Awesome

Data Transform Spec

dtspec is an API for specifying and testing data transformations.

Introduction

Testing data transformations is hard. So hard that a lot of ETL/ELT processes have little or (more often) no automated tests. dtspec aims to make it easier to write and run tests for very complicated data transformations typically encountered in ETL/ELT.

With dtspec, we imagine a data transformation process that takes a set of data sources and transforms them into a set of data targets. dtspec is primarily concerned with structured data sources, like Pandas dataframes or database tables. A user of dtspec defines data factories that generate source data, and a set of expectations that describe how the data should look after it's been transformed.

While dtspec is written in Python, it is intended to be used as more of a language-agnostic API. A dtspec user writes a test spec, which is then passed to dtspec. dtspec processes that spec and then returns to the user test data for all of the source specific in the spec. The user then feeds that test data into their data transformation system, collects the output, and sends it back to dtspec. dtspec compares the actual results of the data transformations with the expected results specific in the spec and reports on any discrepancies.

Tutorial

Let's see this all at work with some examples.

Hello World!

Let's suppose we have a dataset containing student records. Our data transformation simply reads in that data, and returns a new dataframe with a "Hello <student>" salutation. We want to test that it says "hello" to everyone. For the purposes of our tutorial, the data transformation will be written in Pandas as

def hello_world_transformer(raw_students):
    salutations_df = raw_students.copy()
    salutations_df["salutation"] = salutations_df['name'].apply(lambda v: 'Hello ' + v)

    return {"salutations": salutations_df}

dtspec is an API that accepts a JSON blob for the transformation spec. However, I strongly prefer to write specs in YAML and then convert them into JSON before passing them on to dtspec. To begin writing our transform spec, we define the dtspec version, a description of the transform spec, and then list out the sources and targets:

---
version: '0.1'
description: HelloWorld - Simplest example of running dtspec

# The names of sources and targets is arbitrary, but it's up to the user to determine
# how they get mapped to/from their data transformation system.
sources:
  - source: raw_students

targets:
  - target: salutations

These define our inputs and outputs. But we also need to define how to generate data for the input(s). For that, we define a factory:

factories:
  - factory: SomeStudents
    description: Minimal example of what some student records may look like

    data:
      - source: raw_students
        # Tables written as a markdown table
        table: |
          | id | name   |
          | -  | -      |
          | 1  | Buffy  |
          | 2  | Willow |

Lastly, we need to describe how we expect the data to look after it has been transformed. To do this, we define scenarios and cases. Scenarios are collections of cases that share some common data factory or describe similar situations. For now, our transform spec will just contain a single scenario and a single case:

scenarios:
  - scenario: Hello World
    description: The simplest scenario
    # All cases in this scenario will use this factory (which may be modified on case-by-case basis)
    factory:
        parents:
          - SomeStudents

    cases:
      - case: HelloGang
        description: Make sure we say hello to everyone
        expected:
          data:
            - target: salutations
              # The actual output may also contain the "name" field, but the expectation
              # will ignore comparing any fields not listed in the expected table.
              table: |
                | id | salutation   |
                | -  | -            |
                | 1  | Hello Buffy  |
                | 2  | Hello Willow |

That's it. See also the full YAML spec.

Now that we've described the full transform spec, we need to use it. The first step is to parse the YAML file, send it to the dtspec api, and have dtspec generate source data:

import dtspec
import yaml

spec = yaml.safe_load(open("tests/hello_world.yml"))
api = dtspec.api.Api(spec)
api.generate_sources()

The specific steps taken at this point are going to be sensitive to the data transformation environment being used, but we'll stick with our Pandas transformations for the sake of this tutorial. Given this, we can define a simple function that converts the source data returned from dtspec into Pandas dataframes:

import pandas as pd

def parse_sources(sources):
    "Converts test data returned from dtspec api into Pandas dataframes"

    return {
        source_name: pd.DataFrame.from_records(data.serialize())
        for source_name, data in sources.items()
    }

We can then run those test Pandas dataframes through our data transformation function.

sources_data = parse_sources(api.spec["sources"])
actual_data = hello_world_transformer(**sources_data)

Next, we need to convert the output dataframes of the transformations, actual_data, back into a format that can be loaded into dtspec for comparison. For Pandas, this function is:

def serialize_actuals(actuals):
    "Converts Pandas dataframe results into form needed to load dtspec api actuals"

    return {
        target_name: json.loads(dataframe.astype(str).to_json(orient="records"))
        for target_name, dataframe in actuals.items()
    }

