On Wed, Jun 22, 2022 at 08:29:05AM -0700, Quanah Gibson-Mount wrote:
Ignoring the loadbalancer issues, I think you add a race condition when reading possibly older data from your consumers and maybe write them back where newer data may exist already (i.e.: providers). BTW: Is a modify operation a read, or is it a write?
modify ops are always a write operation, since they are doing a modification.
And yes, read-after-write can be a tricky issue to handle.
In "load-balanced" setups, I usually suggest that people maintain two separate endpoints, one for clients that only ever issue reads and the other one for clients that might issue a write operation (potentially mixed in with read traffic).
Regards,