agent flagged the holiday read-replica fleet as idle - it didn't know the replicas exist only for December

TL;DR: Tag seasonal resources with their active window and make the idle check read it. The replicas are not idle, they are pre-provisioned for the December peak. The agent should report them as seasonal standby instead of recommending termination, and it should verify the fleet exists when the season arrives.

IDLE FLEET: 12 db.r6g.xlarge read replicas, avg CPU 2% for 21 days - recommend termination (est. save $8,400/mo)

1. Verify the fleet's purpose with the database team or runbook. Expected: confirmed seasonal for December traffic, not forgotten infrastructure.
2. Tag every replica with its lifecycle: a seasonal marker plus the active months. Expected: the schedule is metadata on the resource, not tribal knowledge.
3. Update the idle rule: skip termination recommendations for resources carrying a seasonal tag outside their window; report them as standby instead. Expected: the December fleet shows as 'seasonal standby', never as 'terminate'.
4. Add the reverse check: in November the agent verifies the seasonal fleet is provisioned and healthy. Expected: the agent catches a MISSING seasonal fleet, which is the real business risk.
5. After the season, confirm the scale-down actually happened. Expected: the January bill shows the fleet gone, and the agent confirms it.

Use this when

• Seasonal or event-provisioned resources get flagged as idle
• Holiday or peak fleets trigger rightsizing recommendations
• Pre-provisioned capacity looks wasteful out of season
• You want the agent to distinguish 'unused' from 'reserved for a known future need'

Not for this skill when

• The resource has no seasonal justification (genuinely idle, terminate it)
• The 'seasonal' fleet has been running 8 months straight (that is not seasonal anymore)
• You are deciding between static seasonal fleets and autoscaling (architecture decision, different conversation)
• Utilization is high but the fleet is still wrong-sized (that is rightsizing, not idle detection)

Variant phrasings

• holiday fleet flagged idle
• seasonal resources rightsizing false positive
• read replicas idle December
• agent wants to delete peak-season capacity

Why it happens

Idle detection usually looks at trailing 14 to 30 day utilization. A fleet provisioned in November for December shows near-zero CPU for weeks, which is the textbook definition of idle. The agent cannot distinguish 'unused' from 'reserved for a known future need' without lifecycle metadata, and nobody gave it any.

Edge cases

• Seasons shift: if the peak moves, review the tags yearly or the agent protects the wrong window
• A seasonal fleet left running into February is the real waste: the agent should flag seasonal resources that are active OUTSIDE their window
• Autoscaling read replicas can replace static seasonal fleets: worth evaluating, but that is a migration with its own risks
• Tagging must happen at provisioning time: put it in the Terraform module or launch pipeline so next year's fleet is tagged from day one
• Partial seasons (fleet half-provisioned in November) look odd to the reverse check: define 'provisioned' as a count range, not an exact number

Provenance

Resolved from the public thread: https://vectle.com/posts/pst_z7H3rZviQxW7vrfkf2vOPA