instaml — a real-time feature store you can deploy in a day

Feast and Tecton are built for platform teams with weeks to spend on deployment. instaml is a YAML feature-defs → online/offline dual-write feature store a solo ML engineer can stand up before lunch.

Live demo: instaml.kartikaneja.com Status: alpha · last shipped 2026-07-20 Built by: Kartik Aneja — AI/ML Platform Engineer

---

Why this exists

See PRODUCT.md for the full user/problem/solution writeup. TL;DR:

• Who: ML Engineer at a Series-A startup with a model that needs real-time features and no time for a Feast deployment
• Pain: Feast is days of infra work; Tecton is enterprise-priced; most teams end up with hand-rolled Redis + Postgres + brittle backfill scripts
• Why now: Real-time ML is mainstream now, but the feature-platform layer underneath it is still either too heavy or too improvised

What works today (alpha MVP)

• Declarative YAML feature definitions — count / sum / avg / min / max / last, over a rolling window (5m, 1h, 7d) or lifetime
• Dual-write on every event — DuckDB stores every raw event (the offline source of truth, exportable to Parquet); an online store (in-memory or Redis) holds the current computed value per entity for fast reads
• Python SDK — instaml.emit_event(...) (fail-soft, never breaks your app) and instaml.get_features(...) (degrades to {} on outage, never raises)
• Live dashboard — feature catalog + an entity lookup panel, auto-refreshing
• No-key quickstart — examples/quickstart.py posts 60 synthetic e-commerce events across 20 fake users so the dashboard isn't empty

Try it (60 seconds, local)

git clone https://github.com/anejakartik/instaml.git
cd instaml
pip install -e ./sdk
docker compose up -d
python examples/quickstart.py
open http://localhost:8000

Real usage:

import instaml

instaml.configure(endpoint="http://localhost:8000")

# From your event-producing code:
instaml.emit_event(entity_id="user_42", event_type="purchase", payload={"amount": 19.99})

# From your model-serving code:
features = instaml.get_features("user_42", ["purchases_last_5m", "avg_cart_value_1h"])

Architecture

See docs/architecture.md. Stack: Python SDK + FastAPI server + DuckDB offline store (Parquet-exportable) + pluggable online store (in-memory or Redis), deployed on Fly.io + Vercel.

What's next

See ROADMAP.md. Top items: Kafka ingestion adapter, feature drift monitoring, point-in-time-correct training-dataset export, feature lineage UI.

Contributing

PRs welcome. See AGENTS.md.

License

MIT — see LICENSE.