Time series: Add "Primer: Machine Learning for Time Series Data"

Co-authored-by: Andreas Motl <[email protected]>
Co-authored-by: ckurze <[email protected]>

Author: 3 people

docs/_assets/img/ml-timeseries-primer/cratedb-admin-ui-data-imported.png

@@ -93,5 +93,6 @@ and analyzing. Industrial applications.
 Basics <basics>
 Advanced <advanced>
 Connectivity <connect>
-video
+Video Tutorials <video>
+Machine Learning Primer <ml-primer/index>
 :::

docs/_assets/img/ml-timeseries-primer/cratedb-cloud-import-ready.png

@@ -0,0 +1,167 @@
+# Experiment Tracking with CrateDB using SQL only
+
+_Introduction to Time Series Modeling with CrateDB (Part 3)._
+
+This is part 3 of our blog series about "Running Time Series Models in Production using CrateDB".
+
+
+## Introduction
+
+While MLflow is a handy tool, it is also possible to track your experiments exclusively using
+CrateDB and SQL, without using any machine learning framework at all.
+
+Because CrateDB provides storage support for both nested documents, and binary data, you can also
+store parameters, metrics, the model configuration, and the model itself, directly into CrateDB.
+
+The next section demonstrates it on behalf of two corresponding examples.
+
+
+## Storing Experiment Metadata
+
+CrateDB supports you in storing and recording your experiment metadata.
+
+### 1. Deploy database schema
+
+Create database tables in CrateDB, to store metrics and parameters.
+
+```sql
+CREATE TABLE metrics_params (
+  timestamp TIMESTAMP DEFAULT now(),
+  run_name TEXT,
+  metrics OBJECT(DYNAMIC),
+  parameters OBJECT(DYNAMIC)
+);
+```
+
+Using CrateDB's dynamic `OBJECT` column, you can store arbitrary key-value pairs into the `metrics`
+and `parameters` columns. This makes it possible to adjust which parameters and metrics you want to
+add throughout the experiments, and evolve corresponding details while you go.
+
+### 2. Record metrics and parameters 
+
+Instead of recording the metrics and parameters to MLflow, as demonstrated at
+[MLOps powered by CrateDB and MLflow » Experiment Tracking][ml-timeseries-blog-part-2], you will
+record them by directly inserting into the database table.
+
+```sql
+INSERT INTO
+  metrics_params (run_name, metrics, parameters)
+VALUES ('random_run_name',
+  '{"precision": 0.667, "recall": 0.667}',
+  '{"anomaly_threshold": 2.5, "alm_suppress_minutes": 3.5}');
+```
+
+### 3. Read back recordings
+
+To read back individual parameters of your recordings, you can utilize the standard
+[SQL `SELECT` statements].
+
+To retrieve all recorded metrics and parameters after a certain point in time:
+```sql
+SELECT *
+FROM metrics_params
+WHERE timestamp > '2021-01-01';
+```
+
+To retrieve specific parameters or metrics:
+```sql
+SELECT metrics['precision'], parameters['anomaly_threshold']
+FROM metrics_params
+WHERE timestamp > '2021-01-01';
+```
+
+
+## Storing Model Data
+
+CrateDB supports you in storing your model data.
+
+Independently of recording experiment metadata, you may also want to store the model itself into
+CrateDB, by leveraging its [BLOB data type].
+
+In order to store models into CrateDB, you will need two database tables:
+
+- A regular RDBMS database table, storing the model configuration and
+  relevant metadata.
+- A blob database table, storing serialized models in binary format,
+  usually in Python's [pickle format].
+
+### 1. Deploy database schema
+
+Create those tables, again utilizing CrateDB's nested object support for flexible
+schema evolution:
+
+```sql
+CREATE TABLE model_config (
+  timestamp TIMESTAMP DEFAULT now(),
+  digest TEXT, -- this is the link to the model blog
+  run_name TEXT,
+  config OBJECT(DYNAMIC)
+);
+
+CREATE BLOB TABLE models;
+```
+
+### 2. Upload the model
+
+To upload the model, run the following Python program after adjusting the spots
+about the database connection and credentials.
+
+```python
+from io import BytesIO
+import pickle
+from crate import client
+
+file = BytesIO()
+# Serialize the model object and store it in the in-memory file
+pickle.dump(model, file)
+
+conn = client.connect(
+   "https://<your-instance>.azure.cratedb.net:4200",
+   username="admin",
+   password="<your-password>",
+   verify_ssl_cert=True,
+)
+
+blob_container = conn.get_blob_container('models')
+blob_digest = blob_container.put(file)
+```
+
+Make sure to update the model configuration table accordingly:
+
+```python
+cursor = conn.cursor()
+cursor.execute(
+  "INSERT INTO model_config (digest, run_name, config) VALUES (?, ?, ?)", 
+  (blob_digest, "random_run_name", model.config.to_dict()))
+```
+
+CrateDB automatically creates all the model config columns.
+
+![crate model config](/_assets/img/ml-timeseries-primer/cratedb-model-configuration.png)
+
+### 3. Read back the model
+
+To retrieve a model from the blob store table again, you will need to get the digest value
+of the model from the configuration table:
+
+```sql
+SELECT digest FROM model_config WHERE run_name = 'random_run_name';
+```
+
+Then, use this digest, i.e. the blob identifier, to get the blob payload, and
+deserialize it from pickle format:
+
+```python
+blob_content = b""
+for chunk in blob_container.get(digest):
+  blob_content += chunk
+
+model = pickle.loads(blob_content)
+```
+    
+
+[BLOB data type]: https://crate.io/docs/crate/reference/en/latest/general/blobs.html
+[ml-timeseries-blog-part-1]: https://cratedb.com/blog/introduction-to-time-series-modeling-with-cratedb-machine-learning-time-series-data
+[ml-timeseries-blog-part-2]: https://cratedb.com/blog/introduction-to-time-series-modeling-with-cratedb-part-2?hs_preview=uXVBkYrk-136061503799
+[pickle format]: https://realpython.com/python-pickle-module/
+[SQL `SELECT` statements]: https://crate.io/docs/crate/reference/en/latest/sql/statements/select.html

docs/_assets/img/ml-timeseries-primer/cratedb-cloud-import-url.png

@@ -0,0 +1,11 @@
+(timeseries-ml-primer)=
+# Primer: Machine Learning for Time Series Data
+
+Learn how to apply machine learning procedures to time series data.
+
+```{toctree}
+:glob:
+:maxdepth: 2
+
+*
+```