How to robustly call third-party code in the presence of exceptions? - exception

In a language which uses exceptions to signal errors, I want to call some third-party code and, if it fails, run fallback code instead. For example:
try:
result = third_party.fast_calculation()
catch:
result = slower_calculation()
In my experience, it is very rare to know all of the exceptions that could be thrown by the third-party code. Therefore I cannot list these exceptions in the catch clause. On the other hand, I am frequently advised not to catch every possible exception.
How should I write the catch clause in this situation?

You should catch specific exception types only if you have a specific way to handle them. You can (and should) catch as many specific types of exception as needed, in the most appropriate order.
In case you just want to treat every exception the same way, I believe your current, untyped catch is as good as it gets. The real problem, IMO, comes when you leave an empty catch, since client code cannot know if the function actually did what it was supposed to do.

First, check that your third-party code actually throws exceptions. It may not.
Second, check the results returned if no exception is thrown by the third party code. A status value may be returned to indicate if a successful result was achieved. If so, a status check may be needed to determine if recovery actions (e.g. use the "slower_calculation" method) are needed after a bad return status.

Related

Why is it better to throw an exception rather than return an error code?

Legacy error handling tends to follow the method that all functions return a code depending on success/failure. You would check this code and handle (if an error) appropriately.
Modern programming languages however follow the exception model, where if something exceptional happens that cannot be handled properly an exception gets thrown - which keeps bubbling up until its handled.
My question is why have we moved towards the exception model? What are the reasons behind this? Why is it better?
Explanations/links would be much appreciated.
I've written about this at length: Exceptions vs. status returns, but briefly:
Exceptions leaves your code clean of all the checks necessary when testing status returns on every call,
Exceptions let you use the return value of functions for actual values,
Exceptions can carry more information than a status return can,
Most importantly: exceptions can't be ignored through inaction, while status returns can.
To expand on the last point: if you forget to do what you should be doing with status returns, you ignore errors. If you forget to do what you should be doing with exceptions, the exception bubbles up to the outer layer of the software where it becomes visible.
Here are a couple of reasons
Ignoring an exception requires action by the developer while ignoring a bad returning value requires exactly 0 action. This, in theory, makes it more likely a developer will handle an error vs. ignoring it or not even realizing it was happening.
Provides a cleaner separation between the point of an error and the handling. It doesn't force manual propagation of the error at every point in between.
Exceptions can a larger and richer information payload than a simple error code. There are ways to do this with error codes but it's more of an afterthought and is a bit cumbersome.
Status codes are typically better than exceptions in cases where they represent cases that a function's immediate calling code is prepared to handle. The problem with status codes is that if the immediate calling code doesn't handle them, it's likely nothing will. If code throws an exception and the immediate calling code isn't prepared to handle it, the exception will propagate to code which at least claims to be so prepared.
Throwing an exception is being explicit about something went wrong which is harder to be ignored by the programmer (As the program will terminate if there is no catch block).
An exception will terminate the execution of the rest of the function and will pass the control to next catch block (If there is any in the current call stack. If there is not, then the program will terminate)
For Javascript
Yes, it is optional but you should always use a .catch() handler (for Promises).
It means, we can handle many errors in many functions in a single handler like below;
storeData()
.then(changeSomething)
.then(storeData)
.catch((e) => {
// Handle the error.
})
Here we can see how to handle error for the three different functions all in one. Essentially the behavior of catch is as it would wrapping all of your functions in a synchronous try/catch.

How should exceptions be handled when processing a collection?

Let's say you have a function (X) that takes a single object and does some work with it. It can fail in some way:
function X(obj) throws SomeException
And I have a collection of these objects, and want to run X on all of them. So I wrap that up in its own function:
function Y(objs)
foreach obj in objs
X(obj)
end
end
What is the best way of handling the exceptions returned by X?
One obvious choice would be to not catch the exception at all, causing the processing of the collection to stop at the first failure.
Another would be to catch each exception as it occurs and save it aside. At the end, you could then throw a new exception if there were any exceptions at all. You could optionally choose to include these exceptions as nested exceptions of the new exception.
Another would be to ignore the exception completely, returning something like a boolean array that marks success/failure, or even returning the array of exceptions that occurred.
It depends very much on the contxt of what you are doing.
If every item must be processed then breaking at the first exception would may be the best option.
If you are committing a bunch of items which have been parsed from a third party source and you are expecting some failures keeping a note of the failed items and returning a status object may well be the way to go.
I think all three that you listed are valid, and which one you pick depends on the specific use case - that is, which is most appropriate for your application. "Best practices" can't trump application logic, and well-written practices don't dictate logic.
There is no one right answer.
All of what you said is an option. You can do nothing, let it propagate. You can catch and immediately rethrow with more information about which object was being processed when the exception occurred, which is much better. You can catch and throw one exception at the end, optionally with all inner exceptions. You can also catch and log exception and not rethrow at all.
From time to time I've used all of these techniques. It really depends on what your app is specifically doing and your business requirements.
This would completely depend on the state of the list after completion.
If it is acceptable that certain items do not process correctly, then either swallow them or return a boolean indicating success/failure.
If it is required that all items process correctly, I would have the exception bubble up rather.
You might want to avoid creating an exception list, you might rather want to check the items before hand and return a list of items that do not meet the required state.
The answer depends entirely on what you're doing with your objects, and the context in which it happens.
If it's okay that not all are processed, then obviously you don't want to fail hard and terminate execution - allowing the calling code access to the exception details might be helpful, or it might not be.
If it's important that all are processed, then you return some value saying it failed, or let the exception bubble further up if the operation is vital to the program state.
This is a very slippery slope you're riding. Handling an exception requires the program to restore the state of the program to the state it had before the exception was thrown so it can meaningful continue. That's going to be mighty difficult to do, the collection obviously contains an object that cannot be processed. Restoring state would require restoring the collection as well.
Maybe that's possible. But not in the code in your snippet, it doesn't has any code that appears to be responsible for entering objects in the collection. You should not catch an exception if you don't know how to restore state.
Well I guess is pretty more of the same, but it stands a no right or wrong answer, just the more right or worg for some given case.
To capture and continuing rethrowing exceptions can be rather costy to the machine, but it can be your wanted behavior.
I can give two pointers:
- if you intend to discard the collection process if one is wrong just capture the exception and do whatever you may feel like it;
- if you want to assert whether some item is write or wrong, asserting all, you should capture them all, and if you feel like it, either just log it or discard it, or save all those 'inner exceptions' and rethrow or treat them all at the end of the process.
What is exactly the situation?

How do I distinguish between a truly exceptional circumstances and a routine bad condition?

A lot of developers say only throw exceptions in truly exceptional circumstances. One of these would be if an external hard drive I want to write to is not switched on (therefore not a connected/registered drive). However, there are some situations which are difficult to work out whether they are truly exceptional or not.
