Re: Deferred reads

Adrian Chadd wrote:

> One of the problems I had while planning out a more efficient storage manager
> is handling range requests. My opinion for this is that these requests
> shouldn't touch the storage manager for server->client data transfers.
> What I was working toward was a way that the server can serve data to the
> client without going through the storage manager at all.

Hrm.. in the long run we have to change the store manager to handle
partial objects such as ranges, and also varying objects. Don't
overestimate what kinds of request you can bypass the storemanager on
based only on the capabilities of the current store manager. Doing so
might put us in an awkward situation later when the capabilities of the
store manager are to be extended.

> If data is fed to the client from the server by a suitable callback
> mechanism, then the storage manager could slot in between the two in
> the case of a cached reply, and not at all in the case of a non-cacheable
> reply. This would simplify the storage manager code considerably, and
> allow the above to take place in both a poll and event based setup.

Hmm.. what about having the data flow three legged (or more).

Source: The origin server
Destination: Client A
Destination: Store manager (swapout)

One tricky question is how to handle more than one client of different
speeds, but lets for a start assume there is only one client (the most
common case). I see a number of different approaches in how to solve
multiple clients (a: use memory, b: make use of the store manager, c: a
combination, d: possibly something else).

> Since I have some further ideas about making the code copy-friendly,
> I'll continue working this out on paper and post updates on the list.
> I'd really appreciate some feedback here ...

Sure. I'll try to give you any feedback when asked. However, my mind is
a bit buzy on other issues at the moment so you have to drive the
discussion (i.e. no questions, no feedback).

/Henrik
Received on Sun Jun 04 2000 - 15:05:34 MDT