E-commerce analytics workloads have a shape that punishes the wrong pricing model. Traffic is spiky rather than steady, order and clickstream events arrive as nested JSON, and the number of people staring at dashboards jumps the moment a promotion goes live. Amazon Redshift, Google BigQuery and Snowflake can all store your orders and sessions. The difference that matters is what each one puts on the invoice when demand is uneven.
Three genuinely different billing models
These platforms are often described as interchangeable. Their meters say otherwise.
Amazon Redshift: you pay for compute time
Redshift Serverless bills in Redshift Processing Units. In the US East (N. Virginia) region the published rate is $0.375 per RPU-hour, and the smallest base capacity is 4 RPUs, which works out to roughly $1.50 per hour while a workload is actually running. Billing is per second with a 60-second minimum charge, and base capacity scales up from there. AWS also sells one- and three-year serverless reservations for teams with predictable baseline demand. Current rates are on the Amazon Redshift pricing page.
Capacity limits are tied to base RPUs: a 4-RPU configuration supports up to 32 TB of Redshift Managed Storage, while 8 or 16 RPUs support up to 128 TB. Storage is billed separately by GB per month, independent of compute.
Google BigQuery: you pay for bytes scanned, or for slots
BigQuery's default on-demand model charges per tebibyte of data scanned by each query rather than for time a cluster is awake. Google's documentation notes that each project gets a free tier of query processing every month, and long-term storage carries an automatic 50% discount once a table goes 90 days without modification.
Teams with steady query volume can switch to capacity pricing through BigQuery editions. The editions documentation sets out real functional differences, not just price tiers. Standard edition caps reservations at 1,600 slots, carries a 99.9% service level objective, offers autoscaling without a baseline allocation, and excludes BigQuery ML, continuous queries and managed disaster recovery. Enterprise raises the SLO to 99.99% and adds autoscaling with a guaranteed baseline, cross-user result caching, BigQuery ML and continuous queries. Enterprise Plus adds managed disaster recovery, Assured Workloads compliance controls, and export to Bigtable, Spanner or AlloyDB. All three support optional capacity commitments, with one-year and three-year terms discounted 20% and 40% respectively.
Snowflake: you pay for credits, per warehouse
Snowflake separates storage from compute entirely. Compute runs on virtual warehouses sized from X-Small to 6X-Large, and each step up doubles credit consumption: an X-Small consumes 1 credit per hour and a Large consumes 8 on first-generation warehouses. As the Snowflake warehouse documentation states, billing is per second with a 60-second minimum each time a warehouse starts. Auto-suspend and auto-resume are enabled by default, so idle warehouses stop burning credits without anyone intervening.
What a credit costs depends on your edition, cloud provider and region, so confirm the rate for your specific deployment rather than assuming a single global number. The edition choice also gates features: multi-cluster warehouses, the mechanism most relevant to concurrency, require Enterprise edition or above.
What happens during a flash sale
Concurrency is where retail workloads diverge from ordinary reporting. Each platform absorbs a spike differently.
- Snowflake scales horizontally with multi-cluster warehouses, adding clusters instead of enlarging one warehouse. Credits are calculated from both warehouse size and the number of running clusters. Separate warehouses can also isolate workloads, so an analyst's heavy query does not slow the merchandising dashboard.
- BigQuery allocates from a slot pool. On demand, scaling is handled for you with no capacity to manage. On editions, autoscaling with a baseline gives Enterprise customers guaranteed floor capacity plus headroom.
- Redshift Serverless scales above base RPU capacity automatically, and AWS bills base and scaled capacity at the same RPU rate. Worth knowing: capacity can stay elevated for a period after load falls, which is why AWS recommends setting maximum RPU hours in the console to cap unexpected spend.
Cost traps that catch retail teams
The billing mechanics carry practical consequences that rarely appear in feature comparisons.
On Redshift, AWS documents several behaviours that surprise teams. Cancelled queries are still billed for the time they ran. Unclosed transactions keep consuming RPUs, which is why session timeouts default to one hour of idle time and six hours for an open transaction. Most relevant to always-on storefront applications, connection pool health checks count as billable activity: routine SELECT 1 keep-alive queries trigger compute usage, so an open pool can generate charges when no real workload is running. On the credit side, Redshift Spectrum queries against S3 carry no separate charge beyond transaction time, and automatic optimizations run free unless you enable extra compute resources.
On BigQuery, the risk is a careless SELECT * against a large table. Google's cost best-practices guide recommends concrete guardrails: dry-run queries to preview bytes processed before execution, a maximum bytes billed setting that fails a query rather than charging for it, custom daily quotas per project or user, and partitioning and clustering to cut the volume scanned.
On Snowflake, the classic waste is a warehouse sized for the worst query of the month and left running between them. Tight auto-suspend settings and workload-specific warehouses address most of it.
Choosing for your store
If your stack already runs on AWS and orders live in Aurora or RDS, Redshift's zero-ETL integrations shorten the path from transactional database to warehouse considerably. AWS suggests tuning the refresh interval, since shorter intervals raise compute costs while intervals of five minutes or longer reduce charges for reporting that does not need instant freshness.
If your analytics centre on Google Analytics and Google Ads data, BigQuery removes an entire integration layer, and on-demand pricing suits teams whose query volume is genuinely intermittent. Once scanning becomes routine and predictable, editions with commitments usually cost less.
If you run multi-cloud, need to isolate finance, marketing and data science workloads from each other, or work heavily with semi-structured event data, Snowflake's per-warehouse model gives the cleanest separation and the most direct chargeback story.
The most reliable way to decide is to load one month of real order and clickstream data into two candidates, run your actual dashboard queries at realistic concurrency, and compare the bills. Every platform here looks affordable in a vendor calculator and reveals its true cost only under your own query patterns.