2021–2026 DoorDash Machine Learning Engineer & ML Platform Interviews, Experienced Hires: 62 Experiences (Part 2 :)
- Company
- DoorDash
- Reports
- 39 interview reports
- Level
- Full-time
- Stages
- Phone screen → Onsite
In this series
- Part 1: 2025–2026: 23 Machine Learning Engineer reports. Screens were mostly a deep dive into one past project; onsites mixed coding, ML design and ML knowledge, and the newest one added an AI-assisted build.
- Part 2 (current): 2021–2024: 37 Machine Learning Engineer reports from the years when screens were mostly coding (k-means, Single-Threaded CPU, tree and string problems), plus one Senior Data Scientist (ML) take-home and one ML Platform software engineer loop (39 reports).
Before 2025 the screen was usually a coding round: 23 of the 25 screens described here had a coding question, and three candidates between early 2022 and fall 2023 wrote k-means from scratch.
Coverage: DoorDash experienced hires in three roles: Machine Learning Engineer, including two Senior titles (37 reports, Jan 2021 to Feb 2025; Seattle and unstated locations; some name the Ads or the marketplace search and discovery team), Senior Data Scientist, ML (1 report, Mar 2022), and Software Engineer on the ML Platform team (1 report, Feb 2021). Stages: phone and technical screens, onsites and virtual onsites, and one take-home. Newer Machine Learning Engineer reports are in Part 1, and the patterns closing the Machine Learning Engineer section cover both parts.
Other cohorts: DoorDash Machine Learning Engineer intern interviews and DoorDash ML interviews where the hiring level wasn't reported.
Machine Learning Engineer, 2021–2024 (37 reports)
What came up
Coding showed up in 30 of these 37 reports, usually at the screen. A project deep dive came up in 16 and recommendation design in 13; in several of those design rounds the interviewer pushed on serving, deployment or real-time parts rather than the model.
| Topic | Reports | Most recent report |
|---|---|---|
| Deep dive or demo of a past project | 16 of 37 | Feb 19, 2025 (report date) |
| Coding (data structures and algorithms) | 30 of 37 | Nov 13, 2024 (report date) |
| System design (mostly ML) | 14 of 37 | Oct 9, 2024 (report date) |
| ML knowledge round (domain, depth or breadth) | 14 of 37 | Oct 9, 2024 (report date) |
| Recommendation and search-ranking systems | 13 of 37 | Oct 9, 2024 (report date) |
| Behavioral questions | 11 of 37 | Oct 9, 2024 (report date) |
| Graph and tree problems | 9 of 37 | Sep 3, 2024 (report date) |
| Serving and infrastructure for ML (pipelines, features, caching) | 6 of 37 | Aug 9, 2024 (report date) |
| A/B tests and online experiments | 5 of 37 | Jun 23, 2024 (report date) |
| Implementing an ML algorithm from scratch (k-means) | 3 of 37 | Oct 24, 2023 (report date) |
Single reports covered policy-breaking ads, payment fraud, a marketing model for a money-losing campaign, matching Dashers to orders, and an order-visibility API for Dasher tiers. A report can appear in more than one row.
Questions we found
Algorithms / Data Structures
| Question | Frequency | All last seen dates |
|---|---|---|
| Single-Threaded CPU (LeetCode 1834) | 2 reports | Oct 9, 2024 (report date); Jul 26, 2024 (report date) |
| Shortest paths that called for a specific Bellman-Ford implementation (practice) | 1 report | Sep 3, 2024 (report date) |
| Return the k most popular stores, given every store's traffic | 1 report | Sep 3, 2024 (report date) |
| Range sums on a 2D grid, in the style of Range Sum Query 2D (LeetCode 304) | 1 report | Aug 9, 2024 (report date) |
| Largest sum over a fixed-size sliding window, framed as model scores | 1 report | Jun 11, 2024 (report date) |
| Order delivery stops or tasks so every prerequisite comes first (practice) | 2 reports | Reports: May 30, 2024 (report date); Apr 1, 2024 (report date); Additional FastPrep history: Apr 1, 2025; Feb 11, 2021 |
| A 2D version of Range Addition (LeetCode 370) | 1 report | May 29, 2024 (report date) |
| Task scheduler: print when each task runs on one processor | 1 report | Apr 17, 2024 (report date) |
| Write k-means from scratch: fit, predict, tests and trade-offs (practice) | 3 reports | Reports: Oct 24, 2023 (report date); Oct 1, 2023 (report date); Mar 8, 2023 (report date); Additional FastPrep history: Dec 30, 2023; Dec 8, 2023; Sep 22, 2023 |
| A tree traversal | 1 report | Nov 21, 2022 (report date; onsite Nov 2022) |
| An object-oriented question built on data structures, LeetCode-like but not from LeetCode | 1 report | Nov 21, 2022 (report date; screen Oct 2022) |
| Merge Intervals (LeetCode 56) | 1 report | Oct 22, 2022 (report date) |
| Pick k random IP addresses between two IPv4 addresses | 1 report | Sep 20, 2022 (report date) |
| Similar names: does one swap of two letters turn one string into the other? (Buddy Strings, LeetCode 859) | 3 reports | Aug 15, 2022 (report date); Aug 2, 2022 (report date); Nov 23, 2021 (report date) |
| Minimum Number of Steps to Make Two Strings Anagram (LeetCode 1347) | 1 report | Aug 15, 2022 (report date) |
| k-anagrams: equal-length names that become anagrams after replacing no more than k letters | 2 reports | Aug 2, 2022 (report date); Nov 23, 2021 (report date) |
| Nearest city to a given city, from names and x and y coordinates | 1 report | May 16, 2022 (report date) |
