Clickzetta Dbt Modeling

clickzetta-dbt-modeling

See references/materialization-guide.md for materialization type selection, index strategy, partition strategy, and dynamic table configuration. See references/incremental-patterns.md for incremental strategy details. See references/test-strategy.md for test coverage standards and custom test templates. See references/grant-patterns.md for grant configuration (multi-tenant / data security scenarios).

For the complete end-to-end pipeline architecture (ingestion → modeling → Studio publishing), see ../clickzetta-dbt-project-setup/references/elt-standards.md.

Goal

Transform raw Lakehouse data into business-value dbt models, validated by tests. Users don't need to describe table schemas — you discover them, you infer the strategy, users only confirm or adjust.

Workflow

This skill runs as a series of conversation turns. Each turn ends with a question to the user — no files are written and no dbt commands are run until the user responds to that turn's question.

Turn 0 — Discover schemas, ask which to use

Run cz-cli schema list. Show the user the schema list immediately and ask:

"Which schemas contain the raw data you want to model?"

Present the schema names as options. Stop here.

If no schemas exist at all → ask about the data source and route to the appropriate ingestion skill. Tell the user: "Once data is ingested, come back and I'll pick up from here."

Turn 1 — Explore selected schemas, ask about business context

After the user selects schemas, run cz-cli table list on those schemas. Show the table list.

Then ask three things:

1. Business scenario: what does this data represent? (e.g. retail orders, user behavior, finance — helps name models correctly)
2. Target layers: which layers to build? If the project was set up with clickzetta-dbt-project-setup, use the layering mode already chosen. Otherwise ask the user — common options:
   • dbt standard: staging → marts (add intermediate if needed)
   • Medallion: bronze → silver → gold
   • Traditional DW: ODS → DWD → DWS → ADS
   • Custom: specify layer names
3. Data freshness requirement: how fresh does the data need to be? (near-real-time / minute-level, T+1 daily batch, or mixed) — this determines whether to use dynamic_table (continuous refresh) or incremental (scheduled batch), which has significant cost implications

Also branch on data quality:

• No relevant tables found → ask: explore other schemas, explore other workspaces, or specify tables directly?
• Data exists but looks unsuitable (empty tables, no meaningful columns) → show what was found and why it's a concern. Ask: proceed anyway, ingest better data first (stay in conversation), or point to different data?

Stop here. Do not run any COUNT or DESC queries yet.

Turn 2 — Analyze data, present modeling plan

After the user confirms business context and layers, run the analysis (tell the user "Analyzing table structures and data volumes..."):

• DESC TABLE each source table — identify columns, types, primary keys
• SELECT COUNT(*) — determine data volume
• Check growth history (last 7 days new rows)

Use these dimensions plus the Decision Tree in references/materialization-guide.md to infer materialization and strategy for each model.

Default to dynamic_table for most models — it handles T+1 batch, large tables, and near-real-time equally well. Only switch to incremental when dynamic_table genuinely cannot do the job:

• The model needs to process a specific time window (yesterday's data only, last hour only) — dynamic_table always reflects full current state
• The model must run after a specific upstream Studio task (dependency ordering)
• The source is a Table Stream and you need to process changes by type or control offset advancement

The confirmation table format depends on materialization type:

For each model, briefly explain why that materialization was chosen — not just what it is. This helps the user make an informed decision.

Ask: "Confirm all and generate / adjust specific models / partial modeling (subset only)?"

Stop here. Do not write any files.

Turn 3 — Generate files

Only after the user confirms the plan: generate all files in one pass.

Generate the directory structure to match the user's chosen layering mode. Examples:

# dbt standard
models/
├── staging/
│   ├── stg_{source}_{table}.sql
│   └── schema.yml
├── intermediate/          # only if needed for complex JOINs
│   └── int_{entity}__{verb}.sql
└── marts/
    ├── dim_{entity}.sql
    ├── fct_{event}.sql
    └── schema.yml

# Medallion
models/
├── bronze/
│   ├── {source}_{table}.sql
│   └── schema.yml
├── silver/
│   ├── {entity}.sql
│   └── schema.yml
└── gold/
    ├── {metric}.sql
    └── schema.yml

# Traditional DW
models/
├── ods/
│   ├── ods_{source}_{table}.sql
│   └── schema.yml
├── dwd/
│   ├── dwd_{domain}_{table}.sql
│   └── schema.yml
├── dws/
│   └── dws_{domain}_{metric}.sql
└── ads/
    └── ads_{subject}_{metric}.sql

For custom layering, generate directories matching the layer names the user specified.

Templates and schema.yml conventions are in the Generated File Structure section below.

After writing files, tell the user what was created and ask: "Ready to run dbt run and dbt test?"

Turn 4 — Execute

Only after the user says yes, run commands in order:

• dbt seed — only if seeds/ has .csv files
• dbt run
• dbt snapshot — only if snapshots/ has .sql files
• dbt test

Report results after each command. If dbt test fails, diagnose and fix before declaring completion.

