Journal: Low-latency Online Data Store for AI-Driven Financial Insights

Abstract

Purpose / Summary

Build an online data store that provides low-latency reads for AI bots serving customer financial insights. The system uses an HTAP database (TiDB) for transactional + analytical workloads [3], Redis as a low-latency cache [4], Kafka + Kafka Streams for streaming ingestion and materialized views [5], and Spark for batch pre-aggregation of metrics [2]. APIs expose the data to AI services/bots.

Goals / SLAs

Read latency for most bot queries: p95 ≤ 30ms, p99 ≤ 100ms (from API to returned result)

Event ingestion throughput: ≥ 100k events/sec (adjustable by business)

End-to-end freshness for streaming view: < 1s for critical events; batch metrics updated hourly (or as needed)

Consistency: read-after-write for single-user critical ops; eventual consistency acceptable for some aggregates

Security & compliance: PCI/PII controls, encryption in transit & at rest, audit logging