Diagnose London Task Completion Drop

Business Context

You’re the analytics lead for Uber’s last‑mile delivery product in the UK (think: grocery + convenience delivery). The London market averages ~450k weekly active consumers, ~35k weekly active couriers, and processes ~1.8M delivery tasks/week (pickup, in-store substitution, dropoff, photo proof, etc.). Task completion is a critical operational KPI because it drives on-time delivery, refunds, courier earnings, and repeat purchase.

Last week (Mon–Sun), the task completion rate in London dropped from 96.2% to 93.4% (−2.8pp). Other UK cities (Manchester, Birmingham, Glasgow) were flat. Operations leadership is asking within 48 hours:

Is this a real behavioral/operational issue or a measurement/instrumentation problem?
If real, what changed (product release, merchant mix, courier supply, weather, fraud, outages)?
What are the top 1–2 drivers we can act on this week, and what are the risks of doing so?

Assume there was a mobile app release mid-week that changed the “Confirm Pickup” screen for couriers, and London also experienced two days of heavy rain.

Your Task

Provide a structured investigation plan and the initial metric framework you would use.

Requirements (what you must produce)

Define the metric precisely: What counts as a “task,” what counts as “completed,” and what are the inclusion/exclusion rules (e.g., canceled orders, reassigned tasks, retries, offline mode).
Quantify the drop: Show how you’d validate the magnitude and rule out artifacts (logging gaps, schema changes, delayed events).
Decompose the metric into a funnel and identify which step(s) drove the decline (e.g., task created → accepted → started → completed).
Segment the drop across at least 4 dimensions (e.g., courier app version, merchant type, zone, time-of-day, network connectivity).
Propose 3–5 plausible hypotheses and the specific analyses (or quick experiments) you’d run to confirm/deny each.
Recommend concrete actions for the most likely driver(s), including guardrail metrics (e.g., cancellations, delivery time, refunds, courier churn).

Data Available

You have access to a Snowflake warehouse with the following sources:

Table / Log	What it contains	Grain
`delivery_tasks`	task_id, order_id, city, zone, task_type, created_at, scheduled_at, merchant_id, courier_id	task
`task_events`	task_id, event_type (created/accepted/started/completed/failed/canceled/reassigned), event_ts, failure_code	event
`courier_app_sessions`	courier_id, session_start, session_end, app_version, os, device_model, network_type, crash_count	session
`orders`	order_id, basket_value, promised_eta, delivered_at, cancel_reason, refund_amount	order
`merchant_dim`	merchant_id, vertical (grocery/convenience), store_size, avg_prep_time	merchant
`courier_supply_hourly`	city, zone, hour, online_couriers, active_deliveries, surge_multiplier	hour-zone
`incident_log`	service_name, start_ts, end_ts, severity, impacted_region	incident
`weather_hourly`	city, hour, precipitation_mm, temperature_c	hour

Constraints: you have 2 days to deliver a readout. You can request one additional dataset if necessary, but assume engineering bandwidth is limited.

Business Context

Is this a real behavioral/operational issue or a measurement/instrumentation problem?
If real, what changed (product release, merchant mix, courier supply, weather, fraud, outages)?
What are the top 1–2 drivers we can act on this week, and what are the risks of doing so?

Assume there was a mobile app release mid-week that changed the “Confirm Pickup” screen for couriers, and London also experienced two days of heavy rain.

Your Task

Provide a structured investigation plan and the initial metric framework you would use.

Requirements (what you must produce)

Define the metric precisely: What counts as a “task,” what counts as “completed,” and what are the inclusion/exclusion rules (e.g., canceled orders, reassigned tasks, retries, offline mode).
Quantify the drop: Show how you’d validate the magnitude and rule out artifacts (logging gaps, schema changes, delayed events).
Decompose the metric into a funnel and identify which step(s) drove the decline (e.g., task created → accepted → started → completed).
Segment the drop across at least 4 dimensions (e.g., courier app version, merchant type, zone, time-of-day, network connectivity).
Propose 3–5 plausible hypotheses and the specific analyses (or quick experiments) you’d run to confirm/deny each.
Recommend concrete actions for the most likely driver(s), including guardrail metrics (e.g., cancellations, delivery time, refunds, courier churn).

Data Available

You have access to a Snowflake warehouse with the following sources:

Table / Log	What it contains	Grain
`delivery_tasks`	task_id, order_id, city, zone, task_type, created_at, scheduled_at, merchant_id, courier_id	task
`task_events`	task_id, event_type (created/accepted/started/completed/failed/canceled/reassigned), event_ts, failure_code	event
`courier_app_sessions`	courier_id, session_start, session_end, app_version, os, device_model, network_type, crash_count	session
`orders`	order_id, basket_value, promised_eta, delivered_at, cancel_reason, refund_amount	order
`merchant_dim`	merchant_id, vertical (grocery/convenience), store_size, avg_prep_time	merchant
`courier_supply_hourly`	city, zone, hour, online_couriers, active_deliveries, surge_multiplier	hour-zone
`incident_log`	service_name, start_ts, end_ts, severity, impacted_region	incident
`weather_hourly`	city, hour, precipitation_mm, temperature_c	hour

Constraints: you have 2 days to deliver a readout. You can request one additional dataset if necessary, but assume engineering bandwidth is limited.

