Avi Bryant wrote:
> On Thu, Apr 3, 2008 at 12:00 PM, Stephen Pair <stephen at pairhome.net> wrote:
>> If I understand what DabbleDB does, they essentially give each user their
>> own squeak process to run in.  When a session is first established, they
>> start up a new image.  Upgrading would be a matter of dropping a new image
>> somewhere and as old sessions expire and new ones created, users are
>> migrated over to the new code.  With hydraVM, a scheme like that might be
>> doable with just a small bit of session management in the main image.
> 
> That's more or less right, yes.  I've used variations on that strategy
> before - in general, having multiple images going where old sessions
> go to the old image but new sessions get routed to the new image makes
> for very smooth upgrades.

Yes, that's what we're doing too. When we need to upgrade a server we 
simply don't let any new sessions go onto it and once everyone is off 
the server is up for grabs. Unfortunately, some of the sessions last 
quite a while (sometimes for several days because our notion of 
"session" spans multiple users which need to overlap only for a short 
amount of time) which I guess is good business-wise and we're not 
running a lot of servers (which I guess is good too, since we scale) but 
it makes for some rather long delays which could be mitigated by a "hot" 
(or maybe just "warm") upgrade.

Cheers,
   - Andreas