On Nov 25, 2009, at 11:45 PM, Nathan Eisenberg
<[email protected]> wrote:
I apologize if this is not the correct place to ask questions.
Welcome! Yes, this is the perfect place for questions like this.
since it appears that the ‘client’ portion of tahoe does not
maintain cached files in memory for any period of time.
That's correct. At some point, to support random-access reads more
efficiently, we'll probably add a minimal cache for individual
segments of a single file, but so far whole-file caching has not been
on the requirement list.
It's useful to note the distinction between mutable and immutable
files here. Caching immutable files is perfectly safe: it's a simple
tradeoff between local storage consumption and performance, assuming a
given locality-of-reference / reader behavior. Caching *mutable* files
(or directories), on the other hand, is not safe: the tradeoff
includes a correctness aspect, since someone else might change the
contents and you might use your (now stale) cached copy. In general,
we try to avoid putting any sorts of heuristics about correctness into
tahoe itself, so any caching layer that requires a decision on a
correctness-vs-performance policy would need to be placed above the
tahoe node.
My first instinct was to turn to Squid.
Hm, I'm surprised that didn't work. Do you know what caused Squid to
believe the file was changing each time? Maybe a quick peek at the
returned headers would be informative.
Basically, any URL that starts with /uri/URI:CHK: should be immutable
and an ideal choice for caching. We'd planned (although I can't check
right now to see whether we got around to implementing it or not) to
add an ETag: header with the file's UEB header, which is basically a
hash of the contents, and thus an ideal etag. And I know we aren't
intentionally adding any Cache-Control or date headers that might make
the file look uncacheable.
(for mutable files, there is a similar value called the "roothash"
which covers the file contents, and allows If-ETag-Differs: -type
queries to do the right thing, but I don't know if we've
actually implemented that either).
hope that helps,
-Brian
_______________________________________________
tahoe-dev mailing list
[email protected]
http://allmydata.org/cgi-bin/mailman/listinfo/tahoe-dev
