Add relevant justifications to challenge.md

docs/challenge.md

@@ -1,13 +1,13 @@
-# Bugs fixed on `exploration.ipynb`
+# Part I
+## Bugs fixed on `exploration.ipynb`
 - The function `is_high_season` had an issue since it didn't consider the time. For example, `is_high_season("2017-12-31 14:55:00")` returned `0` when in reality it should return `1`.
 - All calls to `sns.barplot` were missing the `x` and `y` definition.
 - To improve visualizations and correctly show the delay rate, the method `get_rate_from_column` was updated. Instead of calculating `rates[name] = round(total / delays[name], 2)`, I think it's best to do `rates[name] = round(100 * delays[name] / total, 2)` to get the ratio of delayed flights to the total number of flights (for a specific column value). This value is now between 0 and 100, where `0` indicates that no flights with that specific column value were delayed and `100` indicates that all flights with that specific column value were delayed. After implementing this change, visualization code had to be updated as well to avoid limiting the y-axis.
 - `training_data` was defined but never used. This cell was deleted.
 - Minor style changes were applied, such as using spaces around operators and removing spaces when defining the value of certain method arguments.
 - `xgboost` was not included under installed dependencies.
 
-
-# Model pick
+## Model pick
 When comparing the different models' performances, I want to focus on the positive (minority) class since that's the class that represents delays and the model is intended to predict the probability of **delay**. For this, I'll focus on the F1-score for the positive class, since it combines precision and recall into a single metric, providing a balance between the two and accounting for false positives and false negatives. It offers a consolidated evaluation of the model's performance in predicting the positive class while factoring in the class imbalance. Let's review the results for each model:
 1. **XGBoost**: 0.00
 2. **Logistic Regression**: 0.06
@@ -18,6 +18,24 @@ When comparing the different models' performances, I want to focus on the positi
 
 With this in mind, the best model picked was the third one: **XGBoost with Feature Importance and with Balance**. XGBoost is an ensemble method that can capture complex relationships in the data and it's more suitable to capture non-linear patterns and interactions between features. Plus, XGBoost is highly scalable and effective for large datasets, while LogisticRegression may struggle with very large datasets.
 
+# Part II
+The model was kept as unchanged as possible. My intention was to avoid hardcoding the top 10 features but the test `test_model.py` asserted that only those columns were present. 
+Some improvements that could be done here are:
+- Have `THRESHOLD_IN_MINUTES` and `MODEL_FILE_NAME` as environment variables. That way, if we want to update either of those values, we can do that without the need to update the code.
+- Upload the trained model directly to Google Cloud Storage (GCS) or to another service that can store the different versions of the trained model. For now, we're saving it locally and manually uploading it to GCS, which is where the Continuous Deployment pipeline gets it from.
 
-# Bugs fixed on `test_model.py`
+## Bugs fixed on `test_model.py`
 - Data could not be loaded because the path was incorrect. After changing `"../data/data.csv"` to `"./data/data.csv"` it worked as expected.
+
+# Part III
+API was deployed by building a Docker image and pushing it to Artifact Registry (GCP). The container was then deployed to CloudRun.
+
+# Part IV
+Continuous Integration workflow runs whenever there is a push or pull request to the main/develop branches.
+Continuous Deployment workflow runs whenever there is a push or pull request to the main/release branches.
+The following secrets were set directly as Repository Secrets on Github:
+- GCP_CREDENTIALS: content of GCP's `credentials.json`.
+- GCP_IMAGE_NAME: name of the image pushed to Artifact Registry on GCP.
+- GCP_PROJECT_ID: name of the project used on GCP (e.g. 'project-id').
+- GCP_REGION: region used on GCP (e.g. 'us-central1').
+- MODEL_VERSION: name of the GCS model that will be downloaded and served (e.g. 'v0').