Tags

v0.16.1

Hotfix release containing #478 and #479, fixing a config field name and
a bug in the plugin's Score method that caused scheduling failures.

v0.16.0

This release contains a couple new features, but mostly groups together
a handful of recent hotfix releases.

No breaking API changes.

Features:

- neonvm/builder: Add flags to enable informant/monitor (#477)
- plugin: Penalize underfull nodes during Score (#476)

Fixes:

- agent: Fix endpointAssignedAt for VMs added with endpoint (#463)
  - Hotfixed as v0.15.1
- neonvm/runner: Handle hybrid cgroup v1/v2 systems (#467)
  - Hotfixed as v0.15.3
- neonvm/builder: Add additional fields to vm-monitor conn str (#469)
  - Hotfixed as v0.15.4
- agent/metrics: Fix runner state metrics deletion hanlding (#471)
  - Hotfixed as v0.15.5
- Disable zap's enabled-by-default log sampling (#475)

No protocol changes.

Other changes:

- agent/metrics: Initialize runnersCount label pairings to 0 (#465)
  - Hotfixed as v0.15.2
- neonvm/builder: Remove cgexec entrypoint prefix (#477)

Upgrade path from v0.15.x:

- No ordering requirements. Note that vm-builder now requires adding
  '-enable-informant' and '-enable-monitor' to keep the previous
  behavior.

v0.15.5

Yet another hotfix, this time to hopefully fix an issue with the agent's
runner state metrics. See #470 and #471 for more.

v0.15.4

Another hotfix release: We weren't providing a host or port in the
vm-monitor's connection string, which the informant didn't need because
they were included by default in Go's pq postgres driver, but they *are*
needed in the monitor.

v0.15.3

Another hotfix release, fixing neonvm-runner for hybrid cgroup v1/v2
setups. The only difference between v0.15.2 and this is #467.

v0.15.2

Another hotfix release, adding better metrics initialization to #459,
from PR #465.

v0.15.1

Hotfix release, fixing incorrectly setting podStatus.endpointAssignedAt
to nil, which caused staging to crash-loop with a nil pointer
dereference on startup. This release contains only #463 on top of v0.15.0

v0.15.0

This release contains a new NeonVM feature, a security improvement,
a new component (vm-monitor), and BREAKING CHANGES TO METRICS!

Breaking API changes:

Features:

- neonvm: Add user-defined additional InitContainers (#433)

No "fixes".

No protocol changes.

Other changes:

- informant: Move scaling logic into (new!) vm-monitor (#362)
- agent: Use fixed port for informant -> agent connections (#447)
- agent: Move "runner" state tracking from GaugeFuncs to GaugeVec (#447)
  - The following metrics were removed:
    - autoscaling_errored_vm_runners_current
    - autoscaling_panicked_vm_runners_current
    - autoscaling_agent_tracked_vms_current
    - autoscaling_vms_unsuccessful_communication_with_informant_current
    - autoscaling_billed_vms_unsuccessful_communication_with_informant_current
  - The following metric was added:
    - autoscaling_agent_runners_current

Upgrade path from v0.14.2:

- No ordering requirements, but all components must be upgraded before
  there's *any* usage of NeonVM's new extraInitContainers.

release workflow fix; see v0.15.0-alpha

v0.15.0-alpha

This is the initial release of the monitor.

Features:
The monitor replaces the informants upscaling/downscaling capabilities.
At the moment, it is started by vm-builder and connects to the
informant. In the future, the monitor will be started by compute_ctl
and will connect to the agent.

No protocol changes besides the addition of the
monitor<->informant/agent protocol.

No other changes.

v0.14.2

(very) Minor release, changing the behavior of VMs on kernel panic.

No breaking API changes.

No features.

Fixes:

- neonvm/runner: Add panic=-1 to kernel command line (#448)
  - panic=-1 triggers restart, but we run with --no-restart, so this
    will cause QEMU to exit, and the VM runner pod to be recreated,
    ultimately meaning that we restart on kernel panic, instead of hanging.

No protocol changes.

No "other" changes.

Upgrade path from v0.14.1:

- No ordering requirements, only NeonVM updated.

v0.14.1

This release is just bugfixes and improvements to existing behavior.

No breaking API changes.

No features.

Fixes:

- plugin: Handle VM 'Buffer' on delete (#438)
- agent/billing: Don't record events if VM watch is failing (#439)
- neonvm/runner: Create QEMU cgroup inside current one (#441)
  - This should fix the cAdvisor issue, and cgroup cleanup issue.

No protocol changes.

Other changes:
- plugin: make Filter unknownPods less verbose (#434)
- agent: "billed stuck" metric takes into account endpoint assignment
  timestamp (#425)

Upgrade path from v0.14.0:

- No ordering requirements.

v0.14.0

This release is *mostly* a collection of bugfixes, some that were
previously backported as hotfixes, and some more notable ones that
haven't yet been released.

No breaking API changes.

No new features.

Fixes:

- Fix errors from `go list` due to cilium replace directives (#417)
- agent: Fix LA-based scaling being 4x too small (#421)
- plugin: Fix 'Buffer' underflow (#424)
  - Backported in v0.13.4
- plugin: Ignore completed pods in Filter (#423)
  - Backported in v0.13.4
- plugin: Fix scoring to use current resources (#426)
  - Backported in v0.13.5
- plugin: Calculate "remaining reservable" resources via SaturatingSub (#428)
- plugin: Fix event queue usage (#430)
  - Backported in v0.13.6

No protocol changes.

Other changes:

- plugin: Log 'Buffer' in more places (#424)
  - Backported in v0.13.4

Upgrade path from v0.13.x:

- No ordering requirements.

v0.13.6

Hotfix release fixing the plugin's incorrect usage of fun/pubsub.Queue
that caused some events to be dropped.

v0.13.5

Hotfix release fixing the plugin's Score method so that it takes into
account actual resource usage. This release only contains a backport of
the fix from #426.

v0.13.4

Hotfix release to fix two issues:

1. Scheduler plugin's Filter logic was incorrectly counting Completed
   pods into the usage calculations.
2. Scheduler plugin's node state 'Buffer' field was always underflowing.

The fixes were in #423 and #424, respectively.

v0.13.3

Another small release, with a minor improvement to the plugin's method
call metrics, so we can avoid tripping alerts for overprovisioning pods.

Change was in #422, nothing else included.

v0.13.2

Another small release, primarily to fix the scheduler's handling of
overprovisioning pods. Also contains a minor improvement.

Fixes:

- plugin: Fix handling of ignored namespaces during Filter (#416, #418)

Other changes:

- plugin: Emit k8s Event on failed ExtractVmInfo (#408)
  - Should help with observability for certain failures.

Upgrade path from v0.12.x / v0.13.x:

- No ordering requirements.