| Most beauty collectable within k steps on a graph of tourist sites, each site counted once | 1 report | May 10, 2022 (report date) |
| Highest score from a round trip out of the hotel within a travel-time limit, each site counted once (close to LeetCode 2065) | 1 report | Feb 15, 2022 (interview date) |
| Serialize and Deserialize Binary Tree (LeetCode 297) | 1 report | Dec 8, 2021 (report date) |
| Largest path sum between marked leaves in a binary tree (practice) | 1 report | Sep 27, 2021 (report date) |
| An order-visibility API for Dasher tiers, with early access for preferred Dashers | 1 report | Sep 15, 2021 (report date) |
| The first two parts of the pickup-and-delivery problem many DoorDash candidates report | 1 report | Jan 28, 2021 (report date; onsite Jan 20 and 22, 2021) |
| A simple BFS problem | 1 report | Jan 28, 2021 (report date; onsite Jan 20 and 22, 2021) |
| A variant of Meeting Rooms (LeetCode 252) | 1 report | Jan 12, 2021 (interview date) |
ML System Design
| Question | Frequency | All last seen dates |
|---|---|---|
| Personalized search, or restaurant ranking with geographic indexing (practice) | 2 reports | Reports: Oct 9, 2024 (report date); Jan 28, 2021 (report date); Additional FastPrep history: Nov 7, 2026; Mar 20, 2026; Feb 7, 2026 |
| Find ads that break the platform's policies | 1 report | Oct 9, 2024 (report date) |
| A recommendation system: restaurants, the home feed, a top 10 at login, grocery stores from restaurant data, or a hiring team's own product (practice) | 11 reports | Reports: Aug 9, 2024 (report date); Jul 21, 2024 (report date); Jul 19, 2024 (report date); Jul 4, 2024 (report date); Jun 23, 2024 (report date); Oct 24, 2023 (report date); Oct 12, 2023 (report date); Mar 8, 2023 (report date); Oct 22, 2022 (report date); Sep 20, 2022 (report date); May 16, 2022 (report date); Additional FastPrep history: Mar 22, 2026; Mar 20, 2026; Dec 19, 2025; Sep 1, 2025; May 13, 2025; Apr 5, 2025; Jan 9, 2025; Jan 24, 2024; Feb 11, 2021 |
| Match Dashers with orders, and A/B test it with time-based switchback splits | 1 report | Jun 23, 2024 (report date) |
| After proposing streaming features: Kafka, database reads and writes, and QPS estimates (the design prompt isn't given) | 1 report | Jun 3, 2024 (report date) |
| Explain each part concretely: how the API answers a request, the SQL query, where tables are preprocessed, and how the queue and pub/sub work (the design prompt isn't given) | 1 report | May 28, 2024 (report date) |
| Build a model that classifies intent (practice) | 1 report | Reports: Oct 24, 2023 (report date); Additional FastPrep history: Mar 20, 2025 |
ML Domain Case
| Question | Frequency | All last seen dates |
|---|---|---|
| Payment fraud detection, run like a data science case (practice) | 1 report | Reports: Jul 19, 2024 (report date); Additional FastPrep history: Jan 24, 2024 |
| A marketing campaign is losing money: work out that it needs a marketing model, then pick its labels and features (practice) | 1 report | Reports: May 28, 2024 (report date); Additional FastPrep history: Dec 19, 2025; Jan 9, 2025 |
| Test whether an incentive program for Dashers works: metrics and experiment design | 2 reports | Oct 24, 2023 (report date); Oct 12, 2023 (report date) |
| Predict a city's order volume: features, loss, model choice, metrics and online tests (practice) | 2 reports | Sep 20, 2022 (report date); May 16, 2022 (report date) |
ML Knowledge
| Question | Frequency | All last seen dates |
|---|---|---|
| What to do once a model is in production | 1 report | Sep 3, 2024 (report date) |
| The bias–variance trade-off | 2 reports | Jul 21, 2024 (report date); Jul 19, 2024 (report date) |
| About 20 quick questions built on a delivery-time prediction example (your own example is allowed), reaching into deep learning and production | 1 report | Jul 19, 2024 (report date) |
| Many A/B testing questions | 1 report | Feb 18, 2024 (report date) |
| Practical questions about day-to-day ML engineering work rather than theory | 1 report | Aug 15, 2022 (report date) |
| About your own project: imbalanced data, the label, feature scaling and encoding, the train/test split and the metrics | 1 report | Nov 23, 2021 (report date) |
Project Deep Dive
| Question | Frequency | All last seen dates |
|---|---|---|
| How and why the modeling techniques in your project work | 1 report | Aug 26, 2024 (report date) |
| Have you put a model into production yourself, or served one online? | 2 reports | Jul 4, 2024 (report date); Apr 17, 2024 (report date) |
| Walk through a project prepared from the recruiter's document or prompt; one interviewer asked many business-impact questions | 2 reports | Jun 11, 2024 (report date); May 30, 2024 (report date) |
| Why did your project use transformers, and why was deploying it hard? | 1 report | Feb 18, 2024 (report date) |
| Your most impactful project | 1 report | Oct 24, 2023 (report date) |
| The highest QPS you've handled, your biggest technical project, and leadership | 1 report | Jan 28, 2021 (report date) |
Behavioral (BQ)
| Question | Frequency | All last seen dates |
|---|---|---|
| Conflicts | 2 reports | Reports: Oct 9, 2024 (report date); Sep 3, 2024 (report date); Additional FastPrep history: Nov 5, 2026; Mar 22, 2026; Mar 24, 2025; Feb 3, 2024 |
