Skip to content

The Data Observation Toolkit (DOT) can be used to monitor data in order to flag problems with data integrity and scenarios that might need attention.

License

Notifications You must be signed in to change notification settings

datakind/Data-Observation-Toolkit

Repository files navigation

The Data Observation Toolkit (DOT)

The Data Observation Toolkit (DOT) can be used to monitor data in order to flag problems with data integrity and scenarios that might need attention. Typical tests include checks for missing/duplicate and inconsistent data, outliers, and domain-specific signals such as a missed follow-up medical treatment after initial diagnosis.

DOT includes a simple user interface for configuring the powerful DBT and Great Expectations libraries, as well as a database for storing and classifying data monitoring results.

The ultimate aim of DOT is to provide easier access to data monitoring without the need for deep technical skills.

ER diagram

In 2019, the United Nations Statistical Commission stated: “Every misclassified or unrecorded death is a lost opportunity to ensure other mothers and babies do not die in the same way. When it comes to health, better data can be a matter of life and death.” DataKind developed DOT in response to a demand for tool to increase trust in public health data which underpins the ability of a health system to provide equitable, data-driven, optimized health service delivery and improve policy response.

DOT was developed with our valuable and globally distributed frontline health partners working to strengthen frontline health systems, expertise from Ministries of Health, frontline health workers and funders.

Read more here:

Pathways to Increasing Trust in Public Health Data

Empowering Health Worker and Community Health Systems: Data Integrity and the Future of Intelligent Community Health Systems in Uganda

Harnessing the power of data science in healthcare

How Data Empowers Health Workers—and Powers Health Systems

Quick start

The fastest way to get started using DOT is to use the Docker environment and demo data provided.

Prerequisites

  1. You will need to install Docker desktop. First make sure you have checked the Docker prerequisites. We recommend using at least 4GB memory which can be set in the docker preferences, but this can vary depending on the volume of data being tested
  2. If running on a Mac M1/M2 chip, install Rosetta and set export DOCKER_DEFAULT_PLATFORM=linux/amd64 in the terminal where you will run the instructions below

Setup

You will need Python and some modules installed to run the DOT quick-start demo:

  1. Install Python 3.8.9

  2. Install the necessary python packages by running the following commands in your terminal (Additional information Mac/Linux terminal, additional information Windows terminal):

    • pip install gdown
    • pip install python-on-whales

Alternatively, you can use the provided environment.yml if you have miniconda installed.

Running the DOT quick-start demo environment

  1. Start Docker
  2. Open a Terminal and cd into the docker sub-folder of where you installed DOT
    • Example Windows: cd C:<PATH TO WHERE YOU INSTALLED DOT>\Data-Observation-Toolkit\docker
    • Example Mac/Linux: cd <PATH TO WHERE YOU INSTALLED DOT>/Data-Observation-Toolkit/docker/
  3. When on Mac/Linux, run python3 run_demo.py or when on Windows, run python3 .\run_demo.py
  4. The script will start DOT and open a browser with the DOT User Interface. See demo video below for how to get started.
  5. Once you are done with the demo, press return in your Terminal, to stop the docker containers

Video Demo of DOT

Note: Be sure to activate audio ...

A.quick.demo.of.DOT.mp4

The not-so-quick start, setting up DOT

Building the DOT Docker environment

If you wish to build DOT yourself ...

  1. Save a postgres password to your local environment export POSTGRES_PASSWORD=<some password you decide>
  2. Open a terminal and change directory into the docker sub-directory of DOT: cd ./docker
  3. Build the docker containers: docker compose build
  4. Start them: docker compose up -d

Your environment is now running!

Running DOT

  1. docker exec -it dot /bin/bash
  2. cd dot
  3. python3 ./run_everything.py --project_id 'ScanProject1'

This will run the sample tests against the demo data and save results to the DOT database.

Setting up the DOT User Interface

The above steps will start the DOT user interface, which will allow you to manage DOT and see results. You need to do a few steps the first time you use it ...

  1. Go to http://localhost:82/
  2. Click the button and register to create a login (keep note of the password)
  3. Once created, click 'Build my own' on the next popup
  4. Click app smith icon top-left to go back tom homepage
  5. Top-right next to the new button, click on the '...' and select import
  6. Select Import from file and navigate to file ./docker/appsmith/DOT App V2.json
  7. You will be prompted to enter details for the database connection. You can set these as required, but if using the DOT dockerized Postgres database, the parameters are:
    • Host address: dot-db
    • Port: 5432
    • Database name: dot_db
    • Authentication > User name: postgres
    • Authentication > Password: <THE PASSWORD YOU USED WHEN BUILDING DOT>

You should now see the DOT user interface in developer mode (ie you could edit it). To run in end-user mode:

  1. Click button top-right click the 'Deploy' button. This should open a new tab with the user interface in all its glory!

