Callers hold connections longer than the pool can recycle them. Wait time climbs first, available connections reach zero, and every upstream caller then times out together.

db.pool.wait_ms · db.pool.available · db.active_connections

A mass eviction empties the working set, every request misses at once, and the traffic that the cache was absorbing lands on the origin in a single wave.

cache.hit_ratio · cache.evictions_per_s · cache.memory_used_ratio

Consumers fall behind producers. Nothing errors, so alerting on error rate sees a healthy system while downstream data quietly goes stale.

stream.consumer.lag · stream.consumer.throughput

Resident memory climbs monotonically across a deploy window. Collection pauses lengthen as the heap fills, and the process is eventually restarted by the platform, which resets the graph and hides the trend.

process.memory.rss_mb · runtime.gc.pause_ms · process.restarts

One slow synchronous dependency occupies caller threads until the caller itself runs out. Retries multiply the load on the already-slow dependency, so the system degrades faster the harder it tries to recover.

http.client.timeout_rate · http.client.p99_ms

A certificate reaches its expiry. Failure is total and instant at the moment of expiry rather than gradual, which is what distinguishes it from a capacity problem on the same graph.

tls.handshake_failures_per_s · tls.certificate_days_remaining

Request volume is unchanged but query volume is not. A change replaced a batched read with a per-item read, so cost scales with result size rather than with traffic.

db.queries_per_request · db.query_rate

Documents concentrate on one shard. Mean latency stays acceptable and the tail does not, so a dashboard built on averages shows nothing wrong.

search.query.p99_ms · search.shard.max_docs_ratio