For example, entering a string for a folder path but it is not found. In this case, if there is any input which cannot be found (like a name in a collection which is not found), is it best to just return an error message and some action?
E.G.
public void Find(string name)
{
if(Names.contains(name)
{
string s = Names.get(name);
}
if(!Names.contains(string name)
{
throw new ???Exception;
}
}
Or do something like display a popup and handle the situation gracefully?
Is it wise to throw an exception in an else or if statement? Looking at a list of code smells regarding exception handling would do me a lot of favours.
Generally, it works like this:
If you can handle the situation without any interruptions, do so. (File doesn't exist, but its input isn't essential to continuing operation [preferences, optional configuration, etc])
If you need user intervention, ask them. (File doesn't exist, but you need it to continue operating)
If it's a problem the user can't fix (out-of-memory, failed hardware, etc), then throw an exception.
Each place has their own standard for the details, but I find the above to work generally.
if your code can recover from the exception, do so
if it would be acceptable to require clients to check a return value for minor expected exceptions, do so - this is a judgement call, see below
in all other cases, throw an exception (or don't catch a called method's exception)
for the second case, minor and expected are highly context-sensitive
IMHO don't use exceptions for control-flow, but in all other cases you are probably safer throwing them
note that some 'optimization' advice tells you to check for conditions instead of relying on exceptions, e.g. file-not-found, but in reality you should still expect the exception because a file may be deleted or moved in between the statement that checks for existence and the statement that tries to open the file!
summary: in general, throwing exceptions is the safest course. Of course, directly warning or asking the user should only be done in the user-interface code!
Exceptions should be used if there is something that can be done about it, but the code that detects it can't known what to do.

Which, and why, do you prefer Exceptions or Return codes?

My question is what do most developers prefer for error handling, Exceptions or Error Return Codes. Please be language(or language family) specific and why you prefer one over the other.
I'm asking this out of curiosity. Personally I prefer Error Return Codes since they are less explosive and don't force user code to pay the exception performance penalty if they don't want to.
update: thanks for all the answers! I must say that although I dislike the unpredictability of code flow with exceptions. The answer about return code (and their elder brother handles) do add lots of Noise to the code.
For some languages (i.e. C++) Resources leak should not be a reason
C++ is based on RAII.
If you have code that could fail, return or throw (that is, most normal code), then you should have your pointer wrapped inside a smart pointer (assuming you have a very good reason to not have your object created on stack).
Return codes are more verbose
They are verbose, and tend to develop into something like:
if(doSomething())
{
if(doSomethingElse())
{
if(doSomethingElseAgain())
{
// etc.
}
else
{
// react to failure of doSomethingElseAgain
}
}
else
{
// react to failure of doSomethingElse
}
}
else
{
// react to failure of doSomething
}
In the end, you code is a collection of idented instructions (I saw this kind of code in production code).
This code could well be translated into:
try
{
doSomething() ;
doSomethingElse() ;
doSomethingElseAgain() ;
}
catch(const SomethingException & e)
{
// react to failure of doSomething
}
catch(const SomethingElseException & e)
{
// react to failure of doSomethingElse
}
catch(const SomethingElseAgainException & e)
{
// react to failure of doSomethingElseAgain
}
Which cleanly separate code and error processing, which can be a good thing.
Return codes are more brittle
If not some obscure warning from one compiler (see "phjr" 's comment), they can easily be ignored.
With the above examples, assume than someone forgets to handle its possible error (this happens...). The error is ignored when "returned", and will possibly explode later (i.e. a NULL pointer). The same problem won't happen with exception.
The error won't be ignored. Sometimes, you want it to not explode, though... So you must chose carefully.
Return Codes must sometimes be translated
Let's say we have the following functions:
doSomething, which can return an int called NOT_FOUND_ERROR
doSomethingElse, which can return a bool "false" (for failed)
doSomethingElseAgain, which can return an Error object (with both the __LINE__, __FILE__ and half the stack variables.
doTryToDoSomethingWithAllThisMess which, well... Use the above functions, and return an error code of type...
What is the type of the return of doTryToDoSomethingWithAllThisMess if one of its called functions fail ?
Return Codes are not a universal solution
Operators cannot return an error code. C++ constructors can't, too.
Return Codes means you can't chain expressions
The corollary of the above point. What if I want to write:
CMyType o = add(a, multiply(b, c)) ;
I can't, because the return value is already used (and sometimes, it can't be changed). So the return value becomes the first parameter, sent as a reference... Or not.
Exception are typed
You can send different classes for each kind of exception. Ressources exceptions (i.e. out of memory) should be light, but anything else could be as heavy as necessary (I like the Java Exception giving me the whole stack).
Each catch can then be specialized.
Don't ever use catch(...) without re-throwing
Usually, you should not hide an error. If you do not re-throw, at the very least, log the error in a file, open a messagebox, whatever...
Exception are... NUKE
The problem with exception is that overusing them will produce code full of try/catches. But the problem is elsewhere: Who try/catch his/her code using STL container? Still, those containers can send an exception.
Of course, in C++, don't ever let an exception exit a destructor.
Exception are... synchronous
Be sure to catch them before they bring out your thread on its knees, or propagate inside your Windows message loop.
The solution could be mixing them?
So I guess the solution is to throw when something should not happen. And when something can happen, then use a return code or a parameter to enable to user to react to it.
So, the only question is "what is something that should not happen?"
It depends on the contract of your function. If the function accepts a pointer, but specifies the pointer must be non-NULL, then it is ok to throw an exception when the user sends a NULL pointer (the question being, in C++, when didn't the function author use references instead of pointers, but...)
Another solution would be to show the error
Sometimes, your problem is that you don't want errors. Using exceptions or error return codes are cool, but... You want to know about it.
In my job, we use a kind of "Assert". It will, depending on the values of a configuration file, no matter the debug/release compile options:
log the error
open a messagebox with a "Hey, you have a problem"
open a messagebox with a "Hey, you have a problem, do you want to debug"
In both development and testing, this enable the user to pinpoint the problem exactly when it is detected, and not after (when some code cares about the return value, or inside a catch).
It is easy to add to legacy code. For example:
void doSomething(CMyObject * p, int iRandomData)
{
// etc.
}
leads a kind of code similar to:
void doSomething(CMyObject * p, int iRandomData)
{
if(iRandomData < 32)
{
MY_RAISE_ERROR("Hey, iRandomData " << iRandomData << " is lesser than 32. Aborting processing") ;
return ;
}
if(p == NULL)
{
MY_RAISE_ERROR("Hey, p is NULL !\niRandomData is equal to " << iRandomData << ". Will throw.") ;
throw std::some_exception() ;
}
if(! p.is Ok())
{
MY_RAISE_ERROR("Hey, p is NOT Ok!\np is equal to " << p->toString() << ". Will try to continue anyway") ;
}
// etc.
