Data Build Tool Summary¶

Dbt core is an open source ELT CLI and database agnostic used to allow data analysts and engineers to build reliable, modular data pipelines, creating "models" (SELECT statements) that are version-controlled, automatically documented, and tested for quality before consumption by analytics tools. Learn more about dbt in the docs.
dbt Cloud: A managed service with a web-based IDE, scheduler, job orchestration, and monitoring

Supported by ISVs in lake house market.

Relation with Flink¶

Confluent has also developed a dbt adapter to deploy Flink SQL statements into Confluent Cloud for Flink.

pip install dbt-confluent
# or
uv add  dbt-confluent

Flink SQLs are defined in Models and dbt processes to the deployment to Confluent Cloud using the REST API. When adopting a 'shift left' strategy of moving part of the star model to real time processing, it makes sense to manage real-time streaming project as data engineers manages datawarehouse or lakehouse projects.

We will first work on a concrete example on a database, and then work on a Flink project.

See Confluent cloud product documentation on dbt.

Use Cases¶

Modelling changes are easy to follow and revert
Explicit dependencies between models
Data quality tests
Incremental load of fact tables
Track history of dimension tables
Support automated testing, document generation, and data lineage visualization

Install¶

We need Python, as dbt should be installed in a virtual environment. See installation instructions. See the supported Python database
Create a $HOME/.dbt folder to let dbt persists the profile.yaml file to keep user and Database credentials.
Start a new python session under your working folder (e.g. dbt)
```
uv venv
source .venv/bin/activate
```
Install dbt, and dbt adapters
```
uv add dbt-duckdb
```

Getting started¶

We will go over the getting the foundations of a project, and then go over the main concepts while implementing the examples.

a dbt project is a directory on the data engineer's machine containing a lot of .sql files (called models) and YAML files for configurations.

A data warehouse example¶

This example is in code/dbt/airbnb.

Create the dbt project
```
dbt init airbnb
```
while running this, it will ask to set the dbt profile for this project (I selected duckdb). A project profile is a YAML file containing the connection details for your chosen data platform. When there is an existing ~/.dbt/profiles.yml, the previous command will add a new stanza to it.
```
airbnb:
  outputs:
    dev:
      type: duckdb
      path: dev.duckdb
      threads: 1

    prod:
      type: duckdb
      path: prod.duckdb
      threads: 4

  target: dev
```
It specifies two configurations, dev and prod, with different data warehouse connections. The path specifies where in the working directory (e.g. airbnd) the database will be created.

This will also create a set of folders to manage all the needed elements of data pipelines:
```
models
analyses
tests
seeds
macros
snapshots
```
and the dbt_project.ymlfile to define dbt settings.

Understand the dbt_project.yml

It refences the profile to use, the different paths to use
```
profile: 'airbnb'
```

and the models may have many resources configured at once:

models:
  models:
    # namespace for model configs matching the project name
    airbnb:
      +materialized: view

Next we will cover the dbt main concepts with concrete examples.

A Confluent Cloud Flink example¶

Confluent dbt adapter aims to support standard dbt commands (init, debug, run, test, docs generate, etc.) against Confluent Cloud for Flink, so teams can manage pipelines end-to-end from dbt rather than using Terraform or confluent cli.

For Confluent the adapter is installed via:
```
uv add dbt-confluent
```
Create the project
```
dbt init flink_workshop
```

The profile may include references to environment variables for API KEY and SECRET.

flink_workshop:
outputs:
  dev:
    cloud_provider: aws
    cloud_region: us-west-2
    compute_pool_id: lfcp-
    dbname: j9r-kafka
    environment_id: env-
    execution_mode: streaming_query
    flink_api_key: "{{ env_var('CONFLUENT_FLINK_API_KEY') }}"
    flink_api_secret: "{{ env_var('CONFLUENT_FLINK_API_SECRET') }}"
    organization_id: 49......44
    statement_label: dbt-confluent
    statement_name_prefix: dbt-
    threads: 1
    type: confluent

