> [...] > > I haven't looked at the implementation, but if your description is correct, > then we're definitely in trouble. > > I would have expected that the unversioned-metadata (i.e., currently only > revprop) generation would always be updated regardless of cache settings, > and never be cached. I presume changing the implementation to behave this > way would fix the problem?
If you mean that we could entirely stop sharing the revprop generation across multiple processes, it would probably work. However, I am not entirely sure what was the reasoning behind that kind of sharing. If this shared memory thing was an attempt to get rid of the file parsing overhead, then I would say it is a good idea to choose correctness in favor of an optimization that breaks things in half of the cases. But there might be more to it. > And by all means, do commit your failing test case. Done in r1619105. Regards, Evgeny Kotkov
