Why have both a return-value and a side-effect when raising an error for a mock? - python-mock

A while back I found the following solution-pattern here on StackOverflow for raising an error from a mocked function (I cannot find the original link anymore, sorry):
def _mock_raises_error(my_mock, error_type, output):
my_mock.return_value = mock.Mock()
my_mock.side_effect = error_type(output)
# for example
with mock.patch('mymodule.interface.function') as mock_function:
_mock_raises_error(mock_function, Exception, 'Some error-message')
This works as intended, so I always have used this pattern.
Recently a colleague of mine asked why this definition both has a return_value as a side_effect and why that was needed. And to my shame I could not come up with the correct answer (only that I copied it from StackOverflow, but that is not an explanation why it is correct).
So now my question (to be able to explain it in the future) is why isn't giving the side_effect enough? What does the return_value add?


Let's start with the documentation for unittest.mock.
side_effect: A function to be called whenever the Mock is called. See the side_effect attribute. Useful for raising exceptions or dynamically changing return values. The function is called with the same arguments as the mock, and unless it returns DEFAULT, the return value of this function is used as the return value.
Alternatively side_effect can be an exception class or instance. In this case the exception will be raised when the mock is called.
return_value: The value returned when the mock is called. By default this is a new Mock (created on first access). See the return_value attribute.
One generally tests either with a return_value or with a side_effect. In my experience, I often test with return_value. This is for my standard case, when I'm trying to mock some functionality (such as obtaining the contents of a directory) without the need to make a function call (such as to the OS and to a directory which might change).
I mock the side_effect less often (usually when I'm testing the function's reaction to an exception). (Example, when I'm handling a file-not-found exception.)
In this example of a function that tests for the existence of a file in a directory, I might want to have both the results (say an empty list of files) and the side effect (say a file-not-found exception). But generally I will test one or the other.
The OP's example is only testing for the exception, so the side-effect is critical (but the return value is not). If you were testing for both, you'd need both.

Related

Is there a way to defer parameter resolution?

(I'm reasonably sure the answer is "no", but I want to make sure.)
In JUnit 5 you can write an extension that is an implementation of ParameterResolver. Before your test runs, if the method in question has parameters, then an extension that implements ParameterResolver can return the object suitable as an argument for that parameter.
You can also write an extension that is an implementation of InvocationInterceptor, that is in charge of intercepting a test method's execution. You can get any arguments (such as those resolved by ParameterResolvers), but it appears you cannot change them.
In terms of execution order, if there are relevant parameters, then a ParameterResolver will "fire" first, and then any InvocationInterceptors will "fire" next.
(Lastly, if your test method declares parameters, but there are no ParameterResolvers to resolve them, everything craps out.)
Putting this all together:
Consider the case when a parameter can't really be properly resolved until the stuff that an interceptor sets up prior to execution is complete:
What is the best way, if there is one, to have all of the following:
A parameter that conceivably only the interceptor could resolve
Deferred resolution of that parameter (i.e. the actual parameter value is not sought by the JUnit internals until interception time so that the interceptor could resolve it just-in-time before calling proceed()
…?
(In my very concrete case, I got lucky: the parameter I'm interested in is an interface, so I "resolve" it to a dummy implementation, and then, at interception time, "fill" the dummy implementation with a delegate that does the real work. I can't think of a better way with the existing JUnit 5 toolkit.)
(I can almost get there if ReflectiveInvocationContext would allow me to set its arguments: my resolveParameter implementation could return null and my interceptor could replace the null reference it found in the arguments with an appropriate non-null argument just-in-time.)
(I also am at least aware of the ExecutableInvoker interface that is reachable from the ExtensionContext, but I'm unclear how that would help me in this scenario, since parameter resolution happens before interception.)

Is there an efficient way to avoid instantiating a class with syntax errors?

