Erlang error handling philosophy - case vs throw - exception

Question

I'm writing a REST service in Erlang and need to verify the received data before passing it to other internal functions for further processing; in order to do that, I'm currently using nested case expressions like this:
case all_args_defined(Args) of
true ->
ActionSuccess = action(Args),
case ActionSuccess of
{ok, _} -> ...;
{fail, reason} -> {fail, reason}
end,
_ ->
{fail, "args not defined"}
end,
...
I realize this is kind of ugly, but this way I can provide detailed error messages. Additionally, I don't think the usual make it crash philosophy is applicable here - I don't want my REST service to crash and be restarted every time somebody throws invalid arguments at it.
However, I'm considering abandoning all those cases in favor of an umbrella try/catch block catching any badmatch errors - would this work?
fun() ->
true = all_args_defined(Args),
{ok, _} = action(Args).
%% somewhere else
catch fun().

Answer 1

Since what you want to achieve is error reporting, you should structure the thing around the execution of actions and reporting of the result. Perhaps something like this:
execute(Action, Args) ->
try
check_args(Args),
Result = action(Action, Args),
send_result(Result)
catch
throw:{fail, Reason} ->
report_error(Reason);
ExceptionClass:Term ->
%% catch-all for all other unexpected exceptions
Trace = erlang:get_stacktrace(),
report_error({crash, ExceptionClass, Term, Trace})
end.
%% all of these throw {fail, Reason} if they detect something fishy
%% and otherwise they return some value as result (or just crash)
action(foo, [X1, X2]) -> ...;
action(foo, Args) -> throw({fail, {bad_arity, foo, 2, Args}});
action(...) -> ...
%% this handles the formatting of all possible errors
report_error({bad_arity, Action, Arity, Args}) ->
send_error(io_lib:format("wrong number of arguments for ~w: "
"expected ~w, but got ~w",
[Action, Arity, length(Args)]));
report_error(...) -> ...;
report_error({crash, Class, Term, Trace}) ->
send_error(io_lib:format("internal error: "
"~w:~w~nstacktrace:~n~p~n",
[Class, Term, Trace])).

Answer 2

I've had this problem while developing an application that create users.
I first come with a solution like this:
insert() ->
try
check_1(), % the check functions throw an exception on error.
check_2(),
check_3(),
do_insert()
catch
throw:Error1 ->
handle_error_1();
throw:Error2 ->
handle_error_2();
_:Error ->
internal_error()
end.
The problem with this solution is that you lose the stack trace with the try...catch block.
Instead of this, a better solution is:
insert() ->
case catch execute() of
ok -> all_ok;
{FuncName, Error} ->
handle_error(FuncName, Error);
{'EXIT', Error} ->
internal_error(Error)
end.
execute() ->
check_1(), % the check functions throw an exception on error.
check_2(),
check_3(),
do_insert().
This way you have the full error stack on Error.

Answer 3

I have faced exactly the same question when writing my own REST services.
Let's start with the philosophy:
I like to think of my applications like a box. On the inside of the box are all of the parts I built and have direct control over. If something breaks here, it's my fault, it should crash, and I should read about it in an error log. On the edge of the box are all of the connection points to the outside world - these are not to be trusted. I avoid exception handling in the inside parts and use it as needed for the outer edge.
On similar projects I have worked on:
I usually have about a dozen checks on the user input. If something looks bad, I log it and return an error to the user. Having a stack trace isn't particularly meaningful to me - if the user forgot a parameter there is nothing in my code to hunt down and fix. I'd rather see a text log that says something like: “at 17:35, user X accessed path Y but was missing parameter Z”.
I organize my checks into functions that return ok or {error, string()}. The main function just iterates over the checks and returns ok if they all pass, otherwise it returns the first error, which is then logged. Inside of my check functions I use exception handling as needed because I can't possibly consider all of the ways users can screw up.
As suggested by my colleagues, you can alternatively have each check throw an exception instead of using a tuple.
As for your implementation, I think your idea of using a single exception handler is a good one if you only have the single check. If you end up needing more checks you may want to implement something like I described so that you can have more specific logging.