Skip to content
Tech Interview Prep
Technical interview guide

Dashboard Design Principles

Building a dashboard that actually gets used and drives decisions, not one that just looks comprehensive.

Read
45 min
Practice MCQs
25
Interview Q&A
25
Edition
v4
Editorial status
Reviewed

Scope: WCAG 2.2 and current W3C WAI, USWDS, Microsoft Power BI, Tableau, Google Looker, Grafana, Vega-Lite, and UK Government Analysis Function guidance reviewed 2026-09-04.

Curated: · Written: · Reviewed:

25 questions

MCQ-4A dashboard compares monthly revenue for two products: Product A at $50k and Product B at $52k. The y-axis is truncated to start at $45k. A reviewer flags the chart. Which statement about it is correct?
MCQ-3A service dashboard shows 10 endpoints' success rates. All rates fall between 99.1% and 99.9%. The current chart is a bar chart with a zero baseline; on-call engineers say the bars all look full and they cannot tell which endpoints are worst. The team wants readers to rank the endpoints and judge the size of the 0.2pp gaps accurately. Which encoding should replace the bar chart?
MCQ-12Post-incident, an SRE must fill in the error-budget sheet: for each of 46 services, the exact millisecond overage of its p99 above the 300 ms SLO, recorded to one decimal place, and then look up that figure for the services named in the writeup. The overages are tightly packed - 31 of the 46 fall between 0.4 ms and 6 ms, the other 15 between 12 ms and 41 ms - while raw p99 across the fleet spans 95 ms to 1,400 ms. The report is read once, on a 900 px-wide screen, by the one engineer filling the sheet. Which presentation fits this task?
MCQ-11A payments team plots throughput (requests/s, left axis) and p99 latency (ms, right axis) on one dual-axis line chart. Both axes are labeled and both series are sampled every 10 seconds from the same store. Over the week the two lines rise and fall together, and the team concludes that latency is driven by load. Which property of the chart makes that conclusion unsafe?
MCQ-13A service's mean request latency is estimated from 100 sampled requests: sample mean 200 ms, sample standard deviation 30 ms. The dashboard must show the range that contains the true mean with 95% confidence, alongside a 205 ms SLO. Which display should it show?
MCQ-24A team plots CPU utilization for 24 hosts over 6 hours as a 24-panel small-multiple grid with a shared 0–100% y-axis; each panel is about 90 px tall. Nineteen hosts sit between 3% and 6% the whole time; five run between 60% and 90%. The team wants to rank hosts by how far each strays from its own baseline. Which change to the grid achieves that?
MCQ-6A service dashboard marks each service's health with a solid dot and nothing else: green #4CAF50 for healthy, amber #F5A623 for degraded, red #EF5350 for down, on white rows. An audit measured the 8px amber dot at 2.03:1 against the white row, below the 3:1 that WCAG 1.4.11 requires for non-text UI components, and a test with 12 deuteranopic engineers found 9 could not tell a degraded dot from a healthy one at a glance. The three status color tokens are frozen until Q3 and the dashboard ships next sprint. Which change resolves both findings within those constraints?
MCQ-7A p99 latency dashboard draws the last 24 hours as an inline SVG with one rect per hour. The SVG root is role='img' aria-label='p99 latency, last 24 hours'. The 24 values live only in a JavaScript array and reach the page through a tooltip div that a mouseover handler creates. An audit finds that VoiceOver in browse mode announces one stop, 'graphic, p99 latency, last 24 hours', then the next heading; no hour's value is reachable. The chart has to stay a single tab stop for sighted keyboard users. Which change gives screen-reader users the 24 values?
MCQ-2A checkout dashboard has one tile: 'Success rate 99.56%' in green, computed over a rolling 30-day window. The SLO is 99.5%, measured over 5-minute windows. Yesterday the tile was the only thing the on-call engineer watched, and at 14:30 he reported 'no degradation'. That afternoon: the trailing 5 minutes (14:25–14:30) had 2,000 attempts and 700 failures; the 15-minute outage (14:20–14:35) had 6,000 attempts and 1,800 failures; the trailing hour (13:30–14:30) had 24,000 attempts and 2,400 failures; the trailing 30 days had 4,320,000 attempts and 19,000 failures. What would the tile have read at 14:30 if it were computed over the window the SLO is defined on?
