Frame server analysis output using USE/RED/Golden Signals vocabulary

[erikdarlingdata]

Summary

PerformanceMonitor already collects the data that maps to the three established monitoring frameworks (USE, RED, Golden Signals), but doesn't organize or present findings using their vocabulary. Framing the analysis output this way would make the tool's insights immediately recognizable to anyone trained in SRE/DevOps practices and would provide a structured way to ensure coverage across all resource and service dimensions.

The Frameworks

USE Method (Brendan Gregg) — for every resource

	CPU	Memory	Disk I/O	Workers	TempDB
Utilization	CPU %	Buffer pool vs max memory	File latency	Active vs max workers	Space used
Saturation	SOS_SCHEDULER_YIELD, runnable tasks	RESOURCE_SEMAPHORE waits, grant queuing	PAGEIOLATCH waits	Thread exhaustion	PAGELATCH allocation waits
Errors	Non-yielding schedulers	OOM, memory broker shrinks	15-second I/O warnings	Timeout errors	Out-of-space

RED Method (Tom Wilkie) — for the database as a service

Signal	SQL Server Metric
Rate	Batch Requests/sec (perfmon)
Errors	Failed queries, deadlocks, timeouts, connection errors
Duration	Query execution time distribution

Four Golden Signals (Google SRE) — the combined view

Signal	SQL Server Metric
Latency	Query duration (distinguish successful vs error latency)
Traffic	Batch Requests/sec, Transactions/sec
Errors	Deadlocks, timeouts, failed queries
Saturation	Wait stats (the entire wait statistics framework is fundamentally a saturation measurement system)

Where This Applies

Analysis Engine Output (both Dashboard and Lite)

The analyze_server inference engine produces findings with severity and evidence. These findings could be tagged or grouped by framework category:

• "CPU Saturation (USE): SOS_SCHEDULER_YIELD elevated, runnable task queue depth > 10"
• "Latency (Golden Signals): p95 query duration increased 3x vs baseline"

Progressive Summary View (issue #689)

The server landing summary could organize its ranked problems using USE categories for resource issues and RED/Golden Signals for workload issues. This gives users a mental model, not just a list.

MCP Tool Output

The analyze_server and get_analysis_findings MCP tools could include framework tags in their output, making AI-assisted investigation framework-aware.

Design Notes

• This is primarily a framing/presentation change, not a new data collection effort
• All underlying data already exists in current collectors
• The frameworks provide a checklist for completeness: if the analysis engine never reports on "Memory Errors," that's a gap worth noticing
• The vocabulary is widely taught in SRE and DevOps contexts — using it makes PerformanceMonitor's output portable to incident reports, postmortems, and runbooks
• Applies to both Dashboard and Lite, plus MCP tool output

References

• Brendan Gregg — The USE Method
• RED and USE Metrics — Better Stack
• Four Golden Signals — Google SRE Book

[erikdarlingdata]

Implementation order: Lite first, then port to Dashboard once the pattern is validated. This avoids the complexity of iterating on both apps simultaneously.

[erikdarlingdata]

Implementation Notes

After working through #681, #682, #683, #684 (slicers, drill-downs, investigate actions), the navigation infrastructure is now in place. #691 and #689 should be implemented together — the USE/RED/Golden Signals framing only has value if there is a UI surface that uses it, and that surface is the progressive summary (#689).

Proposed approach

Combined #691 + #689: Framework-aware server summary as landing view

The new landing view replaces the current Overview tab default. It aggregates existing data into a ranked problem list organized by framework category.

Layout sketch

┌─────────────────────────────────────────────────────────────┐
│  Server Health Summary                          Auto-refresh │
├─────────────────────────────────────────────────────────────┤
│                                                             │
│  ── USE Method (Resource Health) ─────────────────────────  │
│                                                             │
│  CPU                                                        │
│    Utilization: 78% avg (Warning)        [Investigate →]    │
│    Saturation: SOS_SCHEDULER_YIELD 340ms/sec (Critical)     │
│    Errors: None                                             │
│                                                             │
│  Memory                                                     │
│    Utilization: 92% of max memory (Warning)                 │
│    Saturation: 3 pending memory grants (Info)               │
│    Errors: None                                             │
│                                                             │
│  Disk I/O                                                   │
│    Utilization: Avg read latency 4ms (OK)                   │
│    Saturation: PAGEIOLATCH_SH 12ms/sec (Info)               │
│    Errors: 2 x 15-second I/O warnings (Warning)            │
│                                                             │
│  TempDB                                                     │
│    Utilization: 45% space used (OK)                         │
│    Saturation: PAGELATCH_UP 8ms/sec (Info)                  │
│    Errors: None                                             │
│                                                             │
│  Workers                                                    │
│    Utilization: 42/512 active (OK)                          │
│    Saturation: None                                         │
│    Errors: None                                             │
│                                                             │
│  ── RED Method (Service Health) ──────────────────────────  │
│                                                             │
│    Rate: 1,247 batch requests/sec                           │
│    Errors: 14 deadlocks, 3 blocking events (Warning)       │
│    Duration: p95 query duration 340ms (OK)                  │
│                                                             │
│  ── Recent Incidents ─────────────────────────────────────  │
│                                                             │
│  #1 BLOCKING STORM (Critical) 2:00-2:45pm    [Investigate →]│
│     14 events, longest blocked 47s, SPID 128               │
│                                                             │
│  #2 CPU SPIKE (Warning) 2:12pm               [Investigate →]│
│     Peak 96%, top query 0x7A3F... (42% of CPU)             │
│                                                             │
└─────────────────────────────────────────────────────────────┘

Data sources (all existing, no new collectors)

Framework	Signal	Source
USE:CPU:Utilization	CPU %	cpu_utilization_stats
USE:CPU:Saturation	SOS_SCHEDULER_YIELD, runnable tasks	wait_stats, cpu_scheduler_stats
USE:CPU:Errors	Non-yielding schedulers	HealthParser_SchedulerIssues
USE:Memory:Utilization	Buffer pool vs max	memory_stats
USE:Memory:Saturation	RESOURCE_SEMAPHORE, grant queuing	wait_stats, memory_grant_stats
USE:Memory:Errors	OOM, broker shrinks	HealthParser_MemoryConditions
USE:DiskIO:Utilization	File latency	file_io_stats
USE:DiskIO:Saturation	PAGEIOLATCH waits	wait_stats
USE:DiskIO:Errors	15-second I/O warnings	HealthParser_IOIssues
USE:TempDB:Utilization	Space used	tempdb_stats
USE:TempDB:Saturation	PAGELATCH allocation waits	wait_stats
USE:Workers:Utilization	Active vs max workers	cpu_scheduler_stats
RED:Rate	Batch Requests/sec	perfmon_stats
RED:Errors	Deadlocks, blocking	blocking_deadlock_stats
RED:Duration	Query execution time	query_stats

Investigate buttons

Use the same pattern from #684 — click navigates to relevant tab with time range pre-set. Infrastructure already built.

Severity thresholds

Each signal gets a severity (OK/Info/Warning/Critical) based on thresholds. The analysis engine already computes many of these. For signals not yet covered, add threshold checks to the summary aggregation query.

Scope

• Lite first (reads from DuckDB), then Dashboard port
• Replaces the default landing tab (currently Wait Stats in Lite, Resource Overview in Dashboard)
• Existing tabs remain as drill-down destinations
• MCP tools get framework tags in output for AI-assisted investigation

[erikdarlingdata]

Closing — low impact. Labeling existing findings with USE/RED/Golden Signals vocabulary doesn't add meaningful value for the DBA audience.