Re: Squid Log File wish list

Chris Wedgwood wrote:

> > Squid does not yet cache partial objects. This is on the todo list
> > sort of..
>
> I'd be keen to here any suggestion on how this might be accomplished
> efficiently....

Don't know if I have answered you on this yet, but here is what I wrote
in reply to Alex when he asked the same question(s) on squid-bugs

Date: Tue, 27 Oct 1998 03:32:10 +0100
Subject: Re: 2.1.PRE1: Range requests far avay [new limit]
To: Alex Rousskov <rousskov@nlanr.net>
CC: squid-bugs@ircache.net

Alex Rousskov wrote:

> There is no [big] problem with caching partial replies.
> Unfortunately, there is not much use from it either. What we need
> is merging partial replies into one object, and that is very hard,
> IMO.

I don't agree completely here. Merging of sequencial ranges is
not a problem. Merging of unordered ranges may be a problem if
the store can't support random reads.

Doing it most efficient is another issue, but I am not sure
efficiency is a important issue until ranges get heavily used.
Until then avoiding wasting huge amounts of bandwith on a few
requests is more important.

> Can you elaborate more on _when_ 206 cached objects will be
> returned to clients as _HITs_ in your schema?

On the second request for on that object. Lets take a Adboe PDF
document for example. Acrobat first fetches certain ranges (most
notably the end to get the index) and then continues to download
what's required to view the requested page.

First we get only have a fraction of the object, and store this.
Then the user agent requess another range of the same object.
Squid now fetches any part that's missing to complete the
request. For each fraction being requested the cached copy gets
more and more complete.

> How many such cases do you expect to have on average?

Not many today, but new uses are apparing. It is a promising
feature of HTTP to make things feel quicker.

> Are you expecting clients to submit identical Range requests
> for the same object?

No. The size and order depends on the user and client.

> For example, Adobe will probably use the same sequence of
> ranges for the same object most of the time...

Yes, but not all of the time.

/Henrik
Received on Tue Jul 29 2003 - 13:15:55 MDT