On Wed, Mar 21, 2012, at 11:17 AM, Arnt Gulbrandsen wrote:
> As long as no client ever gets to see MODSEQ 1000 for two different
> states (such as another client's half-completed command and the same
> command fully completed), everything is fine.
That's certainly how I see it. In the extreme, you could even lazily
allocate MODSEQ only when a client requests the information (including,
obviously, a MODSEQ enabled client doing a change and expecting a new
HIGHESTMODSEQ in the response)
But any non-MODSEQ based access to the store could not bump it at all,
so long as the next time a MODSEQ enabled client connected, the MODSEQ
got allocated to all changed things since then.
At least, I can't see how any client could know the difference, or be
right in caring. The only points in time it can know about are right
now, and last time it did a read and got a MODSEQ out of it. Ordering
of the MODSEQs within that are unimportant (to this particular client)
Bron.
--
Bron Gondwana
brong@fastmail.fm