Business Context

Is this a real behavioral/operational issue or a measurement/instrumentation problem?
If real, what changed (product release, merchant mix, courier supply, weather, fraud, outages)?
What are the top 1–2 drivers we can act on this week, and what are the risks of doing so?

Assume there was a mobile app release mid-week that changed the “Confirm Pickup” screen for couriers, and London also experienced two days of heavy rain.

Your Task

Provide a structured investigation plan and the initial metric framework you would use.

Requirements (what you must produce)

Define the metric precisely: What counts as a “task,” what counts as “completed,” and what are the inclusion/exclusion rules (e.g., canceled orders, reassigned tasks, retries, offline mode).
Quantify the drop: Show how you’d validate the magnitude and rule out artifacts (logging gaps, schema changes, delayed events).
Decompose the metric into a funnel and identify which step(s) drove the decline (e.g., task created → accepted → started → completed).
Segment the drop across at least 4 dimensions (e.g., courier app version, merchant type, zone, time-of-day, network connectivity).
Propose 3–5 plausible hypotheses and the specific analyses (or quick experiments) you’d run to confirm/deny each.
Recommend concrete actions for the most likely driver(s), including guardrail metrics (e.g., cancellations, delivery time, refunds, courier churn).

Data Available

You have access to a Snowflake warehouse with the following sources:

Table / Log	What it contains	Grain
`delivery_tasks`	task_id, order_id, city, zone, task_type, created_at, scheduled_at, merchant_id, courier_id	task
`task_events`	task_id, event_type (created/accepted/started/completed/failed/canceled/reassigned), event_ts, failure_code	event
`courier_app_sessions`	courier_id, session_start, session_end, app_version, os, device_model, network_type, crash_count	session
`orders`	order_id, basket_value, promised_eta, delivered_at, cancel_reason, refund_amount	order
`merchant_dim`	merchant_id, vertical (grocery/convenience), store_size, avg_prep_time	merchant
`courier_supply_hourly`	city, zone, hour, online_couriers, active_deliveries, surge_multiplier	hour-zone
`incident_log`	service_name, start_ts, end_ts, severity, impacted_region	incident
`weather_hourly`	city, hour, precipitation_mm, temperature_c	hour

Constraints: you have 2 days to deliver a readout. You can request one additional dataset if necessary, but assume engineering bandwidth is limited.

Business Context

Is this a real behavioral/operational issue or a measurement/instrumentation problem?
If real, what changed (product release, merchant mix, courier supply, weather, fraud, outages)?
What are the top 1–2 drivers we can act on this week, and what are the risks of doing so?

Assume there was a mobile app release mid-week that changed the “Confirm Pickup” screen for couriers, and London also experienced two days of heavy rain.

Your Task

Provide a structured investigation plan and the initial metric framework you would use.

Requirements (what you must produce)

Define the metric precisely: What counts as a “task,” what counts as “completed,” and what are the inclusion/exclusion rules (e.g., canceled orders, reassigned tasks, retries, offline mode).
Quantify the drop: Show how you’d validate the magnitude and rule out artifacts (logging gaps, schema changes, delayed events).
Decompose the metric into a funnel and identify which step(s) drove the decline (e.g., task created → accepted → started → completed).
Segment the drop across at least 4 dimensions (e.g., courier app version, merchant type, zone, time-of-day, network connectivity).
Propose 3–5 plausible hypotheses and the specific analyses (or quick experiments) you’d run to confirm/deny each.
Recommend concrete actions for the most likely driver(s), including guardrail metrics (e.g., cancellations, delivery time, refunds, courier churn).

Data Available

You have access to a Snowflake warehouse with the following sources:

Table / Log	What it contains	Grain
`delivery_tasks`	task_id, order_id, city, zone, task_type, created_at, scheduled_at, merchant_id, courier_id	task
`task_events`	task_id, event_type (created/accepted/started/completed/failed/canceled/reassigned), event_ts, failure_code	event
`courier_app_sessions`	courier_id, session_start, session_end, app_version, os, device_model, network_type, crash_count	session
`orders`	order_id, basket_value, promised_eta, delivered_at, cancel_reason, refund_amount	order
`merchant_dim`	merchant_id, vertical (grocery/convenience), store_size, avg_prep_time	merchant
`courier_supply_hourly`	city, zone, hour, online_couriers, active_deliveries, surge_multiplier	hour-zone
`incident_log`	service_name, start_ts, end_ts, severity, impacted_region	incident
`weather_hourly`	city, hour, precipitation_mm, temperature_c	hour

Constraints: you have 2 days to deliver a readout. You can request one additional dataset if necessary, but assume engineering bandwidth is limited.

Interview Guides

Business Context

Your Task

Requirements (what you must produce)

Data Available

Diagnose London Task Completion Drop

Business Context

Your Task

Requirements (what you must produce)

Data Available

Your Answer

Diagnose London Task Completion Drop

Business Context

Your Task

Requirements (what you must produce)

Data Available

Diagnose London Task Completion Drop

Business Context

Your Task

Requirements (what you must produce)

Data Available

Your Answer