On Thu, Feb 10, 2022 at 04:52:15PM +0000, Chuck Lever III wrote: > > > On Feb 10, 2022, at 11:32 AM, J. Bruce Fields <bfields@xxxxxxxxxxxx> wrote: > > > > I was standing in the shower thinking.... > > > > We're now removing the persistent client record early, after the first > > lease period expires, instead of waiting till the first lock conflict. > > > > That simplifies conflict handling. > > > > It also means that all clients lose their locks any time a crash or > > reboot is preceded by a network partition of longer than a lease period. > > > > Which is what happens currently, so it's no regression. > > > > Still, I think it will be a common case that it would be nice to handle: > > there's a network problem, and as a later consequence of the problem or > > perhaps a part of addressing it, the server gets rebooted. There's no > > real reason to prevent clients recovering in that case. > > > > Seems likely enough that it would be worth a little extra complexity in > > the code that handles conflicts. > > > > So I'm no longer convinced that it's a good tradeoff to remove the > > persistent client record early. > > Would it be OK if we make this change after the current work is merged? Your choice! I don't have a strong opinion. --b.
