Resolve cross-team conflict and align incentives

# Example Answer (Teaching-Oriented) ## 1) Scenario Setup (Tight Timeline, Misaligned Incentives) - Goal: Launch a new feed-ranking model that requires a new event log (feature interactions) to train, evaluate, and monitor post-launch. - Dependency: The Logging/Infrastructure team must add a schema and provision capacity. They are in a stability period and have an OKR to reduce write QPS by 10% and avoid schema churn. - Deadline: Model launch in 3 weeks tied to a marketing event; missing it defers launch to next quarter. Why it matters: - Without the new log, we cannot (a) validate offline improvements with online metrics, (b) detect model drift, or (c) compute key features. Launching blind increases risk of regression and post-launch rollback. ## 2) Stakeholder Map: Objectives, Constraints, BATNAs - Product DS (you) - Objectives: Ship on time with measurable uplift; ensure data quality and monitoring; minimize risk. - Constraints: Need the log to compute features and validate impact; credibility on the line. - BATNA: Ship model without the new log using proxy features; reduced expected lift; higher risk and weaker monitoring. - Product Manager (PM) - Objectives: Hit date; achieve engagement/retention uplift; de-risk headline metrics. - Constraints: Fixed marketing slot; limited buffer. - BATNA: De-scope; partial rollout later; lose peak exposure and momentum. - Ranking Engineering Lead - Objectives: Low regression risk; maintain latency budgets; avoid last-minute refactors. - Constraints: Limited engineer bandwidth; on-call risk. - BATNA: Launch a smaller model that doesn’t require new logging; lower expected impact. - Logging/Infrastructure Lead - Objectives: Protect SLAs and error budgets; reduce write QPS; avoid schema churn during stability period. - Constraints: SRE policies; storage/cost constraints; change freeze milestones. - BATNA: Defer schema change to next quarter; keep infra OKRs on track. - SRE/On-Call - Objectives: Avoid incidents; preserve error budget; maintain on-call load. - Constraints: Strict change management; capacity limits. - BATNA: Reject changes that raise risk without mitigation; stick to freeze. - Privacy/Policy (if sensitive data) - Objectives: Data minimization; compliance. - Constraints: Review timelines. - BATNA: Deny or delay until approvals complete. ## 3) Data-Backed Negotiation Plan A. Quantify benefits, costs, risks (with a small numeric example) - Expected benefit from model (from offline validation + analogous A/Bs): +0.5% session time. - User base: 200M DAU; baseline total daily session minutes = 50 minutes/DAU → 10B minutes/day. - Uplift: 0.5% × 10B = 50M incremental minutes/day. - If we value 1M minutes at $10k in long-run revenue proxy, then benefit ≈ $500k/day. - Delay cost: 1-week slip ≈ $3.5M in foregone value. - Infra cost estimate (from perf profiling): new event increases write QPS by +0.7% peak; storage +2 TB/day; operational risk adds 3% probability of a P1 in first week if full-scale. - Incident cost proxy: P1 estimated cost = $1M (engineering time, user impact). Expected loss if full-scale launch: 0.03 × $1M = $30k; with 1% canary: 0.5% incident probability → $5k expected. - Decision framing: - EV(Launch Now) = Benefit − Expected Incident Cost − Extra Infra Cost. - Example (with 1%/10%/100% staged rollout): Day 1 EV ≈ $500k − $5k − storage/network marginal cost (small) → strongly positive with staged rollout + guardrails. B. Options with trade-offs - Option A (Minimal viable logging, sampled 10%, treatment-only) - Pros: 10× smaller QPS; enough data to evaluate model and train next iteration; lower risk. - Cons: Slightly slower learning; need sampling-aware estimators. - Option B (Client-side buffered logging with compression + off-peak flush) - Pros: Reduces peak QPS; infra-friendly. - Cons: Increases client complexity; potential data latency. - Option C (Derived feature from existing logs; no schema change) - Pros: Zero infra change; fastest to implement. - Cons: Lower model lift (estimate +0.2%); noisier metrics. - Option D (Defer to next quarter, maintain full fidelity logging later) - Pros: Lowest immediate risk. - Cons: Foregone value ≈ $3.5M/week; loses momentum. C. Proposed plan - Recommend Option A now, with an evolution path to full fidelity later if guardrails hold. - RICE scoring (example): - A: Reach high, Impact high, Confidence medium-high (0.75), Effort medium → best score. - C: Reach high, Impact medium-low, Confidence high, Effort low → second-best. D. Experiment and guardrails - Rollout: 1% canary (24–48h) → 10% (3–5 days) → 50% → 100% by Day 10 if guardrails pass. - Primary success metrics: +0.5% session minutes; +0.3% retention D7; neutral on creator feedback rate. - Guardrails: error rate, p95 write latency, app crash rate, privacy checks. Kill switch if any exceed thresholds. - Statistical design: CUPED or pre-period adjustment; ensure power at 10% rollout with sampled logs. Correct for sampling in estimators. E. Risk mitigations - Log sampling 10%; TTL 14 days; column-level compression; privacy