}
(I have similar macros that are active only on debug).
Note that on production, the configuration file does not exist, so the client never sees the result of this macro... But it is easy to activate it when needed.
Conclusion
When you code using return codes, you're preparing yourself for failure, and hope your fortress of tests is secure enough.
When you code using exception, you know that your code can fail, and usually put counterfire catch at chosen strategic position in your code. But usually, your code is more about "what it must do" then "what I fear will happen".
But when you code at all, you must use the best tool at your disposal, and sometimes, it is "Never hide an error, and show it as soon as possible". The macro I spoke above follow this philosophy.
I use both actually.
I use return codes if it's a known, possible error. If it's a scenario that I know can, and will happen, then there's a code that gets sent back.
Exceptions are used solely for things that I'm NOT expecting.
According to Chapter 7 titled "Exceptions" in Framework Design Guidelines: Conventions, Idioms, and Patterns for Reusable .NET Libraries, numerous rationales are given for why using exceptions over return values is necessary for OO frameworks such as C#.
Perhaps this is the most compelling reason (page 179):
"Exceptions integrate well with object-oriented languages. Object-oriented languages tend to impose constraints on member signatures that are not imposed by functions in non-OO languages. For example, in the case of constructors, operator overloads, and properties, the developer has no choice in the return value. For this reason, it is not possible to standardize on return-value-based error reporting for object-oriented frameworks. An error reporting method, such as exceptions, which is out of band of the method signature is the only option."
My preference (in C++ and Python) is to use exceptions. The language-provided facilities make it a well-defined process to both raise, catch and (if necessary) re-throw exceptions, making the model easy to see and use. Conceptually, it's cleaner than return codes, in that specific exceptions can be defined by their names, and have additional information accompanying them. With a return code, you're limited to just the error value (unless you want to define an ReturnStatus object or something).
Unless the code you're writing is time-critical, the overhead associated with unwinding the stack is not significant enough to worry about.
Exceptions should only be returned where something happens that you were not expecting.
The other point of exceptions, historically, is that return codes are inherently proprietary, sometimes a 0 could be returned from a C function to indicate success, sometimes -1, or either of them for a fail with 1 for a success. Even when they are enumerated, enumerations can be ambiguous.
Exceptions can also provide a lot more information, and specifically spell out well 'Something Went Wrong, here's what, a stack trace and some supporting information for the context'
That being said, a well enumerated return code can be useful for a known set of outcomes, a simple 'heres n outcomes of the function, and it just ran this way'
Always use exceptions by default, BUT consider providing an additional tester-doer option (TryX)!
For me the answer is really clear. When the context dictates a Try or Tester-Doer pattern (ie cpu intensive or public api), I will ADDITIONALLY provide those methods to the exception throwing version. I think blanket rules of avoiding exceptions are misguided, unsupported, and likely cause far more expense in terms of bugs, than any performance issues they claim to prevent.
No, Microsoft does NOT say to not use exceptions (common misinterpretation).
It says if you're designing an API provide ways to help a user of that API to avoid THROWING exceptions if they need too (Try and Tester-Doer patterns)
❌ DO NOT use exceptions for the normal flow of control, if possible.
Except for system failures and operations with potential race
conditions, framework designers should design APIs so users can write
code that does not throw exceptions. For example, you can provide a
way to check preconditions before calling a member so users can write
code that does not throw exceptions.
What is inferred here is that the non-tester-doer/non-try implementation SHOULD throw an exception upon failure and then the user CAN change that to one of your tester-doer or try methods for performance. Pit of success is maintained for safety and the user OPTS INTO the more dangerous but more performant method.
Microsoft DOES say to NOT use return codes TWICE, here:
❌ DO NOT return error codes.
Exceptions are the primary means of reporting errors in frameworks.
✔️ DO report execution failures by throwing exceptions.
and here:
❌ DO NOT use error codes because of concerns that exceptions might
affect performance negatively.
To improve performance, it is possible to use either the Tester-Doer
Pattern or the Try-Parse Pattern, described in the next two sections.
If you're not using exceptions you're probably breaking this other rule of returning return codes or booleans from a non-tester/non-try implementation.
Again, TryParse does not replace Parse. It is provided in addition to Parse
MAIN REASON: Return codes fail the "Pit of Success" test for me almost every time.
It is far too easy to forget to check a return code and then have a red-herring error later on.
var success = Save()? How much performance is worth someone forgetting an if check here?
var success = TrySave()? Better, but are we going to abuse everything with the TryX pattern? Did you still provide a Save method?
Return codes don't have any of the great debugging information on them like call stack, inner exceptions.
Return codes do not propagate which, along with the point above, tends to drive excessive and interwoven diagnostic logging instead of logging in one centralized place (application and thread level exception handlers).
Return codes tend to drive messy code in the form of nested 'if' blocks
Developer time spent debugging an unknown issue that would otherwise have been an obvious exception (pit of success) IS expensive.
If the team behind C# didn't intend for exceptions to govern control flow, execeptions wouldn't be typed, there would be no "when" filter on catch statements, and there would be no need for the parameter-less 'throw' statement.
Regarding Performance:
Exceptions may be computationally expensive RELATIVE to not throwing at all, but they're called EXCEPTIONS for a reason. Speed comparisons always manage to assume a 100% exception rate which should never be the case. Even if an exception is 100x slower, how much does that really matter if it only happens 1% of the time?
Context is everything. For example, A Tester-Doer or Try option to avoid a unique key violation is likely to waste more time and resources on average (checking for existance when a collision is rare) than just assuming a successful entry and catching that rare violation.
Unless we're talking floating point arithmetic for graphics applications or something similar, CPU cycles are cheap compared to developer time.
Cost from a time perspective carries the same argument. Relative to database queries or web service calls or file loads, normal application time will dwarf exception time. Exceptions were nearly sub-MICROsecond in 2006
I dare anybody that works in .net, to set your debugger to break on all exceptions and disable just my code and see how many exceptions are already happening that you don't even know about.
Jon Skeet says "[Exceptions are] not slow enough to make it worth avoiding them in normal use". The linked response also contains two articles from Jon on the subject. His generalized theme is that exceptions are fine and if you're experiencing them as a performance problem, there's likely a larger design issue.
In Java, I use (in the following order):
Design-by-contract (ensuring preconditions are met before trying anything that might fail). This catches most things and I return an error code for this.
Returning error codes whilst processing work (and performing rollback if needed).
Exceptions, but these are used only for unexpected things.
I dislike return codes because they cause the following pattern to mushroom throughout your code
CRetType obReturn = CODE_SUCCESS;
obReturn = CallMyFunctionWhichReturnsCodes();
if (obReturn == CODE_BLOW_UP)
{
// bail out
goto FunctionExit;
}
Soon a method call consisting of 4 function calls bloats up with 12 lines of error handling.. Some of which will never happen. If and switch cases abound.
