Serving is the reason the other stages exist. Data that nobody queries is a cost, not an asset. The main consumers are:
Analytics and business intelligence. Dashboards and reports read curated tables in a warehouse or lakehouse. A semantic layer defines metrics such as "net revenue" once, so every tool computes them the same way (Performance and Semantics).
Machine learning. Training jobs read large historical snapshots; live models read fresh features, such as a customer's purchases in the last hour, with low latency.
Applications and APIs. Other systems read results directly: a "customers also bought" list on BookNest's product page, or a partner API that returns sales figures.
Reverse ETL. Modeled data is pushed back into operational tools, such as a CRM or an email platform, so sales and marketing teams act on it where they work.
AI assistants. Retrieval pipelines feed documents and governed tables to large language models, which raises the stakes on access control and personal data (Security and Privacy Basics).
Each consumer needs a promise about the data it reads. Write it down as a service-level agreement (SLA) with measurable targets: how fresh the data is ("orders up to 15 minutes old"), how complete, and when it is available ("the finance mart is ready by 06:00 UTC"). Then monitor the targets, not just whether the jobs succeeded: a job can finish green and still load zero rows. Pipeline Testing and Alerts builds that alerting.