Redshift Cost Optimisation

Understanding Redshift pricing

Redshift pricing has two components: compute (the cluster nodes) and storage (the data).

Provisioned cluster pricing is per-node-hour. Common node types:

• dc2.large: $0.25/node-hour — best for small clusters or testing
• ra3.xlplus: $1.086/node-hour — general purpose, separates compute from storage
• ra3.4xlarge: $3.26/node-hour — medium workloads
• ra3.16xlarge: $13.04/node-hour — large enterprise workloads

A 2-node ra3.xlplus cluster running 24/7: 2 × $1.086 × 730 hours = $1,585/month

Redshift Serverless charges per RPU-second (Redshift Processing Unit). The base price is approximately $0.375 per RPU-hour. Serverless automatically scales RPUs from a configured minimum to a maximum.

For a team running queries 4 hours/day with an average of 8 RPUs: 8 RPUs × 4 hr × 30 days × $0.375 = $360/month

Versus the provisioned cluster at $1,585/month for the same workload. Serverless is dramatically cheaper for low-utilisation workloads.

The break-even is roughly when your cluster is running meaningful query load for more than 8–10 hours per day, at which point provisioned with Reserved Nodes becomes cheaper.

Choosing between Serverless and provisioned

For development and analytics teams with variable query schedules, Serverless is the default recommendation. Migrate to provisioned only once you have 90 days of Serverless cost data showing consistent high utilisation.

Pausing provisioned clusters during off-hours

If you have a provisioned cluster and your team only runs queries during business hours, pausing the cluster at night and on weekends can cut the compute bill by 65–70%.

A ra3.xlplus 2-node cluster running 8 hours/day × 5 days/week: 2 × $1.086 × (8hr × 22 workdays) = 2 × $1.086 × 176 = $382/month

Versus running 24/7: $1,585/month. Saving: $1,203/month.

Storage charges continue during the pause (ra3 nodes use S3 for managed storage, charged at $0.024/GB/month), but compute stops.

To schedule pausing with the AWS CLI:

# Pause the cluster
aws redshift pause-cluster --cluster-identifier my-cluster

# Resume the cluster
aws redshift resume-cluster --cluster-identifier my-cluster

Use an EventBridge scheduled rule with a Lambda function (or AWS Systems Manager Automation) to pause and resume automatically at specified times. For example, pause at 7pm weekdays and resume at 7am weekdays.

Reserved Nodes for consistent production workloads

For provisioned clusters running 8+ hours/day consistently, Reserved Nodes provide significant savings:

• 1-year Reserved: approximately 40% discount over On-Demand
• 3-year Reserved: approximately 75% discount over On-Demand

For a 2-node ra3.xlplus cluster:

• On-Demand: $1,585/month
• 1-year Reserved: ~$950/month
• 3-year Reserved: ~$400/month

The commitment is per-node, not per-cluster. You buy 2 reserved nodes of type ra3.xlplus, and the reservation applies to any cluster using those node types in that region.

Before purchasing Reserved Nodes, run the cluster for at least 60–90 days On-Demand to confirm the node count and type are appropriate. Changing node type after buying reserved nodes is not straightforward.

For Redshift pricing in detail, see the Redshift pricing explained page.

Redshift Spectrum: query S3 instead of loading data

Redshift Spectrum extends Redshift queries to data stored in S3 without loading it into the cluster. This is significant for cost because:

• S3 storage costs $0.023/GB/month
• Redshift managed storage (ra3) costs $0.024/GB/month
• But data in Redshift must be queried within Redshift (compute cost); data in S3 is accessed via Spectrum at $5/TB scanned

The cost-saving use case: historical data queried infrequently. If you have 5TB of transaction records from 2022 that are queried once a month, keeping them in Redshift costs: 5,000GB × $0.024 = $120/month in storage, plus the compute cost of having a cluster large enough to hold it.

Moving that 5TB to S3 (as Parquet or ORC) costs: 5,000GB × $0.023 = $115/month in storage, plus $5 per TB scanned when queried. Monthly query cost: ~$25 (assuming 5TB scanned in the monthly query).

Net saving: minimal on storage, but the cluster can be smaller (fewer nodes) because it doesn’t need to hold 5TB of cold data, which saves significantly on compute.

Partitioning for Spectrum cost reduction — Spectrum charges for data scanned. If your S3 data is partitioned by date (e.g., s3://bucket/data/year=2022/month=03/), a query filtering to WHERE year = 2022 AND month = 03 only scans that partition — dramatically reducing the per-query cost.

Data compression and storage reduction

Redshift uses columnar storage with compression encodings to reduce the physical size of data on disk. Efficient compression reduces storage costs and improves query performance (less data to read = faster queries).

Redshift automatically applies compression encoding when you load data using COPY. You can also run ANALYZE COMPRESSION on a table to see what encoding Redshift recommends:

ANALYZE COMPRESSION my_table;

Common encoding types:

• RAW — no encoding, largest storage footprint
• AZ64 — AWS proprietary, works well for numeric and date columns
• ZSTD — general-purpose compression, often best for string columns
• Runlength — efficient for low-cardinality columns (e.g., status codes, boolean flags)

For a typical analytics dataset, applying proper compression reduces storage by 50–70%. On an ra3 cluster with 2TB of data, that’s the difference between $48/month and $14–24/month in storage.

Materialized views to reduce repeated query cost

Some queries run frequently and are expensive — full scans of large tables, complex joins, aggregations. Materialized views pre-compute the result and store it, so subsequent queries hit the cached result instead of re-scanning the underlying data.

CREATE MATERIALIZED VIEW monthly_revenue_summary AS
SELECT
    date_trunc('month', order_date) AS month,
    product_category,
    SUM(revenue) AS total_revenue
FROM orders
GROUP BY 1, 2;

Refreshing the materialized view on a schedule (e.g., nightly) means dashboards hit the pre-computed summary instead of scanning the orders table. For a 500GB orders table queried 100 times per day by a BI tool, materialized views can reduce the scan volume from 50TB/day to near zero.

This reduces both query execution time and WLM (Workload Manager) queue pressure, which can allow a smaller node count — directly reducing compute cost.