Interview guides · ACV Auctions

ACV Auctions Machine Learning Engineer Interview Questions & Guide 2026

ACV Auctions Machine Learning Engineer interview guide: reported rounds, practice questions with worked solutions and a 7-day prep plan.

The 60-second brief

As a Machine Learning Engineer at ACV Auctions, you are at the intersection of high-stakes automotive commerce and advanced data science. You play a critical role in building the intelligent systems that power the ACV Auctions digital marketplace, from vehicle inspection automation and condition reporting to complex pricing models that drive real-time bidding decisions.

Ask whether any round happens inside an existing repository instead of a blank file. Reading unfamiliar code, isolating a fault and making the smallest correct change is a different skill from writing a function from scratch, and it needs its own practice.

Focus areas

Choose indexes from the query's access path
Paginate large result sets with keyset cursors
Detect concurrent edits instead of losing writes

Overview

Your work directly impacts the liquidity and transparency of the automotive wholesale market. By deploying scalable machine learning models, you help bridge the gap between physical vehicle inspections and digital trust, ensuring that thousands of transactions occur seamlessly every day. This role requires a balance of rigorous engineering practices, such as building robust inference APIs, and a deep understanding of the lifecycle of machine learning models in a production environment.

The ACV Auctions interview process

ACV Auctions candidates report 4 rounds · ≈ 3-5 weeks. The stages below are what candidates describe, not a published process.

Recruiter Screen

The title covers product work, platform work, infrastructure, mobile and frontend, and those are different jobs with different loops behind them. A screening call is the cheapest place to find out which one the seat is, and asking reads as experienced rather than fussy. The questions that separate them: what the team is on call for, what the last three projects were, and whether any round happens inside an existing repository instead of a blank file. Then say which of that you have done and which you have not. Claiming the whole posting is the fastest way to be found out one round later.

What to demonstrate

Whether you can locate your experience inside one flavour of the role honestly instead of claiming the entire requirements list
Whether you name what you have not done, which an experienced screener reads as a level signal and can plan the loop around
Whether what you want next matches what the seat is: someone who wants greenfield work landing on a team that mostly operates an existing system is a hire that leaves within the year

How to prepare

Mark every line of the posting as done, adjacent or new, and write one sentence for each adjacent line naming the closest thing you actually built
Split your last two years into rough percentages across feature work, operating and debugging live systems, and design or review, so a question about scope gets numbers rather than adjectives
Bring three questions that discriminate between seats: what the team is paged for, how much of the work is changing existing code versus standing up something new, and what shipped in the last quarter

Technical Interviews

Input bounds are the part of the prompt most often skimmed, and they usually contain the answer. They tell you which complexity class is admissible, which narrows the search before you have thought about the problem itself. As a rough planning figure, a compiled language does on the order of 10^8 simple operations per second and an interpreted one roughly an order of magnitude less. So n up to about twenty admits enumerating subsets, a few thousand admits a quadratic pass, and a million admits neither: you need near-linear, or linear with a log factor. If the bounds are missing, ask for them.

What to demonstrate

Whether the approach is justified by the stated input size rather than by whichever pattern you recognised first
Whether you ask about the properties that change the algorithm: whether the input arrives sorted, whether duplicates occur, whether values are bounded integers, whether it all fits in memory
Whether you can name the bottleneck in your own solution and what would remove it, even when you deliberately leave it in place
Whether a claimed speedup is real, since memoising a recursion only helps when subproblems genuinely overlap and the state can be keyed cheaply

How to prepare

For each algorithm you rely on, write down the largest n it handles in roughly a second, then check two of those figures by timing them in the language you will actually type in
For two weeks, write one line naming your target complexity and the bound that justifies it before you write any code, then compare that line with what you ended up submitting
Practise the conversion backwards: given a required O(n log n), list the mechanisms that get you there (sorting, a heap, an ordered map, divide and conquer) and choose by what the problem needs to query, not by what you used last

Take-Home Assignment

A test suite is read as a statement about what you believe can break, which is why the selection matters far more than the count. Assertions that repeat the happy path with different numbers tell a reviewer nothing. Cases at the boundaries tell them where you looked: empty input, one element, duplicates, a malformed row, a value at the edge of the accepted range. The sharpest question to ask of a test you have just written is whether it would fail if the logic were wrong. If deliberately breaking the function leaves the suite green, that test is documentation rather than a check.

