Migrate to Credible

How to move business logic from an existing BI tool or semantic layer into a Malloy data model with the modeling agent, what carries over, which sources have a dedicated skill, and how to prove the numbers match, for whoever owns the migration, such as an analytics engineer, a BI developer or a data lead

What Migration Means in Credible

Migrating means rebuilding the business logic from another tool (tables, joins, dimensions, measures, descriptions, access rules) as a Malloy data model in a Credible package. The modeling agent does the drafting. It reads the exported definitions as reference material, proposes Malloy sources and fields, and checks its proposals against your data when a connection is available. You review the proposals, compare the numbers against the old tool, and publish.

The agent does not transpile line by line. It keeps the logic, skips what only exists for the old tool's interface or performance, and flags anything that needs a decision from you.

What Carries Over

Translated by the agentRe-created on purpose
Table references and grainPre-aggregation, extract and cache settings, as materialization chosen for the new model
Joins and their cardinalityRow-level security from Looker, and Power BI roles that don't fit access rules
Dimension and measure logic, including filtered and ratio measuresTemplating and tool-specific parameters, which the agent notes
Descriptions, which become #(doc) tags
Number formats, which become render tags such as # currency and # percent
Visibility decisions, as proposals for you to confirm
Row-level security from Power BI roles with simple filters, as access rules

Dashboards, drill paths and visual styling, usage history, schedules and alerts are not converted. After the model is published, rebuild the dashboards people still use as reports or data apps on top of it.

Supported Sources

The modeling agent migrates from any of these tools: it reads the exported definitions, maps them to Malloy and checks the numbers with you. For Looker and Power BI it also loads a dedicated skill that inventories the project, flags risky constructs and produces a coverage report. Both skills are open source in Malloy Publisher (opens in a new tab). Each guide below lists what to export, how the tool's concepts map to Malloy, and what to review.

SourceWhat the agent readsDedicated skill
LookerLookML project (.lkml files)malloy-lookml-review
Power BISemantic model as TMDL, a PBIP project, or a .pbixmalloy-powerbi-review
TableauPublished data sources and workbook XML
SigmaData model definitions and workbook formulas
OmniModel YAML (views, topics, relationships)
CubeData model YAML or JavaScript
dbt Semantic LayerSemantic model and metric YAML
Snowflake semantic viewsCREATE SEMANTIC VIEW DDL or Cortex Analyst YAML
Databricks metric viewsMetric view YAML

Using another tool? Export its definitions as text and give them to the agent. For a large migration, get in touch and we'll help plan it.

Prerequisites

  • A Credible organization with an environment, and a connection to the warehouse the old tool reads. The agent can draft a model without a connection, but it marks every proposal as unvalidated until it can query the data.
  • The old tool's definitions exported as text files. Each source guide says how.
  • A short list of the dashboards and numbers people rely on. These decide what is in scope and which numbers you check first.
  • For a whole project, a local modeling directory: Cred CLI sets one up with cred init, including the modeling skills and the connection to your environment. In the Credible App you can paste smaller pieces, such as a single view or a handful of measures, into the chat.

Steps

  1. Export the definitions. Follow the export section of your source's guide. Put the files in a subfolder of your modeling directory, for example lookml/ or export/.
  2. Keep the export out of the package. Add the subfolder to .credignore (one path or glob per line, next to publisher.json) so cred publish leaves it out.
  3. Ask the agent to build the model. Name the folder and the domain, for example "Build a model of our orders domain using the definitions in export/ as prior art." The agent inventories the files, confirms it should use them, explores the tables, and proposes sources, joins and fields with the reasoning behind each.
  4. Review the proposals. Confirm, rename or drop fields. Decide every item the agent flags, such as a join whose cardinality it could not confirm or a filter that might be a security rule.
  5. Compare the numbers. Run the key measures in both tools with the same filters and compare them. See Prove Parity.
  6. Publish. Publish the package as described in Publishing and Versions. Then rebuild the reports people need on top of it.

Scope the Migration

A model that has run for years exposes far more fields than anyone queries. Start from the dashboards that matter and work outward:

  1. Fields the tiles use: every field in a tile's selections, filters and sorts.
  2. Calculations defined on the tiles: measures written in the report rather than in the model.
  3. Fields people query directly: anything used outside dashboards, plus whatever those fields depend on.

Leave the rest out, or mark it deferred so it can be added later. Where the old tool records its own usage (Looker's System Activity is one example), use that history to decide. The Looker Migration Guide describes one customer migration where a certified dashboard of 28 tiles over three explores exposed roughly 6,500 fields, and 377 of them made it into the rebuilt model.

Organize the result by subject area rather than by dashboard: one source per table, then sources that join them for each domain, then named views for the recurring questions.

Prove Parity

Matching definitions is not the same as matching numbers. Before you switch anyone over:

  • Pick the numbers that matter. Start with the measures on the dashboards people open every day, not the easy ones.
  • Compare with more than one filter. Some differences only appear when a filter is active, so a match at the grand total is not enough.
  • Run both sides against the same data. Either get the value from the old tool, or run the equivalent SQL directly against the same warehouse and compare it with the Malloy result. When both sides read the same tables, a difference is a difference in logic.
  • Record the result. Keep a table of measure, filter, old value, Malloy value, and whether they match. Investigate every mismatch; sometimes the old tool is the one that is wrong.

The Looker and Power BI skills also produce a coverage report listing every source and field as modeled, renamed, deferred or skipped, with a reason. For other sources, ask the agent for the same report.

Next Steps

On this page