Dataviz

Data Visualization for Dashboards

A dashboard is a tool for fast, accurate judgement under operational stress. Every visual choice supports that judgement or undermines it.

Scope

• In scope: chart selection, encoding accuracy, color, layout and hierarchy, time-series conventions, observability frameworks (RED, USE, Four Golden Signals, SLO/SLI), anti-patterns, review criteria.
• Out of scope:
  • Grafana JSON, panel options, transformations, variables, library panels — dashboards sibling owns mechanics.
  • Alert rule definitions — alerting sibling. This skill covers SLO/SLI dashboard structure, not rule firing.
  • Query syntax — promql / metricsql / logsql siblings.

References

• Perception and encoding — [${CLAUDE_SKILL_DIR}/references/perception.md] Cleveland & McGill hierarchy, pre-attentive attributes, Tufte data-ink and chart junk, Stephen Few counterpoint.
• Color — [${CLAUDE_SKILL_DIR}/references/color.md] Semantic conventions, sequential/diverging/categorical palettes, Viridis, colorblind safety, WCAG contrast, dark/light themes, Grafana scheme mapping.
• Layout — [${CLAUDE_SKILL_DIR}/references/layout.md] F/Z patterns, tiering, grid discipline, progressive disclosure, RED/USE/Four Golden Signals, SLO/SLI dashboard structure.
• Time-series specifics — [${CLAUDE_SKILL_DIR}/references/time-series.md] Aspect ratio, baseline, interpolation, aggregation, dual y-axes, stacking.

Encoding hierarchy

Cleveland & McGill (1984), most → least accurate:

1. Position on a common scale
2. Position on non-aligned scales (small multiples)
3. Length (bar charts)
4. Angle / slope (pie slices)
5. Area (bubble plots, treemaps)
6. Volume (3D)
7. Color saturation / luminance
8. Color hue (categorical only)

Pick encodings as high as the data shape allows. Color hue distinguishes categories; never magnitude.

Chart-type selection

Flows from data shape and the question being answered. dashboards skill has the panel-type catalog; this skill has the rationale.

• Magnitudes over time → line chart (Time series). Default for observability metrics.
• Magnitudes across discrete categories → bar chart for ranking. Never pie.
• Parts of a whole → stacked bar or stacked area, ≤ 5 components, only when total is meaningful. No pie above 5.
• Distribution at a point in time → histogram. Over time, heatmap.
• One headline number → stat panel; sparkline when trend context matters.
• One value against a target → bar gauge over radial gauge. Radial only for single prominent numbers.
• Relationships between two metrics → scatter (XY chart).
• Tabular detail → table with sparkline cells for per-row trends.
• Status across many entities → state timeline or status history. Pair colored cells with text or icons.

Pie rule: if you must, ≤ 5 segments, parts-of-a-whole framing, meaningful total. Otherwise bar chart.

Pre-attentive attributes

Color hue, intensity, position, length, size, shape, orientation, motion — processed in under 200 ms. The unique attribute "pops out" against the field.

• One pre-attentive attribute per signal. Combining collapses pop-out. If color encodes severity, don't reuse it for category.
• Reserve motion for active anomalies. Decorative blinking trains viewers to ignore motion.
• Group with enclosure (rows, tabs) and proximity (grid), not panel borders.

Color rules

Color is redundant — pair with text, position, or shape.

• Fixed semantic mapping. Green = OK, amber = warning, red = critical, blue = informational, grey = no-data. One mapping per dashboard. Reusing red for "high traffic" and "error rate" defeats the convention.
• Match palette to data:
  • Sequential (single hue or Viridis) for magnitudes.
  • Diverging (two ends, neutral middle) for anomalies around a baseline.
  • Categorical (distinct hues, no order) for series identity. Cap at ~10 hues.
• Colorblind safety: 8% of men have color-vision deficiency. Avoid unmodified red-green as sole signal. Use Viridis/Cividis for sequential data.
• Contrast: WCAG 4.5:1 for value text on stat panels. NOC TVs at half brightness amplify low-contrast failures.
• No rainbow scales. Perceptually non-uniform.

Grafana color scheme defaults

• Status panels (stat, gauge, bar gauge) → From thresholds (by value). Base = green, amber warn, red crit.
• Multi-series time series → Classic palette (by series name) — stable across reruns.
• Heatmap → sequential continuous (Blues / Greens) or Viridis-equivalent. Diverging only when zero/center is operational.
• Non-status numeric metrics → Single color. Position does the work.

Layout and information hierarchy

Users scan dashboards in F-pattern (horizontal sweeps, then vertical down the left) or Z-pattern (top-left → top-right → bottom-left → bottom-right). Top-left has the highest attention; bottom-right the lowest.

Tier the dashboard

• Tier 1 — primary KPIs. Visible without scrolling. 2–4 panels max.
• Tier 2 — supporting trends and breakdowns.
• Tier 3 — drilldown, references, raw tables.

Aim for ≤ 12 panels in the initial viewport.

Grid discipline

• Pick a column unit and stick to it. 6, 8, or 12. Mixed widths in a row are noise.
• Match panel heights within a row. Mixed heights force the eye to re-anchor.
• Group with rows / tabs / collapsed sections, not decorative borders.
• Repeat panels and rows for symmetric per-entity dashboards.

