Disaster Tweets API — Production

Model (HF) · Demo (Space)

One-line

BERT-based binary classifier (Disaster vs Not Disaster) with a production-ready FastAPI service, Docker image, CI, and deployment notes.

---

Table of contents

• Quick local smoke test (Docker)
• Google Cloud Run (recommended production flow)
• GitHub Actions / CI
• API spec
• Environment variables & secrets
• Testing
• Production sizing & ops
• Observability [and miscellaneous]
• License & contact

---

Quick local smoke test (Docker)

Build and run locally to verify the API:

# build
docker build -t disaster-tweets-api:local .

# run (replace HF_TOKEN)
docker run --rm -p 8080:8080 \
  -e HF_TOKEN="hf_xxx" -e HF_REPO_ID="sakibalfahim/disaster-tweets-bert" \
  disaster-tweets-api:local

# test
curl -X POST "http://localhost:8080/predict" \
  -H "Content-Type: application/json" \
  -d '{"text":"Huge explosion reported near the financial district — multiple injuries."}'

If you have GPU access in your environment and want GPU in Docker, add --gpus all and ensure appropriate base image.

---

Google Cloud Run (recommended production flow)

Build and push with Cloud Build and deploy with Cloud Run. Replace PROJECT_ID, REGION, and hf_xxx.

gcloud auth login
gcloud config set project PROJECT_ID

# Build container in GCP
gcloud builds submit --tag gcr.io/PROJECT_ID/disaster-tweets-api:latest

# Deploy (recommended sizing)
gcloud run deploy disaster-tweets-api \
  --image=gcr.io/PROJECT_ID/disaster-tweets-api:latest \
  --platform=managed \
  --region=REGION \
  --allow-unauthenticated \
  --set-env-vars HF_REPO_ID="sakibalfahim/disaster-tweets-bert" \
  --memory=4Gi --cpu=2 --concurrency=1

Secrets: store HF_TOKEN as a Cloud Run secret and mount it into the service; do not put the token directly in CLI history.

---

GitHub Actions / CI

Current workflow .github/workflows/ci.yml:

• Builds Docker image using docker/build-push-action
• Pushes image to GHCR (configured)
  Next CI improvement: run pytest inside a lightweight test container to validate quick smoke tests (tests included in tests/test_basic.py).

---

API spec

Base: GET /health and POST /predict

• GET /health

Returns service readiness and device info.

{"status":"ok","device":"cuda"|"cpu"}

• POST /predict

Request:

{
  "text": "Single tweet string"
}

or

{
  "text": ["tweet1", "tweet2"]
}

Response:

{
  "predictions": ["Disaster"],
  "confidences": [{"Disaster":0.987,"Not Disaster":0.013}],
  "latency_ms": 123.45
}

---

Environment variables & secrets

• HF_TOKEN — Hugging Face token (READ permission). Store as secret.
• HF_REPO_ID — model repo id, default sakibalfahim/disaster-tweets-bert.

Do not commit tokens. Use cloud secret managers or GitHub Secrets for CI.

---

Testing

A lightweight test file tests/test_basic.py is included that fake-injects a minimal transformers shim so CI can run fast without downloading the real model. Run locally:

pip install -r requirements.txt
pip install pytest httpx
pytest -q

CI will run container build; next step is to run tests inside the image before pushing.

---

Production sizing & ops

• Model artifact ~400+MB (safetensors). Memory usage depends on batch size and device.
• Start with: --memory=4Gi, --cpu=2, --concurrency=1. Increase memory to 8Gi if using larger batches.
• Use concurrency=1 to avoid memory competition inside a single instance.
• If you expect sustained high QPS, use autoscaling with a reasonable min instance count to mitigate cold starts.

---

Observability [and miscellaneous]

• /metrics exposes basic in-memory metrics (requests, avg latency). Add Prometheus exporter or forward logs to Cloud Monitoring.
• Add health-checks & readiness for autoscalers.
• Implement request rate limiting + authentication for production.

---

Security

• Keep HF_TOKEN secret and rotate periodically.
• Consider private network / API gateway for production.
• Validate and sanitize inputs; set a max length for text to avoid resource abuse.

---

Next enhancements (roadmap)

• CI: run pytest inside built image; fail fast for regressions.
• Add async batching & queueing to maximize GPU throughput.
• Add model warmup & quantization options (ORT, bitsandbytes) to reduce memory and improve latency.
• Add monitoring (Prometheus), tracing, and alerting.

---

Files of interest

• app/main.py — FastAPI app and inference logic
• Dockerfile — production image
• .github/workflows/ci.yml — CI build & push
• tests/test_basic.py — smoke tests
• README.md — this file

---

License & contact

MIT License

Author: sakibalfahim — contact via mail, LinkedIn, GitHub, or Hugging Face profile.