Major Concepts¶

Batching in dbt with a data warehouse, like DuckDB or Snowflake, is primarily managed through different materialization strategies:
- Table: Replaces the entire target table with each run. Ideal for smaller datasets or full-refresh batches.
- Incremental: Updates only new or changed data using append, merge, or delete+insert.
- Microbatch: Breaks massive datasets into smaller, time-based segments (e.g., daily) that process independently.
- External: Reads from and exports results directly to files (Parquet, CSV, JSON) on local storage or S3.
dbt encourages building complex transformations in smaller, reusable SQL steps, reducing repetitive code.
dbt uses a template mechanism (jinja), functions and a set of features to organize SQL and cross reference them.
The mandatory file for a project is the dbt_project.yml file as it contains information that tells dbt how to operate the project. dbt demarcates between a folder name and a configuration by using a + prefix before the configuration name.
Models: are the basic building blocks of the business logic. They includes materialized tables and views, and SQL files. Models can reference each others and use templates and macros.
Resources types includes models, seeds, snapshots, tests, sources
Properties describe resources
Configurations control how dbt builds resources in the warehouse. Could be set cross resources in dbt_project.yml, in a properties.yml under a folder, config() in a sql or resource file.

Models¶

Models are built in logical layers to keep the pipeline clean and scalable. They are dependent on each other, forming a Direct Acyclic Graph.

Staging (stg_): Clean and rename the raw data (e.g., lowercase names, fix boolean types).
Intermediate (int_): Perform complex joins, aggregations, and business logic here.
Marts (fct_ or dim_): The final, analytics-ready models

The table below lists when to use View vs Table:

	View	Table
Purpose	Use for minor transformation	For intensive transformation
Execution	At runtime and when referenced	Pre-executed, with the results saved in tables
Storage	None	Need Storage space for materialized tables
Performance	Lot of steps leads to slower performance	Chained processes get improved perf.

dbt provides built-in testing (e.g., uniqueness, non-null checks) to catch broken logic
schema is the data contract of elements of the model, and defined in a separate yaml file.
There are two macros to cross reference tables: {{ ref() }} used to reference a table within a model and {{source() }} to reference external data sources.

Create the first model

See the duckdb example. Let start by seeding reference data into the warehouse from a csv file. The seeds/ folder includes all the csv to map to table inside duckdb. The following command let dbt to look inside the seeds/ folder of your project, find any .csv files, and upload them as physical tables into your database.

dbt seed --target dev

Raw data are created in tables in a raw schema, as data landing zone.

Once created, the second step, is to add dbt sources

Create sources¶

Sources are defined using a YAML file sources.yml inside your models/ directory.

sources:
  - name: airbnb  # This is the internal dbt name for the source
    schema: raw   # The actual schema name in your warehouse
    tables:
      - name: listings # This is the name of the raw table in the data warehouse
        identifier: raw_listings

When in sources, it is possible to reference the table with: select * from {{ source('airbnb', 'listings') }}. One first examples is to implement deeduplication, filtering and column renaming. This can be done by adding a models/sources folder and add a SQL like:

WITH ranked_hosts AS (
    SELECT
        *,
        ROW_NUMBER() OVER (
            PARTITION BY id
            ORDER BY updated_at DESC, created_at DESC
        ) AS row_num
    FROM
        {{ source('airbnb', 'hosts') }}
    WHERE id IS NOT NULL
)
SELECT
    id AS host_id,
    name AS host_name,
    is_superhost,
    created_at,
    updated_at
FROM
    ranked_hosts
WHERE
    row_num = 1

So running the following command will create the new table for cleaned records:

dbt run --target dev
# returns
Found 1 seed, 3 sources, 474 macros

which can be validated by opening the duckdb database:

duckdb data/airbnb.duckdb
# in the shell
select * from main.src_hosts;

Same may be done for raw_reviews, raw_listings

For source freshness, we need to consider one DATE column and add a config element to the table to define refreshness condition:

- name: reviews
  identifier: raw_reviews
  config:
    loaded_at_field: date
    freshness:
        warn_after: {count: 1, period: hour}
        error_after: {count: 24, period: hour}

Run the command: dbt source freshness to validate the data freshness.

