Diagnosing Discloud Auto-Scaling Failures During CPU Spikes

The Problem: Threshold Breached, No Scale-Out

You have a Discloud scaling policy configured with a CPU upper threshold of 70%. Your VM instance sustains 85% CPU utilization for over five minutes, yet no scale-out event occurs. The policy is active and the instance is healthy, but the system fails to add capacity. In these scenarios, the scaling logic is usually functional, but the metric bridge—the component forwarding CPU samples from the instance agent to the scaling engine—is delivering stale or missing data.

Diagnostic Summary

Observation	Likely Cause	Primary Check	Resolution
Sustained CPU spike; no event logged	Metric bridge latency or stale samples	Review metric collection logs for gaps	Align aggregation interval with evaluation window
Intermittent scaling triggers	Aggregation interval exceeds evaluation window	Compare window vs. interval settings	Lower aggregation interval to < 30s
Policy fails after platform update	Metric naming convention change (v2.3)	Verify metric selector names in policy	Update selector to current naming convention
Inconsistent behavior across regions	Region-specific defaults or quotas	Check regional documentation for thresholds	Adjust thresholds to regional norms

Step-by-Step Diagnostic Sequence

1. Inspect the Scaling Events Log: Access the Discloud console for the specific instance. Look for entries during the CPU spike. If the log is empty, the scaling engine never registered a breach. If an event exists but no capacity was added, the issue is likely resource quotas or host capacity, not metrics.
2. Verify Interval Alignment: Compare the policy's evaluation window (the look-back period) with the metric aggregation interval (how often samples are bundled). If the aggregation interval is longer than the evaluation window, the engine may evaluate a period before the latest samples have arrived.
3. Check Agent Heartbeats: Review the metric collection logs for heartbeat failures. The monitoring agent should transmit data every 30 seconds by default. Gaps here indicate the agent is failing to push data to the bridge.
4. Validate Metric Names (Version 2.3+): If the instance was upgraded to or deployed on version 2.3, check the metric selector. Discloud changed naming conventions in this version; policies using legacy names will fail silently without triggering an error.

Fixes Based on Findings

• For Interval Mismatches: Lower the aggregation interval to 30 seconds. This ensures that a standard 5-minute evaluation window contains multiple fresh data points. Note: Record the original value before changing to allow for rollback.
• For Agent Failures: Restart the monitoring agent or correct the transmit frequency to 30 seconds. Confirm the heartbeat recovers in the collection logs before re-testing load.
• For Naming Mismatches: Update the policy selector to the current naming convention. Do not guess the names; verify the exact strings in the Discloud console or official documentation for version 2.3.

Verification and Testing

To confirm the fix, monitor the scaling events log during a subsequent spike. For a more rigorous check, use an external monitoring tool like Prometheus to compare the scaling engine's view with actual host data. Run the following query on your Prometheus instance (requires node_exporter on the VM):

100 - avg(rate(node_cpu_seconds_total{mode="idle"}[5m])) * 100

If Prometheus shows 85% CPU with current timestamps but the Discloud scaling engine shows nothing, the metric bridge is still malfunctioning. If Prometheus also shows gaps, the issue lies with the instance agent.

Escalation Criteria

Escalate to the infrastructure team only after confirming that:

• The policy is enabled and using current v2.3+ metric names.
• The aggregation interval is shorter than the evaluation window.
• Agent heartbeats are consistent (every 30s) in the logs.

Provide the team with the Policy ID, Instance ID, the specific time window of the spike, and the metric logs. They will need to investigate hypervisor-level metrics, underlying host load, or global resource quotas.

Limitations

Scaling behavior and default thresholds vary by Discloud region. Always cross-reference regional documentation before assuming a 70% threshold is the standard. Additionally, custom agent deployments may ignore the 30-second default transmit frequency, requiring manual adjustment of the aggregation intervals.