BTW, http://www.sqlite.org/cvstrac/chngview?cn=4599 is the final fix
to this.  It's different from the patch I posted.  The patch did the
job, but felt wrong to me.  This has the same performance
characteristics, but feels ... well, less wrong.

This change should apply cleanly to fts2.c, if anyone is still using fts2.

-scott


On Dec 5, 2007 2:18 AM, Ingo Godau-Gellert <[EMAIL PROTECTED]> wrote:
> Hi Scott!
>
> You're great! I checked the attached modification and found no search
> taking longer than 20s now! It's a great improvement. I didn't find any
> other problems, so I will leave the modification in my FTS3 compilation.
>
> Many thanks!
>
> Ingo
>
>
> Scott Hess schrieb:
>
> > 2007/12/4 Scott Hess <[EMAIL PROTECTED]>:
> >
> >> This seems a little excessive, though.  I do see that there's an
> >> O(N^2) path in the prefix-searching (loadSegmentLeavesInt()'s call to
> >> docListUnion()).  I can reasonably make that O(logN), which might help
> >> a great deal, if you're hitting it.  Not really sure how to tell if
> >> you're hitting it, but I'll experiment at my end and see whether I can
> >> improve things there.
> >>
> >
> > With the attached patch, the time to match against 't*' with the rfc
> > dataset goes from 1m16s to 5s.
> >
> > It passes the tests, but I'll not guarantee that this is what I'll
> > check in.  I want to think on it.  But let me know if this doesn't
> > help.
> >
> > -scott
> >
> > ------------------------------------------------------------------------
>
> >
> > -----------------------------------------------------------------------------
> > To unsubscribe, send email to [EMAIL PROTECTED]
> > -----------------------------------------------------------------------------
>
> -----------------------------------------------------------------------------
> To unsubscribe, send email to [EMAIL PROTECTED]
> -----------------------------------------------------------------------------
>
>

-----------------------------------------------------------------------------
To unsubscribe, send email to [EMAIL PROTECTED]
-----------------------------------------------------------------------------

Reply via email to