On Thu, May 08, 2008 at 09:17:50PM -0500, Shawn Walker wrote:
> The developer I talked to suggested that instead of doing a tar stream
> from the server, we could simply allow the client to perform HTTP/1.1
> pipeline requests for each individual file.
So I'm on record (at least amongst the folks in MPK) as wanting precisely
this. I think it's the most RESTful approach, and, ultimately, the
simplest.
The reason we went with filelist as is, is explained by an internal mail
from K, back in November (in a thread about doing something similar,
publication-side):
I think you'll find that performing multiple adds over the same HTTP
connection is slower than what we do today. W3 published a paper on
HTTP pipelining that Stephen and I looked at before filelist was
written:
http://www.w3.org/Protocols/HTTP/Performance/Pipeline.html
Essentally, these experiments showed no speed improvement for persistent
connections, unless the client batched up multiple requests and sent
them together.
Since the support for pipelined HTTP requests isn't really that great
right now, we wrote filelist so that we would imitate the behavior of
HTTP pipelining without having to rely on support from the HTTP server.
We've all been thinking at least a little about how to improve the current
filelist. Dan and I are both strong proponents of making it more RESTful,
and all of us want it to be more observable (right now the logs are
completely opaque wrt the filelist ops). We've been talking about
decorating the filelist ops to improve it in both aspects, but I still
think that individual file ops are the way to go, if we can get the
performance to match what we've got. Perhaps CherryPy can give us that.
Danek
_______________________________________________
pkg-discuss mailing list
[email protected]
http://mail.opensolaris.org/mailman/listinfo/pkg-discuss