Note: If you want to remove Appsmith information on the deployed app, add ?embed=True at the end of the deployed app URL.

Viewing test results

Your test results will be in the dot-db container. You can view the results by opening a shell in the dot-db container:

docker exec -it dot-db /bin/bash

Then running the psql client locally in that container:

psql -U postgres -d dot_db

Or if you prefer a database client (like DBeaver), you can use their settings:

host=localhost
port=5433 
database=dot_db
user=postgres
password=<the POSTGRES_PASSWORD you set above>

Note: The host and port are set in the docker-compose.yml

To see some raw results you can run SELECT * from dot.test_results LIMIT 100;. Some more advanced queries are provided below.

Some useful DOT database queries

The following queries might be useful for visualizing the DOT DB schema and data:

Statistics on failed tests

SELECT
   tr.run_id,
   ct.test_type,
   COUNT(*)
FROM
   dot.test_results tr,
   dot.configured_tests ct 
WHERE
   tr.test_id = ct.test_id
GROUP BY
   tr.run_id,
   ct.test_type
ORDER BY
   ct.test_type

Same, but adding in test description and entity names in grouping ...

SELECT
   tr.run_id,
   tr.view_name, 
   ct.test_type,
   ct.description,
   ce.entity_name,
   COUNT(*)
FROM
   dot.test_results tr,
   dot.configured_tests ct,
   dot.configured_entities ce 
WHERE
   tr.test_id = ct.test_id AND
   ce.entity_id = ct.entity_id
GROUP BY
   tr.run_id,
   tr.view_name, 
   ct.test_type,
   ct.description,
   ce.entity_name
ORDER BY
   tr.run_id,
   tr.view_name, 
   ct.test_type,
   ct.description,
   ce.entity_name

Linking DOT Data scenarios with configured tests

SELECT 
   s.*,
   ct.*
FROM
   dot.scenarios s,
   dot.configured_tests ct 
WHERE 
   s.scenario_id=ct.scenario_id

Linking configured tests to test results

SELECT 
   s.*,
   ct.*,
   ce.entity_name,
   tr.*
FROM
   dot.scenarios s,
   dot.configured_tests ct,
   dot.test_results_summary tr,
   dot.configured_entities ce
WHERE 
   s.scenario_id=ct.scenario_id AND
   tr.test_id=ct.test_id AND
   ce.entity_id = ct.entity_id
LIMIT 10

Deactivating all tests except one dbt and one great expectation

update dot.configured_tests set test_activated=false where test_id not in ('7db8742b-c20b-3060-93e2-614e35da2d4b','0f26d515-a70f-3758-8266-8da326d90eb6')

Seeing the raw data for failed tests

dot.test_results column view_name provides the name of the DB view which holds the data for failed tests. Additionally, columns id_column_name and id_column_value provide the columns to match in the DB entity view, ie the data that was tested. Finally, you can query to get the underlying data for each test using function get_dot_data_record

SELECT 
   tr.test_id,
   tr.status,
   dot.get_test_result_data_record(ce.entity_name, tr.id_column_name, 
   tr.id_column_value,'public_tests')
FROM
   dot.scenarios s,
   dot.configured_tests ct,
   dot.configured_entities ce,
   dot.test_results tr
WHERE 
   s.scenario_id=ct.scenario_id AND
   tr.test_id=ct.test_id and 
   ce.entity_id=ct.entity_id
LIMIT 10

Where the function parameters are:

  • Test entity name
  • Test result ID column name (in entity view)
  • Test Result ID column value
  • Test results schema name

This returns a json record for the data that was tested. Note: If using the airflow environment, change public_tests to the schema where the data is, for example data_flights_db.

Configuring DOT

The following sections provide instructions for adding entities and tests to DOT directly in the DOT database. You can also use the DOT user interface for tests, for more details please see section see section The DOT User Interface. If you want to test DOT with health related data, please feel free to use this [fabricated dataset] (https://docs.google.com/spreadsheets/d/1l6inpa6ykgUewC-MJkgrQwwc8HjUiTLLJFEDBtAMDB4/edit#gid=31188808) which resembles appointment information and contains some quality issues which DOT can highlight.

