Latency belongs to a request path
Capture the actual route/workload, percentile latency, concurrency and response behavior. A machine load average alone does not identify the slow boundary.
Find expensive work before adding cache
Inspect query frequency, examined rows, locks and indexes. Cache only when it has a clear invalidation model and does not hide correctness problems.
Bound external latency
Use explicit connect/read timeouts and decide what can safely become asynchronous. Provider slowness should not indefinitely consume application workers.
Compare before and after using the same workload
Keep evidence for the baseline and result. If correctness, error rate or reconciliation degrades, a faster median is not a successful optimization.