Core data concepts · Lesson 4

Who owns which decision in the data chain?

8 min read

Lesson objectives

  • Assign a data incident among database administrator, data engineer and analyst with escalation boundaries
Reading tools

DP-900 names foundational data roles, not rigid employment contracts. A database administrator focuses on database operation and health: access, backup and recovery, monitoring, performance, and indexes within the organization's responsibilities and service model. A data engineer designs ingestion, transformation, and movement into an analytically useful store, maintaining flow, data quality, and scheduling. A data analyst defines the question and measure, checks field and denominator meaning, and presents an understandable result. One person may perform several roles in a small team, while a large organization separates them. The learning question is where technical investigation begins during a problem, not the title on a badge. Do not grant an analyst production database privileges merely to obtain a report, and do not blame every numeric discrepancy on the database administrator without tracing the data.

Consider a report showing 180 purchases when the application logged 200. Do not immediately guess that twenty records disappeared. Fix the time window and transaction identifier first: were there 200 attempts including twenty failures, or 200 successes? Then compare checkpoints: correct source records, rows entering the transport pipeline, rows reaching the analytical model, and rows shown after visual filters. If the live database cannot accept writes or returns an access error, its operator begins the operational diagnosis. If 200 successful records exist in the source and 180 in the analytical copy, the data engineer investigates filters and scheduling. If 200 reach the model but a regional filter shows 180, the analyst checks the visual definition. This maps the first investigation; it does not prohibit collaboration between roles.

Responsibilities vary with service model. A managed database reduces the customer's operating-system patching and storage administration compared with a self-managed server, but it does not remove the need to define access, understand queries, and plan business-specific recovery. Azure manages everything is not a useful answer. A data engineer cannot silently rewrite field meaning: if status=returned is incorrectly called completed, the metric owner and source team must agree on the correction. Likewise, an analyst publishing a percentage must state numerator, denominator, and period rather than merely format a chart. For each handoff, write a small data contract: field name, type, definition, tolerable delay, integrity check, and alert owner. That contract narrows the dispute when the source schema changes.

Do not confuse the first investigator with the only person involved in a repair. In the 90-payment case, the analyst can show that a web filter explains 72, while the engineer may document mobile channel values in the model and the administrator may confirm the meaning of successful payment status. A useful handoff table names evidence at each boundary: a source query keyed by transaction ID, pipeline input and output counts, unique transaction IDs in the model, and active visual filters. If any checkpoint is missing, mark that location unverified rather than claiming records were lost or safe. Distinguish an access-control issue from a semantic issue. Restricting personal data from analysts may be correct, yet an authorized aggregate can let them validate a measure without bypassing permissions. If a source schema renames channel to sales_channel, the engineer handles the mapping and the analyst checks that historical categories remain comparable. If 72 appears at the source itself, investigation starts there, not with the chart. Role selection means tracing the first divergence, documenting its limit, and handing a cause to the appropriate team.

A manager asks who owns a defect visible only in reporting. First record where counts agree and where they diverge. Give each colleague a question their records can resolve instead of blaming the previous team. The database administrator needs writing and availability evidence. The data engineer needs a batch identifier and transformation step. The analyst needs metric definitions and filters. A business decision may still belong to the report owner, such as whether returned orders belong in sales. Name one follow-up owner to collect findings while each team remains responsible for its repair. A handover should include a specific sample, reproduction steps and expected output. This reduces repeated referrals and permits testing the same case before closure.

Lesson summary

Assign a data incident among database administrator, data engineer and analyst with escalation boundaries.