URL Shortener & Link Management Service  ·  View 20 of 23  ·  Operations

Observability

Five signal classes across five pipeline stages, including two probes that test claims rather than resources.

Editable source SVG draw.io All views
Edge Resolution Propagation Management Analytics Golden metrics Redirects/s per PoP Cache hit rate ≥ 97% Resolve p99 ≤ 40 ms L1/L2/L3 hit mix Deny-set age ≤ 5 s Projection lag ≤ 5 min Create p99 Claim conflict rate Click freshness ≤ 60 s p95 Correctness checks Terminal-state mix 404 · 410 · block Wrong-destination probe synthetic, zero tolerance Revocation end-to-end probe every 60 s Reserved-alias audit Count reconciliation ±0.5% daily Saturation PoP egress Redis CPU and evictions Coalescer queue depth Change feed backlog Cosmos RU headroom write region Event Hubs lag Cost signals Cost per million redirects ≤ $0.05 Miss cost share RU spend per tenant Raw retention spend Bot traffic share Alarms that page Redirect error rate any region p99 breach 5 min Revocation probe fails the 60 s promise Uniqueness conflict spike Click loss > 0.05% Observability — Signals by Pipeline Stage One row is different in kind: the wrong-destination and revocation probes test claims the platform sells, not resources it consumes. v 1.0 · owner Platform Architecture · date 2026-09

The unusual row

  • The wrong-destination probe asserts the platform's one zero-tolerance guarantee continuously, by resolving known codes and comparing against expected targets.
  • The revocation probe kills a synthetic link every 60 seconds and measures when it actually stops resolving globally — testing the published promise rather than the mechanism's health.
  • Both would pass unnoticed in a conventional golden-signals dashboard while the platform silently broke what it sells.

Thresholds

  • Cache hit rate ≥ 97% (alarm below 94%), resolve p99 ≤ 40 ms, deny-set age ≤ 5 s, projection lag ≤ 5 min.
  • Click freshness ≤ 60 s p95, click loss ≤ 0.05%, count reconciliation within ±0.5%.
  • Cost per million redirects ≤ $0.05, alarm above $0.08 — a cost signal, alarmed like a reliability one.

Saturation to watch

  • Redis evictions are the leading indicator of a hit-rate collapse, which is also a cost event.
  • Change feed backlog is the leading indicator of a missed edit propagation bound; Cosmos RU headroom in the write region is the leading indicator of a creation outage.
◀  Release Pipeline All views Link Lifecycle  ▶
Solution Architecture v1.0 · Microsoft Azure · Platform Architecture · 2026-09 · 23 views · 16 architecture decision records20-observability