Knowledge Base

Redis memory exhaustion is a common pattern in high-traffic services. Key indicators include OOM errors, increasing latency, and sudden cache miss rate spikes. Resolution: (1) Check maxmemory-policy, (2) Identify keys without TTL, (3) Flush stale data, (4) Scale memory allocation.

OOMKilled pods indicate a container exceeding its memory limit. Steps: (1) Check resource limits in deployment spec, (2) Heap dump analysis, (3) Profile allocations, (4) Check for unbounded caches or buffers, (5) Review recent code changes for leak patterns.

Connection exhaustion in Aurora occurs when application connections are not properly released. Root causes: missing connection.close() in error paths, unbounded connection growth, PgBouncer misconfiguration. Fix: Audit connection lifecycle, enable connection pooling properly.

Spot instances can be interrupted with 2-minute notice. Best practices: Use Pod Disruption Budgets, maintain 30% on-demand baseline, implement graceful shutdown handlers, use mixed instance types, enable Karpenter or Cluster Autoscaler.

S3 supports 3,500 PUT and 5,500 GET requests per second per prefix. To avoid rate limits: use date-based prefix partitioning, implement exponential backoff, use Kinesis Firehose for buffering, distribute writes across multiple prefixes.

ES enters flood stage when disk >95% full, making indices read-only. Recovery: (1) Free disk space immediately, (2) Reset read-only blocks, (3) Implement ILM policies for automatic cleanup, (4) Add disk usage CloudWatch alarms at 70/80/90%.

Stale CDN content after deployments is a common issue. Best practices: Use versioned asset filenames for JS/CSS, implement proper cache-control headers, automate invalidation in CI/CD with path wildcards, verify invalidation completion before announcing deploy success.