On Wed, 2005-06-08 at 22:21 +0100, robert burrell donkin wrote:
> > ===
> > Maven reports:
> >
> > I would suggest disabling this report. Firstly, a log of the last 30
> > days isn't of much use. And secondly, due to the import into SVN of
> > back-dated CVS changes, date-based selection on the apache subversion
> > repository is broken, so the report is not just useless but actively
> > WRONG.
> >
> > I suggest that "Developer Activity" and "File Activity" reports are also
> > useless, and (if based on SVN date selection) also wrong.
>
> AIUI the problem occurs only with the dates on the imported data. new
> data is fine. i've checked the results and they look about right. i do
> agree that they aren't all that useful but i know some users like them
> so i'm inclined to keep them...
Unfortunately the date problem is repository-wide, and long-lasting.
When subversion is passed a date, it immediately converts this into a
revision-number. And it does this by performing a binary search on its
revisions.
Assuming there are 1000 revisions currently in the repository, it
* checks whether revision 500 is earlier or later than the desired date
* 500 is earlier, so revision 750 is checked
* 750 is earlier, so 875 is checked
* 875 is later, so 812 is checked, etc
This process is based on a fundamental assumption that revision X has an
earlier date than revision X+1. When this assumption is broken, the
binary search can go off in the wrong direction. And it looks to me like
after a "problem" import date selections will continue to be broken
until the revision# has at least doubled in size.
Whether you actually get bitten by the problem for a particular
date-based select is a bit of a lottery; if the binary search happens to
hit "valid" nodes all the way down, the search will work correctly. But
hit the wrong node and the select can be a long way out.
Alas, it is just not possible to "insert" revisions into an existing
repository; when importing data it can only be added as new revisions.
So the current choice when importing CVS history is
(1) stuff up all date-based selections *repository-wide*, or
(2) discard all date information associated with imported CVS history,
and put "current" dates against the revisions.
The subversion developers are aware of this, and Ben Collins-Sussman has
said that he would "love to add an index" to fix this. But I don't think
it has very high priority.
Regards,
Simon
---------------------------------------------------------------------
To unsubscribe, e-mail: [EMAIL PROTECTED]
For additional commands, e-mail: [EMAIL PROTECTED]