As you may know, it is pretty easy to have active code of a class containing syntax errors (someone activated the code ignoring syntax warnings or someone changed the signature of a method the class calls, for instance).
This means that also dynamic instantiation of such a class via
CREATE OBJECT my_object TYPE (class_name).
will fail with an apparently uncatchable SYNTAX_ERROR exception. The goal is to write code that does not terminate when this occurs.
Known solutions:
Wrap the CREATE OBJECT statement inside an RFC function module, call the module with destination NONE, then catch the (classic) exception SYSTEM_FAILURE from the RFC call. If the RFC succeeds, actually create the object (you can't pass the created object out of the RFC because RFC function modules can't pass references, and objects cannot be passed other than by reference as far as I know).
This solution is not only inelegant, but impacts performance rather harshly since an entirely new LUW is spawned by the RFC call. Additionally, you're not actually preventing the SYNTAX_ERROR dump, just letting it dump in a thread you don't care about. It will still, annoyingly, show up in ST22.
Before attempting to instantiate the class, call
cl_abap_typedescr=>describe_by_name( class_name )
and catch the class-based exception CX_SY_RTTI_SYNTAX_ERROR it throws when the code it attempts to describe has syntax errors.
This performs much better than the RFC variant, but still seems to add unnecessary overhead - usually, I don't want the type information that describe_by_name returns, I'm solely calling it to get a catchable exception, and when it succeeds, its result is thrown away.
Is there a way to prevent the SYNTAX_ERROR dump without adding such overhead?

Most efficient way we could come up with:
METHODS has_correct_syntax
IMPORTING
class_name TYPE seoclsname
RETURNING
VALUE(result) TYPE abap_bool.
METHOD has_correct_syntax.
DATA(include_name) = cl_oo_classname_service=>get_cs_name( class_name ).
READ REPORT include_name INTO DATA(source_code).
SYNTAX-CHECK FOR source_code MESSAGE DATA(message) LINE DATA(line) WORD DATA(word).
result = xsdbool( sy-subrc = 0 ).
ENDMETHOD.
Still a lot of overhead for loading the program and syntax-checking it. However, at least none additional for compiling descriptors you are not interested in.
We investigated when we produced a dependency manager that wires classes together upon startup and should exclude syntactically wrong candidates.
CS includes don't always exist, so get_cs_name might come back empty. Seems to depend on the NetWeaver version and the editor the developer used.
If you are certain that the syntax errors are caused by the classes’ own code, you might want to consider buffering the results of the syntax checks and only revalidate when the class changed after it was last checked. This does not work if you expect syntax errors to be caused by something outside those classes.

Is overloading a method without throwing an exception an antipattern?

We are currently designing an API for storing settings and we are considering having these two types of methods:
public Object getSetting(String key) {
// return null if key does not exist
}
public Object getSettingExc(String key) {
// throw a runtime exception if key does not exist
}
Somehow I feel that this just isn't right, but I can't think of any disadvantages except for doubling the number of functions in the API and perhaps decreased code readability (if I know the setting MUST exist, I think I should throw an exception explicitly rather than relying on the get method).
What are your opinions on this?

Exceptions are for exceptional occurrences, when the code cannot continue to function according to its advertised function.
Requesting a setting that isn't set is hardly exception-worthy. If "you" (i.e. the calling code) "know" that setting "must" exist, call getSetting(), check the return value for null, and then throw an exception yourself out of the knowledge that it should have been there. Add a meaningful message about what setting in which context wasn't found. (This is something only the caller knows.)
Don't delegate the throwing of the exception to code that doesn't know the context of the query or that the setting should be there, and needs to be told explicitly (by getting called under a different name). Also, getSettingExc() will most likely be only a null-check-and-throw wrapper around getSetting() anyway, so why not do it at a point where you can make the exception message so much more helpful?
IMHO. (And this is the point where I realize I should have voted-to-close instead of writing an answer...)

