Quick Overview

This question evaluates proficiency in time-windowed aggregations, event deduplication, and window-function–based user-level analytics across SQL and Pandas workflows.

Query seven-day conversion with windows and dedupe

Company: Snowflake

Role: Data Scientist

Category: Data Manipulation (SQL/Python)

Difficulty: medium

Interview Round: Technical Screen

Assume all timestamps are UTC. Treat "today" as 2025-09-01. Define "last 7 days" as the inclusive window [2025-08-26 00:00:00, 2025-09-01 23:59:59]. Schema (invented): users(user_id INT, signup_at TIMESTAMP, tz STRING) orders(order_id INT, user_id INT, order_ts TIMESTAMP, amount DECIMAL(10,2), status STRING) -- status in ('completed','canceled') events(event_id INT, user_id INT, event_ts TIMESTAMP, event_type STRING, product_id INT, device_id STRING) Sample rows: users user_id | signup_at | tz 1 | 2025-08-10 09:00:00 | America/Los_Angeles 2 | 2025-08-28 14:00:00 | UTC 3 | 2025-08-30 23:30:00 | America/New_York orders order_id | user_id | order_ts | amount | status 101 | 1 | 2025-08-27 23:55:00| 20.00 | completed 102 | 1 | 2025-08-28 00:05:00| 15.00 | canceled 103 | 2 | 2025-08-31 12:00:00| 9.99 | completed 104 | 2 | 2025-09-01 00:00:10| 49.00 | completed 105 | 3 | 2025-08-26 23:59:59| 5.00 | completed events event_id | user_id | event_ts | event_type | product_id | device_id 1 | 1 | 2025-08-27 23:50:00| view_product | 555 | A 2 | 1 | 2025-08-27 23:50:00| view_product | 555 | A -- duplicate of 1 3 | 1 | 2025-08-28 00:03:00| add_to_cart | 555 | A 4 | 1 | 2025-08-28 00:04:00| purchase | 555 | A 5 | 2 | 2025-08-31 11:58:00| view_product | 777 | B 6 | 3 | 2025-08-26 23:55:00| view_product | 999 | C 7 | 3 | 2025-08-26 23:58:00| purchase | 999 | C 8 | 2 | 2025-09-01 00:00:10| purchase | 777 | B Tasks (write a single Standard SQL query; CTEs allowed): 1) At the day level (UTC days), for each day in the 7-day window, compute: unique_viewers = COUNT(DISTINCT user_id with >=1 view_product event that day after deduplicating exact duplicate events), purchasers = COUNT of completed orders that day (exclude canceled), and conversion_rate = purchasers / NULLIF(unique_viewers,0). 2) At the user level, within the same window, return for every user_id: first_completed_order_ts, last_completed_order_ts (NULL if none), and total_completed_orders. Use window functions, not correlated subqueries. 3) Event deduplication: treat rows with identical (user_id, event_ts, event_type, product_id, device_id) as duplicates; keep only one (e.g., ROW_NUMBER over these keys). 4) Output two result sets: (a) daily metrics ordered by day; (b) user-level metrics ordered by user_id. Bonus: Provide a concise Pandas solution sketch achieving (1)-(3) with correct time windowing and deduplication.

Overview: This question evaluates proficiency in time-windowed aggregations, event deduplication, and window-function–based user-level analytics across SQL and Pandas workflows.

7-day daily conversion metrics (UTC), with exact-event deduplication

Assume all timestamps are stored in UTC. Treat "today" as **2025-06-01**, and define the **last 7 days** as the inclusive window from `2025-05-26 00:00:00` to `2025-06-01 23:59:59`. Using PostgreSQL, return one row **for every UTC calendar day in the 7-day window** (2025-05-26 through 2025-06-01), even on days that have zero events or orders. For each day report: - **`day`** — the UTC calendar day (a `date`). - **`unique_viewers`** — `COUNT(DISTINCT user_id)` of users who had at least one *deduplicated* `view_product` event that day (`0` if none). - **`purchasers`** — the number of completed orders placed that day, i.e. `COUNT(*)` over `orders` where `status = 'completed'` (exclude `canceled`); `0` if none. - **`conversion_rate`** — `purchasers / unique_viewers`, returned as a numeric rounded to 4 decimal places. When `unique_viewers = 0` the rate must be `NULL` (use `NULLIF` to avoid division by zero). **Event-deduplication rule:** within the `events` table, rows that share an identical tuple `(user_id, event_ts, event_type, product_id, device_id)` are duplicates; keep exactly one of them (e.g. via `ROW_NUMBER()` partitioned by those five columns) before counting viewers. The day a row belongs to is the UTC day of its timestamp (`event_ts` for events, `order_ts` for orders). Only rows whose timestamp falls inside the 7-day window count. **Order the final output by `day` ascending.**

Tables

users(user_id INT, signup_at TIMESTAMP, tz VARCHAR(64))

orders(order_id INT, user_id INT, order_ts TIMESTAMP, amount DECIMAL(10,2), status VARCHAR(16))

events(event_id INT, user_id INT, event_ts TIMESTAMP, event_type VARCHAR(32), product_id INT, device_id VARCHAR(64))

Hints

  1. Build a calendar with generate_series(start_day, end_day, INTERVAL '1 day') and LEFT JOIN your daily aggregates onto it so zero-activity days still appear.
  2. Deduplicate events with ROW_NUMBER() OVER (PARTITION BY user_id, event_ts, event_type, product_id, device_id ORDER BY event_id) and keep rn = 1 before counting viewers.

7-day user-level completed-order metrics using window functions

Assume all timestamps are UTC. Treat “today” as 2025-06-01. Define the “last 7 days” as the inclusive window from 2025-05-26 00:00:00 to 2025-06-01 23:59:59. Using Standard SQL (CTEs allowed) and window functions (no correlated subqueries), return one row per user_id (including users with zero completed orders in the window) with: - first_completed_order_ts: earliest completed order timestamp in the window (NULL if none) - last_completed_order_ts: latest completed order timestamp in the window (NULL if none) - total_completed_orders: number of completed orders in the window (0 if none) Exclude canceled orders. Order the output by user_id ascending.

Tables

users(user_id INT, signup_at TIMESTAMP, tz VARCHAR(64))

orders(order_id INT, user_id INT, order_ts TIMESTAMP, amount DECIMAL(10,2), status VARCHAR(16))

events(event_id INT, user_id INT, event_ts TIMESTAMP, event_type VARCHAR(32), product_id INT, device_id VARCHAR(64))

Hints

  1. Filter to completed orders within the timestamp window first, then apply window functions.
  2. Use MIN/MAX/COUNT window functions partitioned by user_id, and pick one row per user with ROW_NUMBER().

Loading coding console...