How to add new entities

The DOT will run tests against user-defined views onto the underlying data. These views are called "entities" and defined in table dot.configured_entities:

Column Description
entity_id Name of the entity e.g. ancview_danger_sign
entity_category Category of the entity e.g. anc => needs to be in dot.entity_categories
entity_definition String for the SQL query that defines the entity

For example, this would be an insert command to create ancview_danger_sign:

INSERT INTO dot.configured_entities (project_id,entity_id,entity_category,entity_definition,date_added,date_modified,last_updated_by) VALUES('Project1', 
'ancview_danger_sign', 'anc', '{{ config(materialized=''view'') }}
{% set schema = <schema> %}

select *
from {{ schema }}.ancview_danger_sign');

All entities use Jinja macro statements - the parts between { ... } - which the DOT uses to create the entity materialized views in the correct database location. Use the above format for any new entities you create.

The SQL for the view definition can be more complex than the example above, combining data from multiple underlying tables or views. For example:

{{ config(materialized=''view'') }}
{% set schema = <schema> %}

select ap.*,
        ap.lmp as lmp_date,
        DATE_PART(''day'', reported - lmp) as days_since_lmp
from {{ schema }}.ancview_pregnancy ap

When you add a new entity to the configuration, take a look to the existing dot.entity_categories to associate your new entity to one of them. If you need to add a new category, the table dot.entity_categories has the following columns:

Column Description
entity_category Category of the entity e.g. anc
description A description of the category

An example of an insert statement would be: INSERT INTO dot.entity_categories VALUES('anc', 'antenatal care');

How to configure tests

Tests are defined in the table dot.configured_tests. Each test has an associated test_type, a list of which can be found in table dot.test_types (see section "DOT Database Schema" below for more details on the full schema).

To use one of these test types for a new test, insert a new row in dot.configured_tests.

Here are the columns included in a test:

Column Description
test_activated Whether the test is activated or not
project_id ID of the project, for example "ScanProject1"
test_id UUID of the test
scenario_id ID of the scenario
priority Priority level
description Description of the test
impact Why the test is important
proposed_remediation How the issue in this test might be solved
entity The entity against which the test runs (check entities below)
test_type Test type
column_name The column in the table against which the test runs
column_description Description of the above column
test_parameters Any parameters the test takes in
date_added Date when the test was added
date_modified Date when the test was last modified
last_updated_by Person who last updated the test

The UUID in the above example will get overwritten with an automatically generated value. Also, your test must be unique for the project. If you get a key violation it's probably because that test already exists.

Test validation

Any insert of update of configured tests will call database function dot.test_validation, as defined in ./db/dot/1-schema.sql. This function performs some basic validation to ensure test parameters are in an expected format. It is not infallable, if there are any issues with test parameters and tests do not execute, you can confirm this by looking at columns test_status and test_status_message in dot.test_results_summary.

Example INSERT statements for adding a new test for each test type:

Note: In all of the following examples, the UUID in the insert statement will be replaced with an automatically generated one.

  1. relationship

    'INSERT INTO dot.configured_tests VALUES(TRUE, 'ScanProject1', '0cdc9702-91e0-3499-b6f0-4dec12ad0f08', 'ASSESS-1', 3, '', '', 
    '', 'ancview_pregnancy', 'relationships', 'uuid', '', 
    $${"name": "danger_signs_with_no_pregnancy", "to": "ref('dot_model__ancview_danger_sign')", "field": "pregnancy_uuid"}$$, 
    '2021-12-23 19:00:00.000 -0500', '2021-12-23 19:00:00.000 -0500', 'your-name');
    
  2. unique

    INSERT INTO dot.configured_tests VALUES(TRUE, 'ScanProject1', '52d7352e-56ee-3084-9c67-e5ab24afc3a3', 'DUPLICATE-1', 3, '', 
    '', '', 'ancview_pregnancy', 'unique', 'uuid', 'alternative index?', '', 
    '2021-12-23 19:00:00.000 -0500', '2021-12-23 19:00:00.000 -0500', 'your-name');
    
  3. not_negative_string_column

    INSERT INTO dot.configured_tests VALUES(TRUE, 'ScanProject1', '8aca2bee-9e95-3f8a-90e9-153714e05367', 'INCONSISTENT-1', 3, 
    '', '', '', 'ancview_pregnancy', 'not_negative_string_column', 'patient_age_in_years', '', 
    $${"name": "patient_age_in_years"}$$, '2021-12-23 19:00:00.000 -0500', '2021-12-23 19:00:00.000 -0500', 'your-name');
    
  4. not_null

    INSERT INTO dot.configured_tests VALUES(TRUE, 'ScanProject1', '549c0575-e64c-3605-85a9-70356a23c4d2', 'MISSING-1', 3, '', 
    '', '', 'ancview_pregnancy', 'not_null', 'patient_id', '', '', '2021-12-23 19:00:00.000 -0500', '2021-12-23 19:00:00.000 -0500', 'your-name');
    
