I wouldn't say it is bad, although I'd almost certainly refactor the second block of code into a second method, so keep it comprehensible. And probably catch something more specific than Exception. A second try is sometimes necessary, especially for things like Dispose() implementations that might themselves throw (WCF, I'm looking at you).

The general idea putting a try-catch inside the catch of a parent try-catch doesn't seem like a code-smell to me. I can think of other legitimate reasons for doing this - for instance, when cleaning up an operation that failed where you do not want to ever throw another error (such as if the clean-up operation also fails). Your implementation, however, raises two questions for me: 1) Wim's comment, and 2) do you really want to entirely disregard why the operation originally failed (the e1 Exception)? Whether the second process succeeds or fails, your code does nothing with the original exception.

Generally speaking, this isn't a problem, and it isn't a code smell that I know of.
With that said, you may want to look at handling the error within your first helper method instead of just throwing it (and, thus, handling the call to the second helper method in there). That's only if it makes sense, but it is a possible change.

Yes, a more general pattern is have the basic method include an overload that accepts an int attempt parameter, and then conditionally call itself recursively.
private void MyMethod (parameterList)
{ MyMethod(ParameterList, 0)l }
private void MyMethod(ParameterList, int attempt)
{
try { HelperMethod(); }
catch(SomeSpecificException)
{
if (attempt < MAXATTEMPTS)
MyMethod(ParameterList, ++attempt);
else throw;
}
}

It shouldn't be that bad. Just document clearly why you're doing it, and most DEFINITELY try catching a more specific Exception type.

If you need some retry mechanism, which it looks like, you may want to explore different techniques, looping with delays etc.

It would be a little clearer if you called a different function in the catch so that a reader doesn't think you're just retrying the same function, as is, over again. If there's state happening that's not being shown in your example, you should document it carefully, at a minimum.
You also shouldn't throw e2; like that: you should simply throw; if you're going to work with the exception you caught at all. If not, you shouldn't try/catch.
Where you do not reference e1, you should simply catch (Exception) or better still catch (YourSpecificException)

If you're doing this to try and recover from some sort of transient error, then you need to be careful about how you implement this.
For example, in an environment where you're using SQL Server Mirroring, it's possible that the server you're connected to may stop being the master mid-connection.
In that scenario, it may be valid for your application to try and reconnect, and re-execute any statements on the new master - rather than sending an error back to the caller immediately.
You need to be careful to ensure that the methods you're calling don't have their own automatic retry mechanism, and that your callers are aware there is an automatic retry built into your method. Failing to ensure this can result in scenarios where you cause a flood of retry attempts, overloading shared resources (such as Database servers).
You should also ensure you're catching exceptions specific to the transient error you're trying to retry. So, in the example I gave, SqlException, and then examining to see if the error was that the SQL connection failed because the host was no longer the master.
If you need to retry more than once, consider placing an 'automatic backoff' retry delay - the first failure is retried immediately, the second after a delay of (say) 1 second, then doubled up to a maximum of (say) 90 seconds. This should help prevent overloading resources.
I would also suggest restructuring your method so that you don't have an inner-try/catch.
For example:
bool helper_success = false;
bool automatic_retry = false;
//run initial process
try
{
//call helper method
HelperMethod();
helper_success = true;
}
catch(Exception e)
{
// check if e is a transient exception. If so, set automatic_retry = true
}
if (automatic_retry)
{ //try catch statement for second process.
try
{
HelperMethod();
}
catch(Exception e)
{
throw;
}
}

Here's another pattern:
// set up state for first attempt
if(!HelperMethod(false)) {
// set up state for second attempt
HelperMethod(true);
// no need to try catch since you're just throwing anyway
}
Here, HelperMethod is
bool HelperMethod(bool throwOnFailure)
and the return value indicates whether or not success occurred (i.e., false indicates failure and true indicates success). You could also do:
// could wrap in try/catch
HelperMethod(2, stateChanger);
where HelperMethod is
void HelperMethod(int numberOfTries, StateChanger[] stateChanger)
where numberOfTries indicates the number of times to try before throwing an exception and StateChanger[] is an array of delegates that will change the state for you between calls (i.e., stateChanger[0] is called before the first attempt, stateChanger[1] is called before the second attempt, etc.)
This last option indicates that you might have a smelly setup though. It looks like the class that is encapsulating this process is responsible for both keeping track of state (which employee to look up) as well as looking up the employee (HelperMethod). By SRP, these should be separate.
Of course, you need to a catch a more specific exception than you currently are (don't catch the base class Exception!) and you should just throw instead of throw e if you need to rethrow the exception after logging, cleanup, etc.

