Missing fallback mechanism for replica selection #736
Description
Bug Report
1. Describe the bug
When there are multiple replica candidates, and the first one has already failed to respond, the caller does not try the next one on the list.
2. Minimal reproduce step (Required)
- Use leader and follower replica selector
- Use follower read
- Kill the store that hosts the target follower before sending the request
3. What did you see instead (Required)
As stated in the bug description.
4. What did you expect to see? (Required)
Try the next replica candidate
5. What are your Java Client and TiKV versions? (Required)
- Client Java: 6.5
- TiKV: 6.5
Ideally, it is expected to see that a request trys to connect to the first replica candidate on the list, retry for several times or a certain period of time on the same target when it is unavailable (considering transient error and rate limiting are quite normal on the TiKV side), then fallback to the next one on the list if the first candidate is still unavailable.