This is introducing a weird kind of coupling between the object structure and the potential error conditions. Regarding your comment:
I'm just trying to gather arguments to persuade other guys in my team.
The onus should be on the proponent of this design to justify it, not on you to justify against it. Nobody else uses this in any design I've ever seen.
This does however remind me of another design that maybe is what your team intended? Consider these two methods:
public Object getSetting(String key) {
// return the setting or throw an exception
}
public Object getSettingOrDefault(String key) {
// return the setting or a pre-determined default
}
This aligns the methods more with the functionality than with the error conditions. getSetting() can advertise that it might throw an exception, whereas getSettingOrDefault() can advertise that it will default to a specific value if none can be found in the settings.
If Java has optional parameters or something akin to that, getSettingOrDefault() might even accept as an argument a default to use in the event of no such setting. That might get a little kludgy for consuming code though, just sort of thinking out loud on that one.
Either way, the API should reflect the functionality and not the error conditions. Ideally there should be only one method, but if there's a noticeable need to differentiate between a method that throws and a method that doesn't (and I could certainly see that being the case in a language with checked exceptions), those two should align with the functionality rather than with the exception.

IMHO having two methods to do precisely the same operation indicates that you as the API designer did not complete the job of 'designing' your API. Choose one way or another, publicize it via the API (javadocs) and then the consumers will be consistent in their usage (one way or the other).

Determine if a function is async-signal-safe (can be called inside a signal handler)

My questions are:
Is there a way to conclusively determine if a function is async-signal-safe if you don't have access to its implementation?
If not, is there a way to test if function would be async-signal-safe enough to call from a signal handler?
If you reads the man pages of signal() or sigaction(), you get a list of async-signal-safe functions (functions that can be safely called inside a signal handler). However, I believe that this list is not exhaustive. For example, the following page http://linux.die.net/man/7/signal, under the Async-signal-safe functions header, reads:
POSIX.1-2004 (also known as POSIX.1-2001 Technical Corrigendum 2) requires an implementation to guarantee that the following functions can be safely called inside a signal handler:
And then it proceeds to list the normal async-signal-safe functions listed in the man pages above. As I read it, it says "it requires", not "these are the only ones".
For example, this site says that back_trace_symbols_fd() is async-signal safe. That function obtains is data from dladdr() and it doesn't use malloc() like back_trace_symbols(), so it looks like it may be safe. Also, I did some testing, and the output struct of dladdr() contains char* variables, but these are NOT malloc'ed at runtime. The char string they point to exists at run-time even before dladdr() is called.
Any thoughts or ideas that can point me in the right direction are appreciated.

If you don't have access to the function's implementation, you can look at the manual page. If the manual page doesn't say it is async-safe, and the POSIX standard doesn't say it is async-safe, the only safe conclusion is "it is not async-safe" (coupled with "do not use it").
There is no 100% reliable way to test whether a function is async-safe. Remember, testing can only show the presence of bugs, not their absence (Dijkstra). The mere fact that you don't manage to tickle the function into misbehaving under test may simply mean that your testing is not adequate (but rest assured, the important customer who you can't afford to offend will immediately and accidentally devise a devastatingly effective test that demonstrates that the function is not async-safe almost as soon as you release the code with the faulty assumption).

What are you hoping to achieve in the signal handler? You should consider whether it is the right place for it. It is probably best to follow the advice of the man page:
In general, signal handlers should do little more
than set a flag; most other actions are not safe.

When should I use Debug.Assert()?

I've been a professional software engineer for about a year now, having graduated with a CS degree. I've known about assertions for a while in C++ and C, but had no idea they existed in C# and .NET at all until recently.
Our production code contains no asserts whatsoever and my question is this...
Should I begin using Asserts in our production code? And if so, When is its use most appropriate? Would it make more sense to do
Debug.Assert(val != null, "message");
or
if ( val == null )
throw new exception("message");

In Debugging Microsoft .NET 2.0 Applications John Robbins has a big section on assertions. His main points are:
Assert liberally. You can never have too many assertions.
Assertions don't replace exceptions. Exceptions cover the things your code demands; assertions cover the things it assumes.
A well-written assertion can tell you not just what happened and where (like an exception), but why.
An exception message can often be cryptic, requiring you to work backwards through the code to recreate the context that caused the error. An assertion can preserve the program's state at the time the error occurred.
Assertions double as documentation, telling other developers what implied assumptions your code depends on.
The dialog that appears when an assertion fails lets you attach a debugger to the process, so you can poke around the stack as if you had put a breakpoint there.
PS: If you liked Code Complete, I recommend following it up with this book. I bought it to learn about using WinDBG and dump files, but the first half is packed with tips to help avoid bugs in the first place.