Project structure¶

dbt recursively scans everything under model-paths (by default models/). dbt run, dbt build, and dbt seed will all find those .sql files and deploy/run them.

The folder structure under the models folder can be a hierarchy based on the kimball architecture or star schema See PM methodology:

models/
  sources/
    hosts/
      src_hosts.sql
    listings/
      src_listings.sql
  dimensions/
    hosts/
      dim_hosts_cleansed.sql
  facts/
    reviews/
      fct_reviews.sql

dbt names a model from the file name, not from the folder path. * Formodels/sources/src_hosts.sqlthe model name issrc_hosts* Useref('src_hosts'), notref('sources.hosts.src_hosts')`

So if two subfolders both contain model.sql, you get a name collision, therefore it is recommended to use distinct filenames (dim_hosts.sql, fct_reviews.sql), or set an explicit alias in config.

For incremental deployment it should be possible to run dbt as:

uv run dbt run --select dimensions.hosts 
# or --select dimensions.*

# Select a unique model
uv run dbt  run --project-dir . --select withdrawals_by_account.sql --target dev --profiles-dir ~/.dbt --full-refresh

Materializations¶

There are four possible materializations for a model:

View: this is a lightweight representation of the data, not reused. no recreation of the table at each execution.
Table: reusable data in external table, recreated at each run
Incremental: fact tables appends to tables - more like event data - table is not recreated each time.
Ephemeral (CTEs): aliasing of the data and filtering data. Not adversitized in the data warehouse. For example all the sql under the sources are becoming CTEs

Materialization can be set globally in the dbt_profile.yaml: all models are views, except in the dimensions folder where there are tables:

models:
  airbnb:
    +materialized: view
    dimensions:
      +materialized: table 
    sources:
      +materialized: ephemeral

Also it is possible to define at the model level the materialization:

Create dimensions folder, add sql with config:

{{
  config(
    materialized = 'view'
    )
}}
SELECT
  listing_id,
  listing_name,
  room_type,
  CASE
    WHEN minimum_nights = 0 THEN 1
    ELSE minimum_nights
  END AS minimum_nights,
  host_id,
  CAST(
    REPLACE(price_str, '$', '') AS DECIMAL(10, 2)
  ) AS price,
  created_at,
  updated_at
FROM    {{ ref('src_listings') }}

Incremental¶

Each time dbt run, the tables will be recreated in the data warehouse. But Fact tables may be kept and the dml can run to continue processing only new records. To do so we need to specify the materialization to be incremental. We also want to stop the dml if schema is changed, for that we add conditions: on_schema_change='fail'.

{{
  config(
    materialized = 'incremental',
    on_schema_change='fail'
    )
}}
WITH src_reviews AS (
  SELECT * FROM {{ ref('src_reviews') }}
)
SELECT * FROM src_reviews
WHERE review_text is not null

{% if is_incremental() %}
  AND review_date > (select max(review_date) from {{ this }})
{% endif %}

Finally we need to specify how dbt should increment records in this fact table: this is done by adding jinja template to test if dbt run in incrementatl mode, therefore condition could be added on one of the column like the review_date of the record. It needs to be after the last record in the fct_reviews table () '{{ this }}'.

dbt run with run the fct_reviews in incremental view. If we add records to the raw_reviews with something like:
```
INSERT INTO raw_reviews VALUES(3176, CURRENT_TIMESTANP(), 'j9r', 'Excellent sejour', 'positive');
```
then rerun dbt, the new records will be added as last record in the fct_reviews.
To rebuild everything, make a full-refresh:
```
dbt run --full-refresh
```

