Dual writes with asynchronous secondary cluster

[lukasz-antoniak]

ZDM forwards all requests to primary and secondary cluster in synchronous manner. Response returned to the client is calculated in a way that any failure (on any of the clusters) causes failure propagated to the driver. This behaviour is desired in majority of use-cases, because all idempotent requests will be automatically retried by the client application, and data automatically synced on both clusters.

Described behaviour may impact however the overall availability of Cassandra seen from client perspective. Depending on use-case, there might be occasions where application does not bother about single write failure, as same data will be frequently overridden with new value. Hereby ticket proposes to introduce a configuration parameter to allowing to send synchronous write to the primary cluster, but asynchronous to secondary.

Comments welcome.

┆Issue is synchronized with this Jira Task by Unito
┆Components: Proxy
┆Issue Number: ZDM-627

[Skunnyk]

Another use-case is during the setup of ZDM: By using an asynchronous write (and not only read), we can set up ZDM in the middle of the infrastructure, connect the applications/drivers to it, letting it write to the ORIGIN, and "maybe" write to the TARGET as a dry-run, before switching to a full synchronous mode for a migration