What to demonstrate

Which cases you chose, and whether the suite reaches empty, single, duplicate and malformed inputs instead of the prompt's example with the numbers changed
Whether the suite is deterministic and runs offline, so it produces the same result on a reviewer's machine as on yours
Whether assertions are written against behaviour rather than internals, so a reasonable refactor does not force the tests to be rewritten
Whether a failure is readable on its own, naming the input and the expected value rather than reporting a bare assertion

How to prepare

Mutate your own code by hand before submitting: change a < to <=, drop the last iteration of a loop, return the input unchanged. Anything that stays green is a missing test, and this takes minutes
Remove nondeterminism before writing assertions. Pass the clock in rather than calling now() inside the logic, seed any random generator explicitly, and do not assert on the iteration order of a Go map or a Python set, since Go randomises that order deliberately and Python salts string hashing per process unless PYTHONHASHSEED is pinned
Time a full run of the suite and trace anything slow to its cause, usually a sleep or a real network call, then replace it with a fake so the reviewer is not waiting

Technical Discussion

What to demonstrate

Whether the approach is justified by the stated input size rather than by whichever pattern you recognised first
Whether you ask about the properties that change the algorithm: whether the input arrives sorted, whether duplicates occur, whether values are bounded integers, whether it all fits in memory
Whether you can name the bottleneck in your own solution and what would remove it, even when you deliberately leave it in place
Whether a claimed speedup is real, since memoising a recursion only helps when subproblems genuinely overlap and the state can be keyed cheaply

How to prepare

For each algorithm you rely on, write down the largest n it handles in roughly a second, then check two of those figures by timing them in the language you will actually type in
For two weeks, write one line naming your target complexity and the bound that justifies it before you write any code, then compare that line with what you ended up submitting
Practise the conversion backwards: given a required O(n log n), list the mechanisms that get you there (sorting, a heap, an ordered map, divide and conquer) and choose by what the problem needs to query, not by what you used last

Where candidates lose points

Choosing an index from the columns a query mentions rather than from how it filters and orders

A composite B-tree index on (a, b, c) can be seeked only as a left prefix: equality on a, then equality on b, then a range or an ordering on c. A query that filters on b alone cannot seek into it at all and at best gets a full scan of the index; a query that filters a and ranges on b gets no benefit from c, because the index is only sorted by c within a fixed (a, b) pair. The practical consequence is that one index per column is close to useless for multi-predicate queries while a single correctly ordered composite index turns a scan into a lookup. The ordering half is what gets missed: if the index cannot satisfy the ORDER BY, the database must read every matching row and sort before the limit can apply, so a LIMIT 20 over a million matching rows still reads a million rows.

Shipping a migration and the code that depends on it as a single change

During any rolling deploy, and for as long as a rollback remains possible, old and new code execute against the same schema at the same time. A migration that drops or renames a column breaks every instance that has not restarted yet, and code that requires a column the migration has not applied breaks every instance that restarted early. The discipline is expand then contract: add the new column nullable, write both shapes, backfill in batches, move reads across once the backfill is verified, and only then stop writing the old shape and drop it - four deploys, usually spread over days. It feels disproportionate until the first rollback, at which point it is the only reason the previous version still runs.

Optimising an axis nobody named

Ask which resource is actually scarce here: wall-clock latency, throughput, memory footprint, cost per request, or engineering time. Shaving a constant factor off an in-memory step is wasted effort when the same function makes a blocking remote call inside the loop.

