kdb+

Use it when

• The workload is financial tick data and the latency bar sits where kdb+ has lived for decades.
• As-of joins, whole-column vector operations, and sub-millisecond in-memory reads define the queries.
• The team is committing to q; its expressiveness over time-series is the payoff for the learning curve.
• A tickerplant, real-time database, and historical database pipeline matches how your market data actually flows.

Think twice when

• The team will not learn q; qSQL and an ANSI subset exist, but the native language is where the performance assumptions and the hiring pool live.
• Budgets favor open source; licensing is commercial, with free editions capped (Community for bounded commercial use, Personal strictly non-commercial).
• The data is not time-series; general OLAP engines cover mixed analytics more cheaply.

How it runs

Self-hosted q processes composed into a pipeline: a tickerplant publishes to an in-memory real-time database, history lives in an on-disk HDB of splayed, date-partitioned columnar files (S3-tierable), and gateways fan queries across processes. One q process serves one query at a time; scaling means more processes.