Expect a pipeline round on computing daily per-merchant settlement summaries from a high-volume UPI transaction stream with exactly-once guarantees, a SQL round on payment-health monitoring (merchants whose success rate drops below a threshold, conditional aggregation, partitioned by state), and a feature-engineering round on low-latency rolling per-user fraud signals. Because money is involved, interviewers weight idempotency, deduplication, and reconciliation against the ledger over generic ETL. Ground every answer in real UPI settlement and fraud-scoring workflows, and be precise about how you avoid double-counting.
Recruitment process
- Recruiter screen and technical pre-screen
- SQL and data-manipulation round
- Payment-stream pipeline and feature-engineering round
- Hiring-manager and behavioural round, then offer
Interview rounds
- Round 1 (45-60 min): SQL round on conditional aggregation and payment-health metrics
- Round 2 (60 min): pipeline design round on exactly-once settlement aggregation
- Round 3 (45-60 min): fraud feature-engineering round with rolling windows
- Round 4 (45 min): behavioural and hiring-manager round
Exactly once and reconciled framework
- Guarantee exactly-once aggregation over payment streams with idempotency and dedup, monitor payment health in SQL, and engineer low-latency fraud features
Practice questions
- 01Practise
On-call has gotten miserable, engineers paged 3-4 times/night, half false alarms. Walk me through the redesign you'd lead.
StandardTech-leadership operational craft. Wants: alert-tier discipline (paging vs warning), runbook hygiene, blameless postmortem culture, on-call-load metric tracked weekly. 'We'd just fix the alerts' fails, wants the system change. - 02Practise
Given a stream of UPI transactions, find the top-K merchants by volume in the last 10 minutes with a sliding window. Walk me through your data structure choice.
IntensePhonePe technical rounds tie DSA to payments reality: expect stream-processing, idempotency, and concurrency framed around real UPI scale rather than abstract puzzles. - 03Practise
Implement a rate limiter for an API that must allow 5000 requests per second per merchant. Which algorithm, and how does it behave at the boundary?
Intense - 04Practise
Find the number of ways to make a payment of amount N using given UPI cashback denominations. Then optimise the space complexity of your DP.
StandardClassic coin-change framed in-domain. PhonePe wants the optimal-substructure reasoning stated aloud and the 2D→1D DP space optimisation. - 05Practise
Design a pipeline that computes daily per-merchant settlement summaries from a high-volume UPI transaction stream. How do you guarantee no transaction is double-counted?
IntensePhonePe data-engineer rounds probe exactly-once aggregation over payment streams: idempotency keys, dedup windows, and reconciliation against ledger. - 06Practise
Write SQL to detect merchants whose daily transaction success rate dropped below 90% on any day in the last week, across all states.
StandardPhonePe tests analytical SQL with conditional aggregation and partitioned thresholds for payment-health monitoring. - 07Practise
How would you build a fraud-signal feature pipeline that computes rolling per-user aggregates (velocity, distinct devices) with low latency for scoring?
IntensePhonePe probes real-time feature engineering: rolling windows, stateful stream processing, and latency budgets for fraud scoring.