Exceptions are cleaner if you use them well... to signal exceptional events .. after which the execution path cannot continue. They are often more descriptive and informational than error codes.
If you have multiple states after a method call that should be handled differently (and are not exceptional cases), use error codes or out params. Although Personaly I've found this to be rare..
I've hunted a bit about the 'performance penalty' counterargument.. more in the C++ / COM world but in the newer languages, I think the difference isn't that much. In any case, when something blows up, performance concerns are relegated to the backburner :)
I wrote a blog post about this a while ago.
The performance overhead of throwing an exception should not play any role in your decision. If you're doing it right, after all, an exception is exceptional.
A great piece of advice I got from The Pragmatic Programmer was something along the lines of "your program should be able to perform all its main functionality without using exceptions at all".
I have a simple set of rules:
1) Use return codes for things you expect your immediate caller to react to.
2) Use exceptions for errors that are broader in scope, and may reasonable be expected to be handled by something many levels above the caller so that awareness of the error does not have to percolate up through many layers, making code more complex.
In Java I only ever used unchecked exceptions, checked exceptions end up just being another form of return code and in my experience the duality of what might be "returned" by a method call was generally more of a hinderance than a help.
I use Exceptions in python in both Exceptional, and non-Exceptional circumstances.
It is often nice to be able to use an Exception to indicate the "request could not be performed", as opposed to returning an Error value. It means that you /always/ know that the return value is the right type, instead of arbitarily None or NotFoundSingleton or something. Here is a good example of where I prefer to use an exception handler instead of a conditional on the return value.
try:
dataobj = datastore.fetch(obj_id)
except LookupError:
# could not find object, create it.
dataobj = datastore.create(....)
The side effect is that when a datastore.fetch(obj_id) is run, you never have to check if its return value is None, you get that error immediately for free. This is counter to the argument, "your program should be able to perform all its main functionality without using exceptions at all".
Here is another example of where exceptions are 'exceptionally' useful, in order to write code for dealing with the filesystem that isn't subject to race conditions.
# wrong way:
if os.path.exists(directory_to_remove):
# race condition is here.
os.path.rmdir(directory_to_remove)
# right way:
try:
os.path.rmdir(directory_to_remove)
except OSError:
# directory didn't exist, good.
pass
One system call instead of two, no race condition. This is a poor example because obviously this will fail with an OSError in more circumstances than the directory doesn't exist, but it's a 'good enough' solution for many tightly controlled situations.
I believe the return codes adds to code noise. For example, I always hated the look of COM/ATL code due to return codes. There had to be an HRESULT check for every line of code. I consider the error return code is one of the bad decisions made by architects of COM. It makes it difficult to do logical grouping of the code, thus code review becomes difficult.
I am not sure about the performance comparison when there is an explicit check for the return code every line.
Exceptions are not for error handling, IMO. Exceptions are just that; exceptional events that you did not expect. Use with caution I say.
Error codes can be OK, but returning 404 or 200 from a method is bad, IMO. Use enums (.Net) instead, that makes the code more readable and easier to use for other developers. Also you don't have to maintain a table over numbers and descriptions.
Also; the try-catch-finally pattern is an anti-pattern in my book. Try-finally can be good, try-catch can also be good but try-catch-finally is never good. try-finally can often times be replaced by a "using" statement (IDispose pattern), which is better IMO. And Try-catch where you actually catch an exception you're able to handle is good, or if you do this:
try{
db.UpdateAll(somevalue);
}
catch (Exception ex) {
logger.Exception(ex, "UpdateAll method failed");
throw;
}
So as long as you let the exception continue to bubble it's OK. Another example is this:
try{
dbHasBeenUpdated = db.UpdateAll(somevalue); // true/false
}
catch (ConnectionException ex) {
logger.Exception(ex, "Connection failed");
dbHasBeenUpdated = false;
}
Here I actually handle the exception; what I do outside of the try-catch when the update method fails is another story, but I think my point has been made. :)
Why is then try-catch-finally an anti-pattern? Here's why:
try{
db.UpdateAll(somevalue);
}
catch (Exception ex) {
logger.Exception(ex, "UpdateAll method failed");
throw;
}
finally {
db.Close();
}
What happens if the db object has already been closed? A new exception is thrown and it has to be handled! This is better:
try{
using(IDatabase db = DatabaseFactory.CreateDatabase()) {
db.UpdateAll(somevalue);
}
}
catch (Exception ex) {
logger.Exception(ex, "UpdateAll method failed");
throw;
}
Or, if the db object does not implement IDisposable do this:
try{
try {
IDatabase db = DatabaseFactory.CreateDatabase();
db.UpdateAll(somevalue);
}
finally{
db.Close();
}
}
catch (DatabaseAlreadyClosedException dbClosedEx) {
logger.Exception(dbClosedEx, "Database connection was closed already.");
}
catch (Exception ex) {
logger.Exception(ex, "UpdateAll method failed");
throw;
}
That's my 2 cents anyway! :)
I generally prefer return codes because they let the caller decide whether the failure is exceptional.
This approach is typical in the Elixir language.
# I care whether this succeeds. If it doesn't return :ok, raise an exception.
:ok = File.write(path, content)
# I don't care whether this succeeds. Don't check the return value.
File.write(path, content)
# This had better not succeed - the path should be read-only to me.
# If I get anything other than this error, raise an exception.
{:error, :erofs} = File.write(path, content)
# I want this to succeed but I can handle its failure
case File.write(path, content) do
:ok => handle_success()
error => handle_error(error)
end
People mentioned that return codes can cause you to have a lot of nested if statements, but that can be handled with better syntax. In Elixir, the with statement lets us easily separate a series of happy-path return value from any failures.
with {:ok, content} <- get_content(),
:ok <- File.write(path, content) do
IO.puts "everything worked, happy path code goes here"
else
# Here we can use a single catch-all failure clause
# or match every kind of failure individually
# or match subsets of them however we like
_some_error => IO.puts "one of those steps failed"
_other_error => IO.puts "one of those steps failed"
end
Elixir still has functions that raise exceptions. Going back to my first example, I could do either of these to raise an exception if the file can't be written.
# Raises a generic MatchError because the return value isn't :ok
:ok = File.write(path, content)
# Raises a File.Error with a descriptive error message - eg, saying
# that the file is read-only
File.write!(path, content)
If I, as the caller, know that I want to raise an error if the write fails, I can choose to call File.write! instead of File.write.
Or I can choose to call File.write and handle each of the possible reasons for failure differently.
Of course it's always possible to rescue an exception if we want to. But compared to handling an informative return value, it seems awkward to me. If I know that a function call can fail or even should fail, its failure isn't an exceptional case.
