▤ Stackers Network

January 03, 2026

Stackers Network Digest — January 03, 2026

12 threads · 19 messages · openstack-discuss

The Big Picture

The first week of 2026 was a quiet one — the deep holiday lull, with just 12 threads and 19 messages on openstack-discuss, no releases, no CVEs, and most project meetings cancelled. Two things stood out: a detailed memory-leak diagnosis in the OpenStack SDK metrics path from an operator running at very large scale, and Alan Bishop's resignation from three Puppet OpenStack core teams. The week's most useful community contribution was a brand-new open-source tool for Cinder snapshot lifecycles. Treat this as a breather issue — cadence picks up sharply next week as contributors return.

Development & Technical Decisions

A real memory leak in the openstacksdk metrics path. Sam Clippinger posted a detailed diagnosis worth attention from anyone running at scale. Every request through openstacksdk's Proxy.request() calls _report_stats(), and the Prometheus path records per-request counters and histograms keyed on the full request URL — with no config switch to disable it. Because nova-scheduler queries Placement using each VM's UUID, every unique URL produces its own counter. In Clippinger's build farm (over 500K VMs/week, Caracal), each of six nova-scheduler copies climbs to ~8 GB until workers get OOM-killed and the new VMs land in error state. He notes prometheus_client can't simply be uninstalled — it's pulled in transitively via oslo.metrics → oslo.messaging — and asks whether keystoneauth1's Adapter/CloudRegion should honor a collect-timing-style off switch. This is a genuine, reproducible production scaling bug, not a configuration mistake, and a worthwhile read for SDK, keystoneauth, oslo, and Nova maintainers.

Nova design seeds from Sean Mooney. Mooney opened several Nova threads over the break that signal active design thinking even if the archive bodies aren't fully readable yet: a cross-project heartbeat improvement to verify the main thread actually responds to a RabbitMQ ping (touching Nova, Cinder, and Neutron); a question about using $aggregates with JsonFilter; and a clarification on unified-limits combined with PCI alias support — Mooney's note there was that the flavor is correct as-is and resources:CUSTOM_NVIDIA_T4='1' should not be hand-added. Karol Klimaszewski separately opened a Nova "ephemeral storage potential features" thread that will continue to develop.

Neutron bug deputy report (week of 2025-12-15). Bence Romsics posted a delayed deputy report flagging two High items: functional tests using SqlFixture are extremely slow (fix in progress in 971472), and the highest-priority chassis gateway port fails to bind after a chassis restart/crash. Several Medium-priority OVN and FT issues are also tracked.

Heads Up / Action Needed

  • Puppet OpenStack loses a core. Alan Bishop announced his resignation from the puppet-cinder, puppet-glance, and puppet-manila core teams due to a change in work status. That's reduced review capacity across three Puppet modules — interested contributors should step up.
  • Designate multi-region NS-record restriction still open. Sylvain's question about "Updating a root zone NS record is not allowed" in PRIMARY/SECONDARY across three independent deployments carries over from December and hasn't been answered yet; he's asking whether the long-standing restriction could be made configurable.
  • Mind the noise. A "SECURITY BUG" message from "Umbrella Corporation" was a bounty-seeking solicitation, not a real advisory.

Community & Events

The week's bright spot was a community contribution: SnapSentry, an MIT-licensed, Go-based open-source tool from Aravindh Murugesan to automate Cinder snapshot lifecycle management — scheduling and rotation, which Cinder leaves to the operator. It's metadata-driven (tag your volumes, no central DB), does "atomic-ish" parallel snapshots of all volumes attached to a VM, uses hybrid concurrency to avoid API throttling, cleans up orphaned snapshots, and supports restricted application credentials for least-privilege ops. Pitched as a lightweight alternative to commercial tools like Trilio. The repo is on GitHub, and the author welcomed feedback.

On the support side, Tarun asked for help validating a production cluster while migrating from VMware to OpenStack 2025.2 — a reminder of how much VMware-migration traffic continues to flow into the lists. Mohd Atif Ansari reported a Watcher continuous audit stuck in PENDING with no action plans generated despite all services running, a recurring Watcher pain point.

Most project teams remained on holiday hiatus and the TC's weekly IRC meeting stayed paused — so expect cadence and volume to pick up sharply in next week's issue once contributors return from the break.