With the sources as ephemeral the output of dbt run becomes:

23:16:16  1 of 4 START sql table model DEV.dim_hosts_cleansed ............................ [RUN]
23:16:18  1 of 4 OK created sql table model DEV.dim_hosts_cleansed ....................... [SUCCESS 14111 in 1.93s]
23:16:18  2 of 4 START sql table model DEV.dim_listings_cleansed ......................... [RUN]
23:16:20  2 of 4 OK created sql table model DEV.dim_listings_cleansed .................... [SUCCESS 17499 in 2.47s]
23:16:20  3 of 4 START sql incremental model DEV.fct_reviews ............................. [RUN]
23:16:23  3 of 4 OK created sql incremental model DEV.fct_reviews ........................ [SUCCESS 0 in 2.37s]
23:16:23  4 of 4 START sql table model DEV.dim_listings_with_hosts ....................... [RUN]
23:16:24  4 of 4 OK created sql table model DEV.dim_listings_with_hosts .................. [SUCCESS 17499 in 1.58s]
23:16:24  
23:16:24  Finished running 1 incremental model, 3 table models in 0 hours 0 minutes and 9.83 seconds (9.83s).

dbt compile does not deploy to the target data warehouse

23:39:49  Running with dbt=1.11.6
23:39:50  Registered adapter: snowflake=1.11.2
23:39:50  Found 8 models, 1 seed, 3 sources, 522 macros
23:39:50  
23:39:50  Concurrency: 1 threads (target='dev')

Incremental strategies¶

mode	pros	cons
append	simplest. fully transparent logic	no duplicate checks
merge	idempotent. updates existing records	slower, require unique key
delete+insert	works for data warehouse without merge.	two operations. require unique key
insert+overwrite	fast for partitioned tables	requires partitioned table. Replaces the whole partitions

This is set in the config() element.

  {{
    config(
      materialized = 'incremental',
      on_schema_change='fail',
      incremental_strategy='merge',

      )
  }}

! STOPPED HERE

Type-2 slowly changing dimensions¶

The goal is to keep history of changes to the records over time and not just the last record per key. dbt adds dbt_valid_from and dbt_valid_to columns to mark each records to be valid time from and to. A current correct records have dbt_valid_to sets to null.

snapshots live in the snapshot folder. There are two strategies for assessing data changes: * Timestamp: a unique key and updated_at fields is defined at the source model. These columns are used for determining changes * Check: any changes in a set of columns (or all columns) will be picked up as an update.

To create snapshots we need a yaml file under the snapshot folder:

snapshots:
- name: scd_raw_listings
    relation: source('airbnb', 'listings')
    config:
    unique_key: id
    strategy: timestamp
    updated_at: updated_at
    hard_deletes: invalidate

the dbt snapshot will create a new table with the columns added for the referenced table.

00:04:36  1 of 1 START snapshot DEV scd_raw_listings ..................................... [RUN] 
00:04:40  1 of 1 OK snapshotted DEV.scd_raw_listings ..................................... [SUCCESS 17499 in 3.44s]

An update to an existing record and a new dbt snapshot will create historical record.

¶

Init a project: This command creates some folders to organize work inside the data project (without modifying user's profile).
```
uv run dbt init --skip-profile-setup airbnb
```

profile.yaml¶

profile.yaml defines the structure of the project, and keeps information to connect to database.