Key Constraints

• database: "{{ target.database }}" in sources.yml — resolves to workspace name at runtime; never hardcode the workspace name
• dynamic_table models do NOT need Studio scheduling — refresh_interval drives automatic refresh; creating a Studio cron task for them is redundant and wasteful
• dynamic_table models do NOT need Studio dependency config — the system tracks upstream dependencies automatically through the DAG
• Only incremental / table / snapshot models need Studio scheduling via clickzetta-dbt-studio-pipeline
• Record incremental fields in each incremental model's schema.yml meta block:

  meta:
    incremental_field: updated_at      # incremental filter field name
    incremental_strategy: merge        # incremental strategy
  

  This is the standard interface read by clickzetta-dbt-studio-pipeline — without it, the agent cannot auto-detect incremental fields during Studio publishing and will need to ask the user.

Generated File Structure

Generate files in the directory structure matching the user's chosen layering mode (see Turn 3 above). The naming conventions below apply regardless of which layering pattern is used:

dbt standard naming (staging/marts):

stg_{source}_{table}.sql    # staging: one per raw table
dim_{entity}.sql            # dimension tables
fct_{event}.sql             # fact tables
int_{entity}__{verb}.sql    # intermediate (double underscore)

Medallion naming (bronze/silver/gold):

{source}_{table}.sql        # bronze: raw, minimal transformation
{entity}.sql                # silver: cleaned, conformed
{metric}.sql                # gold: aggregated, business-ready

Traditional DW naming (ODS/DWD/DWS/ADS):

ods_{source}_{table}.sql    # ODS layer
dwd_{domain}_{table}.sql    # DWD layer
dws_{domain}_{metric}.sql   # DWS layer
ads_{subject}_{metric}.sql  # ADS layer

dynamic_table model template (default for most layers):

{{ config(
    materialized='dynamic_table',
    refresh_interval='5 MINUTE',
    refresh_vc='default'
) }}
select
    ...
from {{ ref('upstream_model') }}

incremental model template (only when time-window control or Studio dependency is required):

Two patterns — choose based on how the model will be triggered:

Pattern A — self-driven (dbt decides what to process based on existing data):

{{ config(
    materialized='incremental',
    incremental_strategy='merge',
    unique_key='order_id',
   
) }}
select order_id, customer_id, amount, status, updated_at
from {{ source('raw', 'orders') }}
{% if is_incremental() %}
where updated_at >= (select max(updated_at) from {{ this }})
{% endif %}

Pattern B — partition overwrite (Studio injects the date parameter at scheduling time):

{{ config(
    materialized='incremental',
    incremental_strategy='insert_overwrite',
    partition_by='dt',
   
) }}
select dt, region, count(*) as order_count, sum(amount) as revenue
from {{ source('raw', 'orders') }}
where status = 'completed'
group by dt, region
-- No is_incremental() filter needed — insert_overwrite replaces the entire partition each run
-- When published to Studio, the WHERE clause is injected by the scheduling parameter (bizdate)

on_schema_change default: always set on_schema_change='append_new_columns' on incremental models — the default ignore silently drops new columns from upstream, causing data loss that is hard to diagnose.

sources.yml goes inside models/staging/schema.yml (not a separate file):

sources:
  - name: {source_name}
    database: "{{ target.database }}"
    schema: {raw_schema}
    tables:
      - name: {table}

Comment persistence: When generating schema.yml, write description for every model and column, and add persist_docs config — this way dbt run automatically writes comments into Lakehouse table and column metadata, visible in Studio, BI tools, and data catalogs:

models:
  - name: fct_orders
    description: "Orders fact table, one row per order, incremental merge strategy"
    config:
      persist_docs:
        relation: true    # write table comment
        columns: true     # write column comments
    columns:
      - name: order_id
        description: "Unique order identifier"

Enable globally in dbt_project.yml to avoid repeating per model:

models:
  {project_name}:
    +persist_docs:
      relation: true
      columns: true

Test coverage standards (add by default when generating schema.yml, see references/test-strategy.md):

Routing

Completion Criteria

• dbt test 0 errors
• dynamic_table models have refresh_interval and refresh_vc set in {{ config() }} and recorded in schema.yml meta:

  models:
    - name: fct_orders
      config:
        meta:
          refresh_interval: "5 MINUTE"
          refresh_vc: default
  

• incremental models have incremental_field and incremental_strategy recorded in schema.yml meta (for use by clickzetta-dbt-studio-pipeline):

  models:
    - name: fct_orders_incremental
      config:
        meta:
          incremental_field: updated_at
          incremental_strategy: merge
  

Next Steps After Modeling (proactively present to user)

Present the following options after modeling is complete — let the user choose the direction:

A. Publish model code to Studio as assets → clickzetta-dbt-studio-pipeline Publish SQL code to Studio for unified code management. Only table / incremental / snapshot models need scheduling — dynamic_table models self-refresh and do not need Studio cron tasks.

B. Connect BI tools and start reporting → lakehouse-doc-en official BI / JDBC / SQLAlchemy connection docs Connect marts layer tables to Tableau / Metabase / FineReport or other BI tools for direct query and analysis.

C. Configure data quality monitoring Add singular tests or integrate DQC for key metrics (daily order count, total revenue) with threshold alerts.

D. Continue expanding models Add new business domain models, or optimize existing models (add indexes, partitioning, adjust VCluster).

E. Verify data first Query marts layer manually with cz-cli sql or BI tools to confirm data looks correct before deciding next steps.

Don't only suggest Studio scheduling — the user may be more urgently interested in connecting BI or verifying data. Let them choose.