Tutorial: Enabling Machine Learning in QuestDB with MindsDB
April 18, 2022
MindsDB
Blog Post
48
48
Introduction
MindsDB enables you to use Machine Learning to ask predictive questions about your data and receive accurate answers from it, all in SQL. With MindsDB:
Developers can quickly add AI capabilities to their applications.
Data scientists can streamline MLOps by deploying ML models as AI Tables.
Data analysts can easily make forecasts on complex data, such as multivariate time-series with high cardinality, and visualize these in BI tools like Grafana, and Tableau.
QuestDB is the fastest open-source, column-oriented SQL database for time-series data. It has been designed and built for massively-parallelized vectorized execution and SIMD, as the de-facto backend for high-performance demanding applications in financial services, IoT, IIoT, ML, MLOps, DevOps and observability. QuestDB implements ANSI SQL with additional extensions for time-specific queries, which make it simple to correlate data from multiple sources using relational and time series joins, and execute aggregation functions with simplicity and speed. In addition, QuestDB is resource efficient (comparatively cheaper than other projects to run in cloud environments), simple to install, manage, use, and stable in all production environments.
Combining both MindsDB and QuestDB gives you unbound prediction ability with SQL. You can perform all the pre-processing of your data inside QuestDB using its powerful and unique extended SQL, and then you can access these data from MindsDB, in its own also unique SQL, to produce powerful ML models.
The main goal of this article is to gently introduce these two deep technologies and give you enough understanding to be able to undertake very ambitious ML projects. To that end we will:
Spawn two Docker containers to run MindsDB and QuestDB.
Add QuestDB as a datasource to MindsDB using a SQL Statement.
Create a table and add data for a simple ML use case using QuestDB's web console.
Connect to MindsDB using mysql client and write some SQL.
Create a predictor for our ML use case.
Make some predictions about our data.
Have fun!
Requirements
docker-compose: To define and run our multi-container Docker application (it is usually installed implicitly when Docker is installed).
MySQL: The client we will use to interact with MindsDB (mysql -h 127.0.0.1 --port 47335 -u mindsdb -p).
Curl: To upload data to QuestDB from a local CSV file.
Software repositories in case you are inclined to look under the hood (Give us a star!):
which allows us to start our two service containers with command:
docker-compose up -d
Container questdb: Creates a local folder qdb_root to store table data/metadata, and the default server configuration => available at: localhost:9000.
Container mindsdb: Creates two local folders mindsdb_store, nltk_data, and uses configuration file mindsdb_config.json.
MindsDB takes about 60-90 seconds to become available, logs can be followed in the terminal:
docker logs -f mindsdb
...
Version 22.3.1.0
Configuration file:
/root/mindsdb_config.json
Storage path:
/root/mindsdb_store
http API: starting...
mysql API: starting...
mongodb API: starting...
✓ telemetry enabled
✓ telemetry enabled
✓ telemetry enabled
mongodb API: started on 47336
mysql API: started on 47335
http API: started on 47334
We can stop the two containers with command:
docker-compose down
We can remove all persisted data and configuration executing:
./remove_persisted_data.sh
Note: Doing this means that the next time you start the containers you will need to add QuestDB as a datasource again, as well as recreate the table, add data, and recreate your ML models.
Adding data to QuestDB
We can access QuestDB's web console at localhost:9000 :
and execute this DDL query to create a simple table (copy this query to the web console, select it and click Run):
Either way, this gives us 100 data points, one every 4 hours, from 2021-01-16T12:00:00.000000Z (QuestDB's timestamps are UTC with microsecond precision), conveniently downloaded to file sample_house_rentals_data.csv.
NOTE: If you tried the last query, you will have 200 rows, you can truncate table house_rentals_data and run the curl command again, in QuestDB data are immutable.
Connecting to MindsDB
We can connect to MindsDB with a standard mysql-wire-protocol compliant client (no password, hit ENTER):
mysql -h 127.0.0.1 --port 47335 -u mindsdb -p
Only two databases are relevant to us, questdb and mindsdb
This is a read-only view on our QuestDB instance. We can query it leveraging the full power of QuestDB's unique SQL syntax because statements are sent from MindsDB to QuestDB without interpreting them. It only works for SELECT statements (it requires activation by means of USE questdb;):
Beyond SELECT statements, for instance when we need to save the results of a query into a new table, we need to use QuestDB's web console available at localhost:9000:
CREATETABLEsample_query_resultsAS (
SELECT
ts,
neighborhood,
sum(days_on_market) DaysLive,
min(rental_price) MinRent,
max(rental_price) MaxRent,
avg(rental_price) AvgRent
FROM house_rentals_data
WHERE ts BETWEEN '2021-01-08'AND'2021-01-10'
SAMPLE BY 1d FILL (0, 0, 0, 0)
) TIMESTAMP(ts) PARTITION BY MONTH;
mindsdb
Contains the metadata tables necessary to create ML models and add new data sources:
We can create a predictor model mindsdb.home_rentals_model_ts to predict the rental_price for a neighborhood considering the past 20 days, and no additional features:
mysql> USE mindsdb;
Database changed
mysql>
CREATE PREDICTOR mindsdb.home_rentals_model_ts FROM questdb (
SELECT
neighborhood,
rental_price,
ts
FROM house_rentals_data
)
PREDICT rental_price ORDER BY ts GROUP BY neighborhood
WINDOW 20 HORIZON 1;
This triggers MindsDB to create/train the model based on the full data available from QuestDB's table house_rentals_data (100 rows) as a timeseries on column ts.
You can see the progress by monitoring the log output of the mindsdb Docker container, and you can ask MindsDB directly:
When status is complete the model is ready for use, until then, we simply wait while we observe MindsDB's logs, and repeat the query periodically. Creating/training a model will take time proportional to the number of features, i.e.cardinality of the source table as defined in the inner SELECT of the CREATE PREDICTOR statement, and the size of the corpus, i.e. number of rows. The model is a table in MindsDB:
We can get more information about the trained model, how was the accuracy calculated or which columns are important for the model by executing the DESCRIBE statement.
Additional information about the models and how they can be customized can be found on the Lightwood docs.
Querying MindsDB for predictions
The latest rental_price value per neighborhood in table questdb.house_rentals_data (as per the uploaded data) can be obtained directly from QuestDB executing query:
In this article, we have introduced QuestDB and MindsDB in a hands-on approach. QuestDB can help you store, analyse, and transform timeseries data, while MindsDB can help you make predictions about it. Albeit simple, our use case should have lowered the entry barrier to these two deep technologies, and now you can deepen your knowledge further by undertaking more ambitious ML projects.
Thank you for getting this far!!!, if you liked this content we'd love to know your thoughts, please come and say hello in our welcoming communities:
