Most of the people who describe a data warehouse in these interviews are marketing, ecommerce and analytics leaders who depend on it, and many of them did not choose it. Snowflake, Databricks and BigQuery all rate well above what buyers give software generally. The regrets are about operating it: who decided, what compute costs as use grows, who can reach the data, and how long a move takes. For the tools that sit on the warehouse, see the Snowflake stack; for the two leading platforms side by side, Databricks vs. Snowflake.
What do data warehouse buyers wish they'd known?
Six things recur. Many buyers inherited the warehouse from a data team, IT or a parent company, and with it the constraints of a choice they did not make. Compute is metered, and buyers describe rationing queries once the bill grows. Low ratings are mostly about reach: the data is there, but getting it out takes SQL or an analyst. Several low raters blame their own setup as much as the product. AI assistants are starting to open the warehouse to people without SQL, but what they answer depends on the data, permissions and definitions they are given. And a warehouse migration runs long enough to hold up other purchases.
The six lessons
You may have inherited it — find out who chose it, and why
Asked why their warehouse was selected, buyers in these interviews say it predated them or belonged to another team about twice as often as buyers answering the same question about an email platform, CRM or BI tool. That is a difference in how buyers answer, not a measure of how often companies hand the decision to another team. The decision sat with a data or engineering team, with IT, with corporate, and in one venture-backed DTC brand with an investor that steers the companies it backs onto the same stack. A digital leader at a global hotel group says of Redshift that "there wasn't an active choice. It's just where the data sits." Where buyers do know the reason, they give a handful: bringing many sources into one place, handling large loads on time, leaving an older and more expensive warehouse, a consumption price that suited a startup, and being the standard choice.
The implication — ours, not a buyer's — is that a marketing team buying anything that reads from the warehouse should ask the owning team what the warehouse was bought to do, and what it will be asked to do next, before the shortlist forms. For how warehouse integration then decides other purchases, see the Snowflake stack.
Named in this context: Snowflake · Redshift · BigQuery · Databricks
The meter runs on use, and use is what you bought it for
Buyers describe consumption pricing from both ends. A pet-care startup chose Snowflake partly because "The cost, being consumption-based, especially as we're a startup, was really appealing," and still praises how much visibility it gets into consumption. Further along, the same property becomes a constraint. An online marketplace that rates Snowflake an 8 says "Every time we query the data, we're getting charged for that," and so "we try to limit how often we query the database." That is how the buyer experiences the bill rather than Snowflake's literal unit: a standard virtual warehouse consumes credits by its size and how long it runs, not a fixed charge per query, but more frequent or heavier queries keep compute running longer or need more of it, which is where the pressure to ration comes from. A specialty apparel retailer warns that "costs can run up quickly without proper governance on top of things." Buyers of Databricks and BigQuery raise cost too, if less pointedly: one calls Databricks "very costly", another says BigQuery "is not necessarily cheap".
Each vendor meters compute in its own unit (see what each meter counts, below), so a quote from one is not comparable with a quote from another until both are priced against the same workload.
Named in this context: Snowflake · Databricks · BigQuery
A low score is usually about reaching the data, not the warehouse
When buyers explain a low warehouse rating, the most common reason is that they cannot get to the data without someone else. A health insurer rates BigQuery a 6 because "in order to pull any data out of it, you do have to know how to write a SQL query, which is currently a barrier," while granting in the same answer that it is a good tool that likely has all the information the team needs. A marketing-operations head at a large online retailer says the same of BigQuery: "you always need a layer of analysts behind it." At a media company, the demand-generation lead rating Snowflake does not work in it directly: "we don't even have access to it. Our data teams do."
What these buyers ask for is a layer on top: a BI tool, an activation tool, or increasingly an AI assistant (lesson 05). That is a gap in access for the people rating it, which is a different thing from a gap in what the platform can do.
Named in this context: BigQuery · Snowflake · Databricks
Before blaming the warehouse, check the build
Several buyers who rate their warehouse low hold their own setup partly responsible, without being asked. An online luxury retailer rates Snowflake a 6 "only because we're not using it correctly." A large retail chain rates Databricks "a 6, maybe even a 5," and settles on the 6 "because it's probably partially us." At a healthcare services company, the marketing analytics director says of Redshift, "I don't think it's the tool itself. I think it's just the way in which our data has been structured over time," after acquisitions left the data siloed. An engineering lead at a large healthcare company rates BigQuery a 5 or 6 and wonders whether "it is internally how we have set it up with long data set names and project names and so many permissions" that a simple request takes several times longer than it should.
These are the buyers' own readings, and the interviews cannot separate product from implementation for them. The implication — ours, not a buyer's — is that a structure problem travels with the data to the next platform, so audit naming, permissions and table structure before concluding the warehouse is what needs replacing.