With any decent compiler or runtime environment exceptions do not incur a significant penalty. It's more or less like a GOTO statement that jumps to the exception handler. Also, having exceptions caught by a runtime environment (like the JVM) helps isolating and fixing a bug a lot easier. I'll take a NullPointerException in Java over a segfault in C any day.
I prefer to use exceptions for error handling and return values (or parameters) as the normal result of a function. This gives an easy and consistent error-handling scheme and if done correctly it makes for much cleaner looking code.
One of the big differences is that exceptions force you to handle an error, whereas error return codes can go unchecked.
Error return codes, if used heavily, can also cause very ugly code with lots of if tests similar to this form:
if(function(call) != ERROR_CODE) {
do_right_thing();
}
else {
handle_error();
}
Personally I prefer to use exceptions for errors that SHOULD or MUST be acted upon by the calling code, and only use error codes for "expected failings" where returning something is actually valid and possible.
There is many reason to prefer Exceptions over return code:
Usually, for readibility, people try to minimize the number of return statement in a method. Doing so, exceptions prevent to do some extra work while in a incoorect state, and thus prevent to potentially damage more data.
Exception are generally more verbose arn more easilly extensible than return value. Assume that a method return natural number and that you use negative numbers as return code when an error occurs, if the scope of you method change and now return integers, you'll have to modify all the method calls instead of just tweaking a little bit the exception.
Exceptions allows more easilly to separate error handling of normal behaviour. They allows to ensure that some operations performs somehow as an atomic operation.
I only use exceptions, no return codes. I'm talking about Java here.
The general rule I follow is if I have a method called doFoo() then it follows that if it doesn't "do foo", as it were, then something exceptional has happened and an Exception should be thrown.
One thing I fear about exceptions is that throwing an exception will screw up code flow. For example if you do
void foo()
{
MyPointer* p = NULL;
try{
p = new PointedStuff();
//I'm a module user and I'm doing stuff that might throw or not
}
catch(...)
{
//should I delete the pointer?
}
}
Or even worse what if I deleted something I shouldn't have, but got thrown to catch before I did the rest of the cleanup. Throwing put a lot of weight on the poor user IMHO.
My general rule in the exception vs. return code argument:
Use errorcodes when you need localization/internationalization -- in .NET, you could use these errorcodes to reference a resource file which will then display the error in the appropriate language. Otherwise, use exceptions
Use exceptions only for errors that are really exceptional. If it's something that happens fairly often, either use a boolean or an enum errorcode.
I don't find return codes to be less ugly than exceptions. With the exception, you have the try{} catch() {} finally {} where as with return codes you have if(){}. I used to fear exceptions for the reasons given in the post; you don't know if the pointer needs to be cleared, what have you. But I think you have the same problems when it comes to the return codes. You don't know the state of the parameters unless you know some details about the function/method in question.
Regardless, you have to handle the error if possible. You can just as easily let an exception propagate to the top level as ignore a return code and let the program segfault.
I do like the idea of returning a value (enumeration?) for results and an exception for an exceptional case.
For a language like Java, I would go with Exception because the compiler gives compile time error if exceptions are not handled.This forces the calling function to handle/throw the exceptions.
For Python, I am more conflicted. There is no compiler so it's possible that caller does not handle the exception thrown by the function leading to runtime exceptions. If you use return codes you might have unexpected behavior if not handled properly and if you use exceptions you might get runtime exceptions.
There are some important aspects that remain unmentioned in this - very interesting - discussion so far.
First, it is important to note that exceptions don't apply to distributed computing, but error codes still do. Imagine communicating services distributed over multiple servers. Some communication might even be asynchronous. And the services might even use different technology stacks. Cleary, an error-handling concept is crucial here. And clearly, exceptions can't be used in this most general case, since errors have to be serialized things sent "through the cable", perhaps even in a language-neutral way. From that angle, error codes (really, error messages) are more universal than exceptions. One needs good error-message Kung Fu once one assumes a system-architect view and things need to scale.
The second point is a very different, it is about if or how a language represents discriminated unions. The question was strictly speaking about "error codes". And so were some the answers, mentioning that error codes cannot transport information as nicely as exceptions. This is true if an error code is a number. But for a fairer contrasting with exceptions, one should probably consider error values of discriminated union type. So, the return value of the callee would be of discriminated union type, and it would either be the desired happy-path value or the payload the exception would otherwise have. How often this approach is elegant enough to be preferable depends on the programming language. For example, F# has super elegant discriminated unions and the according pattern matching. In such a language it would be more seductive to avoid exceptions than in, say, C++.
The third and final point is about functional programming and pure functions. Exceptions are (in a practical way and in a theoretical-computer-science way) "side effects". In other words, functions or methods that deal with exceptions are not pure. (One practical consequence is that with exceptions one must pay attention to evaluation order.) By contrast, error values are pure, because the are just ordinary return values, with no side effects involved. Therefore, functional programmers may more likely frown upon exceptions than object-oriented programmers. (In particular if the language also has an elegant representation of the aforementioned discriminated unions.)
I prefer exceptions over return codes.
Consider a scenario when I call a function foo and forget to handle potential errors (exceptions).
If errors in foo are passed via return codes (error codes),
at compile time, I won't be alerted
if I run the code
if the error doesn't happen, I don't notice my mistake
if the error happens but doesn't affect the code after calling foo, I don't notice my mistake
if the error happens and affects the code after calling foo, I notice my mistake, but it may be hard to locate the problem
If errors in foo are passed via exceptions that are thrown,
at compile time, I won't be alerted
if I run the code
if the error doesn't happen, I don't notice my mistake
if the error happens, I'm assured to notice my mistake
From the comparison above, my conclusion is that exceptions are better than error codes.
However, exceptions are not perfect. There are at least two critical problems:
As discussed above, if I forget to handle a potential exception, I won't be alerted until I run the code and the exception is actually thrown.
It's hard to determine all the exceptions foo may throw, especially when foo calls other functions (which may also throw exceptions).
A feature in Java, “checked exceptions”, solves both problems.
In Java, when defining a function foo, I'm required to explicitly specify what exceptions it may throw using the throws keyword. Example:
private static void foo() throws FileNotFoundException {
File file = new File("not_existing_file.txt");
FileInputStream stream = new FileInputStream(file);
}
If I call foo and forget to handle the potential FileNotFoundException, the compiler produces an error. I get alerted at compile time (problem 1 solved). And all possible exceptions are listed explicitly (problem 2 solved).

When to throw an exception?

I have exceptions created for every condition that my application does not expect. UserNameNotValidException, PasswordNotCorrectException etc.
However I was told I should not create exceptions for those conditions. In my UML those ARE exceptions to the main flow, so why should it not be an exception?
Any guidance or best practices for creating exceptions?
My personal guideline is: an exception is thrown when a fundamental assumption of the current code block is found to be false.
