You are operating a Dataflow streaming job to ingest messages from a message bus, which inherently does not support exactly-once delivery semantics. The data undergoes various transformations in the process before being loaded into BigQuery. Given the critical requirement that your data must be streamed into BigQuery with exactly-once delivery semantics, and considering that you anticipate an ingestion throughput rate of approximately 1.5 GB per second, what steps should you take to achieve this?