Prioritize conflicting tasks under shifting deadlines

You have three competing deliverables due this Friday: (A) the CFO wants an updated revenue forecast, (B) the PM needs an experiment readout for Monday’s launch call, and (C) Ops has raised a P0 data-quality problem affecting dashboards. You have an eight-hour window and no backup. Lay out your prioritization framework (for example, cost of delay, blast radius, reversibility), the precise order you would work through, what you would postpone or refuse, and the stakeholder communication plan (including pre-commitments and risk). Give the decision rules and a short sample message you would send to each stakeholder. Overview: This question tests a data scientist’s ability to prioritize, manage stakeholders, weigh trade-offs, and manage time while coordinating competing, time-sensitive requests. Solution

Assumptions (stated explicitly)

• It is Friday morning, with a single eight-hour block left before end of day.
• The PM needs the readout by end of day Friday to prepare for a Monday launch decision.
• The P0 is live and affecting dashboards used right now by several teams.
• No extra staffing is available today; I can timebox and triage, but I cannot fully re-architect. If any of these assumptions is incorrect, I would confirm before starting.

Prioritization Framework (decision rules)

I apply four lenses and a few straightforward rules:

1. Cost of Delay (CoD): the business impact if the work slips by 24 hours.
2. Blast Radius: how many users, teams, or decisions are affected right now.
3. Reversibility: how difficult it is to undo a decision made without this input.
4. Time to Mitigate: whether I can quickly stop the bleeding or offer a safe workaround. Decision rules:

• Rule 1: Triage any high-blast-radius P0 immediately; timebox the first repair attempt to 60–90 minutes. If it is not fixed, put a safe workaround in place and annotate the affected surfaces.
• Rule 2: Next, put the deliverable that gates an irreversible or high-cost decision (the PM’s launch readout) ahead of an executive update that can be reversed (the forecast can be delivered lean and deepened later).
• Rule 3: Reduce each deliverable to a “minimum decisionable product” (MDP); defer non-critical analysis.
• Rule 4: Communicate early, with pre-commitments, risks, and explicit trade-offs; get stakeholder agreement on scope.

Applying the Framework

• (C) P0 data quality:
  • CoD: Extreme; bad data can mislead many decisions today.
  • Blast Radius: High; dashboards across teams.
  • Reversibility: Low; once bad decisions are made, cleanup is expensive.
  • Action: Triage immediately; timebox to 60–90 minutes; if a full fix is not possible, apply a workaround and annotate.
• (B) Experiment readout (launch gating):
  • CoD: High; a delay can push Monday’s launch or risk shipping the wrong change.
  • Blast Radius: Medium–high; customers and operations are affected if the launch is wrong.
  • Reversibility: Medium/low; rollbacks are costly and damage reputation.
  • Action: Do this second; deliver a lean, decision-ready readout.
• (A) CFO forecast update:
  • CoD: Medium; important, but a top-line, range-based update can meet the end-of-day need.
  • Blast Radius: Medium; executive visibility, but less immediate operational damage.
  • Reversibility: High; the numbers can be refined Monday.
  • Action: Do this third; deliver a single-page top-line update with assumptions and risks.

Exact 8-Hour Sequence (time-boxed)

• 0:00–0:15
  • Open an incident channel. Confirm scope and impact with Ops. Post an immediate communication plan and the next update time.
• 0:15–1:30 (75 min)
  • P0 triage: Find the last good timestamp, roll back the offending change if known, disable or flag corrupted metrics, and backfill if possible. Add a dashboard banner: “Data stale from HH:MM; use with caution.”
• 1:30–1:45
  • Send status updates to all stakeholders. Lock in the P0 workaround or hand off the continuing fix details to Ops with clear ownership of next steps.
• 1:45–4:15 (150 min)
  • Experiment readout (MDP): validate data quality; compute the primary metric, effect size, and CI; check guardrails; make a decision recommendation; create a 3–5 slide summary.
• 4:15–4:30
  • Share the draft readout with the PM; collect only must-fix feedback.
• 4:30–7:15 (165 min)
  • Forecast update (MDP): refresh inputs, produce a range-based forecast, run sensitivity on 1–2 key drivers; create a one-pager with assumptions.
• 7:15–8:00
  • Final QA, send the deliverables, confirm Monday follow-ups, and post an end-of-day status in one update to everyone. Contingency: If the P0 runs past 90 minutes without a stable mitigation, freeze the broken tiles, pin a last-known-good snapshot, and move to B and A while Ops continues remediation.

