[ovs-dev] [PATCH v2] ovn: Avoid nb_cfg update notification flooding

[Han Zhou]

On Fri, Apr 6, 2018 at 1:54 PM, Ben Pfaff <blp at ovn.org> wrote:
>
> Thanks for working on making OVN faster and scale better.
>
> I see what you mean about how nb_cfg can be a scale problem.  Really,
> each hypervisor only cares about its own row, but each of them is being
> notified about the current value in every row.  Ideally, we want the HVs
> to be able to say to ovsdb-server, "Give me all the data in the Chassis
> table, except don't bother with nb_cfg in any rows except my own row."
>
> Actually, there's a way for ovn-controller to do that: the OVSDB
> protocol definition of monitor_cond allows the client to specify
> multiple sets of columns to monitor in a table, and each set of columns
> has an independently attached condition.  This can be used to specify
> that most of the columns are monitored unconditionally but nb_cfg is
> monitored only for a particular row.
>
> But there are still problems:
>
> 1. The implementation in ovsdb-server looks broken.  I don't think it
>    implements the spec.  It certainly isn't tested--none of the existing
>    clients ever passes more than a single set of columns.  I think that
>    it will work if all the sets of columns use the same condition, but
>    that doesn't help with this case.  It will need to be fixed.
>
> 2. The client implementation in ovsdb-idl doesn't anticipate difference
>    conditions for different columns.  It will need to be enhanced to
>    support this use.
>
> In the short term this is going to be more work than just creating a new
> table.  In the long term, though, it will allow us to be more flexible
> and more agile, since improvements similar to the one in this patch will
> require only client changes, not database schema changes.
>
> Comments?

Thanks for the valuable information! I didn't even know that ovsdb is
supposed to support multiple set columns with different monitor conditions.
It surely would be the right way to go if it is supported. However, as you
mentioned the mechanism doesn't work yet and I am not sure how much work is
needed to make it work, it seems reasonable to me to "workaround" it at
least for short term until that is ready in the future. Then we could
simply resurrect the old column in the old table with the new monitor
conditions. What do you think?

Thanks,
Han