From mboxrd@z Thu Jan  1 00:00:00 1970
From: bgamari.foss@gmail.com (Ben Gamari)
Date: Thu, 01 Oct 2009 15:01:35 -0400
Subject: [sup-talk] Crash while scrolling
In-Reply-To: <1254421545-sup-8303@masanjin.net>
References: <20090911165830.GA11260@ben-laptop>
	<1252773189-sup-246@masanjin.net>
	<20090916172340.GA20566@ben-laptop>
	<1253975267-sup-8308@masanjin.net> <1254160696-sup-3522@ben-laptop>
	<1254404181-sup-8448@masanjin.net> <1254418089-sup-5800@ben-laptop>
	<1254421545-sup-8303@masanjin.net>
Message-ID: <1254422805-sup-9288@ben-laptop>

Excerpts from William Morgan's message of Thu Oct 01 14:31:20 -0400 2009:
> Reformatted excerpts from Ben Gamari's message of 2009-10-01:
> > It seems that C would definitely be a good start (or perhaps C++ would
> > be a better idea as that is the language in which Xapian is written).
> 
> I was proposing that as a strawman argument. C does not solve my
> problem.

Certainly, you will never be able to write a message indexer fast enough
to index a source instantly. That is why we have an index to begin with.
That being said, I think we can do substantially better than we are
currently doing in a lower-level language. With mutt, I can come close
to saturating my drive bandwidth while loading a folder. While
synchronizing in sup, I am lucky to get a few hundred kilobytes/second.
Certainly a large amount of that difference has to do with the amount of
processing done by each, but even adjusting for this it seems to me that
we should still be I/O bound (or at least close to it).

> 
> > However, I think one of the real issues is the exclusive nature of
> > index access.  In fact, this is one of my primary gripes with the sup
> > workflow. After processing a large number of messages, the write-out
> > time can be quite substantial upon killing the buffer.
> 
> It is possible to address this in a variety of ways, many of which have
> been discussed over the years, but yes, I agree that a delay is
> nonideal.

Glad we agree.

> 
> Making message state saving fast, or backgrounded, is a different beast
> from scanning over a mailstore.

If we are unable to update the index while the client is active,
rescanning sources would destroy usability. I would argue that
asynchronous writeout is very much a prerequisite for mutable sources.

> 
> I have been working on a Sup server for quite some time that would
> address many of these problems, but progress is slow.

This was largely what I had in mind. It seems like moving index
manipulation out-of-process might be best, ultimately.

> 
> > As an aside, it would be quite nice if one could run multiple
> > simultaneous instances of sup. It seems that if one only held write
> > access to the index during writes (is this the case presently?), there
> > should be nothing preventing this from being possible.
> 
> It might be possible to do this with Xapian, especially if there were no
> expectation of updates being transmitted across processes.

I think initially no updates between processes would be fine.
> With Ferret,
> if it is possible, it's only with a tremendous amount of effort.

Is ferret even going to be supported after the Xapian backend
stabilizes?

Thanks for the comments and sup.

- Ben