On 12/8/20 1:56 PM, Ilya Maximets wrote:
> Northbound database has many tables that could contain columns with
> very large sets. For example 'ports' column of the 'Logical_Switch'
> table could contain thousands of ports in a set.
>
> Current strategy of nbctl while adding a new port to the set is to
> copy all existing ports, add one and set the new set of ports.
> Similar behavior is for deletion too.
>
> If we have 1000 ports and want to add one more, resulted transaction
> in current code will look like this:
>
> {transact,
> < wait operation >
> < create new lsp in Logical_Switch_Port table >
>
> where: _uuid == 'logical switch row uuid'
> op : update
> rows : ports ['< list of current 1000 ports + 1 new >']
> }
>
> The code was written before support for partial updates for sets was
> implemented. Now we have it and can replace the old strategy.
> With this change resulted transaction will be:
>
> {transact,
> < wait operation >
> < create new lsp in Logical_Switch_Port table >
>
> where: _uuid == 'logical switch row uuid'
> op : mutate
> mutations : ports insert ['< 1 uuid of a new lsp >']
> }
>
> Unfortunately, for now, this only decreases transaction size in half,
> because '<wait operation>', that is still in the transaction, contains
> the full current list of ports:
>
> where: _uuid == 'logical switch row uuid'
> op : wait
> until: ==
> rows : ports ['< list of current 1000 ports >']
>
> But anyway, beside the overall code cleanup, this reduces transaction
> size in half and should reduce pressure on the ovsdb-server since it
> will not need to parse and process all 1000 ports twice.
>
> This change doesn't affect 'append_request' messages within the raft
> cluster, because ovsdb-server is not smart enough to use mutations
> there, and this will also not affect 'update' messages from the
> ovsdb-server to its clients, because it is smart enough to construct
> 'modify' updates regardless of the original transaction.
>
> One difference between full updates and partial is that partial
> changes are not visible for the idl client until transaction is
> committed and the update received from the server. New switch ports
> are not visible while iterating over 'ports' of the logical switch and
> removed ports are still there. For this reason, we have to maintain
> runtime cache of the mapping between ports and routers/switches they
> attached to.
>
> Signed-off-by: Ilya Maximets <i.maxim...@ovn.org>
> ---
> utilities/ovn-nbctl.c | 348 +++++++++++-------------------------------
> 1 file changed, 87 insertions(+), 261 deletions(-)
>
Oops. I missed one unused variable during rebase, so there is a build
warning with this patch on gcc. It could be fixed by applying this on top:
diff --git a/utilities/ovn-nbctl.c b/utilities/ovn-nbctl.c
index 812e8afab..942dcdb45 100644
--- a/utilities/ovn-nbctl.c
+++ b/utilities/ovn-nbctl.c
@@ -5944,11 +5944,9 @@ cmd_ha_ch_grp_remove_chassis(struct ctl_context *ctx)
const char *chassis_name = ctx->argv[2];
struct nbrec_ha_chassis *ha_chassis = NULL;
- size_t idx = 0;
for (size_t i = 0; i < ha_ch_grp->n_ha_chassis; i++) {
if (!strcmp(ha_ch_grp->ha_chassis[i]->chassis_name, chassis_name)) {
ha_chassis = ha_ch_grp->ha_chassis[i];
- idx = i;
break;
}
}
---
I'd wait for some comments before re-spinning the series.
Best regards, Ilya Maximets.
_______________________________________________
dev mailing list
d...@openvswitch.org
https://mail.openvswitch.org/mailman/listinfo/ovs-dev
