A noisy neighbor is a tenant whose workload consumes enough shared capacity to slow down other customers. The pressure may appear in database connections, CPU, storage throughput, queue workers, cache space, email delivery, or third-party API limits.
Last updated: September 8, 2026.
SELECT tenant_id,
COUNT(*) AS requests,
SUM(duration_ms) AS total_ms,
MAX(duration_ms) AS slowest_ms
FROM request_metrics
WHERE recorded_at >= NOW() - INTERVAL 5 MINUTE
GROUP BY tenant_id
ORDER BY total_ms DESC
LIMIT 20;Aggregate demand by tenant, not only by server. Global averages can look healthy while one customer is slow or one customer is causing everyone else to wait.
Control each constrained resource
- Apply per-tenant rate and concurrency limits before expensive work begins.
- Use fair queues or separate worker pools so one tenant cannot fill every available slot.
- Set query timeouts, pagination limits, export limits, and bounded report ranges.
- Allocate cache and storage quotas by tenant and monitor eviction or growth.
- Limit simultaneous database connections from each tenant or application workload.
Do not rely on billing limits alone. A legitimate paid customer can still create an operational incident, and a failed integration can generate accidental traffic. Alert on sudden changes as well as fixed operational thresholds.
Degrade only the responsible workload
When a limit is reached, delay that tenant’s background work, reduce export size, or return a clear retry response. Avoid disabling unrelated tenants. Provide administrators with the tenant, resource, current usage, threshold, and recent trend.
Know when to isolate
Some tenants consistently require more capacity or stronger guarantees. Move them to another shard, deployment stamp, or dedicated database using the placement strategy from multi-tenant database design. Isolation can also become a paid service tier.
Microsoft’s multitenant storage guidance recommends throttling or rate limiting to reduce noisy-neighbor effects. Combine those controls with tenant-aware logs and metrics so support can explain exactly what happened.
