NServiceBus provides error handling and has immediate retries and delayed retries.
When using the distributor the behavior of retries and the related message flow is different between major versions.
Immediate retries
Immediate retries are always performed on the worker.
Delayed retries
The behavior of delayed retries is different between major versions.
The delayed retries feature has two configurable items TimeIncrease and NumberOfRetries.
The message will be retried on any available worker. Message processing is not sticky to the worker.
The delayed retries policy NumberOfRetries setting is applied on both the distributor and workers, and the TimeIncrease setting is only applied on the distributor.
When an error occurs the delayed retries policy is invoked immediately by the fault manager. The message will not be forwarded to the retries queue which was the behavior prior to version 5.
When the retry limit is reached the message is forwarded immediately to the error queue or else forwarded to the .retries queue and scheduled for retry. If the delayed retries policy output is that it needs to be retried then the message is forwarded to the .retries queue.
Best practice
Due to the behavioral differences between major versions it is advised to have the SecondLevelRetriesConfig exactly the same on both the distributor and the workers even though it could be that settings are ignored.
It is assumed that NumberOfRetries and TimeIncrease are the same on the distributor and the workers.