Credible vs. Snowflake Cortex

A semantic view describes physical objects a pipeline already built. That is the inherited order of the last decade — land the raw tables, transform them layer by layer, and write down what any of it means at the end, if the sprint allows. It was a rational answer when dashboards were the consumer and a person stood between the model and the decision. Agents removed that person. Credible flips the order: the data model is the input, written once in Malloy — entities, metrics, relationships and their cardinality, the business rules and the edge cases — and the engine derives the transformations, the materialized tables, the access enforcement and the context every agent reads. Meaning stops being documentation of the pipeline and becomes the specification for it.

Snowflake is a good answer for the central data team, and that is who a warehouse is built for: centralized, curated, governed by specialists. The work has been moving the other way. Coding agents have put data work in the hands of domain teams — ops, finance, product — who own questions the central backlog was never going to reach. A warehouse is the wrong shape for that: it asks the business domain to land its data centrally first, then wait for someone to model it. Credible inverts the order. The domain writes the data model against the data it already has, and the engine builds and serves everything under it. Centralize the definitions and the access policy; decentralize the rest.

A central warehouse holds what someone decided to load into it. The business domain runs on more than that — the Postgres behind the product, a spreadsheet finance maintains, Parquet sitting in object storage, the second warehouse that arrived with an acquisition. Cortex reaches what Snowflake can reach, which Iceberg and external tables have widened considerably, but the pattern still begins with getting the data in. Credible connects to BigQuery, Snowflake, Postgres, MySQL, Databricks, Trino, DuckDB and MotherDuck, reads CSV, Parquet and JSON in place from S3, GCS or Azure Data Lake, and takes spreadsheets shipped inside a package. One data model spans all of it, and none of it requires a warehouse to start.

Warehouse economics were set when humans asked the questions. Agents ask far more of them, in bursts, and every one is a scan. A warehouse bills a minimum window every time it resumes, so a trickle of small agent queries against an idle one pays that minimum again and again — and Cortex adds AI credits on top, additive across every tool an agent calls. Credible attacks both bills. One annotation materializes a source into the engine’s own storage, another makes a dimension’s values searchable, a third answers coarse questions from a rollup — so queries stop re-scanning the warehouse. And retrieval returns only the slice of the model a question needs, rather than carrying the whole thing in every prompt.

A dashboard tolerates a slow query. A product feature does not. A warehouse gives you variable latency by design: queries queue once concurrency runs out, and a warehouse that has gone idle has to resume before it answers — precisely the traffic shape an embedded feature has. Snowflake’s answer is interactive warehouses, which are genuinely fast and built for exactly this shape of traffic. Credible’s is to serve from materialized storage the engine controls rather than from a warehouse scan, so the latency your customers see does not track your warehouse’s mood. What comes back matters too: results are served over MCP carrying an interactive UI resource, so a chart renders in the client rather than being described to it — where Snowflake’s managed MCP server, real and general as it is, documents no support for MCP resources, prompts or notifications.

Snowflake has built a real lifecycle story, and it deserves saying plainly: Workspaces is a Git-backed IDE, dbt Projects on Snowflake is generally available, a dbt package manages semantic views as models in your repository, and the CLI validates a pull request in an isolated environment before it merges. What it does not have is one unit of release. The transforms version in the dbt project, the semantic view in its own package, the agent is configured apart from both, and the app ships somewhere else again — each with its own cycle, and nothing that makes them true of each other. In Credible the model, its metadata, its access rules, its data apps and its embedded data are one package, published as an immutable version, promoted only once its indexes and materialized tables are actually built, and rolled back by pointing at the previous version.

A semantic view is an object inside a Snowflake account. An agent built in Cortex is configured there, reasons with Snowflake’s orchestration, and bills in Snowflake credits. That is coherent, and worth it if the whole estate is Snowflake and intends to stay. The question worth asking of any vendor here, ourselves included, is what you would still have if you stopped paying tomorrow. Credible’s answer is a Malloy model: an open-source language, in your own repository, served by Malloy Publisher, which we build and maintain in the open, and compiled against seven dialects. Snowflake started Open Semantic Interchange, now incubating as Apache Ossie — a real commitment to portability and worth watching, though no product ships native support for the spec yet.