| name | troubleshoot-nvme |
| description | Use when diagnosing issues with NVMe: availability, health & critical warnings, errors, or endurance & wear degradation. Queries Netdata via MCP for controller state, smart critical warning; read-only mode (bit 3), smart critical warning; nvm subsystem reliability degraded (bit 2), media and data integrity errors, error log entries, applies the diagnostic tree from the Netdata operator playbook, and recommends remediation. |
| version | 0.1.0 |
| author | Netdata |
| license | Apache-2.0 |
| tags | ["netdata","troubleshoot","mcp","nvme"] |
Troubleshoot NVMe
When to use this skill
- Any time the user reports a NVMe service behaving outside its expected envelope (elevated errors,
latency, saturation, resource exhaustion, or unexpected restarts).
- An on-call engineer is paging on a Netdata alert tied to a NVMe instance and wants a structured
triage path.
Key facts
- This skill wraps the Netdata operator playbook for NVMe. It does not replace the playbook; it
routes a coding agent through MCP queries against the same signals the playbook relies on.
- NVMe is a host-to-controller communication protocol that exposes flash storage over PCIe.
Understanding NVMe monitoring requires understanding three layers: the PCIe transport, the NVMe
controller, and the flash media behind it.
- The playbook decomposes NVMe health into 8 signal domains: Availability, Health & Critical
Warnings, Errors, Endurance & Wear, Thermal, Power & Shutdown. Each domain maps to one rule file
in this skill.
- Netdata observes the signals listed in the rule files via its native collectors, plus any
OpenTelemetry-shipped metrics that your NVMe instrumentation adds. Both paths end at the same MCP
query surface.
- Netdata's nvme collector emits 16 context(s) under
nvme.*. The rule files enumerate which
contexts surface which domain; the Verification section below names the load-bearing ones
explicitly.
Step-by-step
- Confirm the NVMe service is up. Query Netdata via MCP with
list_nodes and filter by the host
running the target. A missing node means the symptom is at the network or orchestrator layer, not
inside the service.
- Pull the last 15 minutes of signals for the target. Use
query_metrics against the contexts
listed in the domain rule files. Run find_anomalous_metrics in parallel over the same window;
anomalies frame which rule file to read first.
- Correlate with host-level signals (
system.cpu.utilization, system.memory.usage,
system.disk.io_time). Many service-level failures have a host-resource precursor.
- Apply the remediation hinted at in the matching rule file or the operator playbook. Re-run the
MCP queries from the Verification section to confirm the signals returned to expected ranges. A
fix that does not move the signal back is not a fix.
Handy MCP call templates
# Discover metrics from NVMe
list_metrics with q="nvme"
# Pull a specific context over the last window
query_metrics with context="nvme.device_critical_warnings_state", relative_window=-15m
# Rank anomalies for the service or host
find_anomalous_metrics with node=<host> and context_pattern="nvme.*"
# Correlate a known problem context with others
find_correlated_metrics around the incident window
# Show current alert state
list_raised_alerts scoped to the node
Common mistakes
- Treating NVMe as a generic HTTP or process health check. NVMe has specific failure archetypes (see
Key facts) that generic checks miss.
- Stopping at the first anomalous metric. Several archetypes produce correlated spikes; use
find_correlated_metrics to widen the search before concluding a root cause.
- Quoting percentile latency without the sample count. Low traffic plus a single slow request moves
p99 by seconds.
- Reading dashboards for a window shorter than the failure's fingerprint. Slow-brew failures (queue
growth, bloat, memory fragmentation) need 30+ minutes of data to see the trend.
- Skipping the host-level correlation. A process-level fix for a noisy-neighbour problem does not
hold.
- Assuming alert thresholds are tuned for your workload. Tune against observed NVMe traffic before
escalating an alert configuration issue.
Verification
Run these MCP queries against the Netdata instance that sees the NVMe service. Every context listed
below is a real Netdata chart name; the agent does not need to guess.
1. list_metrics filtered by q="nvme" (returns every nvme.* context Netdata sees)
2. query_metrics with contexts=[nvme.device_critical_warnings_state, nvme.device_unsafe_shutdowns_count, nvme.device_media_errors_rate, nvme.device_error_log_entries_rate, nvme.device_thermal_mgmt_temp1_transitions_rate, nvme.device_thermal_mgmt_temp2_transitions_rate] and relative_window=-30m
3. find_anomalous_metrics filtered by node=<host> and context_pattern="nvme.*"
Load-bearing contexts for this service:
nvme.device_critical_warnings_state: Critical warnings state (state). Dimensions:
available_spare, temp_threshold, nvm_subsystem_reliability,
read_only, volatile_mem_backup_failed,
persistent_memory_read_only.nvme.device_unsafe_shutdowns_count: Unsafe shutdowns (shutdowns). Dimensions: unsafe.nvme.device_media_errors_rate: Media and data integrity errors (errors/s). Dimensions: media.nvme.device_error_log_entries_rate: Error log entries (entries/s). Dimensions: error_log.nvme.device_thermal_mgmt_temp1_transitions_rate: Thermal management temp1 transitions
(transitions/s). Dimensions: temp1.nvme.device_thermal_mgmt_temp2_transitions_rate: Thermal management temp2 transitions
(transitions/s). Dimensions: temp2.
A clean result means every context is within its expected band and the find_anomalous_metrics list
is empty or contains only already-acknowledged items. If the fix was real, re-running the same
queries 10 minutes after applying it will show a clean result. If it does not, revert and look
deeper.
When the fix does not hold
If signals drift back into the anomalous range within 30 minutes of a remediation, the cause was
deeper than the applied change. Typical misdiagnoses for NVMe:
- Host-resource pressure masquerading as application bug.
- Dependent service (DB, cache, upstream) causing a secondary symptom in the instrumented service.
- Configuration change that was never reloaded (some subsystems only pick up config on full
restart).
Escalate by widening the query window: 2-6 hours instead of 15 minutes. Slow-moving causes are
invisible at triage window sizes.
References
