Major Currently northd incremental processing fails when logical switches are added/removed. The I-P node dependency is:
NB.Logical_Switch -> en_northd -> en_lflow
When a new switch is created/deleted the "en_northd" input handler for logical switches is called:
bool
northd_handle_ls_changes(struct ovsdb_idl_txn *ovnsb_idl_txn,
const struct northd_input *ni,
struct northd_data *nd)
{
const struct nbrec_logical_switch *changed_ls;
struct northd_tracked_data *trk_data = &nd->trk_data; NBREC_LOGICAL_SWITCH_TABLE_FOR_EACH_TRACKED (changed_ls,
ni->nbrec_logical_switch_table) {
if (nbrec_logical_switch_is_new(changed_ls) ||
nbrec_logical_switch_is_deleted(changed_ls)) {
goto fail;
}
...
fail:
destroy_northd_data_tracked_changes(nd);
return false;
}
which in turn causes "en_lflow" handler for en_northd data changes to fail:
bool
lflow_northd_handler(struct engine_node *node,
void *data)
{
struct northd_data *northd_data = engine_get_input_data("northd", node);
if (!northd_has_tracked_data(&northd_data->trk_data)) {
return false;
}
...
This triggers a full recompute of the "en_lflow" node which can be costly at scale.