This post is based on 44 firsthand reports from candidates who interviewed at Confluent between 2018 and 2026. Cross-checking accounts of the same questions produced 33 distinct problems: 18 coding, 2 concurrency/low-level design, and 13 system design.

Each FastPrep practice problem follows what candidates remember being asked: from the opening prompt to the constraints, expected APIs, trade-offs, and follow-up questions. When several candidates reported the same problem, their accounts were combined into one fuller version. Any detail added only to make a problem runnable is clearly called out.

Throughout this post, every underlined green problem title is clickable and takes you directly to the corresponding practice problem on FastPrep.

Lastly, hope this post can help anyone who might need it. Happy prep! 🫶

The strongest repeat signals

If preparation time is limited, start here:

1. TinyURL / URL shortener system design — 23 source reports spanning November 2018 to January 2025.
2. Subscription news feed API and data model — 12 dated reports from March 2019 through May 2026.
3. Thread-safe delayed task runner — 9 onsite reports from September 2020 through January 2025.
4. Windowed key-value map — 6 dated onsite reports from March 2021 through November 2024.
5. tail -n over a supplied file API — 5 reports in March–May 2026.
6. Disposable email service — 6 source records across five report dates in 2025–2026.

Those repeats reveal a consistent Confluent preference: stateful APIs, concurrency, storage behavior, precise contracts, and failure-aware design.

Coding and implementation: 18 problems

🪿 Repeated coding questions

1. Windowed Key-Value Map

Evidence: onsite; mainly full-time. Report dates: March 5 and March 8, May 11, and November 9, 2021; August 20, 2022; November 21, 2024.

The recurring contract is a map over a five-minute event window with put, get, delete, and a frequently called average. The important follow-ups are expiration semantics, keeping average better than a full scan, overwrites, and multithreaded access. This is one of the clearest Confluent coding signatures.

2. Implement tail -n

Evidence: onsite; full-time. Report dates: March 4, April 3, April 15, May 2, and May 4, 2026.

Five near-consecutive reports describe the same task. The interviewer supplies file operations such as reading bytes, moving a pointer, and obtaining file size. Candidates are expected to stream the final n lines to stdout and discuss forward buffering versus scanning backward from the end of the file.

3. LRU Cache with Concurrency Follow-Up

Evidence: onsite; new-grad and full-time, with a later staff-level concurrency corroboration. Report dates: November 20, 2018; March 21, 2019; February 18, 2020; June 16, 2022.

The base task is the standard put/get cache using a hash map plus a hand-written doubly linked list. Confluent’s distinctive follow-up is synchronization: get is not really read-only because it mutates recency, so candidates must reason about lock scope and concurrent reads/writes.

4. Wildcard Matching with Stars

Evidence: phone screen; new-grad and full-time. Report dates: February 18 and September 16, 2020; March 8, 2021.

One report explicitly disallowed the usual dynamic-programming shortcut. Another clarified the progression from at most one * to multiple stars. The portable practice version supports the star wildcard only, matching the preserved evidence.

5. Positive Subset Sum

Evidence: onsite; full-time. Report dates: May 3, November 14, and November 21, 2024.

The first question asks whether a subset of positive integers reaches a target; the follow-up reconstructs the selected values. The linked “new question” was originally identified as a Google problem, so the algorithm is not presented as Confluent-authored—only as repeatedly used in Confluent interviews.

6. Find Defenders Against Every Hostile Monster

Evidence: onsite; full-time. Report dates: September 6 and November 21, 2024.

The source models nested monsters with a name, can_defeat descendants, and a hostile flag. Return every friendly monster that can defeat all hostile monsters. The practice interface flattens the same forest into arrays without changing the graph problem.

7. Detect a Silent Sensor

Evidence: onsite; full-time. Firsthand report dates: June 24, 2025 and April 15, 2026.

Given (sensorId, timestamp) pings, determine whether a sensor has three consecutive empty time slots. The reports point toward a hash map plus binary search. The exact five one-minute half-open slots in the practice version are disclosed assumptions. A forum AI-generated box was deliberately excluded from the evidence.

8. Match a Variadic Function Signature

Evidence: phone screen; full-time. Report dates: July 28, 2020 and July 27, 2026.

Match a requested type sequence against registered function signatures. For a variadic signature, the final declared type must occur one or more times. The older source preserved the interface and expected match sets; the 2026 legacy report independently named the same task.

9. Get Best Price

Evidence: coding/virtual onsite. Report dates: April 9, 2024 and April 24, 2025.

Choose up to three menu entries that cover a requested set of unique items at minimum total price. The second firsthand report supplied matching menu/output cases, providing useful confirmation of the earlier prompt.

