Define useful metrics, reconcile the data and build readable dashboard states with explicit freshness, filters and decision context.
Identify the users, decisions and actions the dashboard must support. Define each metric’s numerator, denominator, unit, grain, eligibility, time zone and comparison period. Choose a manageable set from those decisions rather than an arbitrary KPI count or a correlation ranking.
Record tables or APIs, source owners, refresh schedules and transformation versions. Reconcile totals against trusted source reports, check duplicate keys and missing values, and test join and filter behavior. Format copied SQL or JSON for inspection only; the tools do not execute queries or certify the dataset.
Align cohorts, denominators, periods and time zones before comparing values. Use paired correlation only when exploring a defined linear association, retaining the data and outlier context. Use R²/RMSE only for actual numeric prediction evaluation; neither correlation nor regression score establishes that a chosen business KPI is useful.
Choose charts that match the question and preserve units, labels, baselines and relevant uncertainty. Make filter selections and reset behavior visible, and include meaningful empty, loading, error and stale-data states. Check text/background contrast and use labels, symbols or patterns so color is not the sole carrier of meaning.
Compare latest source event, ingestion and dashboard refresh times under the intended time zone and completeness rule. Test filters against independently checked source totals. Measure page/query/request latency and failure outcomes from the actual dashboard path; high sample-processing throughput does not prove fresh or correct data.
Ask representative users to complete the intended decisions using the actual dashboard, noting misread labels or missing context. Save metric definitions, source checks, tested filter states and data-as-of information. Assign owners for refresh failures, definition changes and periodic usefulness review, and retain a versioned release note.
Write and download the task brief, measured results and handoff decisions as Markdown.
Format a copied SQL statement for review; this does not execute the query or verify its result.
Format a copied JSON example for inspection; this does not query, validate a live API contract or configure a server.
Inspect linear association between complete numeric pairs without selecting causal conclusions from correlation.
Evaluate actual/predicted numeric pairs with unweighted R² and RMSE, retaining undefined constant-target cases.
Check explicit text/background colors while preserving labels and non-color ways to read the display.
Summarize actual latency observations or completed-request counts; keep serial stage assumptions separate.
Loading your checklist…
WCAG explains that color cannot be the only visual means of conveying information. Keep understandable labels or other visual cues.
Google SRE describes dashboards and monitoring in relation to questions and actions. Define what a freshness or service failure should cause someone to do.
Keep the following evidence with the actual version used for this task.
| Record | Include | Check |
|---|---|---|
| Definition | Grain, units, denominator, eligibility and time zone | Reconcile with source totals |
| State | Filters, empty/error and data-as-of | Exercise the actual user task |
| Ownership | Refresh, definition and incident owners | Record response and review triggers |
No matching rows and a failed refresh should not produce the same unexplained blank chart.
A percentage change without its population or denominator can hide a different question.