It is loaded into dtspec using:

serialized_actuals = serialize_actuals(actual_data)
api.load_actuals(serialized_actuals)

Finally, dtspec can be called to run all of the expectations:

api.assert_expectations()

Putting all of this together:

spec = yaml.safe_load(open("tests/hello_world.yml"))
api = dtspec.api.Api(spec)
api.generate_sources()

sources_data = parse_sources(api.spec["sources"])
actual_data = hello_world_transformer(**sources_data)
serialized_actuals = serialize_actuals(actual_data)
api.load_actuals(serialized_actuals)

Try running the above code and changing either the YAML spec or the hello_world_transformer function and see how dtspec responds.

Hello World With Multiple Test Cases

Running tests with multiple cases that reference the same data sources introduces a complicating factor. One of the reasons that makes it hard to build tests for ETL/ELT is the fact that many data transformation systems in use today have a high latency for even very small transformations. For example, Redshift is a distributed RDBMS that can process billions of rows in minutes, millions of rows in seconds, thousands of rows in seconds, or 10s of rows in, well, seconds. Given these latency issues, we don't want to have to rely on loading data into our system, running a test, clearing out the data, loading some more, running the next test, and so on as is often done when testing ORM-based applications like Rails or Django.

dtspec seeks to minimize the number of requests on the data transformation system in order to deal with these latency issues. It does this by "stacking" the test data generated in each case and delivering back to the user all of this stacked data. The user then loads this stacked data into their data transformation system once, runs the data transformations once, and then collects the resulting output once.

Let's see how dtspec handles this in action.

First, let's change our hello world data transformation a bit. Instead of just saying hello to our heroes, let's say goodbye to any villians (as identified by a clique data field).

def hello_world_multiple_transformer(raw_students):
    def salutation(row):
        if row["clique"] == "Scooby Gang":
            return "Hello {}".format(row["name"])
        return "Goodbye {}".format(row["name"])

    salutations_df = raw_students.copy()
    salutations_df["salutation"] = salutations_df.apply(salutation, axis=1)

    return {"salutations": salutations_df}

While it would be possible to test saying hello or goodbye in a single case just by adding more records to the source data, we'll split it into two to demonstrate how multiple cases work. Here's how the YAML would look:

scenarios:
  - scenario: Hello World With Multiple Cases
    description: The simplest scenario
    factory:
      parents:
        - SomeStudents

    cases:
      - case: HelloGang
        description: Make sure we say hello to everyone
        expected:
          data:
            - target: salutations
              table: |
                | id | name   | clique      | salutation   |
                | -  | -      | -           | -            |
                | 1  | Buffy  | Scooby Gang | Hello Buffy  |
                | 2  | Willow | Scooby Gang | Hello Willow |

      - case: GoodbyeVillians
        description: Say goodbye to villians
        # For this case, we tweak the factory defined for the scenario.
        factory:
          # The ids here might be the same as above.  However, these are just named
          # references and get translated into unique ids when the source data
          # is generated.
          data:
            - source: raw_students
              table: |
                | id | name     |
                | -  | -        |
                | 1  | Drusilla |
                | 2  | Harmony  |
              # Use values to populate a constant over all records
              values:
                - column: clique
                  value: Vampires

        expected:
          data:
            # Again, the ids here are not the actual ids sent to dtspec after performing
            # the transformations.  They are just named references and dtspec
            # keeps track of the relationship between the actual ids and the named ones.
            - target: salutations
              table: |
                | id | name     | clique   | salutation       |
                | -  | -        | -        | -                |
                | 1  | Drusilla | Vampires | Goodbye Drusilla |
                | 2  | Harmony  | Vampires | Goodbye Harmony  |

This won't quite work as is, because we're missing something. We have two cases that describe variations on the source data raw_students and the output salutations. dtspec collects the source data definitions from each case and stacks them into a single data source. The user then runs the transformations on that source and generates a single target to provide back to dtspec. But dtspec has to know which record belongs to which case. To do this, we have to define an identifier that tells dtspec which columns should be used to identify a record as belonging to a case. A good identifier is often a primary key that uniquely defines a record, but it is not strictly required to be unique across all records.