Put Debug.Assert() everywhere in the code where you want have sanity checks to ensure invariants. When you compile a Release build (i.e., no DEBUG compiler constant), the calls to Debug.Assert() will be removed so they won't affect performance.
You should still throw exceptions before calling Debug.Assert(). The assert just makes sure that everything is as expected while you're still developing.

FWIW ... I find that my public methods tend to use the if () { throw; } pattern to ensure that the method is being called correctly. My private methods tend to use Debug.Assert().
The idea is that with my private methods, I'm the one under control, so if I start calling one of my own private methods with parameters that are incorrect, then I've broken my own assumption somewhere--I should have never gotten into that state. In production, these private asserts should ideally be unnecessary work since I am supposed to be keeping my internal state valid and consistent. Contrast with parameters given to public methods, which could be called by anyone at runtime: I still need to enforce parameter constraints there by throwing exceptions.
Additionally, my private methods can still throw exceptions if something doesn't work at runtime (network error, data access error, bad data retrieved from a third party service, etc.). My asserts are just there to make sure that I haven't broken my own internal assumptions about the state of the object.

From Code Complete (Wikipedia):
8 Defensive Programming
8.2 Assertions
An assertion is code that’s used during development—usually a routine
or macro—that allows a program to check itself as it runs. When an
assertion is true, that means everything is operating as expected.
When it’s false, that means it has detected an unexpected error in the
code. For example, if the system assumes that a customer-information
file will never have more than 50,000 records, the program might
contain an assertion that the number of records is lessthan or equal
to 50,000. As long as the number of records is less than or equal to
50,000, the assertion will be silent. If it encounters more than
50,000 records, however, it will loudly “assert” that there is an
error in the program.
Assertions are especially useful in large, complicated programs and
in high reliability programs. They enable programmers to more quickly
flush out mismatched interface assumptions, errors that creep in when
code is modified, and so on.
An assertion usually takes two arguments: a boolean expression that
describes the assumption that’s supposed to be true and a message to
display if it isn’t.
(…)
Normally, you don’t want users to see assertion messages in
production code; assertions are primarily for use during development
and maintenance. Assertions are normally compiled into the code at
development time and compiled out of the code for production. During
development, assertions flush out contradictory assumptions,
unexpected conditions, bad values passed to routines, and so on.
During production, they are compiled out of the code so that the
assertions don’t degrade system performance.

Use asserts to check developer assumptions and exceptions to check environmental assumptions.

If I were you I would do:
Debug.Assert(val != null);
if ( val == null )
throw new exception();
Or to avoid repeated condition check
if ( val == null )
{
Debug.Assert(false,"breakpoint if val== null");
throw new exception();
}

If you want Asserts in your production code (i.e. Release builds), you can use Trace.Assert instead of Debug.Assert.
This of course adds overhead to your production executable.
Also if your application is running in user-interface mode, the Assertion dialog will be displayed by default, which may be a bit disconcerting for your users.
You can override this behaviour by removing the DefaultTraceListener: look at the documentation for Trace.Listeners in MSDN.
In summary,
Use Debug.Assert liberally to help catch bugs in Debug builds.
If you use Trace.Assert in user-interface mode, you probably want to remove the DefaultTraceListener to avoid disconcerting users.
If the condition you're testing is something your app can't handle, you're probably better off throwing an exception, to ensure execution doesn't continue. Be aware that a user can choose to ignore an assertion.

Asserts are used to catch programmer (your) error, not user error. They should be used only when there is no chance a user could cause the assert to fire. If you're writing an API, for example, asserts should not be used to check that an argument is not null in any method an API user could call. But it could be used in a private method not exposed as part of your API to assert that YOUR code never passes a null argument when it isn't supposed to.
I usually favour exceptions over asserts when I'm not sure.