🐼 Single-source coding questions

10. Serialize a Binary Search Tree

Evidence: onsite; full-time. Report date: March 21, 2019.

The post explicitly identified LeetCode 449, expected preorder serialization, asked for a recursive implementation, and then discussed an iterative version.

11. Knight Dialer Sequences

Evidence: Security-team onsite; full-time. Report date: June 10, 2021.

The source names Knight Dialer but omits a full contract. The practice version uses the conventional keypad graph and dynamic-programming formulation rather than inventing a Confluent-specific variation.

12. Least-Strong Common Defeater

Evidence: phone screen; full-time. Report date: October 18, 2022.

In a directed acyclic “can defeat” graph, find the least-strong node that can defeat every target. The original example makes Zombies the answer for Snakes and Goblins because Dragons is a stronger common defeater.

13. Validate a Nested Object Against a Schema

Evidence: Full Stack onsite. Report date: October 18, 2022.

Recursively validate an object against a nested schema containing string and number descriptors. The source preserved the name plus nested location.x/y example and a successful result.

14. LFU Cache with Thread-Safety Follow-Up

Evidence: senior onsite; full-time. Report date: December 17, 2024.

Implement an LFU-like cache; thread safety is the stated follow-up. Because the post omitted the tie-break rule, the practice problem uses the conventional LRU tie-break and labels that addition.

15. Word Search Across Documents

Evidence: onsite; full-time. Report date: September 4, 2026.

Find a word across a collection of documents and return document identifiers. Tokenization, ordering, and duplicate behavior were missing from the report and therefore made explicit in the practice contract.

16. Sequential Phrase Search Across Documents

Evidence: onsite; full-time. Report date: September 4, 2026.

Extend document search from one word to an ordered phrase using positional indexes. The preserved follow-ups mention deduplication and layered compression.

17. Sudoku Board Validation

Evidence: onsite; full-time. Report date: September 4, 2026.

Validate a Sudoku board first. The report emphasized that the solver must call the validator rather than bypassing it.

18. Sudoku Puzzle Solver

Evidence: onsite; full-time. Report date: September 4, 2026.

Solve only after invoking the separate validation contract. Treating validation and solving as two APIs is the interview signal, not merely implementing standard backtracking.

Concurrency and low-level design: 2 problems

19. Thread-Safe Delayed Task Runner

Evidence: onsite; mostly full-time. Report dates: September 16 and November 19, 2020; March 8, May 11, June 29, and July 19, 2021; June 16 and August 20, 2022; January 16, 2025.

This is the deepest repeat after TinyURL and news feed. Use a deadline-ordered priority queue, block without spinning, wake consumers when a newly inserted task becomes the earliest deadline, and evolve from one consumer to multiple workers. Correct waiting, notification, interruption, shutdown, and use of a monotonic clock matter more than naming a library class.

20. Random-Access FIFO Queue

Evidence: onsite; full-time/senior. Report dates: November 20, 2024; July 20, 2025; April 3, 2026.

Support expected O(1) add, poll, and uniform getRandom. Follow-ups ask how to compare two queues for equality, handle a run-length-encoded representation, and make the structure safe under concurrency.

System design: 13 problems

Confluent-specific or strongly Confluent-shaped versions

21. Idempotent URL Shortening Service

Evidence: onsite; new-grad through staff. 23 source reports, from November 20, 2018 through January 16, 2025. There is at least one dated sighting in every year from 2018 through 2025; the 2023 evidence is a single October report.

The recurring Confluent twist is not merely “design TinyURL.” The same normalized long URL must converge on the same short code under retries and concurrent requests. Reports probe ID generation, uniqueness, collisions, storage sizing, cache, sharding, DNS/regional routing, and whether a database uniqueness boundary is sufficient before reaching for a distributed lock.

22. Subscription News Feed API and Data Model

Evidence: onsite; new-grad through senior. Report dates: March 21, 2019; February 18, 2020; March 5 and July 19, 2021; August 20 and September 22, 2022; October 10 and November 20, 2024; April 3, April 15, May 2, and May 4, 2026.

Interviewers repeatedly go deep on REST methods and URIs, REST versus RPC, mobile/client differences, relational tables, foreign keys, one-to-many and many-to-many relationships, and feed queries. Several reports explicitly say no architecture diagram was required; API and database precision were the center.

23. Crash Recovery for a Single-Node Key-Value Store

Evidence: onsite; full-time. Report dates: March 8 and May 11, 2021; August 20, 2022; January 16, 2025.

