SLA.md

Service Level Agreement (SLA)

Metric	Target	Alert Threshold	Window
P50 search latency	< 50 ms	—	5 min
P99 search latency	< 500 ms	> 500 ms for 2 min	5 min
P99.9 search latency	< 2 000 ms	—	—
Redis cache hit rate	≥ 85%	< 70% for 5 min	5 min
HTTP 5xx error rate	< 0.1%	> 1% for 1 min	5 min
Circuit breaker open	0 occurrences	> 0 for 30 s	—

Severity	Definition	Response Time	Resolution Time
P0 — Critical	Search API fully unavailable OR circuit breaker open > 5 min	15 min	2 hours
P1 — High	P99 latency > 1 s sustained OR error rate > 5%	30 min	4 hours
P2 — Medium	Cache hit rate < 70% OR P99 > 500 ms for > 10 min	2 hours	8 hours
P3 — Low	Non-critical degradation, single transient errors	Next business day	48 hours

Four Prometheus alerts are defined in infra/docker/alert_rules.yml:

Alert	Expression	Duration	Severity
`CircuitBreakerOpen`	`vector_catalog_circuit_breaker_state > 0`	30 s	critical
`P99LatencyHigh`	`histogram_quantile(0.99, ...) > 500`	2 min	warning
`CacheHitRateLow`	`vector_catalog_cache_hit_rate < 0.70`	5 min	warning
`ErrorRateHigh`	5xx rate > 1% (5-min rolling)	1 min	critical

Tool	URL
Live health check	`https://vector-catalog-api.politefield-8fe8e6a2.eastus.azurecontainerapps.io/health`
Live Prometheus metrics	`https://vector-catalog-api.politefield-8fe8e6a2.eastus.azurecontainerapps.io/metrics`
Local Prometheus	`http://localhost:9090`
Local Grafana dashboard	`http://localhost:3000` (admin / admin)
GitHub Actions (CI)	`https://github.com/ritunjaym/vectorscale/actions`
Trivy security scan	`https://github.com/ritunjaym/vectorscale/security`

With a 99.9% uptime target the monthly error budget is 43.8 minutes.

Burn rate	Meaning	Alert within
1×	Consuming budget at exactly the SLO rate	—
5×	Budget exhausted in ~6 days	P2
14.4×	Budget exhausted in ~2 hours	P1
36×	Budget exhausted in ~1 hour	P0

Scheduled maintenance windows (announced ≥ 24 hours in advance, max 4 hours/month)
Force majeure (Azure region outage, DNS provider failure)
Client-side network issues outside Azure