> From: Philip Martin [mailto:philip@codematters.co.uk]
>
> "Bill Tutt" <rassilon@lyra.org> writes:
>
> > > > No, we know (or think we can figure out) a way to unwedge a
*known
> > > > wedged* repository.
> > > >
> > > > The think is, that's the same as lock-stealing, if you're not
> > > > sure that no one else is using it.
> > >
> > >
http://subversion.tigris.org/servlets/ReadMsg?list=dev&msgId=176926
> > >
> > > And there were more threads on this subject IIRC. One
specifically
> > > dealt with the situation you sketch.
> > >
> >
> > In fact, let me be even clearer in case what I've been saying hasn't
> > come across correctly. The issue outlined by the URL that Sander
> > mentioned above IMNSHO is a 1.0 ship stopping bug.
> >
> > We must be atomic, we must be consistent, we must be isolated, and
we
> > must be durable. Additionally, we must also be reliable. There is no
> > compromise possible. Ctrl-C is but the smallest of the annoying
problems
> > that data stores encounter when trying to ensure durability and
> > reliability. There are hundreds of others.
>
> Maybe I'm missing something, but neither the email referenced above
> nor the Berkeley DB documentation it refers to, appear to solve the
> lock-stealing problem. The suggestion, to always run recover when
> opening the DB is not valid. It requires the Subversion clients to be
> either "a single, usually multithreaded, process" or "a set of
> cooperating processes" and neither of those apply. The Subversion
> clients are a set of independent processes.
>
> It boils down to "the Berkeley DB library itself cannot determine
> whether recovery is required" and so neither can a single Subversion
> client. To make the decision to run recovery requires clients either
> to communicate and interact, or to assume that they are the only
> client.
>
> It looks like compromise is required ;-)
>
No, either we must achieve that goal, or we must re-evaluate our design.
There is no compromise. Even if this means that we need to complete
ra_pipe, so that we have a command line client that uses ra_pipe to talk
to a svnd that then uses BDB.
We must ship something that is completely and utterly robust, if we
don't we've failed. There is no compromise.
The referenced email does explain a possible direction for ensuring that
recovery is always run by the first process that wants to open the BDB
data store.
There are other possible ways for the multiple libsvn_fs clients in
different processes to communicate with each other. I'm not going to go
into any of those possible ways now as I need to head on home.
Again, no compromise is possible here,
Bill
---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscribe@subversion.tigris.org
For additional commands, e-mail: dev-help@subversion.tigris.org
Received on Thu Sep 19 02:47:07 2002