What I Defer or Say No To (by deliverable)

• P0 incident:
  • Defer: Full root-cause analysis, long-term fixes, and the retrospective until next week.
  • Yes now: Triage, rollback/flag, user-facing annotations, and a safe state.
• Experiment readout:
  • Defer: Nice-to-have cuts, non-gating secondary metrics, and polished visuals.
  • Yes now: Primary KPI, CI/effect size, guardrails, power/SRM checks, and a clear ship/no-ship call with risks.
• Forecast update:
  • Defer: Model re-spec, new features, category deep dives, and a full slide deck.
  • Yes now: Top-line range, key drivers, scenario band, assumptions, and risks.

Guardrails and Mini-Checklists

• P0 triage quick checks:
  • When did the anomaly begin? Which metrics or dashboards are affected? Was there a deployment or event around that time? Can I revert or hotfix quickly? If not, disable the corrupted tiles and annotate.
• Experiment readout quick checks:
  • Run the Sample Ratio Mismatch (SRM) test.
  • Check data freshness given the P0; exclude the corrupted window if needed.
  • Check primary metric effect size and 95% CI; compare harm probabilities against guardrails.
  • Example: $\text{Uplift} = +2.5\%$ (95% CI: $+0.4\%$ to $+4.6\%$), $P(\text{harm} > 0) < 5\%$; recommend a staged rollout with monitoring.
• Forecast quick checks:
  • Use the last-known-good data cut. Produce a range with sensitivity to 1–2 drivers.
  • Example: Q4 revenue estimate $124M–$129M (base $126.5M), ±$2.5M driven by order volume and AOV; risks: promo depth and seasonality variance.

Example Stakeholder Messages (concise)

• To Ops (P0) at start:
  • “I’m on the P0 now. I’ll send a first update in 60 minutes. I’m timeboxing triage to 90 minutes; if it is not fully resolved, I’ll freeze the affected tiles to the last-good state and post a banner. Can you confirm the earliest anomaly time and any recent ETL jobs or deploys?”
• To PM (experiment) after P0 mitigation begins:
  • “Heads-up: I’m triaging a P0 until about 1:30. I will deliver a decision-ready experiment readout by 4:30 that covers the primary KPI, CI, guardrails, and recommendation. I’ll defer secondary cuts and polish to Monday. If the P0 affects data freshness, I’ll note any exclusions. Risk: medium that we use a slightly smaller window; mitigation: guardrails and a staged rollout.”
• To CFO (forecast) after P0 mitigation begins:
  • “Plan: I’ll send a one-page revenue forecast update by 7:15 with a top-line range, key drivers, and assumptions. I’ll defer deep dives and model changes to Monday. If today’s P0 affects any inputs, I’ll use the last-known-good cut and flag assumptions. Please confirm this scope covers today’s need.”
• End-of-day status (all):
  • “End-of-day status: (1) P0 mitigated: dashboards annotated; last-good snapshot pinned; full RCA on Monday. (2) Experiment readout sent: primary KPI $+2.5\%$ (95% CI $+0.4\%$ to $+4.6\%$); recommend staged rollout with guardrails; details attached. (3) Forecast update sent: $124M–$129M with assumptions and risks. Monday: finalize RCA, deep-dive experiment cuts, and forecast refinements.”

Pre-commitments and Risk Management

• Pre-commitments:
  • Deliver P0 mitigation within 90 minutes (rollback/annotation) or freeze to a safe state.
  • Provide the PM a decisionable readout by 4:30 with a clear recommendation and risks.
  • Provide the CFO a one-pager by 7:15 with a defensible range and assumptions.
• Risks and mitigations:
  • If the P0 data gap overlaps experiment or forecast windows, use the last-good window and document exclusions; avoid over-precision.
  • If the experiment is borderline, recommend a staged rollout and define kill-switch metrics.
  • If forecast uncertainty is elevated, widen the interval and call out the top risk drivers.

Why this order

• It reduces immediate organizational harm (the P0), protects a time-critical product decision (the experiment readout), and still gives executive visibility with acceptable uncertainty (the forecast). Seniority alone does not set priority; cost of delay, blast radius, and reversibility do.