6

I'm trying to understand whether and under what circs one should use Python classes and/or Java ones.

If making a specialist dictionary/Map kind of class, should one subclass from Python's dict, or from Java's HashMap or TreeMap, etc.?

It is tempting to use the Python ones just because they are simpler and sexier. But one reason that Jython runs relatively slowly (so it appears to me to do) seems to have something to do with the dynamic typing. I'd better say I'm not that clear about all this, and haven't spent nocturnal hours poring over the Python/Jython interpreter code, to my shame.

Anyway it just occurs to me that the Java classes might possibly run faster because the code might have to do less work. OTOH maybe it has to do more. Or maybe there's nothing in it. Anyone know?

mike rodent
  • 14,126
  • 11
  • 103
  • 157
  • Is it faster to use Python loops or Java ones? What about if tests? Or variable declarations? – duffymo Jun 24 '12 at 14:26
  • umm... not clear whether this is a hostile comment. My question is about Jython, so doesn't seem that silly... to me. – mike rodent Jun 24 '12 at 14:29
  • @katrielalex no, that is not a valid point: in Jython you have choices between two different language structures, and it is non-trivial to ask whether and under what circs you should choose one or the other – mike rodent Jun 24 '12 at 14:33
  • @mikerodent sorry, deleted comment. I agree with what hexparrot said. – Katriel Jun 24 '12 at 14:35

2 Answers2

5

The point of using Jython is that you can write Python code and have it run on the JVM. Don't ruin that by making your Python into Java.

If -- if -- it turns out that your data structure is too slow, you can drop-in replace it with a Java version. But that's for the optimisation stage of programming, which comes later.

I guess I should try to answer your question. I would guess that using native Java structures will be faster (because the JVM can infer more about them than the Python interpreter can), but that might be counterbalanced by the extra processing needed to interface with Jython. Only tests will tell!

Katriel
  • 120,462
  • 19
  • 136
  • 170
  • perhaps you haven't quite understood the question: I'm not talking about integrating Jython code and Java code, which I know about. I'm talking about *within* the body of the Jython code: subclass from dict or subclass from HashMap. – mike rodent Jun 24 '12 at 14:34
  • 1
    @mikerodent No, I have understood the question. I think you should subclass from `dict`, because you should write Python code. Is inheriting from `HashMap` not "integrating Jython code and Java code"? – Katriel Jun 24 '12 at 14:36
  • Er, no it's not. I refer you to Chapter 10 of the Jython Definitive Guide: integrating Java is about calling compiled Java modules from within the Jython code. If you don't know this, why are you choosing to try and answer a question about Jython? – mike rodent Jun 24 '12 at 14:38
  • @mikerodent that's a very narrow definition of "integrating"! But now we're arguing semantics and it doesn't matter anyway. I rest my opinion =) – Katriel Jun 24 '12 at 14:40
4

Generally, the decision shouldn't be one of speed - the Python classes will be implemented in terms of Java classes anyway, even if they don't inherit from them. So, the speed should be roughly comparable, and at most you would save a couple of method calls per operation.

The bigger question is what you plan on doing with your class. If you're using it with Python APIs, you'll want to use the Python types, or something that behaves like them so that you don't have to do the work of implementing the entire Mapping protocol (only the bits your class changes). If you're using Java APIs, you will certainly need to meet the static type checks - which means you'll need to inherit from Java's classes.

If this isn't easy to answer in your situation, start with the Python ones, since you (correctly ;-) find them "simpler and sexier". If your class doesn't pass outside the boundaries of your project, then this should be trivial to change later if the speed really becomes an issue - and at that point, you might also be thinking about questions like "could it help to implement it entirely at the Java level?" which you've hopefully recognised would be premature optimisation to think about now.

lvc
  • 34,233
  • 10
  • 73
  • 98