On 26 janv. 2010, at 22:03, Simon Denier wrote:
On 26 janv. 2010, at 21:07, Alexandre Bergel wrote:
I do not have a solution, but this is really annoying. In addition to the time taken by loading Moose, it makes 10 minutes just to check what are are dependencies Mondrian has.
Well, temporary solutions sometimes become less than temporary... I fear this is the case.
Maybe we can do something with an incremental, on-demand cache: whenever a class or method is imported, we check whether the cache has been built for it - if not, we built it.
After playing a bit with the idea tonight, now I remember why I choose the global cache solution. So I explain how it works in the current system:
To build the system cache, I go over all classes and all class extensions of PackageInfos and register their most specific package.
Since I go over the whole system, I assume I know all class extensions in all classes.
Now when I ask the package for a method, there are two cases:
- either I can find it in the cache of class extensions and I return its package
- either I can't find it, so I assume it is a normal method and retrieve the package of its class (from the cache).
Now in an incremental cache system, I can cache a class when I import it, I can even cache a package so that its class extensions are imported, as above. The big problem is then that I can't assume that I know all class extensions within the class, because other packages (which I don't necessarily import) may do some. To do that, I should ask all methods within each imported class for its most specific package. This was more or less the situation we had before the cache, which brought the system to its knees on large projects. Perhaps I will try tomorrow but I don't have high hope. We need something better than that.