You could emulate C#'s TryParse method signatures:
class Program
{
static void Main(string[] args)
{
Exception ex;
Console.WriteLine("trying 'ex'");
if (TryHelper("ex", out ex))
{
Console.WriteLine("'ex' worked");
}
else
{
Console.WriteLine("'ex' failed: " + ex.Message);
Console.WriteLine("trying 'test'");
if (TryHelper("test", out ex))
{
Console.WriteLine("'test' worked");
}
else
{
Console.WriteLine("'test' failed: " + ex.Message);
throw ex;
}
}
}
private static bool TryHelper(string s, out Exception result)
{
try
{
HelperMethod(s);
result = null;
return true;
}
catch (Exception ex)
{
// log here to preserve stack trace
result = ex;
return false;
}
}
private static void HelperMethod(string s)
{
if (s.Equals("ex"))
{
throw new Exception("s can be anything except 'ex'");
}
}
}

Another way is to flatten the try/catch blocks, useful if you're using some exception-happy API:
public void Foo()
{
try
{
HelperMethod("value 1");
return; // finished
}
catch (Exception e)
{
// possibly log exception
}
try
{
HelperMethod("value 2");
return; // finished
}
catch (Exception e)
{
// possibly log exception
}
// ... more here if needed
}

An option for retry (that most people will probably flame) would be to use a goto. C# doesn't have filtered exceptions but this could be used in a similar manner.
const int MAX_RETRY = 3;
public static void DoWork()
{
//Do Something
}
public static void DoWorkWithRetry()
{
var #try = 0;
retry:
try
{
DoWork();
}
catch (Exception)
{
#try++;
if (#try < MAX_RETRY)
goto retry;
throw;
}
}

In this case you know this "exception" probably will happen so I would prefer a simple approach an leave exceptions for the unknown events.
//run initial process
try
{
//initial information used in helper method
string s1 = "value 1";
//call helper method
if(!HelperMethod(s1))
{
//backup information if first process generates an exception in the helper method
string s2 = "value 2";
if(!HelperMethod(s2))
{
return ErrorOfSomeKind;
}
}
return Ok;
}
catch(ApplicationException ex)
{
throw;
}

I know that I've done the above nested try catch recently to handle decoding data where two third party libraries throw exceptions on failure to decode (Try json decode, then try base64 decode), but my preference is to have functions return a value which can be checked.
I generally only use the throwing of exceptions to exit early and notify something up the chain about the error if it's fatal to the process.
If a function is unable to provide a meaningful response, that is not typically a fatal problem (Unlike bad input data).
It seems like the main risk in nested try catch is that you also end up catching all the other (maybe important) exceptions that might occur.

Enforcing required function call

I have a "Status" class in C#, used like this:
Status MyFunction()
{
if(...) // something bad
return new Status(false, "Something went wrong")
else
return new Status(true, "OK");
}
You get the idea.
All callers of MyFunction should check the returned Status:
Status myStatus = MyFunction();
if ( ! myStatus.IsOK() )
// handle it, show a message,...
Lazy callers however can ignore the Status.
MyFunction(); // call function and ignore returned Status
or
{
Status myStatus = MyFunction();
} // lose all references to myStatus, without calling IsOK() on it
Is it possible to make this impossible? e.g. an throw exception
In general: is it possible to write a C# class on which you have to call a certain function?
In the C++ version of the Status class, I can write a test on some private bool bIsChecked in the destructor and ring some bells when someone doesn't check this instance.
What is the equivalent option in C#?
I read somewhere that "You don't want a destructor in your C# class"
Is the Dispose method of the IDisposable interface an option?
In this case there are no unmanaged resources to free.
Additionally, it is not determined when the GC will dispose the object.
When it eventually gets disposed, is it still possible to know where and when you ignored that specific Status instance?
The "using" keyword does help, but again, it is not required for lazy callers.

I know this doesn't answer your question directly, but if "something went wrong" within your function (unexpected circumstances) I think you should be throwing an exception rather than using status return codes.
Then leave it up to the caller to catch and handle this exception if it can, or allow it to propogate if the caller is unable to handle the situation.
The exception thrown could be of a custom type if this is appropriate.
For expected alternative results, I agree with #Jon Limjap's suggestion. I'm fond of a bool return type and prefixing the method name with "Try", a la:
bool TryMyFunction(out Status status)
{
}

If you really want to require the user to retrieve the result of MyFunction, you might want to void it instead and use an out or ref variable, e.g.,
void MyFunction(out Status status)
{
}
It might look ugly but at least it ensures that a variable is passed into the function that will pick up the result you need it to pick up.
#Ian,
The problem with exceptions is that if it's something that happens a little too often, you might be spending too much system resources for the exception. An exception really should be used for exceptional errors, not totally expected messages.