Work on Models¶

Add Kimball structure as sources, dimensions, facts under the models folder
Add SQL materialized view using SELECT .... Do not use INSERT INTO, as it will be added automatically by dbt

Validate each new SQL creation: within the folder with the dbt_profile.yaml, to build a view in Snowflake for example

dbt run

Example of output:

22:20:56  Found 1 model, 522 macros
22:20:56  
22:20:56  Concurrency: 1 threads (target='dev')
22:20:56  
22:20:57  1 of 1 START sql view model DEV.src_listings ................................... [RUN]
22:20:58  1 of 1 OK created sql view model DEV.src_listings .............................. [SUCCESS 1 in 1.17s]
22:20:59  
22:20:59  Finished running 1 view model in 0 hours 0 minutes and 2.78 seconds (2.78s).
22:20:59  
22:20:59  Completed successfully
22:20:59  
22:20:59  Done. PASS=1 WARN=0 ERROR=0 SKIP=0 NO-OP=0 TOTAL=1

and within Snowflake:

dbt run creates final sql queries under the target folder. This run command can also apply to a specific table:
```
dbt run --select +models/facts/fct_reviews.sql
```

The + in front of the name, specifies to deploy parents tables too.

Tests¶

Two types of tests:
- Unit Tests
- Data Tests: run on actual data
There are two types of data tests: singular(SQL queries stored in tests) and generic
Defining test is by adding a schema.yaml with conditions on table. See example in models folder

To run the tests

dbt text --target duckdb -x
# -x it to continue even if one test fails

To debug a test, we can always look at where the dbt created the SQL to run, and execute this SQL in a SQL client, like duckdb cli.
By setting in the dbt_project.yaml
```
data_tests:
  _store_failures: true
```

Then any test failures will be saved in a new schema with table in the datawarehouse.

02:48:41  Failure in test accepted_values_dim_listings_cleansed_room_type__Private_room__Entire_home_apt__Shared_room__Hotelroom (models/schema.yaml)
02:48:41    Got 1 result, configured to fail if != 0
02:48:41  
02:48:41    compiled code at target/compiled/airbnb/models/schema.yaml/accepted_values_dim_listings_c_72f6cd1e8c350657dd7c7e44ed95fd70.sql
02:48:41  
02:48:41    See test failures:
---------------------------------------------------------------------------------------------------------------
select * from "airbnb"."main_dbt_test__audit"."accepted_values_dim_listings_c_72f6cd1e8c350657dd7c7e44ed95fd70"
---------------------------------------------------------------------------------------------------------------

See elementary-data.com
For unit test, we can define a yaml file at the same level as the SQL under validation. See unit_test.yaml, then run:
```
dbt test -s mart_fullmoon_reviews
```
Example to validate consistency among the created_at fields of the listings and reviews, we may want to add a singular test under the tests folder in the form of SQL:

Other Databases¶

Using `dbt` with DuckDB¶

Install dbt-duckdb python module: uv add dbt-duckd
Import raw data to the Duckdb table or use Airflow to ETL such data. For example in code/dbt/airbnb/, bootstrap the DuckDB raw tables with:
```
export DBT_DUCKDB_PATH=./data/airbnb.duckdb
duckdb "${DBT_DUCKDB_PATH:-./data/airbnb.duckdb}" < scripts/bootstrap_duckdb_raw.sql
```

Run dbt against DuckDB:

cd airbnd
dbt seed --target duckdb
dbt run --target duckdb
dbt test --target duckdb

Use duckdb query engine to look at the tables: First duckdb ./data/airbnb.duckdb command. See the dot commands

.open data/airbnd.duckdb
.databases
.schema
.read FILENAME  Read and execute SQL from an external file
select * from raw.raw_hosts;

As duckDb can be embedded into Python code, it is possible to load a Pandas dataframe from a table in duckdb.