MCQ-1A payments team runs one dashboard for two decisions. The on-call engineer must decide within two minutes whether to roll back a deploy; the Monday business review decides which region gets the next marketing budget. The dashboard refreshes every 15 minutes and plots a 7-day rolling average of checkout error rate plus a regional conversion breakdown. Last Tuesday a deploy held the error rate at 6% for 11 minutes; the 7-day average moved from 0.40% to 0.41%, the panel showed nothing unusual, and the rollback was triggered by a customer tweet. Which change fixes the failure?
MCQ-8A payments team's on-call dashboard renders 42 tiles on one page (every endpoint's p50/p95/p99, error rate and saturation at 1-minute resolution) with the four signal tiles (error-budget burn, p99, saturation, last deploy) already in the top row. In a timed drill across 12 incidents, 8 on-call engineers took a median 3 min 40 s to decide whether to page, and in 6 of the 12 they scrolled past the signal tiles before finding the endpoint that was actually failing. The runbook requires the aggregate and the endpoint driving it to be readable in the same viewport, with no navigation away. Which redesign fits the drill data?
MCQ-15An executive dashboard places ten KPI cards in a single top row on a 1920×1080 canvas. Thirty days of session analytics show a median session of 26 seconds, 84% of sessions never scrolling past the first viewport, and card clicks split by position as 1st–3rd: 58%, 4th–7th: 9%, 8th–10th: 2%. The metric the CFO opens the dashboard for, 'Net Revenue vs Plan', is the 9th card. Which change is best supported by these numbers?
MCQ-16A dashboard opens with 24 tiles, each firing its own query, and it counts as loaded only once every tile has painted. All 24 sit on the first screen of the default layout. Twenty-three tiles read pre-aggregated daily rollups and return in about 200 ms; the last tile builds a 30-day distinct-user histogram that the warehouse answers by scanning 4 TB of raw events, taking about 40 s. The warehouse allows 10 concurrent queries per user, and the measured p95 time-to-all-tiles-painted is 41 s on cold loads with no warm cache. Which change cuts that p95 the most?
MCQ-10A payments dashboard reads from a warehouse table loaded by an ETL job every 15 minutes. The 14:45 UTC run succeeded and wrote rows through 14:41 UTC because the source system has a 4-minute replication delay. The dashboard page rendered at 15:02 UTC. An operator asks what 'as of' time the dashboard should display. Which timestamp should it show?
MCQ-14The exec dashboard's weekly-active-users line drops from 412k to 338k on March 2. That day, the WAU job's identity key was switched from device_id to user_id, so a person active on two devices is now counted once instead of twice. The drop is concentrated in accounts with more than one device. What should happen to the chart?
MCQ-9A revenue dashboard's filter bar has two selectors: Region (default 'All regions') and Date range (default 'Last 30 days'). It renders one chip per selector whose displayed value differs from its default, plus a Reset link. A saved view named 'EMEA board pack' is configured with Region = 'EMEA' and Date range = 'Last 90 days'. Opening it, an analyst sees a single chip reading 'Last 90 days' and a revenue tile reading $4.8M. Company-wide revenue for the last 90 days is $12.0M; EMEA revenue for the last 90 days is $4.8M. Setting the Region selector to 'All regions' by hand and reloading makes the same tile read $12.0M. What explains the single chip and the $4.8M tile?
MCQ-19The orders table holds 1,284,400 rows; status is NOT NULL and takes four values (open, packed, shipped, cancelled), and 321,100 rows are open. The Fulfilment Ops dashboard renders an 'Orders in scope' summary tile above an order grid, both reading only that table. This week the shared status filter's default was cleared to blank so reps would see everything: the summary now reads 1,284,400 while the grid returns 0 rows. The generated SQL shows why. With the widget empty the summary's builder emits SELECT count(*) FROM orders, and the grid's builder emits SELECT ... FROM orders WHERE status = ''. With any status selected, both builders emit the same status = <value> predicate. Reps want a blank filter to mean all statuses on both tiles. Which change delivers that?
