Pinterest Software Engineer Interview Experience — Onsite With Two System Design Rounds and Escalating Follow-ups

Round 1 — Behavioral

A standard set of leadership questions, one by one:

• Self-introduction
• Pick a project where you had the strongest ownership, or learned the most from, and go deep on it
• How did you measure success for this project?
• What's the strongest team environment you've been part of?
• A time you had to make a fast decision and live with the consequences
• A time you got constructive feedback from a manager or coworker — how did you handle it, what did you learn
• What's your biggest strength?
• Ever had a conflict with a coworker? What was the outcome, what did you learn
• Questions for the interviewer

Round 2 — Coding #1

Q1: Given a list of numbers and a target, insert + or * between the numbers (you can't reorder them) and ask whether you can reach the target — return a bool. Evaluate left to right first, ignore operator precedence.
DFS / backtracking, follow-up asked about time/space complexity (2^n, stack depth n)

Q2 (follow-up): Keep the function above as is, write a new function with the same signature, but this time respect operator precedence (multiplication before addition), and also return all expressions that reach the target.

Round 3 — Coding #2

The interviewer worked on ad retrieval / candidate generation. This round was clearly a three-part sweep-line / difference-array follow-up chain:

Q1: Given a bunch of logs, each line is (pin_id, start_time, end_time), meaning this pin was being interacted with by a user during that time window. Output how many pins are active during each time slice. First constrain the time window to just 60 seconds (timestamps 0–59).
Difference map: +1 at start, −1 at end, sweep left to right accumulating.

Q2: Remove the "only a 60-second window" constraint — timestamps can be any value. Does the same approach still work? What's the complexity now?
You can't iterate 0 to 60 anymore, instead sort the timestamps that actually appear and sweep → O(k log k)

Q3: Remove the assumption that "pins don't repeat" — the same pin can be interacted with by multiple users in overlapping time windows, and the output wants the deduplicated pin count.
A plain counter isn't enough anymore — you need to maintain, for each time point, a "set of pins to add / set of pins to remove," plus a map from pin_id to its current active count; when the count hits zero, delete the key, and use the number of keys in the map as the current unique pin count.

Round 4 — System Design: Search Typeahead / Autocomplete

Problem: given a prefix, return the top 10 suggestions ranked by search popularity; exact prefix match; requires real-time freshness (right after a World Cup goal, the autocomplete needs to reflect it immediately). Opened by saying "let's assume it fits on a single machine for now, we'll scale later." Given scale: 1 billion users x 10 searches per day each.

Follow-ups (roughly in order):

• Why use a trie instead of a direct KV store (prefix → top 10)?
• How often does the offline job run?
• After the offline job finishes and swaps in the new trie, don't you lose the real-time updates from the hours in between?
• So what's actually different between offline batch processing and pure streaming?
• Does accuracy matter in this system?
• What if it doesn't fit on one machine, or QPS is too high?
• How do you handle uneven sharding (entries starting with "A" are way more common than "Z")?
• The distribution shifts over time and one small shard suddenly blows up — how do you handle that dynamically? Walk through it step by step (they wanted new-node warmup plus routing switchover covered)
• What if the tree itself isn't big but QPS is too high? How does a replica sync with the leader?

Round 5 — System Design: Bulk Update of Merchant Product Attributes

The interviewer was a senior full-stack engineer from the Relevance Measuring team.

Problem: design a system for merchants to upload/update product attributes (price, stock status, description, images) to annotate a product pin. For images, merchants upload the original and the server resizes it; the system needs to support both single-item updates and bulk file updates for an entire catalog. Image updates are lower priority.

Given numbers: 50,000 merchants, 1 billion products; write spikes 0–100k QPS, reads a fairly steady 100k QPS; roughly 100 batch jobs; a normal attribute payload is ~1KB, images average 500KB; price and stock need to take effect quickly, images can lag.

Follow-ups:

• What database would you choose, and why
• If a single-item update and a batch update hit the same product at the same time, how do you handle the conflict