For this example, we'll define an identifier called "students" with a single identifier attribute called id that is a unique integer:

identifiers:
  - identifier: students
    attributes:
      - field: id
        generator: unique_integer

We tell dtspec that this identifier is associated with the id columns of both the source and the target via:

sources:
  - source: raw_students
    identifier_map:
      - column: id
        identifier:
          name: students
          attribute: id


targets:
  - target: salutations
    identifier_map:
      - column: id
        identifier:
          name: students
          attribute: id

With the sources and targets with identifiers, the values we see in the source factories and target expectations are not the values that are actually used in the data. Instead, they are simply named refereces. For example, in the "HelloGang" case, id=1 belongs to Buffy and id=2 belongs to Willow. But when dtspec generates the source data, the actual values may be 3 and 9, or 4 and 7, or something else. Unique values are not generated in any deterministic manner -- each run of dtspec can give a diferent set. dtspec only guarantees that the each named reference will be a unique integer (via the generator defined in the identifier section).

Futhermore, in the second case called "GoodbyeVillians", we see that id=1 belongs to Drusilla and id=2 belongs to Harmony. dtspec will generate unique values for this case as well, and they will not conflict with the values generated for the first case. So dtspec will pass back to the user 4 total records (Buffy, Willow, Drusilla, Harmony) with 4 different ids

With the full YAML spec defined, we can run the assertions in the same fashion as the the earlier example

spec = yaml.safe_load(open("tests/hello_world_multiple_cases.yml"))
api = dtspec.api.Api(spec)
api.generate_sources()

sources_data = parse_sources(api.spec["sources"])
actual_data = hello_world_multiple_transformer(**sources_data)
serialized_actuals = serialize_actuals(actual_data)
api.load_actuals(serialized_actuals)

api.assert_expectations()

Embedded Identifiers

It is also possible to embed identifiers in the value of a particular column. For example, suppose our salutation column said hello to the id instead of the name of the person. To make this work, we have to put a particular string pattern in the column that indicates the name of the identifier, the attribute, and the named id - {identifier.attribute[named_id]}. The yaml spec would look like:

      - case: HelloGang
        description: Make sure we say hello to everyone
        expected:
          data:
            - target: salutations
              table: |
                | id | name   | clique      | salutation             |
                | -  | -      | -           | -                      |
                | 1  | Buffy  | Scooby Gang | Hello {students.id[1]} |
                | 2  | Willow | Scooby Gang | Hello {students.id[2]} |

The realistic example discussed below has another example of using embedded identifiers.

Note that embedded identifiers cannot be used to associate records with cases. A target must have at least one column listed in the identifier_map section.

A More Realistic Example

Finally, let's example a more realistic example that one might encounter when building a data warehouse. In these situations, we'll have multiple sources, targets, scenarios, and cases. Now suppose we have a students table, where every student belongs to a school and takes 0 to many classes. Our goal is to create one denormalized table that combines all of these data sources into one table. Additionally, we want to create a table that aggregates all of our students to give a count of the students per school. In Pandas, the data transformation might look like:

def realistic_transformer(raw_students, raw_schools, raw_classes, dim_date):

    student_schools = raw_students.rename(
        columns={"id": "student_id", "external_id": "card_id"}
    ).merge(
        raw_schools.rename(columns={"id": "school_id", "name": "school_name"}),
        how="inner",
        on="school_id",
    )

    student_classes = student_schools.merge(
        raw_classes.rename(columns={"name": "class_name"}),
        how="inner",
        on="student_id",
    ).merge(
        dim_date.rename(columns={"date": "start_date"}), how="left", on="start_date"
    )

    student_classes["student_class_id"] = student_classes.apply(
        lambda row: "-".join([str(row["card_id"]), str(row["class_name"])]), axis=1
    )

    students_per_school = (
        student_schools.groupby(["school_name"])
        .size()
        .to_frame(name="number_of_students")
        .reset_index()
    )

    return {
        "student_classes": student_classes,
        "students_per_school": students_per_school,
    }

Given the full YAML spec defined, we can again run the data assertions using a familiar pattern:

spec = yaml.safe_load(open("tests/realistic.yml"))
api = dtspec.api.Api(spec)
api.generate_sources()

sources_data = parse_sources(api.spec["sources"])
actual_data = hello_world_multiple_transformer(**sources_data)
serialized_actuals = serialize_actuals(actual_data)
api.load_actuals(serialized_actuals)

api.assert_expectations()

dbt support

dtspec also contains a CLI tool that can facilitate using it with dbt. The CLI tools helps you set up a test environment, run dbt in that environment, and execute the dbt tests. The CLI tool currently only works for Postgres and Snowflake dbt projects.

