Core answer
Start here. This is the direct spoken answer to practice first.
Why this question matters
A brief failover or throttling event can resemble a release regression, while retries can temporarily hide a slow query, lock, or capacity problem. Error codes, affected query patterns, retry outcomes, deployment timing, and database waits separate the two.
I would look at error codes, timing, affected queries, deployment correlation, retry success rate, and database metrics. A small burst that succeeds on limited retry may be transient. A specific endpoint or query that keeps failing after a release suggests a regression, lock, missing index, or changed data access pattern. I would keep retries bounded with backoff and jitter so the API does not create a retry storm.