Named in this context: Snowflake · BigQuery · Databricks · Redshift
The AI layer depends on what it can access and understand
The newest material in these interviews is AI that lets people ask questions of warehouse data in plain language: Databricks Genie, Snowflake Intelligence (which Snowflake renamed Snowflake CoWork in June 2026), and outside tools such as Claude or ChatGPT connected to the warehouse. Snowflake Cortex AI is the company's broader suite of AI products, not another name for the assistant. For some buyers it reduces the need for SQL, though what an assistant answers depends on the data it can reach, its permissions and the business definitions it can use. A large insurer says Databricks needs technical skills not everyone who touches the data has, but "the Genie functionality, is lowering the bar"; a marketer at a software company now queries Databricks "without having to learn SQL or any of those languages."
Others find the built-in assistant the weak part. The online marketplace that rationed its queries says "Their built-in agent isn't very good" and connects Claude to Snowflake instead, and the specialty apparel retailer finds "ChatGPT is a lot better at writing SQL than Snowflake." One of the lowest scores in the set comes from an airline analyst who rates Snowflake Intelligence a 2 because only "one or two schemas" had been loaded into it, and who is not sure "if it is the tool itself that is lacking or how we ingested the data." That score describes the airline's limited deployment rather than a test of what the product can reach: schema coverage, permissions, business definitions and configuration all shape what an assistant answers correctly. The pet-care startup names the work underneath: getting an assistant to use the company's own terms needs "a really robust semantic layer". Databricks' own product page sets similar conditions for Genie: the data must already be managed in Unity Catalog, and subject-matter experts add example queries and instructions. These are individual deployments, not a ranking of the assistants.
Named in this context: Databricks Genie · Snowflake Intelligence · Snowflake Cortex · Claude · ChatGPT
A warehouse migration sets the calendar for the marketing stack
Many of the moves between warehouses in these interviews leave older platforms, such as Teradata, Redshift and Azure Synapse. Snowflake is the most common destination, with Databricks and BigQuery also named, and some moves run between the newer platforms in both directions. Some are complete, others are still under way, and a few are being evaluated. The ones under way take long enough to schedule other work around. A healthcare services company moving from Redshift to Snowflake is going step by step: "Certain campaigns have already been migrated." An apparel retailer expects Redshift to stay in use "for the next maybe six to nine months" while it moves to Databricks.
Marketing waits on it. An online marketplace part-way through a move from BigQuery to Snowflake is keeping its CMS for now, because the product team lacks bandwidth for a second migration "given that we are already amidst a migration from BigQuery to Snowflake." A membership-services company finished its move onto Snowflake ahead of an email-platform switch, so the new platform's data share was ready before the old one was switched off. The implication — ours, not a buyer's — is to put any planned warehouse move on the marketing calendar first. For which workloads move between Snowflake and Databricks, see Databricks vs. Snowflake.
Named in this context: Snowflake · Redshift · Databricks · BigQuery · Teradata · Azure Synapse
How buyers rate the platforms
Every figure is the average across verified buyer interviews, on published transcripts. All three well-sampled platforms sit close together and well above what buyers give software generally, so the spread between them is small next to the spread in how buyers experience any one of them.
| Platform | Rating | What buyers describe |
|---|---|---|
| 8.3 | The most-rated warehouse here. Praised for bringing sources together and for reach; the recurring complaint is the consumption bill. | |
| 8.1 | Flexibility and scale for data science and pipelines. Its lower raters are non-technical users who need everything built for them; some credit Genie with closing that gap. | |
| 8.0 | Scale and a natural fit with Google's tools. Lower raters cite the SQL barrier and integrating with systems outside Google. | |
 Redshift |
around 7 | Too few buyers for a decimal. Described as a standard tool and as where the data sits; several companies describe moving off it or evaluating alternatives, and one chose it over Snowflake on its own cost and performance tests. |
| not rated | Too few buyers rated it to publish a number. Named by companies building or consolidating data on Microsoft's stack, including one planning to move off BigQuery onto it. | |
| not rated | Too few buyers rated it to publish a number. Appears mostly as the older warehouse companies are moving off. |
What each warehouse's meter counts
The units below are not interchangeable: a credit, a DBU, a terabyte scanned and a capacity unit each count something different, so a price per unit says little until it is multiplied by your own workload.
| Platform | What compute is billed in | What it counts | Storage |
|---|---|---|---|
| Snowflake | Credits | Standard virtual warehouses consume credits by size, number of clusters and running time, billed per second after a one-minute minimum each time a warehouse starts or resumes. Other compute, such as serverless features and container compute pools, follows its own billing rules. Credits are bought on demand or as pre-paid capacity. | A separate monthly fee based on average storage used. |
| Databricks | DBUs | A Databricks Unit is a normalized unit of processing power, metered per second on pay-as-you-go. How many a workload consumes depends on the compute and data it processes, and the price per DBU varies by product and cloud. | Databricks says storage, networking and related costs vary with the services and cloud provider chosen, and cloud infrastructure prices vary by region and provider. |
| BigQuery | Bytes processed, or slots | On-demand pricing charges for the bytes each query processes; capacity pricing charges for slots, in slot-hours, through reservations in three editions, with optional commitments. | Priced separately from compute. |
| Amazon Redshift | Node-hours, or RPU-hours | Provisioned clusters bill per node for as long as the cluster runs; Serverless bills RPU-hours per second, with a 60-second minimum. | Redshift Managed Storage, billed per GB-month regardless of node count. |
| Microsoft Fabric | Capacity units | A shared pool of capacity used across Fabric's workloads, pay-as-you-go or reserved for one or three years. | OneLake storage, priced separately. |
Want this checked against your own stack?
