Hi, I feel it's a nice update and agree the change.
- first it creates unnecessary load on the neutron api and neutron db/network backend in creasing latency and reducing overall performance.
The periodic task causes huge API workload not only for Neutron API but for the Keystone API too, because of token issue in nova-compute and token validation in neutorn-server. best regards, Masahito -----Original Message----- From: "Sean Mooney"<smooney@redhat.com> To: "openstack-discuss@lists.openstack.org"<openstack-discuss@lists.openstack.org>; Cc: Sent: 2025/01/22(水) 22:24 (GMT+09:00) Subject: [nova][neutron][ops] disabling the heal instance info cache periodic hi folks o/ This is a short PSA and request for comments The nova and neutron teams have discussed the existence and importantly the default behavior of nova's "heal instance info cache periodic" on an off for the better part of the last 4+ years but we never had time to take any action on it until now. Context for those who are not aware of it, nova has an instance info cache which we use to cache networking information such as ip address related to ports associated with an instance. This was introduced in the Essex release when nova networks was being replaced with quantum/neutron. Its rater fitting but over due that in the Epoxy release we are now turning this off. for more context i have captured some of the important point in the release note of the proposed change https://urldefense.com/v3/__https://review.opendev.org/c/openstack/nova/*/93... But ill provide a TL;DR here. By default, as of essex, [compute]heal_instance_info_cache_interval defaulted to 60. it has been possible to disable this by setting the value to <=0 (recommended to use -1 by convention). when the periodic is enabled, every nova compute agent loop over the instance on the current host, retrieves all there ports form neutron and fully rebuilds the cache(as of stein). on a small cloud (less then 50 computes) this has minimal impact on the performance of the neutron server, on moderate to large clouds this creates 10s of MB of constant db read load. This has several side-effects: - first it creates unnecessary load on the neutron api and neutron db/network backend in creasing latency and reducing overall performance. - second, it create unnecessary load on rabbitmq as each compute service need to do an RPC to update the info cache(every 60 seconds). - third both of the first two waste power doing extra work that is not needed. - finally as of Icehouse this periodic has been optional because we added the os-server-external-events api so that neutron can tell nova if there has been a change to refresh the cache as needed. With that in mind hte nova team is proposing that we disable this going forward, for this release i have not deprecated the periodic so operators can re-enable it if they find its required for any reason. in the 2026.1 release we will likely revisit this based on operator feedback/bug reports. In this weeks team meeting we agree to wait 2 more weeks to get feedback from OPS if they had any concerns with this change hence this email. If anyone knows of a reason why we should not proceed with this default change feel free to let us know via this email thread or the gerrit review https://urldefense.com/v3/__https://review.opendev.org/c/openstack/nova/*/93... regards sean