| Strengths and weaknesses | 2 reports | Reports: Oct 9, 2024 (report date); Sep 3, 2024 (report date); Additional FastPrep history: Feb 11, 2021 |
| Constructive or negative feedback you received (practice) | 2 reports | Reports: Oct 9, 2024 (report date); May 16, 2022 (report date); Additional FastPrep history: Sep 16, 2026; May 13, 2025 |
| Handling ambiguity | 1 report | Oct 9, 2024 (report date) |
| The project you're proudest of (practice) | 3 reports | Reports: Sep 3, 2024 (report date); Aug 9, 2024 (report date); Jul 19, 2024 (report date); Additional FastPrep history: Sep 16, 2026; Feb 11, 2021 |
| Blockers you ran into | 1 report | Sep 3, 2024 (report date) |
| A mistake you made | 1 report | Reports: Jul 19, 2024 (report date); Additional FastPrep history: Mar 22, 2026; Feb 3, 2024 |
| How would you cut down on orders that arrive with wrong items? (practice) | 1 report | Reports: Jul 19, 2024 (report date); Additional FastPrep history: Sep 16, 2026; Mar 20, 2026; Dec 19, 2025 |
| Teamwork | 1 report | May 16, 2022 (report date) |
37 detailed candidate timelines, loops, and screens
24: A background-only screen that ended early
Report date: Feb 19, 2025 (interview Oct–Dec 2024) · Role: Machine Learning Engineer · Background: PhD, currently employed; applied online · Result: Rejected the next day
Technical screen (45 minutes booked, about 30 used)
- Background and projects only, with no concerns raised during the call.
25: Personalized search, policy-breaking ads and a scheduling problem
Report date: Oct 9, 2024 (interviews Jul–Sep 2024) · Role: Machine Learning Engineer · Background: PhD, currently employed; came through a recruiting agency · Result: Did not pass
Onsite
- ML practitioner round: personalized search. Closest practice: Design Personalized Restaurant Search.
- ML system design: find ads that break the platform's policies.
- Behavioral: conflict, strengths and weaknesses, handling ambiguity, and constructive feedback.
- Coding: Single-Threaded CPU (LeetCode 1834), not solved; the candidate hadn't practiced it.
- ML concepts: many basic questions, none hard.
26: Four-years-experience MLE who lost the coding round to Bellman-Ford
Report date: Sep 3, 2024 (onsite by early Sep 2024) · Role: Machine Learning Engineer · Background: 4 years of experience, working at Meta · Result: Not reported (ML system design still to come)
Project deep dive (screen)
- The candidate says the interviewer may dig into why and how your modeling techniques work, and advises preparing from the document the recruiter sends.
Onsite: coding
- A problem that needed a very particular Bellman-Ford implementation (a shortest-path method that works with negative edge weights). The candidate couldn't solve it and asked whether that alone means a rejection. Closest practice: Bellman-Ford Route Costs With Credits.
Onsite: ML depth (domain knowledge)
- The prompt isn't given. The candidate's advice is to ask as many clarifying questions as you can and approach it like one more ML system design round.
Onsite: ML tech breadth
- Basic ML questions of the kind found online, plus what to do after a model is in production.
Onsite: behavioral
- Standard questions on conflicts, strengths, best projects, and blockers you ran into.
27: A top-k stores screen
Report date: Sep 3, 2024 (interview Jul–Sep 2024) · Role: Machine Learning Engineer · Background: Master's, currently employed; came through a recruiting agency
Technical phone screen
- Coding: given traffic counts for every store on DoorDash, return the k most popular stores. The candidate rated the experience well.
28: Senior MLE screen on Single-Threaded CPU
Report date: Jul 26, 2024 (interview Jul 2024) · Role: Senior Machine Learning Engineer · Background: Master's, currently employed; came through a recruiting agency · Result: Did not pass
Technical phone screen
- Coding: Single-Threaded CPU (LeetCode 1834). Not solved; no hints were offered.
29: An onsite with no LeetCode, then a verbal offer
Report date: Jul 21, 2024 (interview the week before) · Role: Machine Learning Engineer · Background: Master's, currently employed; referral · Result: Passed; verbal offer from HR, with results two days after the onsite
Onsite (four rounds, no LeetCode-style coding)
- Format: two ML system design rounds, an ML fundamentals round and a deep dive.
- ML fundamentals: included the bias–variance trade-off, which the candidate got the wrong way round.
- ML system design: restaurant recommendation, weighted toward backend topics such as serving and databases. Closest practice: Model Restaurant Recommendations.
30: Values, rapid-fire ML, a deployed-system diagram and payment fraud
Report date: Jul 19, 2024 (interviews Jul 2024) · Role: Machine Learning Engineer · Background: PhD, currently employed, a first job change; past work in ML data science, current work closer to software engineering; referral · Result: Did not pass
Onsite (four rounds)
- Engineering values and hiring manager: behavioral questions on a mistake you made and the project you're proudest of, and a short case on reducing orders with wrong items, answered as an approach rather than a model. Closest practice: the wrong-delivery complaints question.
- ML technical breadth: about 20 rapid questions tied to a delivery-time prediction example (your own example is allowed), from bias–variance to deep learning and production, with drilling where the interviewer has expertise.