See the dbt-container-skeleton for a working example.

dtspec CLI Config

All of the dtspec files should be placed in a subdirectory of your dbt project: dbt/dtspec. The first thing to set up for the dtspec CLI is the configuration file, which should be placed in dtspec/config.yml. The configuration file tells dtspec how to recreate the table schemas in a test environment, where to recreate the table schemas, and where to find the results of a dbt run. Here is an example:

# A target environment is where the output of data transformations appear.
# Typically, there will only be on target environment.
target_environments:
  # The target environment IS NOT your production environment.  It needs to be a separate
  # database where dbt will run against the test data that dtspec generates.  The name
  # of this environment needs to be the same as a target defined in dbt profiles.yml (in this case `dtspec`)
  dtspec:
    # Field names here follow the same conventions as dbt profiles.yml (https://docs.getdbt.com/dbt-cli/configure-your-profile)
    type: postgres
    host: "{{ env_var('POSTGRES_HOST') }}"
    port: 5432
    user: "{{ env_var('POSTGRES_USER') }}"
    password: "{{ env_var('POSTGRES_PASSWORD') }}"
    dbname: "{{ env_var('POSTGRES_DBNAME') }}_dtspec"

# A source environment is where source data is located.  It may be in the same database
# as the target environment or it may be different if the data warehouse supports it (e.g., Snowflake).
# It is also possible to define several source environments if your source data is spread
# across multiple databases.
source_environments:
  raw:
    # Use `tables` to specify source tables that need to be present to run tests.
    tables:
      # `wh_raw` is the name of a namespace (aka schema) in the `raw` source environment
      wh_raw:
        # tables may be listed indivdually (or, use `wh_raw: '*'` to indicate all tables within the `wh_raw` namespace)
        - raw_customers
        - raw_orders
        - raw_payments

    # In order to run tests, we need to replicate the table schemas in the test environment.
    # The schema section here contains credentials for a database where those tables are defined.
    # This is likely a production database (in your warehouse), or is a production replica.
    # dtspec only uses this database to read reflect the table schemas (via `dtspec db --fetch-schemas`).
    schema:
      type: postgres
      host: "{{ env_var('POSTGRES_HOST') }}"
      port: 5432
      user: "{{ env_var('POSTGRES_USER') }}"
      password: "{{ env_var('POSTGRES_PASSWORD') }}"
      dbname: "{{ env_var('POSTGRES_DBNAME') }}"
    # The test section contains credentials for a database where test data will be created.
    # Data in this database is destroyed and rebuilt for every run of dtspec and SHOULD NOT be
    # the same as the schema credentials defined above.
    test:
      type: postgres
      host: "{{ env_var('POSTGRES_HOST') }}"
      port: 5432
      user: "{{ env_var('POSTGRES_USER') }}"
      password: "{{ env_var('POSTGRES_PASSWORD') }}"
      dbname: "{{ env_var('POSTGRES_DBNAME') }}_dtspec"

  # Pretending snapshots are in a different database because Postgres doesn't support cross-db queries.
  # This is how you would do it if snapshots were in a different database than other raw source data.
  snapshots:
    tables:
      snapshots: '*'
    schema:
      type: postgres
      host: "{{ env_var('POSTGRES_HOST') }}"
      port: 5432
      user: "{{ env_var('POSTGRES_USER') }}"
      password: "{{ env_var('POSTGRES_PASSWORD') }}"
      dbname: "{{ env_var('POSTGRES_DBNAME') }}"
    test:
      type: postgres
      host: "{{ env_var('POSTGRES_HOST') }}"
      port: 5432
      user: "{{ env_var('POSTGRES_USER') }}"
      password: "{{ env_var('POSTGRES_PASSWORD') }}"
      dbname: "{{ env_var('POSTGRES_DBNAME') }}_dtspec"

Test environment setup

Once the configuration file has been defined, the next step is to fetch/reflect schemas for the source tables. From the dbt directory, run the following CLI command:

dtspec db --fetch-schemas

This will query all of the databases defined in the schema section of the source environments defined in dtspec/config.yml, and create table schema files in dtspec/schemas. The files in this directory should be committed to source control and updated whenever your source data changes (in so much as it would affect the dtspec tests).

Next, initialize the test databases defined in the test section of the source environments defined in dtspec/config.yml with the CLI command

dtspec db --init-test-db

This will create empty source tables in your test databases, ready to be loaded with test data.