Start my checkup →The stories behind the lessons
The counter-current: the warehouse is rarely the part buyers want to replace
For all the complaints above, few buyers in these interviews describe the warehouse as the problem to solve. The complaints that recur are about the bill, about who can reach the data, and about the setup the company built, and buyers generally rate the platform itself well. Many of the moves that do happen are away from older platforms. Our reading, not a buyer's: a team unhappy with what it gets from the warehouse should look at access, governance and the tools on top before looking for a new warehouse.
What this means for your warehouse decisions
Most readers of this page will not be choosing a warehouse but living with one, so the checks below are about running it as much as buying it. Our recommendations, not practices buyers describe. Before a purchase that reads from the warehouse, find out who owns it, what it was bought for and whether a migration is planned, and schedule around that. Before opening it to more users, tools or AI assistants, decide who can run what, and price the busiest month rather than the average one. When comparing warehouses, run two or three representative workloads through each vendor's pricing model rather than comparing unit prices, counting compute size and running time, concurrency, storage, data movement, any cloud infrastructure charges and the people or automation managing consumption, at both routine and peak demand. Test access with the people who will actually use the data: have a marketer answer a real question, first through the BI or activation layer and then through the assistant, against the full set of tables they need. And before replacing the warehouse over a low internal rating, audit naming, permissions and structure, because those move with the data.
For the tools buyers run on the warehouse, see the Snowflake stack; for the BI layer on top of it, what BI buyers wish they'd known; for whether the warehouse can replace a customer data platform, what CDP buyers wish they'd known; and for the two leading platforms side by side, Databricks vs. Snowflake.
What these interviews do not show
The ratings measure how satisfied each platform's own users are, not a head-to-head test, and most raters here are marketing and analytics leaders rather than the engineers who run the platform. The comparison of who chose the warehouse shows how often buyers say the decision was made elsewhere, not how warehouses are bought, and it cannot say whether the original decision was a good one. The moves described are a set of individual company accounts, not a migration rate or a market share. The AI accounts come from early deployments and describe individual setups, so they do not show which assistant works best. Pricing units come from vendor pages and can change; nothing here is a price.
Common questions
Which data warehouse do buyers rate highest?
In Alium's verified interviews Snowflake is the most-rated warehouse and the highest-rated in this set at 8.3 out of 10, followed by Databricks at 8.1 and BigQuery at 8.0. Amazon Redshift rates around 7 on a smaller sample, and Microsoft Fabric and Teradata have too few ratings to publish a number. All three leaders sit well above what buyers give software generally. A rating reflects how satisfied each platform's own users are, not a head-to-head test.
How do Snowflake, Databricks, BigQuery and Redshift charge for compute?
Each uses its own unit, so quotes are not directly comparable. Snowflake's standard virtual warehouses consume credits by size, cluster count and running time, per second after a one-minute minimum, rather than a charge per query. Databricks meters DBUs, its normalized unit of processing power, per second, and how many a workload uses and what each costs depend on the product and compute chosen. BigQuery charges either for bytes processed by queries or for reserved slot capacity. Redshift bills node-hours on provisioned clusters or RPU-hours on Serverless, with managed storage billed separately. Microsoft Fabric sells a shared pool of capacity units. Buyers describe the consumption models as a reason they chose a platform and as a risk once more people query it.
Why do some marketers rate their data warehouse low?
When buyers explain a low warehouse rating, the most common reason is that they cannot reach the data without SQL or an analyst. A health insurer calls the SQL requirement a barrier, and a large online retailer says it always needs a layer of analysts behind the warehouse. Several low raters also hold their own setup partly responsible: data structure, naming and permissions rather than the product.
Can AI let marketers query a data warehouse without SQL?
For some buyers it already does. Buyers describe Databricks Genie lowering the bar for people without SQL, and others connect Claude or ChatGPT to the warehouse to write their queries. The accounts are mixed: some find the warehouse's built-in assistant weaker than an outside model, and one low score came from an assistant loaded with only part of the data. Buyers describe the groundwork as loading the right tables and defining the company's own terms in a semantic layer, and what an assistant answers correctly depends on the data it can access, its permissions and the business definitions it is given.
Does a data warehouse replace a CDP?
Sometimes it replaces part of one. Some companies in these interviews run customer data directly from the warehouse with an activation tool on top, while many still run a packaged CDP beside it. The gap buyers name is usually activation and identity rather than storage. Alium's pages on CDP buyers and on the Snowflake stack cover that decision in detail.
Get a checkup on your tech stack
A checkup for your marketing stack, built from hundreds of verified buyer interviews: what matters most to you, a shortlist for your stack, and your next steps.
Start my checkupFree · no password · your answers are anonymized before they inform research like this.