In Short
Asserts are used for guards and for checking Design by Contract constraints, i.e. to ensure that the state of your code, objects, variables and parameters is operating within the boundaries and limits of your intended design.
Asserts should be for Debug and non-Production builds only. Asserts are typically ignored by the compiler in Release builds.
Asserts can check for bugs / unexpected conditions which ARE in the control of your system
Asserts are NOT a mechanism for first-line validation of user input or business rules
Asserts should not be used to detect unexpected environmental conditions (which are outside the control of the code) e.g. out of memory, network failure, database failure, etc. Although rare, these conditions are to be expected (and your app code cannot fix issues like hardware failure or resource exhaustion). Typically, exceptions will be thrown - your application can then either take corrective action (e.g. retry a database or network operation, attempt to free up cached memory), or abort gracefully if the exception cannot be handled.
A failed Assertion should be fatal to your system - i.e. unlike an exception, do not try and catch or handle failed Asserts - your code is operating in unexpected territory. Stack Traces and crash dumps can be used to determine what went wrong.
Assertions have enormous benefit:
To assist in finding missing validation of user inputs, or upstream bugs in higher level code.
Asserts in the code base clearly convey the assumptions made in the code to the reader
Assert will be checked at runtime in Debug builds.
Once code has been exhaustively tested, rebuilding the code as Release will remove the performance overhead of verifying the assumption (but with the benefit that a later Debug build will always revert the checks, if needed).
... More Detail
Debug.Assert expresses a condition which has been assumed about state by the remainder of the code block within the control of the program. This can include the state of the provided parameters, state of members of a class instance, or that the return from a method call is in its contracted / designed range.
Typically, asserts should crash the thread / process / program with all necessary info (Stack Trace, Crash Dump, etc), as they indicate the presence of a bug or unconsidered condition which has not been designed for (i.e. do not try and catch or handle assertion failures), with one possible exception of when an assertion itself could cause more damage than the bug (e.g. Air Traffic Controllers wouldn't want a YSOD when an aircraft goes submarine, although it is moot whether a debug build should be deployed to production ...)
When should you use Asserts?
At any point in a system, or library API, or service where the inputs to a function or state of a class are assumed valid (e.g. when validation has already been done on user input in the presentation tier of a system, the business and data tier classes typically assume that null checks, range checks, string length checks etc on input have been already done).
Common Assert checks include where an invalid assumption would result in a null object dereference, a zero divisor, numerical or date arithmetic overflow, and general out of band / not designed for behaviour (e.g. if a 32 bit int was used to model a human's age, it would be prudent to Assert that the age is actually between 0 and 125 or so - values of -100 and 10^10 were not designed for).
.Net Code Contracts
In the .Net Stack, Code Contracts can be used in addition to, or as an alternative to using Debug.Assert. Code Contracts can further formalize state checking, and can assist in detecting violations of assumptions at ~compile time (or shortly thereafter, if run as a background check in an IDE).
Design by Contract (DBC) checks available include:
Contract.Requires - Contracted Preconditions
Contract.Ensures - Contracted PostConditions
Invariant - Expresses an assumption about the state of an object at all points in its lifespan.
Contract.Assumes - pacifies the static checker when a call to non-Contract decorated methods is made.

Mostly never in my book.
In the vast majority of occasions if you want to check if everything is sane then throw if it isn't.
What I dislike is the fact that it makes a debug build functionally different to a release build. If a debug assert fails but the functionality works in release then how does that make any sense? It's even better when the asserter has long left the company and no-one knows that part of the code. Then you have to kill some of your time exploring the issue to see if it is really a problem or not. If it is a problem then why isn't the person throwing in the first place?
To me this suggests by using Debug.Asserts you're deferring the problem to someone else, deal with the problem yourself. If something is supposed to be the case and it isn't then throw.
I guess there are possibly performance critical scenarios where you want to optimise away your asserts and they're useful there, however I am yet to encounter such a scenario.

According to the IDesign Standard, you should
Assert every assumption. On average, every fifth line is an assertion.
using System.Diagnostics;
object GetObject()
{...}
object someObject = GetObject();
Debug.Assert(someObject != null);
As a disclaimer I should mention I have not found it practical to implement this IRL. But this is their standard.

