You must deploy a new streaming Dataflow pipeline version (reads Pub/Sub, writes BigQuery) without losing data, avoiding inconsistencies, and keeping added latency under 10 minutes. What deployment sequence do you follow?

Draining an old pipeline ensures all in-flight data is processed and written to the sink (BigQuery in this case) before the pipeline stops. This prevents data loss and maintains consistency. Once drained, the new pipeline can be started, picking up from where the old one left off, minimizing downtime and latency. Updating the old pipeline directly can lead to inconsistencies if the new code has breaking changes or alters state. Snapshotting and then stopping/starting from a snapshot is not applicable for updating a streaming pipeline with new code; snapshots are for recovery or creating new pipelines from a specific state. Canceling the old pipeline immediately stops processing, leading to data loss and inconsistencies.