recent updates
Jeff Rogers <dvrsn <at> DIPHI.COM>
2011-10-19 05:45:30 GMT
Hey all, me again.
For those of you who don't monitor the commits list I wanted to share a
few changes I've recently made as well as a few I'm still thinking about.
- implemented native decoding of strings in ns_returnfile. This allows
filenames that are not utf-8 to be passed, similar to tcl core functions.
- committed a long-outstanding patch to address a fastpath caching bug
where incorrect file content could be delivered in a system when inodes
are rapidly being reused. Added a new config option "cacheminage" to
control this behavior.
- implemented a suggestion to re-run pre-auth and post-auth filters on
internal redirects. As this could break assumptions in unusual
situations, there is a config option "filterredirect" to disable it.
Among other things, this addresses an issue in openacs that made it
complicated to handle 404 errors. Also fixed a server crash when there
is an error in the configured error page.
- added flags to insert a filter at the beginning of the list of
registered filters instead of only at the end, and to order filters by
priority.
- exposed the gzip flag to tcl scripts (ns_conn gzip)
- added another new filter, "pre-write" which runs after content has
been returned but before it has been sent over the network. To make
this useful, there is a new method to ns_conn to examine and set the
response content that is about to be sent. This could be useful for
implementing something like apache's mod_pagespeed, which rewrites
content to do things like merge javascript and css files to reduce
external requests, or just strip extraneous whitespace from a page.
- added a tuning knob "maxaccept" to let the driver accept more than one
connection per spin. The SF RFE for this referenced a paper that found
this type of change could improve performance and reduce dropped
connections under load; I unfortunately don't have a network capable of
simulating enough load to stress this so I cannot claim any measured
improvement.
I also have a few other ideas I haven't gotten around to adding (or even
necessarily thinking through completely) yet:
- removing encoding and gzip compression from the normal response
writing codepath (Ns_ConnFlush) and making them into pre-write filters.
- adding an api to name C-coded filters and handler procs such that they
could be used by the tcl ns_register_filter/ns_register_proc commands.
This could eliminate a few special-purpose commands that do this same
thing for particular operations (ns_register_fastpath and
ns_register_filter_shortcut).
- add support for chunked encoding for requests, digest authentication,
and Expect: 100-continue. I've heard that these are already supported
in NaviServer, so this may be a fairly simple cut&paste enhancement.
- replace c-coded versions of many file operations with equivalent
tcl-coded versions. Many of these functions (e.g., ns_mkdir, ns_cp)
come from before tcl had equivalent functionality, but tcl now has that
functionality, and it's better (native decoding).
- replace ns_share with tcl version built around nsv, if not eliminating
it entirely.
- remove other really old code like tclx keylist support. I honestly
don't know if this is still being used, perhaps it could be kept around
in a "nscompat" module along with other things like ns_var and
ns_share). I'm a bit reluctant to break compatibility without an easy
switch to un-break it.
- make content sending event based, by pushing it back to the driver
thread, a task thread, or a completely new writer thread, rather than
blocking a conn thread if the content is big and the network is slow.
There is support for doing this at the tcl level, but it would be nice
to have it implicitly happen everywhere.
- let queue-waits be set up from a conn thread and the conn re-queued in
the driver thread. This would let you do database operations before
making network calls and still wait for those calls to complete without
tying up conn threads. This would mean a rather esoteric coding style
but I think it would make for a really powerful mashup engine.
Some of these are pretty disruptive (especially the last two), and
introducing incompatibilities (even ones with an easy off switch) is not
something to be taken lightly; does anyone else think a new 4.5.2
release with the recent bugfixes and non-disruptive new features could
be done, and then slightly disruptive stuff (i.e., moving c-coded
commands to tcl) could be targeted at 4.6?
Naturally, I'm also interested in any feedback on anything I've done or
suggested also.
Cheers,
-J
RSS Feed