Tags

v0.18.0-patch1

A few fixes for issues noticed during staging deploy of v0.18.0. This
release contains only the changes from #551, #552, and #553; refer to
those PRs for more info.

v0.18.0

This release contains breaking changes and significant refactoring in
the autoscaler-agent, and little else.

No breaking API changes (but, check "Protocol changes").

Features:

- billing: Record push error causes in agent's metrics (#369)

Fixes:

- ARCHITECTURE.md: update vm-monitor link (#537)
- plugin/ARCHITECTURE.md: Remove misleading section on runtime config
  changes (#538)

Protocol changes:

- agent: Remove vm-informant support (#506)
  - Also RENAMES ALL METRICS WITH 'informant' IN THE NAME TO 'monitor'

Other changes:

- billing: Change idempotency key format (#545)
- agent/runner: Simplify getMetricsLoop [small refactor] (#548)
- agent: Scaling logic refactor [big! refactor] (#371)
- agent/runner: Add temporary monitor state dump (#549)

Upgrade path from v0.17.x:

- No ordering requirements, though care should be made to ensure that
  all current and future VMs exclusively use the vm-monitor instead of
  the vm-informant.

v0.17.12

(yet another) small release, this time just with #534 and #533.
The two PRs are unrelated; #533 is a relatively minor rbac change, #534
is another vm-builder change to process lifecycle management.

v0.17.11

Relatively small release, containing fixes around VM shutdown,
implemented both in neonvm-runner and in vm-builder.

This release only contains the changes from #525 and #528; refer to
those PRs for more details.

v0.17.11-pre1

Micro-release 1 of 3 for fixing VM shutdown behavior. This release
contains only the changes from #528.

v0.17.10

Another small release, two changes this time:

- Add 'sleep 1' before VM shutdown to hopefully allow TCP connections to
  finish their FIN/RST packets (#523)
- "state dump" missing field fix in autoscaler-agent (#522)

Added ability to specify serviceAccountName for runner pod

v0.17.8-patch1

Follow-up fix to the hotfix earlier, because it wasn't working in the
way that was intended. This release contains only the contents of #519.

v0.17.8

Small hotfix release affecting NeonVM only, making EmptyDisk mounts more
permissive.

v0.17.4-patch1

Backport of #515 without the other changes involved before it.

v0.17.7

Small release to fix an issue that can leave VMs with "autoscaling
stuck" around failed downscaling and autoscaler-agent restart.

v0.17.6

Another small update, with notable changes in NeonVM.

Features:

- neonvm: Add EmptyDisk.Discard option (#507)
  - It's opt-in, allows extra mounts to release filesystem usage on the
    host when they shrink.

Fixes:

- agent/dispatcher: Close websocket if NewDispatcher fails (#511)

Other changes:

- neonvm/runner: Add overcommit factor for CPU limits (#509)
  - Should fix some disk speed issues.
- neonvm/runner: Switch to zap for logging (#460)
  - Should help make it more clear where logs are coming from.
- pull vm-monitor from neon repository (#505)
  - completes the transition away from neondatabase/vm-monitor

Upgrade path from v0.17.5:

- Make sure all autoscaling components have been updated to v0.17.6
  before setting EmptyDisk.Discard on VMs in the cluster.

v0.17.5

Minor change to vm-builder, from PR #502.

v0.17.4

Hotfix-ish release containing a few improvements, from all of the
following PRs: #481, #496, #499, #500, and #501.

Except for #481 (randomize plugin scores), these are all improving the
behavior around agent <-> monitor communication.

v0.17.3

Hotfix release containing a number of fixes, primarily for the agent's
vm-monitor connection handling, but also a couple for the neonvm-controller.
The PRs contributing to this release are all of:
  #461, #491, #495, #494, #493, #497, and #498

v0.17.2

Another hotfix release, fixing a false positive in the "stuck VMs"
metric & associated alert, when communicating directly with the
vm-monitor. This release contains a single commit, from #489.

v0.17.1

Hotfix release for some agent/monitor communication issues that popped
up on staging, with only the changes from #487 and #488.

v0.17.0

This release contains bugfixes, breaking changes, features, and
quality-of-life improvements. Previous hotfix or single-PR releases
since v0.16.0 are also included here (notably v0.16.3, which shouldn't
have been just a patch release).

Breaking API changes:

- neonvm: Make spec.enableAcceleration a pointer (#474)
  - This should be backwards compatible at the k8s API server level (and
    with existing clients), but may cause compilation failures for usage
    of the Go client.

Features:

- neonvm/builder: add flag to configure uid of cgroup (#480)
  - already released in v0.16.2

Fixes:

- plugin: Fix nodeConfig.ScorePeak JSON name (#478)
  - already released in v0.16.1
- plugin: Fix Score scale usage (#479)
  - already released in v0.16.1
- neonvm/controller: Handle pod termination during Scaling phase (#483)
- agent: harden dispatcher concurrency (#484)

Protocol changes:

- agent: establish connection with monitor instead of informant (#442)
  - NOTE: this removes the vm-informant from everywhere.
  - already released in v0.16.3

Other changes:

- plugin: Add availability zone label to node metrics (#482)

Upgrade path from v0.16.x:

- No ordering requirements for cluster-wide components (i.e.
  neonvm/scheduler/agents)
- autoscaler-agents *must* be updated before using the updated
  vm-builder
- #474 may require updating usage of the EnableAcceleration field.

v0.16.3

Another single-PR release, this time for #442, which is a significant PR
that adds handling in the autoscaler-agent to connect directly to the
vm-monitor, if it's listening on the port that the vm-informant *was* on.

v0.16.2

Tiny release just adding #480, to allow configuring what vm-builder sets
as the admin user for the neon-postgres cgroup.