On Dec 13, 2014 12:33 PM, "Aaron Halfaker" <ahalfaker@wikimedia.org> wrote:
> 1. It turns out that generating diffs is computationally complex, so generating them in real time is slow and lame. I'm working to generate all diffs historically using Hadoop and then have a live system listening to recent changes to keep the data up-to-date[2].

IIRC Mako does that in ~4 hours (maybe outdated and takes longer now) for all enwiki diffs for all time. (don't remember if this is namespace limited) But also using an extraordinary amount of RAM. i.e. hundreds of GB

AIUI, there's no dynamic memory allocation. revisions are loaded into fixed-size buffers larger than the largest revision.

-Jeremy