# Memory pressure
Symptom: latency climbs under load, throughput drops, but CPU is not saturated.
## Diagnose
Atlas Metrics: cache eviction rate, page faults, and resident memory vs the tier's RAM. High evictions with a hot working set means MongoDB keeps throwing useful data out of cache to make room.
## Confirm
Correlate with query patterns: unindexed queries that scan large ranges churn the cache. Check the slow query list; a few cache-busting shapes are often the whole story.
## Fix, in order
1. Index the cache-busting queries so they touch pages instead of scanning them.
2. Reduce document bloat the hot paths fetch (project only needed fields).
3. Scale the tier up so RAM covers the working set. There is no tuning that substitutes for enough RAM when the hot data genuinely exceeds it.
## Verify
Eviction rate drops to near zero under the same load, p95 latency recovers, and the improvement survives a full traffic cycle.