- ML system design: restaurant recommendations, graded mainly on a diagram of the deployed system rather than a recommender flow; deployment got the most time and data processing was skipped. Closest practice: Model Restaurant Recommendations.
- Domain knowledge: payment fraud detection as a data science case, with features and data preparation in depth, then modeling and post-launch monitoring. Closest practice: Model Order Fraud for Review.
31: Range sums at the screen, then an onsite that probed real-time serving
Report date: Aug 9, 2024 (interviews Apr–Jun 2024) · Role: Machine Learning Engineer · Background: Master's, currently employed; applied online · Result: Did not pass
Phone screen
- Coding: 2D range sums in the style of Range Sum Query 2D (LeetCode 304), reached from the 1D prefix-sum idea with a hint; tests passed.
- A short project chat.
Onsite
- Domain knowledge: a standard recommendation system design, end to end.
- ML depth: rapid fundamentals that filled about 30 minutes of a 60-minute slot.
- ML system design: a recommendation system, with pressure on real-time serving, which the candidate hadn't built. Closest practice: Model Restaurant Recommendations.
- Behavioral: the project you're proudest of became a second deep dive on what was new about it, then standard questions.
Before the onsite, the recruiter had said online deployment experience wasn't a hard requirement.
32: Two-day onsite: production ownership and grocery ranking with few labels
Report date: Jul 4, 2024 (interviews Apr–Jun 2024) · Role: Machine Learning Engineer · Background: Master's, currently employed; came through a recruiting agency · Result: Rejected before the last two rounds
Onsite, split over two days
- Hiring manager: resume and ML fundamentals, focused on whether the candidate had personally put a model into production (their models had been deployed by engineers).
- ML design with a staff engineer: adapt a restaurant ranker to grocery and convenience stores that have few labels; then end-to-end home feed ranking, post-launch model upkeep and ML infrastructure choices. Closest practice: Model Restaurant Recommendations.
33: Two coding problems at the screen, design-only onsite
Report date: Jun 23, 2024 (interviews Apr–Jun 2024) · Role: Machine Learning Engineer · Background: Master's, currently employed; came through a recruiting agency · Result: Rejected; the feedback pointed to the behavioral round, which the candidate took to mean the answers didn't sound senior enough
Phone screen
- Coding: two problems (not named).
Onsite (no coding)
- Behavioral with the hiring manager.
- Design: matching Dashers with orders, plus A/B testing it; marketplace tests need time-based (switchback) splits rather than user splits.
- Design: restaurant recommendations. Closest practice: Model Restaurant Recommendations.
- ML breadth: general ML questions.
34: A sliding-window screen that ended over work authorization
Report date: Jun 11, 2024 (interview the week before) · Role: Machine Learning Engineer · Background: PhD, currently employed; referral · Result: Passed the screen, then the process stopped over visa timing
Phone screen
- Coding: the largest sum over a fixed-size sliding window, framed as an ML model's predicted probabilities.
- Hiring manager: a project prepared from the recruiter's document, with many questions on business impact.
35: Onsite design round on Kafka, storage and QPS
Report date: Jun 3, 2024 (interviews Apr–Jun 2024) · Role: Machine Learning Engineer · Background: PhD, currently employed in a scientist-style modeling role; came through a recruiting agency · Result: Not final when written; the candidate expected a rejection
Onsite
- ML breadth: ordinary ML questions.
- Domain knowledge: in practice a normal ML design round.
- Behavioral: standard questions.
- ML system design, focused on pipelines and deployment: proposing streaming features led to questions on Kafka, database reads and writes, and QPS (queries per second) estimates, which a modeling-only background didn't cover.
36: A two-part screen with the prerequisite-ordering problem
Report date: May 30, 2024 (interviews Apr–May 2024) · Role: Machine Learning Engineer · Background: PhD, currently employed; came through a recruiting agency · Result: Passed
Phone screen
- Coding: n delivery stops with ordering rules (for example, stop 2 only after stop 1, stop 3 only after stop 2). Return an order that keeps every rule, or -1 if none exists, as in Course Schedule. Practice: Task Order With Prerequisites.
- Project deep dive prepared from the recruiter's prompt, plus a few behavioral questions.
37: A two-part screen: 2D range additions and a deep dive
Report date: May 29, 2024 (interviews Apr–May 2024) · Role: Machine Learning Engineer · Background: PhD, currently employed; applied online · Result: Did not pass
Phone screen
- Coding: a 2D version of Range Addition (LeetCode 370).
- Project deep dive (no details given).
38: Half an onsite, from a data scientist moving into MLE
Report date: May 28, 2024 (interviews Apr–May 2024) · Role: Machine Learning Engineer · Background: Master's; currently a data scientist working on ML, also interviewing for applied and research scientist roles; came through a recruiting agency
Before the onsite
- Recruiter guidance: ML system design is about the system, not the model; domain knowledge is about choosing models and metrics.
Onsite (first half)
- Domain knowledge: a product question about a money-losing marketing campaign that turned into building a marketing model, with heavy pushback on labels and features and an expectation of the standard marketing setup. Modeling methods and metrics never came up. Closest practice: Set Discounts by Their Extra Margin.
- ML system design: every component explained concretely, from how the API answers a request and the actual SQL query to where tables get preprocessed and how the queue, pub/sub and publisher work. High-level answers weren't accepted.