review pre-approved schema. - Pre-canary load test to confirm <0.2% impact on peak write QPS. - DS provides on-call monitoring dashboard; Eng sets feature flag and automated rollback. F. Incentive alignment - Joint success metric across teams: "Launch with <0.2% QPS delta at p95 and +0.5% session uplift at p<0.05 or stop." Both teams get OKR credit. - DS/PM offer to fund infra tickets (1 eng-week) to retire debt created by the change. ## 4) Exact Meeting Agenda and Decision Log A. Meeting agenda (45 minutes) - Pre-reads (sent 24h prior): 1-page proposal, risk matrix, perf/load test plan, experiment design. - Attendees: PM (chair), DS (presenter), Ranking Eng Lead, Infra Lead, SRE rep, Privacy (if needed). - Purpose (5m): Decide whether/how to implement logging to meet launch date. - Context recap (5m): Deadline, expected value, constraints, decision criteria. - Options review (10m): A/B/C/D with data (benefit, cost, risk, effort). - SRE/Infra risk assessment (10m): Error budget, capacity, freeze policies; discuss mitigations. - Decision (10m): Select option, rollout plan, guardrails, owners, timeline. - Next steps and comms (5m): Owners, docs, Slack channel, checkpoints. Decision criteria: Net expected value positive, risk within error budget, compliance approved, rollout + kill-switch in place. B. Decision log template - Title: - Decision: [e.g., Adopt Option A: 10% sampled logging, treatment-only] - Date/Time: - Approvers: PM, Infra Lead, Eng Lead, DS - Participants: - Context: - Problem, deadline, impacted metrics, dependencies. - Options considered: - A, B, C, D (with brief pros/cons, data). - Data snapshot: - Expected benefit, QPS delta, storage delta, incident probability, EV, load test results. - Decision: - Chosen option, rollout plan, guardrails, success criteria, kill-switch owner. - Dissent/concerns and how addressed: - Action items: - Owners, due dates, checkpoints. - Review date: - Date to re-evaluate or expand to full fidelity. Example entry (condensed) - Decision: Proceed with Option A. 1%→10%→50%→100% rollout over 10 days, if guardrails pass. - Data: +0.5% session minutes; QPS delta +0.1% at 10% sample; expected incident cost <$10k; EV strongly positive. - Guardrails: p95 write latency <+2%; error rate <+0.05%; crash rate unchanged; privacy sign-off complete. - Owners: Eng Lead (flagging), DS (monitoring), Infra (capacity watch), SRE (alerts). ## 5) Escalation Criteria and Relationship Preservation Escalation criteria (objective triggers) - Blocked >48 hours on a decision critical to hitting the date, with no viable alternative. - Infra/SRE risk assessment exceeds thresholds despite mitigations, and PM/DS disagree on trade-offs. - Data shows expected value >3× expected risk, but decision stalls or lacks an owner. - Legal/privacy approval cannot meet deadline and no compliant alternative exists. Escalation path - Level 1: Team leads (PM, Eng, Infra) meet with Directors for tie-break, using the decision doc. - Level 2: If still unresolved within 24–48h, escalate to cross-org GM/VP. Present 1-page: options, EV math, risks, what you tried, recommended path. Preserving relationships if blocked - Separate people from the problem: acknowledge Infra’s OKRs and constraints upfront. - Offer help: commit DS/Eng time to mitigate infra debt; co-own success metrics. - Use "yes-and" language: "We need X by date; we can reduce risk via sampling, canary, and rollback. If this still violates freeze, can we align on Option C now and pre-commit to A by [date]?" - Keep a blameless written record; thank contributors; follow up with shared learnings and recognition. ## 6) Measuring Success Beyond the Immediate Deliverable Process/relationship metrics - Cross-team lead time: time from dependency request → decision → first canary; target −30% quarter over quarter. - On-time dependency rate: % of cross-team dependencies delivered by planned date. - Escalation count and time-to-resolution: aim to reduce frequency and shorten cycle time. - Stakeholder trust/NPS: quarterly pulse ("I can get timely support from X"), target +10 points. - Pre-read engagement: open rate and comments before meetings; aim >80% opens and substantive comments. Reliability/quality metrics - SLO adherence/error budget burn pre/post change. - Incident rate/severity during rollout; time to rollback if triggered. - Data quality: missingness, latency, schema change stability; sampling-bias diagnostics. Business/learning metrics - Realized uplift vs forecast; calibration error of impact estimates. - Time-to-learning: days to stable readout; speed of iteration to full fidelity logging. - Reuse: number of teams adopting the logging pattern or shared library. Validation and guardrails - Pre-commit thresholds for go/no-go and rollback; publish a dashboard visible to all stakeholders. - Postmortem with actions and owners; track action closure within 30 days. This approach shows you can translate misaligned incentives into a structured negotiation with quantified trade-offs, a clear decision process, respectful escalation, and durable improvements to how teams work together.