Hi Mike - I have sent several emails about this, but haven''t heard back yet. I saw you committed a new HLD. Did you cover in some detail the use case of - reconnecting, dealing with old and new gaps? - assigning versions upon replay I believe the proposed epoch-transno versioning solves this issue, but the tricky part is the following. When a server reboots and not all clients reconnect, the server doesn''t know what last transactions (not committed) used as their transno. This is one of the reasons that it is much better to have epoch-transno as a version. The other related use case that we need to understand really in detail is reconnecting and gaps. Again, reconnecting in the next epoch is I think fairly trivially covered, but reconnecting later leads to more questions about gaps in earlier epochs (which may lead to cache entries that need to be purged from the client). What do you think about these three? Have you thought of other tricky cases? - Peter - Mike Pershin wrote:> hello, Peter > > are there any news from Lee? You said that need to clarify some things > with cray before we can proceed. > > --Mike Pershin > >-------------- next part -------------- An HTML attachment was scrubbed... URL: http://mail.clusterfs.com/pipermail/lustre-devel/attachments/20070324/cabfc412/attachment.html