Example 1: say I have a function which is supposed to examine an arbitrary class and return true if that class inherits from List<>. This function asks the question, "Is this object a descendant of List?" This function should never throw an exception, because there are no gray areas in its operation - every single class either does or does not inherit from List<>, so the answer is always "yes" or "no".
Example 2: say I have another function which examines a List<> and returns true if its length is more than 50, and false if the length is less. This function asks the question, "Does this list have more than 50 items?" But this question makes an assumption - it assumes that the object it is given is a list. If I hand it a NULL, then that assumption is false. In that case, if the function returns either true or false, then it is breaking its own rules. The function cannot return anything and claim that it answered the question correctly. So it doesn't return - it throws an exception.
This is comparable to the "loaded question" logical fallacy. Every function asks a question. If the input it is given makes that question a fallacy, then throw an exception. This line is harder to draw with functions that return void, but the bottom line is: if the function's assumptions about its inputs are violated, it should throw an exception instead of returning normally.
The other side of this equation is: if you find your functions throwing exceptions frequently, then you probably need to refine their assumptions.
Because they're things that will happen normally. Exceptions are not control flow mechanisms. Users often get passwords wrong, it's not an exceptional case. Exceptions should be a truly rare thing, UserHasDiedAtKeyboard type situations.
My little guidelines are heavily influenced by the great book "Code complete":
Use exceptions to notify about things that should not be ignored.
Don't use exceptions if the error can be handled locally
Make sure the exceptions are at the same level of abstraction as the rest of your routine.
Exceptions should be reserved for what's truly exceptional.
It is NOT an exception if the username is not valid or the password is not correct. Those are things you should expect in the normal flow of operation. Exceptions are things that are not part of the normal program operation and are rather rare.
I do not like using exceptions because you can not tell if a method throws an exception just by looking at the call. Thats why exceptions should only be used if you can't handle the situation in a decent manner (think "out of memory" or "computer is on fire").
One rule of thumb is to use exceptions in the case of something you couldn't normally predict. Examples are database connectivity, missing file on disk, etc. For scenarios that you can predict, ie users attempting to log in with a bad password you should be using functions that return booleans and know how to handle the situation gracefully. You don't want to abruptly end execution by throwing an exception just because someone mistyped their password.
Others propose that exceptions should not be used because the bad login is to be expected in a normal flow if the user mistypes. I disagree and I don't get the reasoning. Compare it with opening a file.. if the file doesn't exist or is not available for some reason then an exception will be thrown by the framework. Using the logic above this was a mistake by Microsoft. They should have returned an error code. Same for parsing, webrequests, etc., etc..
I don't consider a bad login part of a normal flow, it's exceptional. Normally the user types the correct password, and the file does exist. The exceptional cases are exceptional and it's perfectly fine to use exceptions for those. Complicating your code by propagating return values through n levels up the stack is a waste of energy and will result in messy code. Do the simplest thing that could possibly work. Don't prematurely optimize by using error codes, exceptional stuff by definition rarely happens, and exceptions don't cost anything unless you throw them.
I think you should only throw an exception when there's nothing you can do to get out of your current state. For example if you are allocating memory and there isn't any to allocate. In the cases you mention you can clearly recover from those states and can return an error code back to your caller accordingly.
You will see plenty of advice, including in answers to this question, that you should throw exceptions only in "exceptional" circumstances. That seems superficially reasonable, but is flawed advice, because it replaces one question ("when should I throw an exception") with another subjective question ("what is exceptional"). Instead, follow the advice of Herb Sutter (for C++, available in the Dr Dobbs article When and How to Use Exceptions, and also in his book with Andrei Alexandrescu, C++ Coding Standards): throw an exception if, and only if
a precondition is not met (which typically makes one of the following
impossible) or
the alternative would fail to meet a post-condition or
the alternative would fail to maintain an invariant.
Why is this better? Doesn't it replace the question with several questions about preconditions, postconditions and invariants? This is better for several connected reasons.
Preconditions, postconditions and invariants are design characteristics of our program (its internal API), whereas the decision to throw is an implementation detail. It forces us to bear in mind that we must consider the design and its implementation separately, and our job while implementing a method is to produce something that satisfies the design constraints.
It forces us to think in terms of preconditions, postconditions and invariants, which are the only assumptions that callers of our method should make, and are expressed precisely, enabling loose coupling between the components of our program.
That loose coupling then allows us to refactor the implementation, if necessary.
The post-conditions and invariants are testable; it results in code that can be easily unit tested, because the post-conditions are predicates our unit-test code can check (assert).
Thinking in terms of post-conditions naturally produces a design that has success as a post-condition, which is the natural style for using exceptions. The normal ("happy") execution path of your program is laid out linearly, with all the error handling code moved to the catch clauses.
Exceptions are a somewhat costly effect, if for example you have a user that provides an invalid password, it is typically a better idea to pass back a failure flag, or some other indicator that it is invalid.
This is due to the way that exceptions are handled, true bad input, and unique critical stop items should be exceptions, but not failed login info.
I would say there are no hard and fast rules on when to use exceptions. However there are good reasons for using or not using them:
Reasons to use exceptions:
The code flow for the common case is clearer
Can return complex error information as an object (although this can also be achieved using error "out" parameter passed by reference)
Languages generally provide some facility for managing tidy cleanup in the event of the exception (try/finally in Java, using in C#, RAII in C++)
In the event no exception is thrown, execution can sometimes be faster than checking return codes
In Java, checked exceptions must be declared or caught (although this can be a reason against)
Reasons not to use exceptions:
Sometimes it's overkill if the error handling is simple
If exceptions are not documented or declared, they may be uncaught by calling code, which may be worse than if the the calling code just ignored a return code (application exit vs silent failure - which is worse may depend on the scenario)
In C++, code that uses exceptions must be exception safe (even if you don't throw or catch them, but call a throwing function indirectly)
In C++, it is hard to tell when a function might throw, therefore you must be paranoid about exception safety if you use them
Throwing and catching exceptions is generally significantly more expensive compared to checking a return flag
In general, I would be more inclined to use exceptions in Java than in C++ or C#, because I am of the opinion that an exception, declared or not, is fundamentally part of the formal interface of a function, since changing your exception guarantee may break calling code. The biggest advantage of using them in Java IMO, is that you know that your caller MUST handle the exception, and this improves the chance of correct behaviour.
Because of this, in any language, I would always derive all exceptions in a layer of code or API from a common class, so that calling code can always guarantee to catch all exceptions. Also I would consider it bad to throw exception classes that are implementation-specific, when writing an API or library (i.e. wrap exceptions from lower layers so that the exception that your caller receives is understandable in the context of your interface).
Note that Java makes the distinction between general and Runtime exceptions in that the latter need not be declared. I would only use Runtime exception classes when you know that the error is a result of a bug in the program.
