A retry is another request to a service that may already be struggling. That sounds obvious until the browser, API and SDK all helpfully decide to try again. If each layer makes three attempts in total, one original request can produce as many as 27 attempts at the bottom of the stack. The individual settings look modest. Their combined behaviour is rather less modest.
The AWS article on timeouts, retries and backoff with jitter explains how retries can amplify load and why spreading them out matters. I would start by finding which layers already retry, including the client library. Adding an application retry loop without checking the SDK can change the amount of work much more than the code suggests.
Then give the operation an overall time budget. A five-second deadline is not useful if every new attempt gets a fresh five seconds and there are several pauses in between. Count the initial attempt, cap subsequent attempts and account for the time already spent. Backoff and jitter help spread the load, but neither makes an unlimited retry loop sensible. The caller still needs a clear outcome when the budget is exhausted.
Decide which failures are worth another attempt and whether repeating the operation is safe. Invalid input is unlikely to improve after a short sleep. A temporary connection failure may be different, but writes still need to account for work that could already have completed. I would test a persistently failing dependency and count actual outbound calls, rather than only checking that the final error looks right. A retry policy should make a temporary problem easier to survive without turning a longer outage into a traffic generator.