The current Redshift configuration uses eight ra3.4xlarge nodes with on-demand pricing and is paused outside an 8-hour daily window. The ingestion workload runs only 3 hours per day and has about 85% CPU utilization during ingestion, which indicates that the existing node size and count are adequate for performance. The main opportunity is cost optimization for a workload that is time-bound and does not require a continuously running provisioned cluster.

Amazon Redshift Serverless is designed for such intermittent or variable workloads. With Redshift Serverless, you configure compute capacity in terms of Redshift Processing Units (RPUs) and pay per use based on the RPU-hours consumed when queries and data ingestion jobs are actually running. When there is no activity, there is no compute charge. Redshift Serverless can also be created from an existing provisioned Redshift snapshot, allowing an easy migration path from a node-based cluster to a serverless endpoint without re-ingesting or manually moving data.

Option C creates a new Redshift Serverless endpoint with 64 RPUs from the most recent snapshot. This takes advantage of snapshots that already exist on the current cluster. The internal applications are then updated to read from the new serverless endpoint, and the existing cluster is deleted to stop incurring node-based charges. For a workload that runs only a few hours per day, moving to serverless compute is usually more cost-effective than maintaining a provisioned cluster, even if that cluster is paused outside the ingestion window, because serverless eliminates the need to manage cluster sizing and pausing and optimizes billing around actual usage.

Option A proposes staying on provisioned RA3 nodes and additionally purchasing 1-year Reserved Nodes. Reserved Nodes apply discount to node hours whether or not the cluster is paused or fully utilized, and they are designed to be cost-effective for steady, predictable, long-running workloads. For an 8-hour-per-day workload, especially with only 3 hours of intensive ingestion, Reserved Nodes are likely to be less cost-optimal than an on-demand, usage-based serverless model. Also, concurrency scaling is primarily for handling bursty query concurrency, not for cost reduction of a regular ingestion workload.

Option B replaces the existing eight ra3.4xlarge nodes with six ra3.16xlarge nodes and enables auto scaling. ra3.16xlarge provides significantly more capacity per node and is generally more expensive. The original workload already runs at acceptable utilization (about 85% during ingestion), so scaling up to larger nodes is not necessary for performance. Adding auto scaling on top of this likely increases costs and complexity without a clear benefit for a predictable, time-bound daily ingestion pattern.

Option D uses Redshift Spectrum and unloads data from the cluster to Amazon S3 for querying. Redshift Spectrum allows querying data stored in S3 using external tables from within a Redshift cluster, and is charged per amount of data scanned. However, the option still depends on maintaining a Redshift cluster and does not fundamentally optimize the compute cost of the ingestion workload. It also changes query patterns and data access architecture for the internal applications and may increase query cost if large volumes of S3 data are scanned by Spectrum.

Therefore, using Redshift Serverless created from the existing snapshot and paying only for compute usage during ingestion and queries, as described in option C, is the best way to optimize cost for this specific workload pattern.