Offline database vs online feature store
The offline order database keeps full history for training and is slow for point lookups. An online feature store keeps the latest values and gets updates from a change feed. At 10,000 lookups/s, the offline database gives p99 over 6 s and 36 percent errors. The online store gives p99 3.5 ms.
2 workers · 1 TB moves in 1.1 h · 1,000 Mbps per card
Calls per page: 20 · 18% of pages wait 2.0 s
Case 1
A fraud model reads 10,000 user features/s. The order database takes 300 new orders/s. Lookups need p99 below 10 ms. A value may lag new orders by at most 60 s.
| Measure | Goal | Read the offline order database | Online store fed by a change feed |
|---|---|---|---|
| p99 latency (Lookups) | under 10 ms | 6.3 s Fail | 3.5 ms Pass |
| Error rate (Lookups) | under 0.1% | 35.7% Fail | 0% Pass |
| Data age | under 60 s | 0 ms Pass | 100 ms Pass |
| Monthly cost | 24 USD/mo | 44 USD/mo |