MCQ-20A dashboard panel reports p99 = 180 ms for the checkout service, computed over a 5-minute rolling window with the filter region=eu-west and refreshed every 60 s. An engineer exports the panel to CSV during an incident. Two days later a reviewer re-runs the same query with a 1-hour window and no region filter, gets 240 ms, and files a bug that the dashboard under-reports latency. Which export content would have let the reviewer reconcile 180 ms with 240 ms without re-deriving the query?
MCQ-5A checkout service exports spans to a collector. During a 10-minute window, the collector's queue fills and 30% of spans from shard B are dropped; shard B carries the slowest checkout calls. The dashboard's p95 latency line falls from 180 ms to 45 ms. The dashboard has a coverage metric but no annotation, and its recovery policy requires 95% coverage. The on-call sees the drop and is about to close the incident. What should the dashboard show for that window?
MCQ-18A dashboard's time-range chips each fire `GET /api/metrics?range=<r>` and, on resolve, call `setSeries(json.series)`: no request id, no abort. Server latency measured over the last hour: `1h` ≈ 60 ms, `24h` ≈ 80 ms, `7d` ≈ 900 ms (the 7d query scans seven times the partitions). A user clicks `7d` at t=0 and `24h` at t=150 ms, then stops. The chip highlight follows the click immediately. What does the chart show once both responses have landed?
MCQ-17An incident dashboard has 12 tiles and a 1 s page SLA. Tile 7's inventory query exceeds its 800 ms deadline; the other 11 tiles return in 250 ms. Tile 7 has a cached snapshot 4 minutes old. The contract: each tile's freshness SLA is 60 s; if the live query fails, a cached snapshot up to 5 minutes old must be shown with a stale badge rather than as an error, and any rollup that consumes it must be marked stale. The total-units rollup sums tiles 1-12. What should the dashboard render?
MCQ-25A reviewer signs in to a pre-launch revenue dashboard with a `viewer` account scoped to EMEA. The 'Q3 Revenue' tile reads $4.2M with the region filter set to 'All regions'; the reviewer's own EMEA-scoped export for Q3 returns $1.1M. The tile's query log shows the query executed as `svc_dashboard`, not the signed-in user. What should the final review conclude?
MCQ-23A single dashboard serves both audiences from one semantic layer. The exec view renders six KPI tiles, each a query against the 2-billion-row fact table at p95 900 ms; the analyst view keeps its 24 slicers on that same table. The warehouse runs at most three queries per user concurrently, and each query beyond that limit adds 1200 ms to every query in flight. The exec page must load at p95 ≤ 2.5 s, KPI values must be no more than 15 minutes old, and the 40 metric definitions must stay single-sourced. Which plan meets all three constraints?
MCQ-21An SRE team runs a task-based usability test of a new incident dashboard with 8 on-call engineers, each working three tasks drawn from real pages from the last quarter. Task A, 'checkout is paging; decide whether to roll back the last deploy', 8/8 correct, median 35 s. Task B, 'find the panel showing Redis memory', 8/8, median 6 s. Task C: 'is payments within its error budget this week?': 3/8 correct, median 95 s; all five who answered wrong read the 7-day error-budget bar as a 30-day bar. Launch is one sprint away. What should the team do next?
MCQ-22A platform team runs an SRE dashboard. Last quarter it logged 1,400 views from 260 unique viewers, and the 12-engineer on-call rotation handled 25 incidents. Post-incident reviews show that a dashboard panel was opened at some point in 10 of those 25 incidents, but only in 3 was a panel opened before the first mitigation command, and in those 3 the responder used the panel to choose the mitigation. In the other 22 incidents the responder ran a runbook CLI command before opening any panel. MTTR was 38 minutes, the same as the previous quarter, when the dashboard logged 900 views. The team wants one number for the quarterly review that shows how often the dashboard changed what responders did. Which figure should they report?