Progressive disclosure

Headline → trend → breakdown → trace. Top answers "is this OK?" in under 3 seconds; lower tiers answer "why?" when something is wrong. Use data links for drilldown preserving time range and variables (see dashboards skill, annotations-links.md).

Time-series rules

• Aspect ratio: wider than tall, 2:1 to 4:1. Cleveland's "banking to 45°."
• Baseline:
  • Zero for ratios, rates, counts, bounded fractions. Anywhere "twice as much" means something.
  • Skip zero only for unbounded magnitudes where variation is the signal (temperature, stock prices).
  • Never truncate silently — most common chart lie.
• Interpolation:
  • Linear — default.
  • Smooth (curved) — invents values between samples. Never on operational dashboards.
  • Step before / after — discrete-state metrics (queue depth, feature flag, node count).
  • Bars — closed-bucket aggregates (requests-per-minute), not continuous samples.
• Aggregation:
  • max for latency, queue depth — preserves spikes.
  • avg for throughput — shows typical level.
  • Overlay both when in doubt.
  • $__rate_interval inside Prometheus rate() (4× scrape-interval rule).
• Dual y-axes:
  • OK for correlated signals on shared time axis (req rate + latency during incidents).
  • Anti-pattern for unrelated metrics or scale-mismatched signals — split panels or normalize.
• Stacking:
  • Default unstacked.
  • Stack only when components sum to a meaningful total and user needs both total and breakdown.
  • Non-bottom series read against non-flat baseline — magnitudes noisy.

Observability frameworks

• Four Golden Signals (Google SRE) — Latency, Traffic, Errors, Saturation. User-facing services.
• RED (Tom Wilkie) — Rate, Errors, Duration. Request-driven services (HTTP/gRPC). Golden Signals minus saturation; better for bursty workloads.
• USE (Brendan Gregg) — Utilization, Saturation, Errors. Resources (hosts, disks, queues).

Pick by entity:

• Customer-facing service → Four Golden Signals or RED.
• Resource → USE.
• Platform overview → mix: Golden Signals on top, USE per resource tier below.

SLO/SLI dashboards

• SLI current value — large stat, threshold = SLO target. Green ≥ target, red below.
• Error budget remaining — 1 - (1 - SLI) / (1 - SLO) as %. Red near zero.
• Burn rate — current_failure_rate / budget_failure_rate. > 1 burns budget faster than allowed.
• SLO compliance over time — SLI series with SLO target as threshold line.
• Burn-down chart — remaining budget over SLO window (typically 28d).
• Multi-window burn-rate alert state — short/long-window burn signals from alert system.

Layout: SLO summary stats top tier. SLO trend + burn-down middle tier. Failure-mode breakdown bottom tier.

Dashboard review checklist

1. Purpose — Can a new on-call describe the dashboard in one sentence?
2. Headline visible without scrolling? 3-second "is this OK?" answer in initial viewport.
3. Hierarchy — Do size, position, and color emphasis agree?
4. Encoding hierarchy — Every magnitude encoded as high as shape allows?
5. Units set everywhere? No bare numbers.
6. Color — Consistent semantic mapping? Stable categorical palette (by series name)? No rainbow sequential?
7. Baseline — Zero on rates/ratios/counts? Truncation marked?
8. Interpolation — Linear or step? No smooth curves?
9. Time range — Default matches dashboard purpose (24h ops, 28d SLO)?
10. Drilldown — Data links preserve time range and variables?
11. Chart junk — Decoration without information removed?
12. Repeating panels and rows for entity symmetry?
13. Time-window consistency — All panels controlled by dashboard time picker?

Application

When designing:

• Start from the question the dashboard answers, not the metrics you have.
• Pick framework (RED, USE, Four Golden Signals, SLO) before panels.
• Tier panels: ≤ 4 headline KPIs, then trends, then breakdowns.
• Apply encoding hierarchy: position > length > angle > area > color.
• Lock semantic color mapping for the whole dashboard.
• Default to unstacked, linear-interpolation, zero-baseline time series.

When reviewing:

• Cite the panel and rule. Show the fix.
• Distinguish "wrong" from "stylistic preference."
• Point and fix; don't lecture.

Integration

• dashboards — Grafana mechanics. This skill picks position-not-angle; that skill sets the scheme.
• alerting — rule definitions. This skill structures SLO/SLI dashboards that visualize burn.
• promql / metricsql / logsql — query syntax. This skill assumes correct queries.

Critical rules

• Position and length carry magnitude. Color hue carries category. Never confuse them.
• Color is redundant — pair semantic color with text, position, or shape.
• Zero baseline for rates, ratios, counts. Never silently truncate.
• Linear or step interpolation on operational dashboards. Smooth curves invent data.
• Default unstacked. Stack only when the total matters.
• Dual y-axes only for correlated signals.
• Top-left is the headline. 3-second "is this OK?" without scrolling.
• One pre-attentive attribute per signal.
• No rainbow palettes, no 3D, no decorative thresholds, no pie above 5 segments.