We've started down two alternative reimplementations of ORES, and I suspect that they might be mutually exclusive. One is to keep ORES as a standalone service, but to implement worker nodes for k8 deployment. The other is to replace ORES with a pure cache-backed service, calculating new scores in a streaming paradigm.
It's possible that backfilling old scores could be done by k8 nodes, making the plans compatible, but I think there may be better options such as distributing the Python workload over the Hadoop cluster and rewriting other features such as our training.
We don't have spare resources for contradictory implementations, so should decide between these approaches soon, or find a way to reconcile.