Connection pool hit zero. Memory leaked to the ceiling. Disk filled with logs overnight. GPU capacity ran out under inference load.
Resource exhaustion (FM-13) looks different on the surface every time — CPU, memory, disk, connections, tokens, GPU compute — but for SRE teams, the shape is always the same: a finite resource consumed faster than it can be replenished. It's 11% of classified unplanned incidents in 2025, and one of the most automatable patterns in the taxonomy.
Full data: stackgen.com/state-of-reliability-2026.
Two Sub-Patterns, Different Responses
Demand Saturation: traffic spike exceeds capacity. The resource is correctly sized; the load grew unexpectedly. Response: scale out, expand capacity, increase quota.
Resource Leak / Runaway Consumption: pathological growth uncorrelated with traffic — memory leak, queue runaway, log disk fill. Response: restart and fix. Confusing these two leads to adding capacity to a leaking system, watching it fill up again.
Where FM-13 Appears in the Data
- Communications / infrastructure services: high connection-count services (Twilio, Bandwidth, Sinch) show FM-13 frequently, particularly connection pool exhaustion under traffic spikes
- AI model providers: GPU capacity exhaustion under inference load is an emerging FM-13 variant. When an AI service can't handle inference load and returns degraded quality, that crosses into FM-17. When it simply fails requests, it stays FM-13.
- Crypto operators: queue exhaustion during high transaction volume periods (major market moves, new chain launches) appears repeatedly in Kraken, Coinbase, and Luno incident histories
Why FM-13 Is SRE’s Highest-Automation Pattern
- Detection: resource utilization chart hits ceiling (CPU 100%, connections 0, disk 100%, OOM events)
- Diagnosis: demand saturation or leak? (monotonic growth vs. traffic-correlated spike)
- Remediation: demand → scale out. Leak → restart + fix.
All three steps are automatable from standard telemetry. No business-logic judgment required for the initial response. This is why FM-13 is Tier C — AI-Closed — in the autonomy framework.
Prevention: Alert Early
- Alert at 75%, not 95%: gives time to respond proactively before the incident
- Track P99 resource consumption, not just mean: peak consumption is what causes exhaustion
- Separate demand saturation and leak alerting: monotonic growth over 24 hours is a different signal from a traffic-correlated spike
Key Takeaways
- 11% of 2025 incidents — high frequency, high automation potential
- Two sub-patterns require different responses: confusing demand saturation with leaks makes incidents worse
- AI inference is an emerging FM-13 vector: GPU capacity exhaustion growing as AI features embed in products
- FM-13 is Tier C (AI-Closed) — the most automatable pattern in the taxonomy
stackgen.com/state-of-reliability-2026 | LinkedIn webinar