Even System.Net.WebRequest throws an exception when the returned HTTP status code is an error code. The typical way to handle it is to wrap a try/catch around it. You can still ignore the status code in the catch block.
You could, however, have a parameter of Action< Status> so that the caller is forced to pass a callback function that accepts a status and then checking to see if they called it.
void MyFunction(Action<Status> callback)
{ bool errorHappened = false;
if (somethingBadHappend) errorHappened = true;
Status status = (errorHappend)
? new Status(false, "Something went wrong")
: new Status(true, "OK");
callback(status)
if (!status.isOkWasCalled)
throw new Exception("Please call IsOK() on Status").
}
MyFunction(status => if (!status.IsOK()) onerror());
If you're worried about them calling IsOK() without doing anything, use Expression< Func< Status,bool>> instead and then you can analyse the lambda to see what they do with the status:
void MyFunction(Expression<Func<Status,bool>> callback)
{ if (!visitCallbackExpressionTreeAndCheckForIsOKHandlingPattern(callback))
throw new Exception
("Please handle any error statuses in your callback");
bool errorHappened = false;
if (somethingBadHappend) errorHappened = true;
Status status = (errorHappend)
? new Status(false, "Something went wrong")
: new Status(true, "OK");
callback.Compile()(status);
}
MyFunction(status => status.IsOK() ? true : onerror());
Or forego the status class altogether and make them pass in one delegate for success and another one for an error:
void MyFunction(Action success, Action error)
{ if (somethingBadHappened) error(); else success();
}
MyFunction(()=>;,()=>handleError());

I am fairly certain you can't get the effect you want as a return value from a method. C# just can't do some of the things C++ can. However, a somewhat ugly way to get a similar effect is the following:
using System;
public class Example
{
public class Toy
{
private bool inCupboard = false;
public void Play() { Console.WriteLine("Playing."); }
public void PutAway() { inCupboard = true; }
public bool IsInCupboard { get { return inCupboard; } }
}
public delegate void ToyUseCallback(Toy toy);
public class Parent
{
public static void RequestToy(ToyUseCallback callback)
{
Toy toy = new Toy();
callback(toy);
if (!toy.IsInCupboard)
{
throw new Exception("You didn't put your toy in the cupboard!");
}
}
}
public class Child
{
public static void Play()
{
Parent.RequestToy(delegate(Toy toy)
{
toy.Play();
// Oops! Forgot to put the toy away!
});
}
}
public static void Main()
{
Child.Play();
Console.ReadLine();
}
}
In the very simple example, you get an instance of Toy by calling Parent.RequestToy, and passing it a delegate. Instead of returning the toy, the method immediately calls the delegate with the toy, which must call PutAway before it returns, or the RequestToy method will throw an exception. I make no claims as to the wisdom of using this technique -- indeed in all "something went wrong" examples an exception is almost certainly a better bet -- but I think it comes about as close as you can get to your original request.

Using Status as a return value remembers me of the "old days" of C programming, when you returned an integer below 0 if something didn't work.
Wouldn't it be better if you throw an exception when (as you put it) something went wrong? If some "lazy code" doesn't catch your exception, you'll know for sure.

Instead of forcing someone to check the status, I think you should assume the programmer is aware of this risks of not doing so and has a reason for taking that course of action. You don't know how the function is going to be used in the future and placing a limitation like that only restricts the possibilities.

That would sure be nice to have the compiler check that rather than through an expression. :/
Don't see any way to do that though...

You can throw an exception by:
throw MyException;
[global::System.Serializable]
public class MyException : Exception
{
//
// For guidelines regarding the creation of new exception types, see
// http://msdn.microsoft.com/library/default.asp?url=/library/en-us/cpgenref/html/cpconerrorraisinghandlingguidelines.asp
// and
// http://msdn.microsoft.com/library/default.asp?url=/library/en-us/dncscol/html/csharp07192001.asp
//
public MyException () { }
public MyException ( string message ) : base( message ) { }
public MyException ( string message, Exception inner ) : base( message, inner ) { }
protected MyException (
System.Runtime.Serialization.SerializationInfo info,
System.Runtime.Serialization.StreamingContext context )
: base( info, context ) { }
}
The above exception is fully customizable to your requirements.
One thing I would say is this, I would leave it to the caller to check the return code, it is their responsability you just provide the means and interface. Also, It is a lot more efficient to use return codes and check the status with an if statement rather than trhowing exceptions. If it really is an Exceptional circumstance, then by all means throw away... but say if you failed to open a device, then it might be more prudent to stick with the return code.

#Paul you could do it at compile time with Extensible C#.

GCC has a warn_unused_result attribute which is ideal for this sort of thing. Perhaps the Microsoft compilers have something similar.

One pattern which may sometimes be helpful if the object to which code issues requests will only be used by a single thread(*) is to have the object keep an error state, and say that if an operation fails the object will be unusable until the error state is reset (future requests should fail immediately, preferably by throwing an immediate exception which includes information about both the previous failure and the new request). In cases where calling code happens to anticipate a problem, this may allow the calling code to handle the problem more cleanly than if an exception were thrown; problems which are not ignored by the calling code will generally end up triggering an exception pretty soon after they occur.
(*) If a resource will be accessed by multiple threads, create a wrapper object for each thread, and have each thread's requests go through its own wrapper.
This pattern is usable even in contexts where exceptions aren't, and may sometimes be very practical in such cases. In general, however, some variation of the try/do pattern is usually better. Have methods throw exception on failure unless the caller explicitly indicates (by using a TryXX method) that failures are expected. If callers say failures are expected but don't handle them, that's their problem. One could combine the try/do with a second layer of protection using the scheme above, but I'm not sure whether it would be worth the cost.