39: Senior MLE tech screen with a prerequisites ordering question
Report date: Apr 1, 2024 · Role: Senior Machine Learning Engineer · Location: Seattle · Background: data scientist · Result: No offer
HR screen
- A 30-minute call; no details given.
Tech screen (1 hour 45 minutes)
- Questions about projects the candidate had done.
- A LeetCode-medium coding question: given the prerequisites, work out what order the tasks should run in. The candidate stumbled, putting it down to a data science background, but reached the answer in the end. Practice: Task Order With Prerequisites.
40: Ads team screen: a task scheduler and an online-serving question
Report date: Apr 17, 2024 (interviews Jan–Mar 2024) · Role: Machine Learning Engineer, Ads team · Background: Master's, currently employed; referral · Result: Did not pass
Phone screen
- Coding: tasks with processing times and earliest start times on one processor; print the span in which each task runs (the Single-Threaded CPU family).
- Hiring manager: resume questions, including online serving experience, which the candidate didn't have (their models were offline).
41: An "ML coding" interview that was ordinary data structures
Report date: Mar 5, 2024 (interview late Feb or early Mar 2024) · Role: Machine Learning Engineer, ML coding interview · Background: working at BNY Mellon
ML coding interview
- Afterward the candidate described it as regular data-structures-and-algorithms work dressed up as ML: say, you get a model and its output and must work something out from it. No ML algorithm to implement.
42: Project deep dive that revealed a data science role
Report date: Feb 18, 2024 (interview the week before) · Role: ML Engineer · Background: works on MLOps tooling and software engineering
Coding
- LeetCode easy-to-medium questions. The candidate felt coding didn't matter much to the interviewers.
Project deep dive
- The candidate had trouble explaining why the project used transformers and why deploying it was hard, given the distributed inference pipeline behind it.
The candidate also warned of heavy A/B testing questions, without naming the round. The recruiter never said which role this was. Only in the deep dive did the hiring manager describe it as closer to product data science, a poor fit for the candidate's engineering background.
43: Two coding problems at the screen
Report date: Nov 13, 2024 (interview period given as Oct–Dec 2023) · Role: Machine Learning Engineer · Background: Master's, currently employed; applied online · Result: Passed; onsite invite three days later
Technical phone screen
- Coding: two unnamed problems; the second needed an optimization hint.
44: Promised ML-from-scratch, got LeetCode-style questions
Report date: Nov 8, 2023 (interview on or before that date) · Role: Machine Learning Engineer, ML coding interview · Background: working at Goldman Sachs · Result: No reply after the interview
ML coding interview
- The recruiter said, by phone and email, to expect ML algorithms from scratch and no LeetCode. The interview asked LeetCode-type questions anyway (not named).
Asked about it, the recruiter said the ML coding questions had leaked and the format had changed. Then the recruiter stopped replying.
45: k-means from scratch, checked against prepared tests
Report date: Oct 1, 2023 · Role: Machine Learning Engineer · Background: Master's, currently employed; came through a recruiting agency · Result: Passed
Phone screen (two rounds back to back)
- Coding: implement k-means by hand and pass the interviewer's prepared test cases. Practice: K-Means Delivery Zones.
- Hiring manager (45 minutes): about 30 minutes on resume projects, then the candidate's questions.
46: A full loop with k-means, an intent classifier and a Dasher incentive test
Report date: Oct 24, 2023 (interviews Jul–Sep 2023) · Role: Machine Learning Engineer · Background: Master's, currently employed; applied online · Result: Did not pass
Loop
- Coding: implement k-means. Practice: K-Means Delivery Zones.
- Project deep dive: your most impactful project.
- System design: restaurant recommendations. Closest practice: Model Restaurant Recommendations.
- ML design: an intent classifier. Closest practice: Detect Intent and Order Details in Help Chats.
- Technical breadth: when Dashers are scarce, how would you test whether an incentive program for them works?
- Hiring manager: a project and behavioral questions.
47: Four rounds and no coding: a recommender and a Dasher bonus case
Report date: Oct 12, 2023 (interviews Jul–Sep 2023) · Role: Machine Learning Engineer · Background: Master's, currently employed; applied online · Result: Passed
Virtual onsite
- ML domain knowledge (60 minutes): resume projects, then ML fundamentals growing out of them.
- ML system design (60 minutes): a restaurant recommendation system end to end, without deep modeling questions. Closest practice: Model Restaurant Recommendations.
- Technical breadth (45 minutes): a data-science-style case on metrics and the experiment for paying Dashers incentive bonuses.
- Hiring manager (45 minutes): standard behavioral questions.
48: OOP screen, then a tree traversal at the onsite
Report date: Nov 21, 2022 (tech screen Oct 2022, onsite Nov 2022) · Role: Machine Learning Engineer · Background: working at Roku · Result: Rejected after the onsite
Tech screen
- No ML, only coding: an object-oriented question built on data structures, LeetCode-like but not from LeetCode. Out of time, the candidate talked through the rest.
Onsite
- Easy, the candidate said; the coding was a tree traversal. It went well but ended in rejection.
49: Recommending grocery stores from restaurant data
Report date: Oct 22, 2022 (interview Oct 2022) · Role: Machine Learning Engineer · Background: PhD, currently employed; came through a recruiting agency · Result: Did not pass
Technical phone screen
- Coding: Merge Intervals (LeetCode 56); the candidate's solution wasn't optimal.
- ML: using what DoorDash knows about users and restaurants, recommend grocery stores. The candidate's approach was user and restaurant embeddings carried over to grocery with transfer learning. Closest practice: Model Restaurant Recommendations.
