Comment 13 for bug 118625

Some more thoughts: it's probably easy enough to scale loggerhead horizontally by running it on several machines behind a load balancer. But then they will not share the revision graph caches they build. Swapping out the sqlite cache for something that can talk over the network -- couchdb springs to mind as being entirely suitable, although any key-value store will do -- would make sense for this, and probably isn't very hard.

For idea 2) in comment #12, a trick that would probably work is to insert a special value into the cache (like None) when any thread starts to compute a graph; a thread that finds this value would take it as an indication that another thread is computing the value and wait, rather than computing the graph itself. It would probably have to poll, unless we can do something fancy with triggers.