144 lines
6.2 KiB
Markdown
144 lines
6.2 KiB
Markdown
# data-platform
|
|
|
|
[](https://github.com/Stijnvandenbroek/data-platform/actions/workflows/ci.yml)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
A personal data platform for ingesting, transforming and analysing external data sources. Built with
|
|
[Dagster](https://dagster.io/), [dbt](https://www.getdbt.com/) and
|
|
[PostgreSQL](https://www.postgresql.org/), managed with [uv](https://github.com/astral-sh/uv) and
|
|
deployed via Docker Compose.
|
|
|
|
## Stack
|
|
|
|
| Layer | Tool |
|
|
| -------------- | ------------------------------------------------- |
|
|
| Orchestration | Dagster (webserver + daemon) |
|
|
| Transformation | dbt-core + dbt-postgres |
|
|
| ML | LightGBM, MLflow, scikit-learn |
|
|
| Storage | PostgreSQL 16 |
|
|
| Notifications | Discord webhooks |
|
|
| Observability | Elementary (report served via nginx) |
|
|
| CI | GitHub Actions (Ruff, SQLFluff, Prettier, pytest) |
|
|
| Package / venv | uv |
|
|
| Infrastructure | Docker Compose |
|
|
|
|
## Data pipeline
|
|
|
|
### Ingestion
|
|
|
|
Python-based Dagster assets fetch data from external APIs and load it into a raw PostgreSQL schema
|
|
using upsert semantics. Assets are prefixed with `raw_` to clearly separate unmodelled data from
|
|
transformed outputs. They are chained via explicit dependencies and run on a recurring cron
|
|
schedule.
|
|
|
|
### Transformation
|
|
|
|
dbt models follow the **staging → intermediate → marts** layering pattern:
|
|
|
|
| Layer | Materialisation | Purpose |
|
|
| ---------------- | --------------- | ---------------------------------------------------------------- |
|
|
| **Staging** | view | 1:1 cleaning of each raw table with enforced contracts |
|
|
| **Intermediate** | view | Business-logic joins and enrichment across staging models |
|
|
| **Marts** | incremental | Analysis-ready tables with derived metrics, loaded incrementally |
|
|
|
|
Mart models use dbt's **incremental materialisation** — on each run only rows with a newer ingestion
|
|
timestamp are merged into the target table. On a full-refresh the entire table is rebuilt.
|
|
|
|
All staging and intermediate models have **dbt contracts enforced**, locking column names and data
|
|
types to prevent silent schema drift.
|
|
|
|
### Data quality
|
|
|
|
- **dbt tests**: uniqueness, not-null, accepted values, referential integrity and expression-based
|
|
tests run as part of every `dbt build`.
|
|
- **Source freshness**: a scheduled job verifies raw tables haven't gone stale.
|
|
- **Elementary**: collects test results and generates an HTML observability report served via nginx.
|
|
|
|
### Machine learning
|
|
|
|
An **ELO rating system** lets you rank listings via pairwise comparisons. An ML pipeline then learns
|
|
to predict ELO scores for unseen listings. The ranking UI and model logic originate from
|
|
[house-elo-ranking](https://github.com/Stijnvandenbroek/house-elo-ranking).
|
|
|
|
| Asset | Description |
|
|
| ---------------------- | ----------------------------------------------------------------------------------- |
|
|
| `elo_prediction_model` | Trains a LightGBM regressor on listing features → ELO rating. Logs to MLflow. |
|
|
| `elo_inference` | Loads the best model from MLflow, scores all unscored listings, writes to Postgres. |
|
|
| `listing_alert` | Sends a Discord notification for listings with a predicted ELO above a threshold. |
|
|
|
|
All three are tagged `"manual"` — they run only when triggered explicitly.
|
|
|
|
### Notifications
|
|
|
|
The `listing_alert` asset posts rich embeds to a Discord channel via webhook when newly scored
|
|
listings exceed a configurable ELO threshold. Notifications are deduplicated using the
|
|
`elo.notified` table.
|
|
|
|
## Scheduling & automation
|
|
|
|
Ingestion assets run on cron schedules managed by the Dagster daemon. Downstream dbt models use
|
|
**eager auto-materialisation**: whenever an upstream raw asset completes, Dagster automatically
|
|
triggers the dbt build for all dependent staging, intermediate and mart models.
|
|
|
|
Assets tagged `"manual"` are excluded from auto-materialisation and only run when triggered
|
|
explicitly. A guard condition (`~any_deps_in_progress`) prevents duplicate runs while upstream
|
|
assets are still materialising.
|
|
|
|
## Getting started
|
|
|
|
```bash
|
|
# 1. Install uv (if not already)
|
|
curl -Lsf https://astral.sh/uv/install.sh | sh
|
|
|
|
# 2. Clone and enter the project
|
|
cd ~/git/data-platform
|
|
|
|
# 3. Create your credentials file
|
|
cp .env.example .env # edit .env with your passwords
|
|
|
|
# 4. Install dependencies into a local venv
|
|
uv sync
|
|
|
|
# 5. Generate the dbt manifest (needed before first run)
|
|
uv run dbt deps --project-dir dbt --profiles-dir dbt
|
|
uv run dbt parse --project-dir dbt --profiles-dir dbt
|
|
|
|
# 6. Start all services
|
|
docker compose up -d --build
|
|
|
|
# 7. Open the Dagster UI
|
|
# http://localhost:3000
|
|
```
|
|
|
|
## Local development
|
|
|
|
```bash
|
|
uv sync
|
|
source .venv/bin/activate
|
|
|
|
# Run the Dagster UI locally
|
|
DAGSTER_HOME=$PWD/dagster_home dagster dev
|
|
|
|
# Useful Make targets
|
|
make validate # Check Dagster definitions load
|
|
make lint # Ruff + SQLFluff + Prettier
|
|
make lint-fix # Auto-fix all linters
|
|
make test # pytest
|
|
make reload-code # Rebuild + restart user-code container
|
|
```
|
|
|
|
## Services
|
|
|
|
| Service | URL |
|
|
| ----------------- | --------------------- |
|
|
| Dagster UI | http://localhost:3000 |
|
|
| MLflow UI | http://localhost:5000 |
|
|
| pgAdmin | http://localhost:5050 |
|
|
| Elementary report | http://localhost:8080 |
|