50: Virtual onsite: random IP addresses, a city's order count and a top-10 at login
Report date: Sep 20, 2022 (interviews Jul–Sep 2022) · Role: Machine Learning Engineer · Background: Master's, currently employed; came through a recruiting agency · Result: Rejected three days later
Virtual onsite
- Coding: pick k random IP addresses between two given IPv4 addresses (treat each address as a base-256 number, sample k values in the range, convert back).
- ML system design: predict how many orders a city will get. Closest practice: Forecast Hourly Orders for Each City.
- System design: at each login, push the 10 restaurants the user is most likely to want; the interviewer gave little feedback. Closest practice: Model Restaurant Recommendations.
- Hiring manager (no details given).
51: 40 minutes of string coding and 20 of ML chat
Report date: Aug 15, 2022 (interview Jul–Aug 2022) · Role: Machine Learning Engineer · Background: Master's, currently employed; came through a recruiting agency
Technical phone screen
- Coding (40 minutes): Buddy Strings (LeetCode 859) and Minimum Number of Steps to Make Two Strings Anagram (LeetCode 1347).
- ML chat (20 minutes): practical questions about day-to-day ML engineering work rather than theory.
52: Buddy strings with one swap, then k-anagrams at scale
Report date: Aug 2, 2022 (interview Jul–Aug 2022) · Role: Machine Learning Engineer · Background: Master's, currently employed; came through a recruiting agency · Result: Waitlisted
Technical phone screen
- Coding: Buddy Strings (LeetCode 859) limited to at most one swap; a small bug cost time.
- Follow-up: k-anagrams with a character-count map, then how to split the work MapReduce-style for very long inputs.
53: Tourist-site paths at the screen, then nearest cities and order volume at the onsite
Report date: May 16, 2022 (screen write-up May 10, 2022, screen period Jan–Mar 2022; onsite Apr–May 2022) · Role: Machine Learning Engineer · Background: Master's, currently employed; applied online · Result: Passed the screen; the onsite result isn't given
Technical phone screen
- Coding: tourist sites form a graph with two-way paths and possible cycles, and each site has a beauty score. From a given start, find the most beauty collectable within k steps, counting each site once. DFS is enough because the step limit k bounds the search even with cycles; the candidate needed a hint on that point.
Onsite
- Coding: given cities with a name and x and y coordinates, answer queries for the city nearest to a given one.
- ML domain knowledge: how to predict order volume, covering feature design, the loss function, model choice, evaluation metrics and online experiments, with no design or code. Closest practice: Forecast Hourly Orders for Each City.
- HR and behavioral: an introduction, then negative feedback and teamwork stories.
- System design: a recommendation system, first scalability and availability, then features, candidate selection, prediction and evaluation. Closest practice: Model Restaurant Recommendations.
54: k-means and a hiring manager's recommendation design, back to back
Report date: Mar 8, 2023 (interview period given as Jan–Mar 2022) · Role: Machine Learning Engineer · Background: Master's, currently employed; applied online · Result: Did not pass
Phone screen 1: ML coding, with an interviewer from outside the hiring manager's team
- Explain how k-means works.
- Turn a given outline of the algorithm into a model class, then write its fit and predict methods and pass a set of tests (about 40 minutes).
- Time and space complexity, then performance concerns such as how the starting centers change the outcome.
- Where k-means is strong and weak: its running time, the data structures behind it, and clusters with curved shapes. Practice: K-Means Delivery Zones.
Phone screen 2: hiring manager
- Design, from the ground up, a recommendation system for the hiring manager's own team. Closest practice: Model Restaurant Recommendations.
- Follow-ups: which model families make sense with user history but no existing model; the loss and objectives, and whether to train several tasks at once; how predictions get used in production; ranking a page that mixes carousels and lists; context and session features; keeping results diverse and checking for bias; finding and testing new items; and using what each item actually is, such as a product taxonomy.
55: Touring a city's sites on a time budget
Report date: Feb 16, 2022 (interview Feb 15, 2022) · Role: Machine Learning Engineer · Background: Bachelor's, currently employed; came through a recruiting agency
Technical phone screen
- Coding: you're touring a city. Each site has a score, and your hotel counts as site 0. You get the scores, the direct routes as two parallel lists (from[i] → to[i]) with a travel time for each, and a maximum travel time. Start and end at the hotel, keep the total travel time within the limit, and count each site's score once even if you pass it again. Return the highest total score. The question came with a small worked example, and it is close to Maximum Path Quality of a Graph (LeetCode 2065).
56: Tree serialization at the screen
Report date: Dec 8, 2021 (interview Oct–Dec 2021) · Role: Machine Learning Engineer · Background: Master's, currently employed; referral · Result: Did not pass
Technical phone screen
- Coding: Serialize and Deserialize Binary Tree (LeetCode 297), on nodes with a value and left and right children. The candidate considered preorder plus inorder or level order and coded level order.
57: Similar restaurant names, k-anagram names and ML basics
Report date: Nov 23, 2021 (interview Oct–Nov 2021) · Role: Machine Learning Engineer · Background: PhD, currently employed; referral · Result: Did not pass
Technical phone screen
- Coding: call two restaurant names similar when swapping one pair of letters in the first gives the second. Given a name and a list, return the similar names (with a worked example).
- Follow-up: given a name, a number k and a list, return the names that are k-anagrams of it: equal length, and the two become anagrams once you replace no more than k letters in one of them (also with an example).
