Word Search II (LeetCode 212): Given an m x n board of characters and a list of words, find all words that exist in the board by moving to adjacent cells (horizontally or vertically). Each cell may be used only once per word.

The key insight is to build a Trie from the word list, then run DFS from each cell, pruning branches not present in the Trie. Mark cells visited during DFS with a sentinel and restore afterward. Remove words from the Trie as they are found to avoid duplicates and prune dead branches. Time complexity is O(M * 4 * 3^(L-1)) where M is board cells and L is max word length; space is O(total characters in dictionary). A common follow-up is to optimize for large boards with sparse dictionaries, or to incrementally add/remove words. Trie construction correctness, DFS backtracking discipline (restoring visited state), recognizing that deleting matched nodes prunes future work — signs of experience with multi-pattern search problems rather than rote memorization.

Maximum Profit in Job Scheduling (LeetCode 1235): Given n jobs each with a start time, end time, and profit, schedule jobs (no two overlapping) to maximize total profit.

Sort jobs by end time. Use binary search to find the latest job that ends before the current job starts, then apply DP: dp[i] = max(dp[i-1], profit[i] + dp[last_non_conflict(i)]). Binary search makes this O(n log n) time with O(n) space. A likely follow-up is supporting a fixed number of machines (k-machine variant) or streaming job arrivals, which pushes toward a segment tree or priority queue approach. Clean DP formulation, recognition that binary search on the sorted-by-end-time structure is essential for the optimal complexity, and ability to extend the solution to variants without starting over.

Non-Overlapping Intervals (LeetCode 435): Given an array of intervals, find the minimum number of intervals to remove so the rest are non-overlapping.

Sort by end time. Greedily keep the interval with the earliest end time that does not overlap the previously kept interval; count every skipped interval. This greedy choice is optimal because finishing earlier leaves more room for subsequent intervals. Time O(n log n), space O(1). Follow-ups often involve returning the actual removed intervals or extending to a weighted variant where each interval has a value (which reduces to the job-scheduling DP). Immediate recognition of the greedy approach, correct sort key (end time, not start time), and clean loop without off-by-one errors. Interviewers want to see concise, confident greedy reasoning.

Design an audit event/log system for Snowflake that tracks every query executed: who ran it, when, against which objects, with what result, at petabyte scale with sub-second query lookup by user or resource (directly reported in a Feb 2024 Snowflake Senior SWE phone screen on LeetCode Discuss thread 4727339).

Start with requirements: write-heavy (every query emits an event), reads are sparse but latency-sensitive for compliance. Use an append-only write path (Kafka or Kinesis ingest, then flush to columnar object storage in micro-batches). For fast lookup, maintain a secondary index on (user_id, timestamp) and (resource_id, timestamp) backed by a distributed key-value store or a metadata service. Discuss retention tiers (hot/warm/cold), encryption at rest, role-based access to audit records, and immutability guarantees. Follow-ups typically ask about handling clock skew, cross-region replication for global compliance, and query-time masking of sensitive columns. Security-first thinking (immutability, encryption, RBAC from the start), recognition of the write-heavy vs. read-sparse access pattern, and practical discussion of storage tiers. Snowflake is a regulated enterprise product, so interviewers penalize designs that bolt on security as an afterthought.

Design a query result cache for Snowflake that serves 1 million queries per second across tens of thousands of customer accounts, invalidating results correctly when underlying data changes.

The key challenge is multi-tenant cache isolation combined with staleness detection. Use a deterministic hash of the normalized query text plus a version token derived from the micro-partition metadata of the referenced tables as the cache key. On any write to a table, increment that table's version token, which automatically invalidates all cache entries referencing it without a scan. For storage, a distributed in-memory tier (Redis Cluster or a custom shard ring) with an LRU eviction policy per tenant prevents noisy-neighbor issues. Follow-ups ask about handling schema changes, partial result caching for large result sets, and budgeting memory fairly across tenants. The insight that cache keys must embed table version metadata — not just query text — is the crux. Candidates who miss this will propose TTL-based invalidation, which interviewers at Snowflake treat as a red flag because TTL can silently return stale data under compliance workloads.

Design the metadata service for a columnar storage engine: track micro-partitions (50-500 MB immutable blobs) for petabyte-scale tables, support concurrent DML without locking entire tables, and enable partition pruning at query time using min/max statistics stored per partition.