Start with a single-node store, then explain WAL, snapshots, recovery, memtables, SSTables, Bloom filters, compaction, tombstones, cache, and read/write/space amplification. The January 2025 report specifically framed the store as a Kafka-oriented KV component and asked about WAL/snapshot recovery.

24. Leader-Based Distributed Key-Value Store

Evidence: onsite; full-time. Report date: November 21, 2024.

Extend the local LSM-style store into a distributed one: consistent hashing, strong versus eventual consistency, leader/follower topology, how many leaders exist, leader failure, and service discovery for routing clients to the correct owner.

25. Durable Disposable Email Service

Evidence: onsite; full-time/senior. Dated sightings: June 18, June 24, and July 20, 2025; April 3 and April 15, 2026. Six source records support the consolidated prompt.

Design temporary mail for roughly 100 million active users. The difficult parts are generating human-looking usernames, quickly checking uniqueness, never reusing an issued address even after expiration, and supporting mail receive/send flows. The Bloom-filter discussion is useful, but durable exact uniqueness remains necessary.

26. Podcast Subscription Feed API and Data Model

Evidence: onsite; full-time/senior. Report dates: June 16, 2022; October 20, 2023; June 18 and July 20, 2025.

Model subscriptions, refresh, pagination, storage, and read/write fanout. This is separated from the broader news-feed problem because the reports preserve podcast/RSS-specific behavior.

27. Priority-Aware Distributed Worker Platform

Evidence: onsite; full-time. Report date: January 15, 2022.

Design workers and queues with priorities, leases, retries, idempotency, backpressure, and observability. The Confluent-shaped focus is operational correctness rather than a decorative microservice diagram.

28. Kubernetes-Managed Kafka Service

Evidence: onsite; full-time. Report date: May 5, 2022.

Design broker lifecycle and control-plane behavior for Kafka on Kubernetes: partitions, durable storage, rebalancing, recovery, rollout safety, and tenant isolation.

29. Resilient Infinite-Scroll Social Feed

Evidence: Full Stack onsite. Report date: October 18, 2022.

The frontend-heavy round focused on infinite scrolling. The practice version makes the hidden systems questions explicit: stable cursors, duplicate/gap prevention during concurrent writes, retry behavior, and recovery after partial page loads.

30. Diagnose and Scale a Multi-Region Cloud Service

Evidence: senior/staff system-design discussion. Report date: August 14, 2024.

Begin with evidence about traffic and bottlenecks, then evolve the topology. Discuss read/write paths, region and availability-zone failure, replication, migration, RPO/RTO, capacity, and operational signals. The point is diagnosis before architecture expansion.

Shared canonicals, because the source was too thin for a Confluent-only contract

31. Centralized Log Ingestion and Search Platform

Evidence: 5 source URLs on four distinct report dates: June 10, 2021; August 4, August 21, and August 25, 2026.

The firsthand reports named a logging system but did not preserve enough unique requirements to justify inventing a Confluent-specific prompt. We therefore linked the evidence to the shared canonical: ingestion, buffering, storage tiers, indexing, search, retention, and failure handling.

32. CI/CD Pipeline Orchestration System

Evidence: report dates November 19, 2020 and July 24, 2026.

The sources named CI/CD discussion but did not preserve a stable company-specific contract. The shared canonical covers pipeline DAGs, workers, artifact flow, isolation, retries, cancellation, and observability.

33. Ticket Booking System

Evidence: dated sightings July 30, 2024; July 15 and July 27, August 9, September 6, and September 8, 2026.

The stable topic is global consistency around inventory. The shared canonical covers holds, expiration, payment state, idempotency, and oversell prevention; source-specific details that could not be verified were not added.

What the bank says about Confluent

Across formats, four themes recur:

1. Concurrency changes ordinary data structures. A cache, queue, map, or scheduler is rarely finished when the single-threaded code works.
2. Contracts matter. Interviewers probe exact API methods, URI design, failure semantics, idempotency, and schema relationships.
3. Storage internals matter. WAL, snapshots, LSM components, compaction, caching, and amplification appear repeatedly.
4. Operational reasoning matters. Backpressure, failover, service discovery, rebalancing, retries, and observability show up across system-design prompts.

Suggested preparation order

1. Windowed Map → Delayed Task Runner → tail -n.
2. TinyURL → News Feed API/Data Model → Single-Node KV Recovery.
3. LRU/LFU concurrency → Random-Access Queue.
4. Disposable Email → Distributed KV → Kafka on Kubernetes.
5. Use the single-source questions for breadth after the repeat set is solid.

Come hangout with us on Experience~Sharing whatever is on your mind and have fun together! 🦉

Like always, the sources come from anywhere across the internet where useful interview information can be found.