All asserts should be code that could be optimised to:
Debug.Assert(true);
Because it's checking something that you have already assumed is true. E.g.:
public static void ConsumeEnumeration<T>(this IEnumerable<T> source)
{
if(source != null)
using(var en = source.GetEnumerator())
RunThroughEnumerator(en);
}
public static T GetFirstAndConsume<T>(this IEnumerable<T> source)
{
if(source == null)
throw new ArgumentNullException("source");
using(var en = source.GetEnumerator())
{
if(!en.MoveNext())
throw new InvalidOperationException("Empty sequence");
T ret = en.Current;
RunThroughEnumerator(en);
return ret;
}
}
private static void RunThroughEnumerator<T>(IEnumerator<T> en)
{
Debug.Assert(en != null);
while(en.MoveNext());
}
In the above, there are three different approaches to null parameters. The first accepts it as allowable (it just does nothing). The second throws an exception for the calling code to handle (or not, resulting in an error message). The third assumes it can't possibly happen, and asserts that it is so.
In the first case, there's no problem.
In the second case, there's a problem with the calling code - it shouldn't have called GetFirstAndConsume with null, so it gets an exception back.
In the third case, there's a problem with this code, because it should already have been checked that en != null before it was ever called, so that it isn't true is a bug. Or in other words, it should be code that could theoretically be optimised to Debug.Assert(true), sicne en != null should always be true!

Use assertions only in cases where you want the check removed for release builds. Remember, your assertions will not fire if you don't compile in debug mode.
Given your check-for-null example, if this is in an internal-only API, I might use an assertion. If it's in a public API, I would definitely use the explicit check and throw.

I thought I would add four more cases, where Debug.Assert can be the right choice.
1) Something I have not seen mentioned here is the additional conceptual coverage Asserts can provide during automated testing. As a simple example:
When some higher-level caller is modified by an author who believes they have expanded the scope of the code to handle additional scenarios, ideally (!) they will write unit tests to cover this new condition. It may then be that the fully integrated code appears to work fine.
However, actually a subtle flaw has been introduced, but not detected in test results. The callee has become non-deterministic in this case, and only happens to provide the expected result. Or perhaps it has yielded a rounding error that was unnoticed. Or caused an error that was offset equally elsewhere. Or granted not only the access requested but additional privileges that should not be granted. Etc.
At this point, the Debug.Assert() statements contained in the callee coupled with the new case (or edge case) driven in by unit tests can provide invaluable notification during test that the original author's assumptions have been invalidated, and the code should not be released without additional review. Asserts with unit tests are the perfect partners.
2) Additionally, some tests are simple to write, but high-cost and unnecessary given the initial assumptions. For example:
If an object can only be accessed from a certain secured entry point, should an additional query be made to a network rights database from every object method to ensure the caller has permissions? Surely not. Perhaps the ideal solution includes caching or some other expansion of features, but the design does not require it. A Debug.Assert() will immediately show when the object has been attached to an insecure entry point.
3) Next, in some cases your product may have no helpful diagnostic interaction for all or part of its operations when deployed in release mode. For example:
Suppose it is an embedded real-time device. Throwing exceptions and restarting when it encounters a malformed packet is counter-productive. Instead the device may benefit from best-effort operation, even to the point of rendering noise in its output. It also may not have a human interface, logging device, or even be physically accessible by human at all when deployed in release mode, and awareness of errors is best provided by assessing the same output. In this case, liberal Assertions and thorough pre-release testing are more valuable than exceptions.
4) Lastly, some tests are unneccessary only because the callee is perceived as extremely reliable. In most cases, the more reusable code is, the more effort has been put into making it reliable. Therefore it is common to Exception for unexpected parameters from callers, but Assert for unexpected results from callees. For example:
If a core String.Find operation states it will return a -1 when the search criteria is not found, you may be able to safely perform one operation rather than three. However, if it actually returned -2, you may have no reasonable course of action. It would be unhelpful to replace the simpler calculation with one that tests separately for a -1 value, and unreasonable in most release environments to litter your code with tests ensuring core libraries are operating as expected. In this case Asserts are ideal.

