Right now switching over the current processing is mostly automatic - when the DC switchover would happen events would start flowing from new active DC and a change-prop instance in an active DC will start picking them up - no intervention needed here.
However, in a real DC failover if eqiad goes offline for real, we also want to reprocess the backlog that's been non processed in eqiad. We have mirrored topics in codfw, but there's no way to instruct change-prop to reprocess something. We need to find a way to do that.
Update 2017-03-15:
- Timestamp-based indexing is now available in Kafka 0.10.1 and librdkafka.
- As far as we are aware, the node bindings don't expose this functionality yet.
Next steps towards timestamp support:
- Upgrade to Kafka 0.10.1 (or 0.10.2)
- Add support for calling rd_kafka_offsets_for_times in node-rdkafka.