  5. accepted_values

    INSERT INTO dot.configured_tests VALUES(TRUE, 'ScanProject1', '935e6b61-b664-3eab-9d67-97c2c9c2bec0', 'INCONSISTENT-1', 3, 
    '', '', '', 'ancview_pregnancy', 'accepted_values', 'fp_method_being_used', '', 
    $${"values": ['oral mini-pill (progestogen)', 'male condom', 'female sterilization', 'iud', 'oral combination pill', 'implants', 'injectible']}$$, 
    '2021-12-23 19:00:00.000 -0500', '2021-12-23 19:00:00.000 -0500', 'your-name');
    
  6. possible_duplicate_forms

    INSERT INTO dot.configured_tests VALUES(TRUE, 'ScanProject1', '7f78de0e-8268-3da6-8845-9a445457cc9a', 'DUPLICATE-1', 3, '', 
    '', '', 'ancview_pregnancy', 'possible_duplicate_forms', '', '', 
    $${"table_specific_reported_date": "delivery_date", "table_specific_patient_uuid": "patient_id", "table_specific_uuid": "uuid"}$$, '2021-12-23 19:00:00.000 -0500', '2021-12-23 19:00:00.000 -0500', 'your-name');
    
  7. associated_columns_not_null

    INSERT INTO dot.configured_tests VALUES(TRUE, 'ScanProject1', 'd74fc600-31c3-307d-9501-5b7f6b09aff5', 'MISSING-1', 3, '', 
    '', '', 'ancview_pregnancy', 'associated_columns_not_null', 'diarrhea_dx', 'diarrhea diagnosis', 
    $${"name": "diarrhea_dx_has_duration", "col_value": True, "associated_columns": ['max_symptom_duration']}$$, 
    '2021-12-23 19:00:00.000 -0500', '2021-12-23 19:00:00.000 -0500', 'your-name');
    
  8. expect_similar_means_across_reporters

    INSERT INTO dot.configured_tests VALUES(TRUE, 'ScanProject1', '0cdc9702-91e0-3499-b6f0-4dec12ad0f08', 'BIAS-1', 3, 
    'Test for miscalibrated thermometer', '', '', 'ancview_pregnancy', 'expect_similar_means_across_reporters', 
    'child_temperature_pre_chw', '', '{"key": "reported_by","quantity": "child_temperature_pre_chw",
    "form_name": "dot_model__iccmview_assessment","id_column": "reported_by"}', '2022-01-19 20:00:00.000 -0500', 
    '2022-01-19 20:00:00.000 -0500', 'your-name');
    
  9. expression_is_true

    INSERT INTO dot.configured_tests VALUES(TRUE, 'ScanProject1', '3081f033-e8f4-4f3b-aea8-36f8c5df05dc', 'INCONSISTENT-1', 3, 
    'Wrong treatment/dosage arising from wrong age of children (WT-1)', '', '', 'ancview_pregnancy', 
    'expression_is_true', '', '', 
    $${"name": "t_under_24_months_wrong_dosage", "expression": "malaria_act_dosage is not null", "condition": "(patient_age_in_months<24) and (malaria_give_act is not null)"}$$, 
    '2022-02-14 19:00:00.000 -0500', '2022-02-14 19:00:00.000 -0500', 'your-name');
    
  10. custom_sql

    Custom SQL queries require special case because they must have primary_table and primary_table_id_field specified within the SQL query as shown below:
    INSERT INTO dot.configured_tests VALUES(TRUE, 'ScanProject1', 'c4a3da8f-32f4-4e9b-b135-354de203ca90', 'TREAT-1', 6, 'Test for new family planning method (NFP-1)', '', '', 'ancview_pregnancy', 'custom_sql', '', '',
    