Quote Taken from The Pragmatic Programmer: From Journeyman to Master
Leave Assertions Turned On
There is a common misunderstanding about assertions, promulgated by
the people who write compilers and language environments. It goes
something like this:
Assertions add some overhead to code. Because they check for things
that should never happen, they'll get triggered only by a bug in the
code. Once the code has been tested and shipped, they are no longer
needed, and should be turned off to make the code run faster.
Assertions are a debugging facility.
There are two patently wrong assumptions here. First, they assume that
testing finds all the bugs. In reality, for any complex program you
are unlikely to test even a miniscule percentage of the permutations
your code will be put through (see Ruthless Testing).
Second, the optimists are forgetting that your program runs in a
dangerous world. During testing, rats probably won't gnaw through a
communications cable, someone playing a game won't exhaust memory, and
log files won't fill the hard drive. These things might happen when
your program runs in a production environment. Your first line of
defense is checking for any possible error, and your second is using
assertions to try to detect those you've missed.
Turning off assertions when you deliver a program to production is
like crossing a high wire without a net because you once made it
across in practice. There's dramatic value, but it's hard to get life
insurance.
Even if you do have performance issues, turn off only those
assertions that really hit you.

You should always use the second approach (throwing exceptions).
Also if you're in production (and have a release-build), it's better to throw an exception (and let the app crash in the worst-case) than working with invalid values and maybe destroy your customer's data (which may cost thousand of dollars).

You should use Debug.Assert to test for logical errors in your programs. The complier can only inform you of syntax errors. So you should definetely use Assert statements to test for logical errors. Like say testing a program that sells cars that only BMWs that are blue should get a 15% discount. The complier could tell you nothing about if your program is logically correct in performing this but an assert statement could.

I've read the answers here and I thought I should add an important distinction. There are two very different ways in which asserts are used. One is as a temporary developer shortcut for "This shouldn't really happen so if it does let me know so I can decide what to do", sort of like a conditional breakpoint, for cases in which your program is able to continue. The other, is a as a way to put assumptions about valid program states in your code.
In the first case, the assertions don't even need to be in the final code. You should use Debug.Assert during development and you can remove them if/when no longer needed. If you want to leave them or if you forget to remove them no problem, since they won't have any consequence in Release compilations.
But in the second case, the assertions are part of the code. They, well, assert, that your assumptions are true, and also document them. In that case, you really want to leave them in the code. If the program is in an invalid state it should not be allowed to continue. If you couldn't afford the performance hit you wouldn't be using C#. On one hand it might be useful to be able to attach a debugger if it happens. On the other, you don't want the stack trace popping up on your users and perhaps more important you don't want them to be able to ignore it. Besides, if it's in a service it will always be ignored. Therefore in production the correct behavior would be to throw an Exception, and use the normal exception handling of your program, which might show the user a nice message and log the details.
Trace.Assert has the perfect way to achieve this. It won't be removed in production, and can be configured with different listeners using app.config.
So for development the default handler is fine, and for production you can create a simple TraceListener like below which throws an exception and activate it in the production config file.
using System.Diagnostics;
public class ExceptionTraceListener : DefaultTraceListener
{
[DebuggerStepThrough]
public override void Fail(string message, string detailMessage)
{
throw new AssertException(message);
}
}
public class AssertException : Exception
{
public AssertException(string message) : base(message) { }
}
And in the production config file:
<system.diagnostics>
<trace>
<listeners>
<remove name="Default"/>
<add name="ExceptionListener" type="Namespace.ExceptionTraceListener,AssemblyName"/>
</listeners>
</trace>
</system.diagnostics>

I don't know how it is in C# and .NET, but in C will assert() only work if compiled with -DDEBUG - the enduser will never see an assert() if it's compiled without. It's for developer only. I use it really often, it's sometimes easier to track bugs.

I would not use them in production code. Throw exceptions, catch and log.
Also need to be careful in asp.net, as an assert can show up on the console and freeze the request(s).