append_entries handling of multiple log records as batch

[kishorekrd]

end_of_append_batch() is called for disk flush after processing all the log records passed to append_entries(). I see that append_entries takes vector of log records. Is this operation atomic, as all or none? If I submit 10 records to it, what happens if 5 log records get appended and then 6th one fails? How is this handled?

[greensky00]

From NuRaft's point of view, appending multiple logs doesn't need to be atomic as long as succeeded logs are consecutive.

"All or nothing" is up to you; if you want to make 10 logs atomic, you should make append_entries fail on the 6th log failure. Otherwise, it is OK for NuRaft to move forward with the succeeded 5 logs.

[kishorekrd]

Thanks for the response. Irrespective of the number of logs submitted to append_entries, handle_result() is called once. If partial number of logs append fails, How to handle it? What is the error that handle_result() gets in this case?

[greensky00]

If failure is detected so that you want to make the call fail, you can return ReturnNull in the callback AppendLogs. https://github.com/eBay/NuRaft/blob/7ce0ef89775c0976c98381b3d5eda93cea6dca37/src/handle_client_request.cxx#L146-L147 then it will make append_entries call fail with BAD_REQUEST code.

[kishorekrd]

In case of raft_params::async_handler, // Async mode: // append_entries returns immediately. // handle_result will be invoked asynchronously, // after getting a consensus.

While appending 10 (multiple) logs, after appending 5 logs successfully, if leader change happens, remaining 5 logs will fail. How do the caller come to know this failure, so that he can either

1. make first 5 logs also as failed and retry all again.
2. retry the failed 5 logs again.

I see that caller comes to know result with handle_result(), which shows whether it succeeded or failed? https://github.com/eBay/NuRaft/blob/7ce0ef89775c0976c98381b3d5eda93cea6dca37/examples/calculator/calc_server.cxx#L47

Is there any other way to find this and return ReturnNull in the callback AppendLogs as you mentioned?

[greensky00]

• If the leader changes before committing 5 appended logs, those logs are invalid and will be rolled back by the new leader.
• YOU are the one who implements log store, so you are responsible for detecting such errors, not NuRaft. So any error happens in YOUR LOG STORE, you should remember it, and in YOUR callback function for AppendLogs, you should return ReturnNull. Then handle_result will check it by result.get_result_code() != cmd_result_code::BAD_REQUEST.

[kishorekrd]

OK that means, out of 10 logs, if leader change happens after appending 5 logs, logs 6 to 10 get the leader change error in the handle_result(), and also the new leader will rollback logs 1 to 5. Is this correct? If so, All I need to do is tell the caller that logs 1 to 10 failed and it can retry.

AppendLogs callback is for having custom checks on the appended logs.

Thanks for the help.

[greensky00]

OK that means, out of 10 logs, if leader change happens after appending 5 logs, logs 6 to 10 get the leader change error in the handle_result(), and also the new leader will rollback logs 1 to 5. Is this correct?

Yes, that's correct.

[kishorekrd]

Thanks a lot for your help