format('{%s: %s}', to_json('query'::text), to_json($query$ select a.patient_id, a.reported, a.fp_method_being_used, 'dot_model__fpview_registration' as primary_table, 'patient_id' as primary_table_id_field from {{ ref('dot_model__fpview_registration') }} a inner join ( select distinct patient_id, max(reported) reported from {{ ref('dot_model__fpview_registration') }} where fp_method_being_used in ('vasectomie','female sterilization') group by patient_id ) b on a.patient_id = b.patient_id and a.reported > b.reported and fp_method_being_used not in ('vasectomie','female sterilization') and fp_method_being_used not like '%condom%' $query$::text) )::json,'2021-12-23 19:00:00.000 -0500', '2021-12-23 19:00:00.000 -0500', 'Leah');
```

Activating/Deactivating existing tests

Note that it's possible to deactivate tests, which can sometimes be useful for testing. To do this simply set test_activated=False in table dot.configured_tests. For example:

update dot.configured_tests set test_activated=false where test_id not in ('7db8742b-c20b-3060-93e2-614e35da2d4b','0f26d515-a70f-3758-8266-8da326d90eb6');

Developing the Appsmith UI

Clicking 'Edit' on the imported app gives you the ability to edit the app. See the appsmith help site for more information on this.

Key things to note:

  • All validation, defaulting, tool tips are controlled via the Database as much as possible. See the 'Queries' section
  • Of particular note is that the test_parameters JSON generates a form. This was done once, be aware it will remove many settings if you click regenerate. I have a feature request open with appsmith to create fields with default types, values and validation functions, but for now we are adding new fields manually. Note that you might be able to reinstate the logic en masse by looking at the raw appsmith app JSON file

Updating to latest version of Appsmith

Appsmith release bug fixes and enhancements. By default the DOT docker build will do this automatically. To do this manually:

  1. In ./docker run docker compose -f docker-compose-with-appsmith-ui.yml stop
  2. docker rm appsmith
  3. In a clean directory: curl -L https://bit.ly/32jBNin -o $PWD/docker-compose.yml
  4. docker-compose pull && docker-compose rm -fsv appsmith && docker-compose up -d
  5. Back in ./docker run docker compose -f docker-compose-with-appsmith-ui.yml up -d

Note, these will remove your saved configuration and app, so be sure to make note of these before doing the above steps.

Advanced topics

The DOT DB Schema

The DOT Database schema is defined as follows:

ER diagram

Tables are defined as follows:

Table Description
projects Defines the projects the DOT can be run for. Typically a project is associated with a single source database where the data resides
configured_entities Holds entities, which are the views onto underlying data the DOT will test
entity_categories Category of each entity, for example "Pre-natal care". Useful for segmenting DOT results by category.
configured_tests the table which defines what tests to run for each project
scenarios The DOT Taxonomy scenario for each test
test_types The available test types for each test, eg null, unique, custom_sql
scenario_test_types The test types that apply for any given scenario
test_parameters_interface JSON object defining Interface parameters required for each test type. Note: Not currently used but will be in future
run_log Log of DOT runs, with stop/start times and failure message
test_results Main test results table, indicating test fails
test_results_summary Aggregated test results for each run
remediation_log Record of remediation actions. (Note: Not yet implemented)

The test_results table includes fields to link back to the data tested. This can be a fixed table or view, but also a custom SQL query. Given this, there is a useful Postgres function which will retreive a JSON record for the data tested, see 'Seeing the raw data for failed tests' above.

Please refer to CONTRIBUTING.md for information on more complex configuration options.

How to visualize the results using Superset

At time of writing Superset setup isn't fully automated, so once your Docker containers are running the first time, run these commands:

docker exec -it superset superset fab create-admin  --username admin  --firstname Superset  --lastname Admin   --email admin@superset.com  --password admin
docker exec -it superset superset db upgrade
docker exec -it superset superset init

You can ignore the warnings about CACHE_TYPE being null.

Once complete, open the following URL in a browser:

http://localhost:8080/login/

Note: Some of the team have observed that Google Chrome works a better than Safari.

... and login with admin/admin.

Once in, connect to the DOT database:

  1. In super select data > databases from top menu
  2. On the right click the '+ Database' button
  3. Choose PostgresSQL
  4. Use these DB parameters:
    • Host: dot_db
    • Database: dot_db
    • Display name: dot_db
    • Port: 5432
    • User: postgres
    • Password:

Once set up, you can create datasets for any of the DOT tables ...

  1. Select *Data > Datasets+
  2. Click '+ DATASET' button
  3. Select the tables(s) you want from the DOT database, for example test_results

Now you have datasets, you can create charts

IMPORTANT! If superset dashboards and using Docker, be sure to export your dashboard so you don't lose work if rebuilding your Docker environment.

Deploying in Airflow

The following setup is a proof-of-concept to illustrate how Airflow might be configured to run the DOT on multiple databases. It is configured to copy data from the data DB back into the DOT DB, in production the flow would be to copy data from the source production db into the DOT DB data_ schema.

Configuring/Building Airflow Docker environment

  1. cd ./docker 2export POSTGRES_PASSWORD=<**Some password you will use to access DOT DB**> 3docker compose -f docker-compose-with-airflow.yml build 4docker compose -f docker-compose-with-airflow.yml up airflow-init 5docker compose -f docker-compose-with-airflow.yml up -d 6docker exec -it docker-airflow-worker-1 /bin/bash 7cd /app/dot && ./install_dot.sh

Note: If using Docker on AWS, you might need to use docker-compose instead of docker compose.

Now, to set up DB connection for the DOT ...

  1. To test: Connect to the DOT DB as mentioned above in section 'Testing your database connection'
  2. Go to: http://localhost:8083/ and log in with airflow/airflow
  3. Next, create a copy of the DOT DB to be used as the source database. Open a SQL session and run ...

SELECT pg_terminate_backend(pg_stat_activity.pid) FROM pg_stat_activity WHERE pg_stat_activity.datname = 'dot_db' AND pid <> pg_backend_pid();

CREATE DATABASE dot_data WITH TEMPLATE dot_db OWNER postgres;

  1. In top menu select Admin > Connections, click +
  2. Enter the Docker DOT DB details as follows:
  • Conn Id: dot_db
  • Conn Type: Postgres
  • Host: dot_db
  • Schema: dot
  • Login: postgres
  • Password: Whatever you used when building the DOT docker environment
  • Port: 5432
  1. Set up new connection in Airflow as follows:
  • Conn Id: dot_data
  • Conn Type: Postgres
  • Host: dot_db
  • Schema: public
  • Login: postgres
  • Password: Whatever you used when building the DOT docker environment
  • Port: 5432

See section below for how to run DOT

Starting Airflow docker environment

NOTE: You might need to use docker-compose on some hosts.

docker compose -f docker-compose-with-airflow.yml up -d

Stopping Airflow docker environment

docker compose -f docker-compose-with-airflow.yml stop

Removing Airflow docker environment

docker compose -f docker-compose-with-airflow.yml down -v

Running the DOT in Airflow (Demo)

A DAG has been included which copies data from the uploaded DB dump into the DOT DB 'data_ScanProject1' schema, and then runs the toolkit against this data. To do this ...

  1. Go to http://localhost:8083
  2. Log in with airflow/airflow
  3. Click on 'DAGs' in top menu
  4. Search for 'dot' in top-right search box (be sure to hit return twice)
  5. Toggle DAG 'run_dot_project' to Active (far left toggle on row)
  6. Click play icon under 'Actions', this will start the run

To see progress, and/or logs ...

  1. Click on one of the circles under 'Runs', eg 'running', or 'completed'
  2. Click on top 'Dag Id'
  3. Click on task and then 'Log', you should now see logs

To use Airflow CLI for debugging a task (assuming you are getting object ancview_danger_sign):

  1. docker exec -it docker-airflow-worker-1 /bin/bash
  2. airflow tasks test run_dot_project sync_object_flight_data 2022-03-01

Or to run just DOT stage ...

airflow tasks test run_dot_project run_dot 2022-03-01

Running the DOT in Airflow (Connecting to external databases)

The following instructions illustrate how to use a local airflow environment, connecting with external databases for the data and DOT.

NOTE: These are for illustrative purposes only. If using Airflow in production it's important that it is set up correctly and does not expose a http connection to the internet, and also has adequate network security (firewal, strong password, etc)

  1. Edit [./dot/dot_config.yml] and set the correct parameters for your external dot_db
  2. Create a section for your data databases and set connection parameters
  3. If you have a DAG json file dot_projects.json already, deploy it into ./airflow/dags
  4. Run steps 1-11 in [Configuring/Building Airflow Docker environment](#Configuring/Building Airflow Docker environment)
  5. Run steps 12 and 13, but use the values for your external databases you configured in dot_config.yml

You will need to configure DOT tests and the DAG json file appropriately for your installation.

Adding more projects

If configuring Airflow in production, you will need to adjust ./docker/dot/dot_config.yml accordingly. You can also add new projects by extending the projects array in the Airflow configuration JSON

How the DOT works

The Data Observation Toolkit (DOT) uses the open source projects DBT and Great Expectations as a framework for the definition, development and execution of the data tests.

Most tests are developed on DBT, leveraging its support for base tests for column not nulls, foreign key checks, and a few more, all of which can be expressed as entries on a YAML file. Additionally, making use of its custom tests, many complex tests involving a set of columns of one or two tables can be expressed using DBT, written with SQL/Jinja statements.

Great Expectations adds the possibility of implementing tests in python, and are used for checks that cannot be expressed in SQL, for example tests that inspect the statistical distribution of the data.

DBT

Directory structure

The DBT folder contains DBT models, macros, custom tests, and project files. Most of these files are actually written by the DOT itself from the configured entities and tests in the dot schema, but note that any DBT command needs the actual files as inputs.

  • Core subdirectories

    • macros: Jinja macros to support organized code and custom schema tests
    • models: This is the core of the DBT project; it contains both SQL for models and yml files for schema tests, which are written by the tool from dot.configured_entities, for the SQL for models, and from dot.configured_tests, for the tests' yml files
    • tests: Custom SQL tests, written by the tool from dot.configured_tests, for the test of type "custom_sql"
  • Non-version controlled directories (that are created via the dbt runs)

    • target: Compiled sql used to create models and tests
    • dbt_modules: DBT plugins (e.g. dbt_utils) installed via dbt deps
    • logs: Logs created by dbt (not used elsewhere in this stack)
  • Scripts

    • dbt.py: Utilities that manage the run of DBT jobs, which create views containing rows flagged by failing tests, creates a test coverage reports of the DBT side of the stack and updates tests results to the DB
  • Config files

    • dbt_project.yml: Defines paths to important DBT assest (including those listed above) and project-specific variables -not in version control, it is either generated by the tool or copied from the project-specific configuration
    • packages.yml: Indicates DBT plugins used

Running tests

Commands for reference -these are actually run by the tool.

  • Run all tests: dbt test -m core
  • Run all tests for patients only: dbt test -m patient
  • Run schema tests for patients only dbt test --schema -m patient
  • Run data tests for patients only dbt test --data -m patient

Updating models

The models in the core folder are a dump of the DB dot.configured_entities. Right now, most of the entites are just full table copies of the raw data. Reference commands:

  • Update all models: dbt run -m core

Writing new tests

Since we'll have lots of tests we want to run, it's important to keep them organized. Wherever possible, schema tests should be implemented with simple test types in dot.configured_tests (such as not_null, unique, relationships, accepted_values, etc, that go to a yml associated to the the schema.yml) because it minimizes the need for adding lots of additional SQL. You may find the (already set up on this project) dbt_utils extension useful for writing more complex versions of these kinds of tests.

For tests requiring custom SQL, a new custom_sql row should be created in dot.configured_tests, which in turn creates a file in tests/core.

For more details on implementing tests in dbt please refer to the dbt section in the DOT.

Making sense of failed test results

The DOT runs a number of steps for generating test results:

  • it runs dbt test to run the tests
  • it autogenerates model code for only the tests that have failed; these models will produce views when run
  • the end results will be a new set of data sources (prefixed with "tr_") containing all the rows violating a test
  • these failing rows are also loaded into the table dot.test_results

You should use run_everything.sh as a quick way to run the right dbt commands in sequence and guarantee that the right dependencies/assumptions are in place to make things work.

Note that the code generated for these models is gitignored because they're based on compiled SQL and contain specific database names instead of dbt references.

It should also be noted that DBT does support auto-generated documentation out of the box. HOWEVER, these docs are fairly limited in their usefulness for this project, especially since they don't provide a way to find the specific records that are causing a test to fail. This is why we have written new code to generate report that better address the community health DOT use case.

In the event that you do want to explore the native DBT docs, they can be created via dbt docs generate. To view them, running dbt docs serve will fire up a local web server so you can explore them in a browser.

Boilerplate Resources

These were listed automatically by dbt init and are helpful general resources for learning more about DBT

  • Learn more about dbt in the docs
  • Check out Discourse for commonly asked questions and answers
  • Join the chat on Slack for live discussions and support
  • Find dbt events near you
  • Check out the blog for the latest news on dbt's development and best practices

Great Expectations

Directory structure

The great_expectations folder is a Great Expectations project.

  • Core directories

    • expectations: JSON files indicating the tests to be run (similar to yml files containing schema tests in DBT); these are not in version control, but generated by the tool from dot.configured_tests
    • notebooks: Jupyter notebooks automatically created by Great Expectations during setup to allow for a more convenient front-end to edit the JSON files in expectations (check out the flow described by the Great Expectations documentation)
    • plugins: Additional code for customizing this Great Expectations project. Most important in here is custom_expectations.py, which is where tests requiring arbitrary python should be added as methods under the CustomSqlAlchemyDataset class (somewhat similar to the custom SQL tests in DBT, except written in python).
  • Non-version controlled directories (automatically created by Great Expectations)

    • uncommitted: Generic place to capture all Great Expecations files that should not be version controlled, including logs and database connection details

    • Scripts

      • great_expectations.py: Utility script to automatically run GE tests and create coverage reports. GE's CLI commands for this are much more verbose and more difficult to remember than DBTs.
    • Config files -not in version control, these are either managed by the tool or set in the project-dependent config

      • batch_config.json: Defines datasources, test suites, and tables to be included when running Great Expectations
      • great_expectations.yml: Main config file for Great Expectations (similar to dbt_project.yml)

Terminology

  • An expectation is a particular function accepting one or multiple parameters (defined in Python)
  • A test is an instance of an expectation, with a specific set of parameters (defined in a JSON file)
  • Out Of The Box (OOTB) expectations are provided by Great Expectations and built-in with the library's codebase.

Structuring tests

  • Tests are defined in JSON, akin to dbt schema tests in yaml

    • These live in a suite called great_expectations/expectations/<FILE>.json
  • Tests could be used OOTB, but most of the time are written custom