If it's code running inside a loop that will likely cause an exception over and over again, then throwing exceptions is not a good thing, because they are pretty slow for large N. But there is nothing wrong with throwing custom exceptions if the performance is not an issue. Just make sure that you have a base exception that they all inherite, called BaseException or something like that. BaseException inherits System.Exception, but all of your exceptions inherit BaseException. You can even have a tree of Exception types to group similar types, but this may or may not be overkill.
So, the short answer is that if it doesn't cause a significant performance penalty (which it should not unless you are throwing a lot of exceptions), then go ahead.
Exception classes are like "normal" classes. You create a new class when it "is" a different type of object, with different fields and different operations.
As a rule of thumb, you should try balance between the number of exceptions and the granularity of the exceptions. If your method throws more than 4-5 different exceptions, you can probably merge some of them into more "general" exceptions, (e.g. in your case "AuthenticationFailedException"), and using the exception message to detail what went wrong. Unless your code handles each of them differently, you needn't creates many exception classes. And if it does, may you should just return an enum with the error that occured. It's a bit cleaner this way.
the rule of thumb for throwing exceptions is pretty simple. you do so when your code has entered into an UNRECOVERABLE INVALID state. if data is compromised or you cannot wind back the processing that occurred up to the point then you must terminate it. indeed what else can you do? your processing logic will eventually fail elsewhere. if you can recover somehow then do that and do not throw exception.
in your particular case if you were forced to do something silly like accept money withdrawal and only then check user/pasword you should terminate the process by throwing an exception to notify that something bad has happened and prevent further damage.
I agree with japollock way up there--throw an acception when you are uncertain about the outcome of an operation. Calls to APIs, accessing filesystems, database calls, etc. Anytime you are moving past the "boundaries" of your programming languages.
I'd like to add, feel free to throw a standard exception. Unless you are going to do something "different" (ignore, email, log, show that twitter whale picture thingy, etc), then don't bother with custom exceptions.
I'd say that generally every fundamentalism leads to hell.
You certainly wouldn't want to end up with exception driven flow, but avoiding exceptions altogether is also a bad idea. You have to find a balance between both approaches. What I would not do is to create an exception type for every exceptional situation. That is not productive.
What I generally prefer is to create two basic types of exceptions which are used throughout the system: LogicalException and TechnicalException. These can be further distinguished by subtypes if needed, but it is not generally not necessary.
The technical exception denotes the really unexpected exception like database server being down, the connection to the web service threw the IOException and so on.
On the other hand the logical exceptions are used to propagate the less severe erroneous situation to the upper layers (generally some validation result).
Please note that even the logical exception is not intended to be used on regular basis to control the program flow, but rather to highlight the situation when the flow should really end. When used in Java, both exception types are RuntimeException subclasses and error handling is highly aspect oriented.
So in the login example it might be wise to create something like AuthenticationException and distinguish the concrete situations by enum values like UsernameNotExisting, PasswordMismatch etc. Then you won't end up in having a huge exception hierarchy and can keep the catch blocks on maintainable level. You can also easily employ some generic exception handling mechanism since you have the exceptions categorized and know pretty well what to propagate up to the user and how.
Our typical usage is to throw the LogicalException during the Web Service call when the user's input was invalid. The Exception gets marshalled to the SOAPFault detail and then gets unmarshalled to the exception again on the client which is resulting in showing the validation error on one certain web page input field since the exception has proper mapping to that field.
This is certainly not the only situation: you don't need to hit web service to throw up the exception. You are free to do so in any exceptional situation (like in the case you need to fail-fast) - it is all at your discretion.
In general you want to throw an exception for anything that can happen in your application that is "Exceptional"
In your example, both of those exceptions look like you are calling them via a password / username validation. In that case it can be argued that it isn't really exceptional that someone would mistype a username / password.
They are "exceptions" to the main flow of your UML but are more "branches" in the processing.
If you attempted to access your passwd file or database and couldn't, that would be an exceptional case and would warrant throwing an exception.
Firstly, if the users of your API aren't interested in specific, fine-grained failures, then having specific exceptions for them isn't of any value.
Since it's often not possible to know what may be useful to your users, a better approach is to have the specific exceptions, but ensure they inherit from a common class (e.g., std::exception or its derivatives in C++). That allows your client to catch specific exceptions if they choose, or the more general exception if they don't care.
Exceptions are intended for events that are abnormal behaviors, errors, failures, and such. Functional behavior, user error, etc., should be handled by program logic instead. Since a bad account or password is an expected part of the logic flow in a login routine, it should be able to handle those situations without exceptions.
The simple answer is, whenever an operation is impossible (because of either application OR because it would violate business logic). If a method is invoked and it impossible to do what the method was written to do, throw an Exception. A good example is that constructors always throw ArgumentExceptions if an instance cannot be created using the supplied parameters. Another example is InvalidOperationException, which is thrown when an operation cannot be performed because of the state of another member or members of the class.
In your case, if a method like Login(username, password) is invoked, if the username is not valid, it is indeed correct to throw a UserNameNotValidException, or PasswordNotCorrectException if password is incorrect. The user cannot be logged in using the supplied parameter(s) (i.e. it's impossible because it would violate authentication), so throw an Exception. Although I might have your two Exceptions inherit from ArgumentException.
Having said that, if you wish NOT to throw an Exception because a login failure may be very common, one strategy is to instead create a method that returns types that represent different failures. Here's an example:
{ // class
...
public LoginResult Login(string user, string password)
{
if (IsInvalidUser(user))
{
return new UserInvalidLoginResult(user);
}
else if (IsInvalidPassword(user, password))
{
return new PasswordInvalidLoginResult(user, password);
}
else
{
return new SuccessfulLoginResult();
}
}
...
}
public abstract class LoginResult
{
public readonly string Message;
protected LoginResult(string message)
{
this.Message = message;
}
}
public class SuccessfulLoginResult : LoginResult
{
public SucccessfulLogin(string user)
: base(string.Format("Login for user '{0}' was successful.", user))
{ }
}
public class UserInvalidLoginResult : LoginResult
{
public UserInvalidLoginResult(string user)
: base(string.Format("The username '{0}' is invalid.", user))
{ }
}
public class PasswordInvalidLoginResult : LoginResult
{
public PasswordInvalidLoginResult(string password, string user)
: base(string.Format("The password '{0}' for username '{0}' is invalid.", password, user))
{ }
}
Most developers are taught to avoid Exceptions because of the overhead caused by throwing them. It's great to be resource-conscious, but usually not at the expense of your application design. That is probably the reason you were told not to throw your two Exceptions. Whether to use Exceptions or not usually boils down to how frequently the Exception will occur. If it's a fairly common or an fairly expectable result, this is when most developers will avoid Exceptions and instead create another method to indicate failure, because of the supposed consumption of resources.