Executing tests

In order to use dtspec with dbt, spec files must make use of the dbt_source and dbt_ref Jinja functions. These are analogous to the dbt source and ref functions. dtspec will compile your dbt project and use the dbt/target/manifest.json file to resolve the names of sources and targets that you want to test. For example, the SomeStudents factory would be written as follows if this were a dbt project:

factories:
  - factory: SomeStudents
    data:
      - source: {{ dbt_source('raw', 'raw_students') }}
        table: |
          | id | name   |
          | -  | -      |
          | 1  | Buffy  |
          | 2  | Willow |

and an expectation would be:

    cases:
      - case: HelloGang
        expected:
          data:
            - target: {{ dbt_ref('salutations') }}
              table: |
                | id | salutation   |
                | -  | -            |
                | 1  | Hello Buffy  |
                | 2  | Hello Willow |

With these references set, dtspec tests can be executed via the CLI command:

dtspec test-dbt

This command will do the following:

1. It will first compile your dbt project. If your dbt code does not change between dtspec tests, you may skip this step by pass the --partial-parse argument.
2. The dtspec spec files are compiled into a single document and dbt references are resolved. The compiled dtspec document is output to dtspec/compiled_specs.yml, which does not need to be saved to source control.
3. Source data is generated and loaded into the test databases.
4. dbt is executed against the test database.
5. The models that dbt built in the target test environment are extracted. These are the "actuals".
6. The actuals are compared with the expected data as specified in the dtspec specs.

The test-dbt command has several options that may be useful. See dtspec test-dbt -h for a full list, but here are some noteworthy options:

• --models specifies the models that dbt should run, using standard dbt model selection syntax.
• --scenarios is used to restrict the number of scenarios that are tested. The argument is a regular expression that will match on the compiled Scenario name. This can be used in combination with the --models command to only run those tests and models that you're concerned with.

Additonal CLI notes

Log level

If you want to see more detailed loggin information, set the DTSPEC_LOG_LEVEL environment variable (options are DEBUG, INFO, WARN, and ERROR). For example:

DTSPEC_LOG_LEVEL=INFO dtspec test-dbt

Project location

If you really don't want to put dtspec in the dbt project directory you can override the default by setting DTSPEC_ROOT and DBT_ROOT environment variables that point to the root path of these projects.

Special Values

When dtspec is run via the CLI, it recognizes nulls and booleans in the spec files. To indicate these kinds of values in a dtspec spec, use {NULL}, {True}, and {False}. For example:

    cases:
      - case: HelloGang
        expected:
          data:
            - target: {{ dbt_ref('salutations') }}
              table: |
                | id | salutation   | is_witch |
                | -  | -            | -        |
                | 1  | Hello Buffy  | {False}  |
                | 2  | Hello Willow | {True}   |
                | 3  | Hello NA     | {NULL}   |

Jinja context

When writing spec files that will be parsed with the dtspec CLI, the following functions are available in the jinja context:

• datetime -- This is the Python datetime.datetime type
• date -- This is the Python datetime.date type
• relativedelta -- This is the Python relativedelta type0
• UTCNOW -- The UTC datetime value at the time the specs are parsed
• TODAY -- The current UTC date value at the time the specs are parsed
• YESTERDAY -- Yesterday's date
• TOMORROW -- Tomorrow's date
• dbt_source -- Used to reference dbt sources
• dbt_ref -- Used to reference dbt models

Some example of using these functions:

- source: raw_products
   table: |
    | export_time                          | file                    | product_id | product_name |
    | -                                    | -                       | -          | -            |
    | {{ YESTERDAY }}                      | products-2021-01-06.csv | milk       | Milk         |
    | {{ TODAY - relativedelta(days=5) }}  | products-2021-01-02.csv | milk       | Milk         |

Additional notes about dtspec

• At the moment, all source data values are generated as strings. It is up to the the user to enforce data types suitable to their data transformation system. Note that the dtspec dbt CLI commands handle this for Postgres and Snowflake warehouses.
• Additionally, data expectations are stringified prior to running assertions.

Contributing

We welcome contributors! Please submit any suggests or pull requests in Github.

Developer setup

Create an appropriate python environment. I like miniconda, but use whatever you like:

conda create --name dtspec python=3.8
conda activate dtspec

Then install pip packages

pip install pip-tools
pip install --ignore-installed -r requirements.txt

run tests via

inv test

and the linter via

inv lint