Hey @dsanader. I was wondering if you see this being dependent or independent of #505?

At first glance #505 includes the same fix. If #505 is merged, you should be able to safely drop this PR. I will build the #505 code just to check though and report back If I find anything wrong on my side.

Well, it happens to be more complicated than I thought. My initial "fix" was in fact ugly. It actually uses an exporter.ClientV2 where Microversion is unset, so the API answers a single page. It should be more explicit that we downgrade Microversion on purpose to fallback to single page mechanism (which has the advantage to work consistently accross versions, see below).

Regarding #505, it does indeed automatically sets the latest supported Microversion, but since hypervisors.ListOpts is missing, limit is not added to the query params. So the Nova API returns here the api.max_limit (default to 1000) objects, without including the link to the next page. If Microversion > 2.33, a default limit should be defined to trigger pagination, maybe like this:

@@ -167,8 +167,17 @@ func ListNovaAgentState(ctx context.Context, exporter *BaseOpenStackExporter, ch
 func ListHypervisors(ctx context.Context, exporter *BaseOpenStackExporter, ch chan<- prometheus.Metric) error {
        var allHypervisors []hypervisors.Hypervisor
        var allAggregates []aggregates.Aggregate
+       var listOpts hypervisors.ListOpts
 
-       allPagesHypervisors, err := hypervisors.List(exporter.ClientV2, nil).AllPages(ctx)
+       apiMv, _ := strconv.ParseFloat(exporter.ClientV2.Microversion, 64)
+       if apiMv >= 2.33 {
+               limit := 1000
+               listOpts = hypervisors.ListOpts{Limit: &limit}
+       } else {
+               listOpts = hypervisors.ListOpts{}
+       }
+
+       allPagesHypervisors, err := hypervisors.List(exporter.ClientV2, listOpts).AllPages(ctx)
        if err != nil {
                return err
        }

Lastly, on older versions of Nova, the hypervisors_links key containing the next page is not always included in the results, which makes pagination not that useful in that case. I haven't sorted that point yet: I don't know what fixed it and when.

The python SDK includes a fallback logic which covers that (see: https://opendev.org/openstack/openstacksdk/src/commit/fb794edea9c1d1e9e4da8397f0dc9adc73bd1de1/openstack/resource.py#L2213). So, that behavior can't be seen with the openstack CLI.

I've changed my code to explicitly set a microversion not supporting pagination for the hypervisors list call for the time being.

Since having >1000 compute nodes is not that common, I guess it can wait a more proper fix when the gophercloud v2 is merged.

Appreciate the detailed breakdown @dsanader . Thank you!

@dsanader i've added your fix to v2.