When SCM polling takes a long time to complete - that is to the order of 30mins. Hudson starts to leak memory.

I stopped polling the diretory tree that had the massive set of changes and this immediately stabilised the memory usage.

see hudson_leak.bmp - the red line shows the time when ClearCase polling (lshist) was taking a long time, the green line shows the time I update the load rules of the snapshot view in order to avoid scanning for the huge change set.

Only a few projects out of say 50 in my config were polling for this change set, but that was enough.