    • Custom expectations can operate on any tables passed as a parameter, but OOTB expectations will only be applied on the selected table in batch_config.json (see extra notes below for details)
    • OOTB tests can use views defined in DBT
    • OOTB tests can be defined directly in the JSON file
    • Custom expectations need to be added as decorated methods in plugins/custom_expectations.py
      • Custom tests can run arbitrary python/pandas (even though this isn't well-documented in the GE published docs)
        • Once added to in custom_expectations.py, tests can be defined in the JSON file similarly to OOTB tests
      • OOTB tests have a variety of outputs and therefore might not conform to the format expected by the Data Integrity framework -whenever possible, use DBT tests instead

If a mix of OOTB and custom expectations are needed, it is suggested to keep them in two suites of tests to manage their differences efficiently

Extra notes

The Data Observation Toolkit (DOT) works with a few assumptions in terms of what an expectation should accept and return.

  1. We create views out of the DOT results with Postgresql-specific syntax. If you're using any other database engine, please adapt the query in great_expectations.py.

  2. An expectation accepts both column names and table names as arguments. Great Expectations generally has table-agnostic suites running on specific single tables, but we're changing this model a bit because data integrity queries often depend on more than one table. Therefore, a default empty dataset is added in the batch_config.json for all custom expectations, and a relevant table name should be passed to the expectation in the suite definition. The default dataset won't be read at all and is used as a placeholder.

  3. Custom expectations are found in custom_expectations.py under plugins, it is recommended to follow their format and to add your own custom expectations as methods of that same class.

  4. The toolkit's post-processing step expects a few specific field in the output of the expectations (refer to example custom expectations to see how they're implemented)

Developing DOT

The DOT is open source so we encourage any development is done as part of that intiative. For reference this would include code quality standards and self-tests, some details of which are provided below.

Self-tests

The DOT runs some self-tests to check that its functions are running fine. If you are contributing to the project by adding a new feature or solving a bug, please follow the following guidelines:

  • find the existing tests applicable to the area that you are modifying
  • add a few more relevant tests for your bugfix/improvement
  • implement your changes until the test pass

Running self-tests

Using Docker

The Docker build provided above to run DOT, also includes self-test. So once you have the container running, all you need to do is ..

  1. exec -it dot /bin/bash
  2. pytest dot/self_tests/unit
On your local machine

Assuming you would like to run the tests locally, as preparation steps, you will need to:

  • Create a local env for python via either venv or conda
  • Make sure your Python version aligns with that in .github/workflows/lint.yml
  • Create a conda environment using the environment.yml file at the root directory
  • Prepare a postgres database that the tests can use (e.g. you can deploy a docker container and use it as a database only, or you could use a local instance of a Postgres DB)
  • Prepare a dot_config.yml at directory dot/self_tests/data/base_self_test with the same structure as the dot_config.yml for the DOT; should look like something as follows (note that the config below points to DB in the docker container):
dot:
  save_passed_tests: False
  output_schema_suffix: tests
dot_db:
  type: postgres
  host: localhost
  user: postgres
  pass: "{{ env_var('POSTGRES_PASSWORD') }}"
  port: 5433
  dbname: dot_db
  schema: self_tests_dot
  threads: 4
ScanProjec1_db:
  type: postgres
  host: localhost
  user: postgres
  pass: "{{ env_var('POSTGRES_PASSWORD') }}"
  port: 5433
  dbname: dot_db
  schema: self_tests_public
  threads: 4

And finally you can run the tests from a terminal as follows:

pytest dot/self_tests/unit

Guidelines for adding new tests

  • Existing tests are at the self-tests folder
  • All tests extend the test base class that
    • facilitates the import of modules under test
    • recreates a directory in the file system for the test outputs
    • provides a number of function for supporting tests that access the database, mocking the config files to point to the the test dot_config.yml, (re)creates a schema for DOT configuration and loads it with test data, etc

Code quality

We have instituted a pair of tools to ensure the code base will remain at an acceptable quality as it is shared and developed in the community.

  1. The formulaic python formatter “black”. As described by its authors it is deterministic and fast, but can be modified. We use the default settings, most notably formatting to a character limit of 88 per line.
  2. The code linter pylint. This follows the PEP8 style standard. PEP8 formatting standards are taken care of in black, with the exception that the default pylint line length is 80. Pylint is also modifiable and a standard set of exclusion to the PEP8 standard we have chosen are found here. We chose the default score of 7 as the minimum score for pylint to be shared.

The combination of black and pylint can be incorporated into the git process using a pre-commit hook by running setup_hooks.sh

About

The Data Observation Toolkit (DOT) can be used to monitor data in order to flag problems with data integrity and scenarios that might need attention.

Resources

License

Code of conduct

Security policy

Stars

Watchers

Forks

Packages

No packages published