Partition metadata — min/max per column, row count, creation timestamp, S3/GCS path — is stored in a distributed key-value store sharded by table ID. MVCC (multi-version concurrency control) handles concurrent reads and writes: writers append a new partition version; readers see a consistent snapshot keyed on transaction ID. The pruning index is an in-memory zone map: a sorted array of (min, max) per partition per column, enabling binary-search-style elimination of partitions before any data is fetched. Discuss shard rebalancing, write amplification from small inserts, and how clustering keys improve pruning hit rates. Follow-up: implement the in-memory pruning filter in code. Familiarity with MVCC as a concurrency mechanism (not just as a term), understanding that zone maps are the core pruning primitive in columnar systems, and the ability to implement the pruning filter efficiently in the in-code follow-up. This question directly probes database internals knowledge.

Tell me about a time when you took full ownership of a project, including parts outside your defined scope. Why did you take that ownership, and what was the outcome?

Use STAR tightly: ground the situation in a concrete project with a named deadline or outcome at risk. The action must center on what you personally decided and did — not what the team did. Quantify the outcome (latency improved by X%, incident resolved in Y hours, feature shipped Z weeks ahead). Snowflake probes specifically for 'why' — candidates who expanded scope for career visibility score lower than candidates who expanded scope because the customer impact required it. Individual agency vs. collective credit — Snowflake's behavioral round penalizes 'we' answers. Interviewers are checking for initiative that is customer- or mission-driven, not self-promotional.

Tell me about a time you made a significant mistake. How did you discover it, and what did you do?

Choose a real mistake with meaningful impact, not a humble-brag ('I worked too hard'). Structure the answer as: how you discovered it (ideally proactively, not from a customer complaint), the immediate mitigation steps you personally executed, the root cause you identified, and the systemic fix you put in place. End with what changed as a result — changed process, runbook, test coverage. Owning the mistake cleanly and showing structural fix is far more important than minimizing the original error. Intellectual honesty and accountability. Snowflake wants engineers who detect and escalate problems quickly rather than hiding them, and who drive post-mortems to durable fixes. Candidates who deflect blame or frame the mistake as a team failure score poorly.

Describe a time when you had to work effectively with a team or organization you had never worked with before. What obstacles came up, and how did you navigate them?

Identify a cross-functional scenario: a partner team, an external vendor, or a newly formed task force. Describe the communication gap or misaligned incentive concretely, then explain what you personally did to build context fast — reading their documentation, attending their standups, framing shared success metrics. The resolution should show a concrete deliverable that required both teams. Follow-ups often probe what you would do differently, so have a reflection ready. Collaboration skills and ability to operate without authority. Snowflake's engineering org spans multiple product teams and the interview evaluates whether the candidate can build credibility and move work forward in ambiguous organizational structures.

Snowflake Interview Questions

Snowflake's SWE interview loop is decentralized and team-specific but generally follows a consistent pattern: a recruiter screen, a 2-hour technical phone screen (one coding problem plus a system design question, run over CoderPad/Zoom), and a virtual onsite of 4-5 rounds covering coding, infrastructure-oriented system design, a behavioral/values round, and — for IC3 and above — a 30-minute past-project presentation with 30 minutes of Q&A on architectural trade-offs. Coding questions skew toward LeetCode medium-hard with a database or data-processing flavor; system design questions are infrastructure-heavy rather than product-design-heavy, reflecting Snowflake's cloud data warehouse business. The process typically takes 2-6 weeks from initial screen to offer.

17 real questions across 3 rounds

Reported via LeetCode Discuss first-hand reports: 'Snowflake Senior Software Engineer Phone Screen' (thread 4727339, Feb 2024) confirmed Word Search II; 'Snowflake | Phone Screen | Patching Array' (thread 424385) confirmed LC 330; 'Snowflake | OA | All Questions | Summer Internship 2023 | Canada' (thread 3169843) confirmed Maximum Profit in Job Scheduling (LC 1235) and Non-Overlapping Intervals (LC 435); 'Snowflake | Software Engineer | Phone/Code Screen' (thread 1965084) confirmed stream-processing class design; Blind post 'snowflake-system-design-and-leetcode-questions-a83swi3f' and interviewing.io Snowflake guide (2024) corroborate Merge k Sorted Lists and concurrent LRU Cache; prachub.com Snowflake coding questions and algo.monster Snowflake guide cross-confirm difficulty tier and concurrency theme.

Run a Snowflake mock with the AI interviewer. It asks questions like these, follows up on your answers, and tells you exactly what to fix. The only score that means anything is the one on your real answers.

Continue with Google

Keep prepping

All company interviews