From the future

Model 04
HTTP/2gRPCgRPCgRPCPartner apps1,000 req/sPartner apps200 req/sAPI server2 vCPU · 2 GBModel (CPU)4 vCPU · 4 GBFeature sto…2 vCPU · 8 GBFeature sto…2 vCPU · 16 GB
HTTP/2gRPCgRPCgRPCgRPCPartner a…1,000 req/sPartner a…200 req/sAPI server2 vCPU · 2 GBModel (CP…4 vCPU · 4 GBFeature s…2 vCPU · 8 GBFeature s…

Fix the board. Turn every goal green.

Traffic
Orders1,000 req/s · reads onlyUpdates200 req/s · writes only
Limit
API server answers about 6,700 req/s
Bill
  • API server8 USD/mo
  • Model (CPU)16 USD/mo
  • Feature store32 USD/mo
  • Feature store64 USD/mo
120 of 200 USD/mo

Goals

Not tested yet

The share of requests that fail or get no answer.What the servers and services cost each month.

A fraud model passed its offline test with 99 percent recall. Live, it catches 79 percent of fraud and flags 21 percent of good orders. Its training rows and live requests read the same features. Keep recall above 90 percent and false alarms under 5 percent. 3 edits.

Day 3 after launch. Offline recall was 0.99; live recall is 0.79 and 21 percent of good orders are flagged. Graph: live recall.

Each worker adds one network cardDrag or use arrow keys to change the workers

2 workers · 1 TB moves in 1.1 h · 1,000 Mbps per card