Feature store
Serves the same features online (low latency) and offline (training) to avoid skew.
Problem it solves: Training/serving skew.
Where it sits: Read by model serving; written by pipelines.
When to use
- Real-time ML features.
When not to use
- Tiny batch-only models.
Failure modes
- Stale features degrade predictions.
Metrics to watch
- Online read latency
- Feature freshness
Configuration in Labs (illustrative defaults)
- Capacity per instance (req/s) = 20000 — Maximum requests one instance completes per second before it saturates.
- Write capacity (primary) (writes/s) = 5000 — Writes the single primary can absorb per shard. Replicas do not add write capacity.
- Read replicas (count) = 1 — Each replica adds one instance of read capacity. Replication lag is not modelled numerically.
- Shards (count) = 2 — Partitions; each multiplies read and write capacity assuming keys spread evenly (no hot key).
- Base service time (ms) = 3 — Processing time per request at zero load. Queueing adds to this as utilisation rises.