Here's an example of avoiding using Exceptions in a scenario like just described, using the Try() pattern:
public class ValidatedLogin
{
public readonly string User;
public readonly string Password;
public ValidatedLogin(string user, string password)
{
if (IsInvalidUser(user))
{
throw new UserInvalidException(user);
}
else if (IsInvalidPassword(user, password))
{
throw new PasswordInvalidException(password);
}
this.User = user;
this.Password = password;
}
public static bool TryCreate(string user, string password, out ValidatedLogin validatedLogin)
{
if (IsInvalidUser(user) ||
IsInvalidPassword(user, password))
{
return false;
}
validatedLogin = new ValidatedLogin(user, password);
return true;
}
}
for me Exception should be thrown when a required technical or business rule fails.
for instance if a car entity is associated with array of 4 tires ... if one tire or more are null ... an exception should be Fired "NotEnoughTiresException" , cuz it can be caught at different level of the system and have a significant meaning through logging.
besides if we just try to flow control the null and prevent the instanciation of the car .
we might never never find the source of the problem , cuz the tire isn't supposed to be null in the first place .
the main reason for avoiding throwing an exception is that there is a lot of overhead involved with throwing an exception.
One thing the article below states is that an exception is for an exceptional conditions and errors.
A wrong user name is not necessarily a program error but a user error...
Here is a decent starting point for exceptions within .NET:
http://msdn.microsoft.com/en-us/library/ms229030(VS.80).aspx
Throwing exceptions causes the stack to unwind, which has some performance impacts (admitted, modern managed environments have improved on that). Still repeatedly throwing and catching exceptions in a nested situation would be a bad idea.
Probably more important than that, exceptions are meant for exceptional conditions. They should not be used for ordinary control flow, because this will hurt your code's readability.
I have three type of conditions that I catch.
Bad or missing input should not be an exception. Use both client side js and server side regex to detect, set attributes and forward back to the same page with messages.
The AppException. This is usually an exception that you detect and throw with in your code. In other words these are ones you expect (the file does not exist). Log it, set the message, and forward back to the general error page. This page usually has a bit of info about what happened.
The unexpected Exception. These are the ones you don't know about. Log it with details and forward them to a general error page.
Hope this helps
Security is conflated with your example: You shouldn't tell an attacker that a username exists, but the password is wrong. That's extra information you don't need to share. Just say "the username or password is incorrect."
I have philosophical problems with the use of exceptions. Basically, you are expecting a specific scenario to occur, but rather than handling it explicitly you are pushing the problem off to be handled "elsewhere." And where that "elsewhere" is can be anyone's guess.
To my mind, the fundamental question should be whether one would expect that the caller would want to continue normal program flow if a condition occurs. If you don't know, either have separate doSomething and trySomething methods, where the former returns an error and the latter does not, or have a routine that accepts a parameter to indicate whether an exception should be thrown if it fails). Consider a class to send commands to a remote system and report responses. Certain commands (e.g. restart) will cause the remote system to send a response but then be non-responsive for a certain length of time. It is thus useful to be able to send a "ping" command and find out whether the remote system responds in a reasonable length of time without having to throw an exception if it doesn't (the caller would probably expect that the first few "ping" attempts would fail, but one would eventually work). On the other hand, if one has a sequence of commands like:
exchange_command("open tempfile");
exchange_command("write tempfile data {whatever}");
exchange_command("write tempfile data {whatever}");
exchange_command("write tempfile data {whatever}");
exchange_command("write tempfile data {whatever}");
exchange_command("close tempfile");
exchange_command("copy tempfile to realfile");
one would want failure of any operation to abort the whole sequence. While one could check each operation to ensure it succeeded, it's more helpful to have the exchange_command() routine throw an exception if a command fails.
Actually, in the above scenario it may be helpful to have a parameter to select a number of failure-handling modes: never throw exceptions, throw exceptions for communication errors only, or throw exceptions in any cases where a command does not return a "success" indication.
You may use a little bit generic exceptions for that conditions. For e.g. ArgumentException is meant to be used when anything goes wrong with the parameters to a method (with the exception of ArgumentNullException). Generally you would not need exceptions like LessThanZeroException, NotPrimeNumberException etc. Think of the user of your method. The number of the conditions that she will want to handle specifically is equal to the number of the type of the exceptions that your method needs to throw. This way, you can determine how detailed exceptions you will have.
By the way, always try to provide some ways for users of your libraries to avoid exceptions. TryParse is a good example, it exists so that you don't have to use int.Parse and catch an exception. In your case, you may want to provide some methods to check if user name is valid or password is correct so your users (or you) will not have to do lots of exception handling. This will hopefully result in more readble code and better performance.
Ultimately the decision comes down to whether it is more helpful to deal with application-level errors like this using exception handling, or via your own home-rolled mechanism like returning status codes. I don't think there's a hard-and-fast rule about which is better, but I would consider:
Who's calling your code? Is this a public API of some sort or an internal library?
What language are you using? If it's Java, for example, then throwing a (checked) exception puts an explicit burden on your caller to handle this error condition in some way, as opposed to a return status which could be ignored. That could be good or bad.
How are other error conditions in the same application handled? Callers won't want to deal with a module that handles errors in an idiosyncratic way unlike anything else in the system.
How many things can go wrong with the routine in question, and how would they be handled differently? Consider the difference between a series of catch blocks that handle different errors and a switch on an error code.
Do you have structured information about the error you need to return? Throwing an exception gives you a better place to put this information than just returning a status.
Some useful things to think about when deciding whether an exception is appropriate:
what level of code you want to have run after the exception candidate occurs - that is, how many layers of the call stack should unwind. You generally want to handle an exception as close as possible to where it occurs. For username/password validation, you would normally handle failures in the same block of code, rather than letting an exception bubble up. So an exception is probably not appropriate. (OTOH, after three failed login attempts, control flow may shift elsewhere, and an exception may be appropriate here.)
Is this event something you would want to see in an error log? Not every exception is written to an error log, but it's useful to ask whether this entry in an error log would be useful - i.e., you would try to do something about it, or would be the garbage you ignore.
"PasswordNotCorrectException" isn't a good example for using exceptions. Users getting their passwords wrong is to be expected, so it's hardly an exception IMHO. You probably even recover from it, showing a nice error message, so it's just a validity check.
Unhandled exceptions will stop the execution eventually - which is good. If you're returning false, null or error codes, you will have to deal with the program's state all by yourself. If you forget to check conditions somewhere, your program may keep running with wrong data, and you may have a hard time figuring out what happened and where.
Of course, you could cause the same problem with empty catch statements, but at least spotting those is easier and doesn't require you to understand the logic.
So as a rule of thumb:
Use them wherever you don't want or simply can't recover from an error.
I would say that exceptions should be thrown if an unexpected behaviour is occuring that wasnt meant to be.
Like trying to update or delete a non existing entity. And it should be catched where the Exception can be handled and has meaning. For working in an alternative way to continue, add logging or returning a specific result on Api level.
If you expect something to be the case, you should build code to check and ensure it.