- ML questions: a project of yours; how you handled imbalanced data; what your label was; feature engineering, such as scaling numeric features and encoding categorical ones; how you split training and test data; and which evaluation metrics you used.
Both coding questions had come up in earlier candidates' accounts; the candidate found them manageable but still did not pass.
58: Max path sum between marked leaves
Report date: Sep 27, 2021 · Role: Machine Learning Engineer · Background: PhD, currently employed; applied online
Technical phone screen
- Coding: in a binary tree, some leaves are marked with an asterisk; find the largest path sum between two marked leaves. The candidate's small example tree has an answer of 47. Practice: Maximum Path Sum Between Marked Nodes.
59: An order-visibility API for Dasher tiers
Report date: Sep 15, 2021 (interview Jul–Sep 2021) · Role: Machine Learning Engineer · Background: Master's, currently employed · Result: Passed
Technical phone screen
- Coding: given a set of scheduled orders, build a Dasher-facing API, getEvents, that takes a datetime and the Dasher's tier and returns orders. Top-tier Dashers see all of a day's orders from 17:00 the day before; regular Dashers see them only from 18:00 the day before.
- Follow-up: a list of (dasher, restaurant_id) pairs marks Dashers as preferred for a restaurant. A preferred Dasher sees that restaurant's orders for the next day from 16:00 the day before.
60: Offer five days after a two-day virtual onsite (search and discovery)
Report date: Jan 28, 2021 (phone screen Jan 12; virtual onsite Jan 20 and 22; offer Jan 27) · Role: Machine Learning Engineer, marketplace search and discovery team · Background: Master's, currently employed; came through a recruiting agency · Result: Offer by email, 15 days after the phone screen
Phone screen
- Coding: a variant of Meeting Rooms (LeetCode 252); not hard.
Virtual onsite (five rounds)
- Coding: the first two parts of the pickup-and-delivery problem often reported by DoorDash candidates, with no tricky follow-up.
- System design: restaurant ranking, half on geographic indexing (geohash, R-tree) and half on the end-to-end ranking pipeline; prior search ranking work made the second half quick. Closest practice: Design Personalized Restaurant Search.
- Domain knowledge: resume and experience, such as the highest QPS you've handled, your biggest technical project, and leadership.
- Coding: a simple BFS problem; a small bug just needed to be found and fixed.
- Engineering manager: behavioral questions, a project and a small case study.
Patterns across the Machine Learning Engineer reports (both parts)
Numbers 01 to 23 are in Part 1 and 24 to 60 are above.
-
The screen moved from coding to a project deep dive. In 2025–2026, 13 of the 15 screens described were a deep dive, demo or resume walk-through with no coding (02, 04, 05, 08, 09, 10, 11, 13, 14, 16, 17, 21, 22); only 07 and 15 had coding. In 2021–2024, 23 of the 25 screens described had coding; 24 and 26 were the exceptions. Two candidates in early 2025 said the screen had recently become a deep dive (21, 22), and one in mid-2025 got a hard coding problem after preparing for a deep dive (15).
-
Single-Threaded CPU, again and again. Six candidates got Single-Threaded CPU (LeetCode 1834) between mid-2024 and early 2026, at a phone screen (07, 28) or at the onsite (09, 17, 18, 25). Report 40 had a close relative that prints when each task runs.
-
k-means from scratch, 2022–2023. Three candidates with interviews between early 2022 and fall 2023 were asked to implement k-means (45, 46, 54). Two of those rounds came with prepared test cases (45, 54), and in 54 the follow-ups covered complexity, initialization and the algorithm's weak spots.
-
Recommendation design, pushed toward serving. A recommendation or search-ranking design came up in 22 reports (01, 06, 07, 09, 10, 12, 17, 18, 23, 25, 29, 30, 31, 32, 33, 46, 47, 49, 50, 53, 54, 60). In eight of them the interviewer cared more about deployment, storage, caching or real-time serving than about the model (09, 10, 12, 17, 18, 29, 30, 31).
-
Production experience gets checked. Five candidates said interviewers probed whether they had put models into production themselves, and each struggled where their work had stopped at modeling: no online deployment (10, 40), no real-time system (31), whether they had shipped a model personally (32), and Kafka, database and QPS details (35).
-
Knowledge rounds that behave like cases. Three candidates said the domain or depth round played out like another ML system design (14, 26, 35). Two more had domain rounds run like a data science or product case: payment fraud (30) and a money-losing marketing campaign (38).
-
Task ordering with prerequisites. Three Machine Learning Engineer candidates got it (19, 36, 39), at a screen or an onsite; the ML Platform loop below had it too (62).
-
Puzzles from 2021–2022. A one-swap string check came up three times (51, 52, 57), twice with a k-anagram follow-up (52, 57). Two screens used a graph of tourist sites where each site's value counts once within a step or time budget (53, 55), with different inputs and limits.
-
Dasher incentives in the breadth round. In 2023 two candidates were asked how to measure or test incentive pay for Dashers (46, 47).
-
Wrong items in the behavioral round. Three candidates were asked about orders arriving with the wrong items: the causes (09), preventing it with a check at pickup (10), and a short case on cutting it down (30).
-
"ML coding" that wasn't. Two interviews billed as ML coding turned out to be standard data-structure questions (41, 44); in 44 the recruiter had promised ML algorithms from scratch. The k-means rounds (45, 54) did ask for a real ML algorithm.
-
Prep documents for the deep dive. Four candidates said the recruiter sent a document or prompt to prepare the project deep dive from (05, 26, 34, 36), and one wrote their own section-by-section outline (22).
