Cache Hit Ratio

Cache hit ratio is the fraction of reads served from memory. It is the clearest early indicator that the working set has outgrown the cache — and it is frequently misread.

Per engine

-- PostgreSQL: buffer cache hit ratio.
SELECT sum(heap_blks_hit) * 100.0 / nullif(sum(heap_blks_hit + heap_blks_read), 0) AS hit_pct
FROM pg_statio_user_tables;
 
-- Per table, to find the one that is missing.
SELECT relname,
       round(100.0 * heap_blks_hit / nullif(heap_blks_hit + heap_blks_read, 0), 1) AS hit_pct,
       heap_blks_read
FROM pg_statio_user_tables
WHERE heap_blks_hit + heap_blks_read > 10000
ORDER BY heap_blks_read DESC LIMIT 20;

-- MySQL / InnoDB.
SELECT
  (1 - reads.value / requests.value) * 100 AS hit_pct
FROM performance_schema.global_status reads,
     performance_schema.global_status requests
WHERE reads.variable_name = 'Innodb_buffer_pool_reads'
  AND requests.variable_name = 'Innodb_buffer_pool_read_requests';

INFO stats      -- Redis: keyspace_hits and keyspace_misses

// MongoDB: WiredTiger cache pressure.
db.serverStatus().wiredTiger.cache;

Reading it correctly

For Redis, the ratio means something different again: it measures whether the application is asking for keys that exist, not whether the data is in memory — everything in Redis is in memory. A falling Redis hit ratio means keys are being evicted, expiring, or were never written. See Redis Monitoring.

What a falling ratio means

• Data growth. The working set genuinely no longer fits.
• A new query pattern. A report scanning a large table evicts the hot pages every time it runs.
• A cache-unfriendly access pattern. Random access over a large key space, or a large sequential scan.
• A restart. A cold cache after a restart looks like a crisis and resolves itself as the cache warms. Distinguish the two before acting.

What to do

1. Confirm the workload changed before assuming the data did. A single new query can explain the whole shift.
2. Reduce the working set — archive, partition, drop unused indexes competing for cache.
3. Isolate scan-heavy work onto a replica so it does not evict the transactional working set.
4. Then consider more memory.