Using `dbt` with postgresql¶

Install Kubernetes Postgresql operator, then a postgres cluster and PGadmin webapp. See the minikube/posgresql folder

Do port forwarding for both Postgresql server and pgadmin webapp

kubectl port-forward service/pg-cluster 5432:5432
kubectl port-forward service/pgadmin-service 8080:80

get user, database name , password and URI from the postgresql secrets (see Makefile)
Create customers and orders tables, insert some records
Define the connection to the database in the .dbt/profiles.yaml
Validate with the connection dbt debug
Write some sql scripts in the models folder, then use dbt run and it will create new views in the default schema and one table. Example of join
The results can be also seen by querying the newly created views or tables.
```
select * from "default".customerorders;
```

Confluent Cloud Flink Specifics¶

In Confluent Cloud for Flink context, the dbt run does not process data; it deploys or updates the definition of a continuous dataflow to the streaming engine. User runs dbt run only when the SQL queries changes. Important chapter from Confluent cloud product documentation on dbt.

A dbt schema is a Flink database, while a dbt database is a Flink Catalog, and finally a dbt identifier is a Flink Table. The schema field in profiles.yml actually refers to a Kafka cluster name. The database field refers to an environment ID. Some error messages reference "schema" when they mean "Kafka cluster/database". It is confusing to set the environment_id where is Confluent Cloud the environment name is used.
The dbt mapping:

Dbt construct	CC Flink	dbt confluent materialization
view	create view .. as select	view
table	snapshot query	streaming_table
incremental	not-supported
ephemeral	not supported
materialized_view	create table ... as select	streaming_table
seeds	CREATE TABLE + INSERT VALUES (point-in-time)	seed (default)

How to¶

Seed reference data on Confluent Cloud for Flink

The airbnb_streaming project loads small reference CSVs into Flink tables with dbt seed. The dbt-confluent adapter infers column types from the CSV (via agate) and issues CREATE TABLE followed by INSERT INTO ... VALUES. Override types in seeds/seeds.yml when inference is too coarse (e.g. map date strings to DATE).

# seeds/seeds.yml (dbt properties format)
version: 2
seeds:
  - name: seed_full_moon_dates
    config:
      alias: raw_full_moon_dates
      column_types:
        full_moon_date: DATE

Set +execution_mode: snapshot on seeds in dbt_project.yml so DDL and INSERT run as point-in-time statements (not streaming queries). Seeds are limited to 10,000 rows per file.

cd code/dbt/airbnb_streaming
dbt seed --target dev

Optionally run make seed to regenerate seeds.yml column types from CSV headers before seeding.

How to specify table properties?

Use the config function, and the with as json.

{{ config(
    materialized = 'streaming_table',
    with         = {
        'changelog.mode: 'upsert'
    }
   ) 
}}

How to define primary key for non source tables?

This is done in the schema.yml file. For a column:

columns:
  - name: account_number
    data_type: varchar(2147483647)
    constraints:
      - type: not_null
      - type: primary_key
        expression: "not enforced"

or as combined keys set with external contraints:

columns:
  - name: account_number
    data_type: varchar(255)
  - name: transaction_type
    data_type: varchar(50)
  - name: total_withdrawn
    data_type: decimal(38, 2)
constraints:
  - type: primary_key
    columns: [account_number, transaction_type]
    expression: "NOT ENFORCED"

How to define materialized table

Add the config element in the model file

{{ config(
    materialized       = 'materialized_table',
    freshness_interval = "INTERVAL '1' MINUTE",
    distributed_by     = "order_id",
    start_mode         = 'RESUME_OR_FROM_BEGINNING'
    with               = {
        'key.format: 'avro-registry',
        'value.format: 'avro-registry'
    }
  ) 
}}

Flink Demos using dbt:¶

Sources of Information¶

Udemy training from Zoltan C. Toth with Git Repo. Example of data from Inside AirBnB.
Dbt core
Learn more about dbt in the docs
Preset is a SaaS for Apache Superset to develop BI dashboard, on cloud with dbt integration. It also includes a SQL Editor.
Snowflake username: jbcodeforce. Using key-pair authentication. Public key in Snowlflake
Youtube tutorial
Patrick Neff's git repo: Stream Processing in Confluent Cloud Flink with data build tool (dbt)
Check out Discourse for commonly asked questions and answers
Check out the blog for the latest news on dbt's development and best practices