Senior Data Scientist, ML (1 report)
What came up
One report: the only technical step was a take-home that predicts food delivery time, judged on the code, data cleaning, features, modeling and the written report.
Questions we found
Take-Home
| Question | Frequency | All last seen dates |
|---|---|---|
| Predict food delivery time; hand in code, a report and results within ten days | 1 report | Mar 5, 2022 (report date) |
1 detailed candidate timeline
61: Take-home only: predict delivery time
Report date: Mar 5, 2022 (take-home late Feb to early Mar 2022) · Role: Senior Data Scientist, ML · Background: PhD, currently employed; came through a recruiting agency · Result: Rejected five days after submitting
Take-home (the only technical step)
- Task: predict food delivery time, delivering code, a written report and results within ten days. The stated effort was five to six hours; it took the candidate four days.
- Grading: one assignment covering coding, data science and ML; a pass would have meant no further technical interviews.
- Feedback on request: strong code, cleaning and write-up; weaker feature engineering (more time-based and advanced features wanted) and too little hyperparameter tuning.
Software Engineer, ML Platform team (1 report)
What came up
One report: a grid question at the phone screen, then task ordering, a restaurant recommendation design and behavioral questions at the onsite.
Questions we found
Algorithms / Data Structures
| Question | Frequency | All last seen dates |
|---|---|---|
| A grid question similar to Number of Islands (LeetCode 200) (practice) | 1 report | Feb 11, 2021 (report date) |
| A question very similar to Course Schedule II (LeetCode 210): order tasks so every prerequisite comes first (practice) | 1 report | Reports: Feb 11, 2021 (report date); Additional FastPrep history: Apr 1, 2025; May 30, 2024; Apr 1, 2024 |
ML System Design
| Question | Frequency | All last seen dates |
|---|---|---|
| Design DoorDash's homepage restaurant recommendations (practice) | 1 report | Reports: Feb 11, 2021 (report date); Additional FastPrep history: Mar 22, 2026; Mar 20, 2026; Dec 19, 2025; Sep 1, 2025; May 13, 2025; Apr 5, 2025; Jan 9, 2025; Aug 9, 2024; Jul 21, 2024; Jul 19, 2024; Jul 4, 2024; Jun 23, 2024; Jan 24, 2024; Oct 24, 2023; Oct 12, 2023; Mar 8, 2023; Sep 20, 2022; May 16, 2022 |
Behavioral (BQ)
| Question | Frequency | All last seen dates |
|---|---|---|
| A time you disagreed with a colleague (practice) | 1 report | Reports: Feb 11, 2021 (report date); Additional FastPrep history: Sep 16, 2026; Mar 20, 2026 |
| A project you take pride in (practice) | 1 report | Reports: Feb 11, 2021 (report date); Additional FastPrep history: Sep 16, 2026; Sep 3, 2024; Aug 9, 2024; Jul 19, 2024 |
| Strengths and weaknesses | 1 report | Reports: Feb 11, 2021 (report date); Additional FastPrep history: Oct 9, 2024; Sep 3, 2024 |
| Your career high and low points | 1 report | Feb 11, 2021 (report date) |
1 detailed candidate timeline
62: ML Platform software engineer with a homepage recommendation design
Report date: Feb 11, 2021 · Role: Software Engineer, ML Platform team
Phone screen
- A grid question close to LeetCode's Number of Islands. Practice: Count Islands in a Grid.
Virtual onsite: coding
- A question very close to Course Schedule II: given tasks and their prerequisites, return an order that respects them. Practice: Task Order With Prerequisites.
- The friendly interviewer helped check the code against sample tests once it was written. The candidate suggests talking through the approach before coding.
Virtual onsite: system design
- Design how DoorDash's homepage recommends restaurants. It felt like a design chat with a colleague, and the candidate enjoyed it. Practice: Model Restaurant Recommendations.
Virtual onsite: behavioral
- Questions about past work, framed around DoorDash's engineering principles, such as a time you disagreed with a colleague and the project you're proudest of.
Informal chat with the hiring manager
- Strengths and weaknesses, and your career high and low points, with time left over for the candidate's own questions.
The candidate says both coding questions were on LeetCode's DoorDash-tagged list.
Keep practicing
- Coding: practice versions of reported questions: Bellman-Ford Route Costs With Credits (report 26), Task Order With Prerequisites (reports 36, 39 and 62), K-Means Delivery Zones (reports 45, 46 and 54), Maximum Path Sum Between Marked Nodes (report 58), and Count Islands in a Grid (report 62). The DoorDash question bank has every DoorDash practice question; it needs sign-in.
- Design and ML cases: Design Personalized Restaurant Search (reports 25 and 60), Model Restaurant Recommendations (reports 29, 30, 31, 32, 33, 46, 47, 49, 50, 53, 54 and 62), Model Order Fraud for Review (report 30), Set Discounts by Their Extra Margin (report 38), Detect Intent and Order Details in Help Chats (report 46), and Forecast Hourly Orders for Each City (reports 50 and 53). More are in the DoorDash system design bank and the DoorDash ML modeling bank; both need sign-in.
- Behavioral: the DoorDash behavioral question bank has the prompts above (sign-in needed).
- Everything DoorDash: the DoorDash interview page and DoorDash hiring insights.
- More firsthand reports: the Experience feed.
In this series
- Part 1
- Part 2 (current)
Sources: Anywhere useful interview info can be found.
Comments
0No comments yet — be the first to share what you know.