Working With Dbt Mesh

Working with dbt Mesh

Core principle: In a mesh project, upstream data comes through ref(), not source(). Every cross-project reference requires the project name. When in doubt, read dependencies.yml first.

When to Use

• Working in a dbt project that references models from other dbt projects
• Resolving ambiguity when multiple upstream projects have similarly-named models (e.g. multiple stg_ models)
• Adding model contracts, access modifiers, groups, or versioning
• Setting up cross-project references with dependencies.yml
• Splitting a monolithic dbt project into multiple mesh projects

Do NOT use for:

• General model building or debugging (use the using-dbt-for-analytics-engineering skill)
• Unit testing models (use the adding-dbt-unit-test skill)
• Semantic layer work (use the building-dbt-semantic-layer skill)

First: Orient Yourself in a Multi-Project Setup

Before writing or modifying any SQL in a project that uses dbt Mesh, follow these steps:

1. Read dependencies.yml

This file at the project root tells you which upstream projects exist:

# dependencies.yml
projects:
  - name: core_platform
  - name: marketing_platform

If this file has a projects: key, you are in a multi-project mesh setup. Every model you reference from those upstream projects must use cross-project ref().

2. Understand how upstream data gets into this project

In a mesh setup, upstream project models replace what would alternatively be sources:

The upstream project has already staged and transformed the raw data. Your project builds on top of their public models, not their raw sources.

3. Disambiguate similarly-named models

When multiple upstream projects have models with the same name (e.g. stg_customers in both core_platform and marketing_platform), you must use the two-argument ref():

-- Correct: explicit project name, no ambiguity
select * from {{ ref('core_platform', 'stg_customers') }}
select * from {{ ref('marketing_platform', 'stg_customers') }}

-- WRONG: dbt cannot determine which project's stg_customers you mean
select * from {{ ref('stg_customers') }}

4. Check existing patterns in the codebase

Before writing new SQL:

• Search for existing two-argument ref() calls to see which upstream projects and models are already in use
• Look at the upstream project's YAML for access: public models — only these are referenceable cross-project
• The first argument of ref() must exactly match the name field in the upstream project's dbt_project.yml (case-sensitive)

5. Know what you can and cannot reference

If you need a model that isn't public, coordinate with the upstream team to widen its access.

Cross-Project Refs Require dbt Cloud Enterprise

Cross-project ref() and the projects: key in dependencies.yml are only available on dbt Cloud Enterprise or Enterprise+ plans. Before setting up any cross-project collaboration, verify plan eligibility:

1. If dependencies.yml already has a projects: key and the project is actively using cross-project refs — Enterprise is already in place. Proceed.
2. Otherwise — ask the user to confirm they are on dbt Cloud Enterprise or Enterprise+ before adding projects: to dependencies.yml or writing new two-argument ref() calls.

If the user cannot confirm the plan level, or confirms they are on a plan below Enterprise, do not set up cross-project refs. Explain that this feature requires upgrading to Enterprise or Enterprise+ and suggest they use the intra-project governance features (groups, access modifiers, contracts) instead.

Cross-Project ref() Syntax

-- Reference an upstream model (latest version)
select * from {{ ref('upstream_project', 'model_name') }}

-- Reference a specific version
select * from {{ ref('upstream_project', 'model_name', v=2) }}

For full cross-project setup details (dependencies.yml, prerequisites, orchestration), see references/cross-project-collaboration.md.

Governance Features

dbt Mesh includes four governance features. These work independently and can be adopted incrementally:

YAML placement rule

In model property YAML files, access, group, and contract are configs and must always be nested under the config: key — never placed as top-level model properties. Placing them at the top level may appear to work in dbt Core but causes parse errors in dbt's Fusion engine.

# ✅ CORRECT — all governance configs under `config:`
models:
  - name: fct_orders
    config:
      group: finance
      access: public
      contract:
        enforced: true
    columns:
      - name: order_id
        data_type: int

# ❌ WRONG — governance configs as top-level properties (breaks Fusion)
models:
  - name: fct_orders
    access: public          # WRONG — not under config:
    group: finance          # WRONG — not under config:
    contract:               # WRONG — not under config:
      enforced: true
    columns:
      - name: order_id
        data_type: int

This applies to property YAML files only. In dbt_project.yml, use the + prefix for directory-level assignment (e.g. +group: finance, +access: private). In SQL files, use {{ config(access='public', group='finance') }}.

Adoption order

1. Groups & Access  →  2. Contracts  →  3. Versions  →  4. Cross-Project Refs
   (organize teams)     (lock shapes)    (manage changes)  (split projects)

• Groups & Access — no schema changes needed, start here
• Contracts — require declaring every column and data type in YAML
• Versions — only needed when a contracted model must introduce a breaking change
• Cross-Project Refs — require dbt Cloud Enterprise or Enterprise+ and a successful upstream production job. Do not set up cross-project refs if you cannot confirm the plan level is Enterprise or higher.

Contracts vs. Tests

Contracts are enforced before tests run. If a contract fails, the model is not built, and no tests execute.

Decision Framework

Should this model have a contract?

Use a contract when:

• The model is access: public (especially if referenced cross-project)
• Other teams depend on this model's schema stability
• The model feeds an exposure (dashboard, ML pipeline, reverse ETL)
• External consumers (other dbt projects, BI dashboards, reverse ETL) query the table directly and would break from column renames or removals

Do NOT add a contract when:

• Staging models (stg_*) — these are internal implementation details, not consumer-facing APIs
• The model is still evolving — if the user says they are iterating on the design, advise waiting until the schema stabilizes
• No external consumers exist — in a single-project setup with no cross-project refs, no BI tools depending on the schema, and no exposures, contracts add maintenance overhead without benefit. Ask about consumers before recommending contracts.
• Dynamic/pivot columns — models that use pivot(), unpivot(), or dynamically generate columns are poor candidates because the column list isn't fixed and the contract will break whenever the dynamic values change
• Ephemeral models — contracts are not supported on ephemeral materializations

If the user asks for a contract on a model that matches the "do NOT add" criteria above, advise against it and explain why. Do not simply comply — the user may not realize the contract is inappropriate. Suggest alternatives (e.g., data tests for staging models, waiting for schema stability, or switching materialization for ephemeral models).

Should this model be versioned?

Version a model when:

• It has an enforced contract AND you need to introduce a breaking change (column removal, rename, type change)
• Downstream consumers need a migration window before the old shape goes away

Do NOT version a model:

• For additive changes (new columns) — these are non-breaking
• For bug fixes — fix in place
• Preemptively "just in case" — version only when a breaking change is actually needed

What access level should this model have?

Is it referenced cross-project?
  └─ Yes → public (with contract recommended)
  └─ No
      Is it referenced outside its group?
        └─ Yes → protected (default)
        └─ No
            Is it internal to a small team?
              └─ Yes → private
              └─ No → protected (default)

Best practice: Default new models to private and widen access only when needed. The default protected is permissive — be intentional.

Common Mistakes