On Feb 20, 2009 at 16:01, Andrei Pelinescu-Onciul andrei@iptel.org wrote:
On Feb 20, 2009 at 15:54, Henning Westerholt henning.westerholt@1und1.de wrote:
On Friday 20 February 2009, Andrei Pelinescu-Onciul wrote:
[..]
(1) - could be changed in some cases (e.g. string case with some int label allowed, which could be automatically converted to string), but I think it would too confusing and I disallowed it (in general having mixed types in a switch() are 99% an error).
I would suggest to convert numbers to strings in this case automatically. For most people things get more confusing with the increasing amount of details they have to remember about the configuration language.
This would only make sense if we use match() for strings and switch() for ints. Otherwise it would be too confusing. Anyway I don't think the amount of details of the configuration language is ever a problem, as long as one gets meaningful error messages when checking the config (and before ser startup).
Hi Andrei,
the amount of detail in the config language is IMHO important. The developers must document and test every single statement. Every user must read, learn and memorize each statement too.
I regulary need to check the documentation for some special cases in the config when i get asked for a review of some changes a co-worker did, because he still feel not completely confortable after years of administrating SER* systems. During all the time i spend on user channels i really rarely heard the that our server is to slow (only when some real bottlenecks were involved, like DB), the common complain is that the learning curve is too steep. Perhaps i'm lazy, but i'd not say that everybody out there is it too. ;-)
Good luck with runtime debugging then :-)
The other approaches trade-off less config details for guessing what the user intended, which IMO is much more dangerous. Is much better to get meaningful errors when running ser -cf ... , then getting unexpected behaviour at runtime (a very good example for this are typed variables vs. untyped ones or operators that try to guess the type and assume the user made the right choice).
I just checked in kamailio, we don't throw an error if one mix strings and ints in a switch case. We also check for a correct type of an integer in a mixed expression, e.g. if its a valid int value. I don't think that we convert them implicitly, as many pseudo-variables hold internally both a string and a integer value, so mixed expression will work just as intended in
Sorry, but that's just broken (mixed expressions).
I don't really understand this argument. Leaving performance and optimizations aside, do you really think is better to guess what the user might have wanted in ambiguous cases, rather then through an error and point him to explain more clearly what he wants (e.g. use match() instead of switch() for strings, or use typed variables)? With the general approach advocated by some people here, a small error when writing the script (e.g. forgetting some "") will be visible only at runtime (and depending where that error is it could take weeks until that branch of the config is reached and the error is triggered). I think this is much worse then having to add a little more "details" to the script (which BTW wouldn't make it any more difficult then any common programming language).
I'd rather have to declare even variable types (e.g. int $a), rather then debugging strange bugs in a script, because I used a var that happens to get a string value at runtime.
In brief I much prefer adding to the script if this helps check its correctness at sr -cf time, rather then trying to be excessively friendly and leaving a lot of potential script bugs pass-through. The ability to be able to tell that the script would behave correctly (as his writer wanted it to) and eliminate ambiguous cases is more important than trying to keep a very reduced script and guessing what the writer really wanted.
Andrei