Never running a concrete value through the code

Trace one small input and one edge input by hand, index by index, out loud. Re-reading your own code catches design mistakes; walking a real value through it catches the off-by-one, the uninitialised accumulator and the loop that never advances.

Questions & practice

Walk me through a challenging machine learning project you led from conception to production.

Medium · machine learning fundamentals

Approach

Pick the metric from the cost of each error type, not from habit.
Name the simplest model that could work and what would make you move past it.
Say how you would validate it, and where leakage could enter the split.

Follow-up

How would you know the model is overfitting?
What changes if the classes are heavily imbalanced?

Find overlapping job attempts and peak concurrency from lease records

Medium · sweep line · intervals · leases

A day of job_run history yields about 50,000,000 attempt records: (job_run_id, job_type, attempt, started_at, finished_at which is NULL when the worker died, lease_expires_at). Leases expire on a clock, so a job that outran its lease ran twice. Produce (a) every job_run_id whose attempts overlapped in wall-clock time and (b) the peak number of simultaneously running attempts per job_type with the minute it occurred. Target O(n log n). State how you treat a NULL finished_at and what clock skew does to your answer.

Approach

Define the interval before sorting anything: an attempt occupies [started_at, COALESCE(finished_at, lease_expires_at)). finished_at is observed and lease_expires_at is only a promise, so every attempt without a finish contributes an estimate and the whole result is a lower bound on overlap rather than an exact count.
For peak concurrency, sweep: emit 2n endpoints, sort by (timestamp, kind) with ends ordered before starts at equal timestamps, then walk the sequence maintaining a counter per job_type and record each type's maximum with its timestamp. O(n log n) dominated by the sort, O(n) space, or O(1) extra if the sort is external and the walk streams.
For overlap detection, do not compare attempts pairwise. A single global sort by (job_run_id, started_at) gives both the grouping and the order; within a group, keep the maximum end seen so far and report an overlap exactly when the next start is less than that running maximum, which is one linear pass after the sort.
Half-open intervals matter and are easy to get wrong: with closed intervals an attempt ending at the same millisecond another begins reads as concurrency two, and across 50,000,000 records that artefact swamps the real signal.
State the clock caveat: started_at and finished_at are written by different workers, so under skew of a few hundred milliseconds an apparent overlap shorter than that bound is not evidence. Filter reported overlaps by a minimum duration, or prefer timestamps written by whichever component heartbeats the lease.
Scale the sort rather than assuming it fits: the sweep emits two endpoints per attempt, so 50,000,000 records become 100,000,000 endpoints, and at roughly 24 bytes each, an 8-byte timestamp plus a 4-byte job_type plus a kind flag padded to alignment, that is about 2.4 GB of sort keys before any scratch space. Either push the ordering into the database behind an index on (job_type, started_at) or run an external merge sort in chunks; the overlap pass sorts n records rather than 2n, so it is the cheaper of the two.

Worked solution (30 min)

Write the interval derivation with the COALESCE and state in one line which of the two end sources is observed and which is assumed.
Write the concurrency sweep: the endpoint tuples, the sort key including the end-before-start tie-break, and the per-job_type counter.
Hand-trace four attempts of one job, two disjoint and two overlapping by three seconds, and confirm the overlap detector fires exactly once.
Add the skew filter as a minimum overlap duration, state the value you chose, and justify it from how the timestamps are written.

Expected result: One global sort by (job_run_id, started_at) driving a running-maximum-end check for overlaps, plus a sweep over 2n endpoints with ends ordered before starts for per-job_type peak concurrency, both O(n log n), with half-open intervals, NULL finished_at falling back to lease_expires_at, and overlaps shorter than the clock-skew bound excluded.

Verify it yourself

Two attempts touching at exactly the same instant report concurrency 1, not 2.
A job with a single attempt never appears in the overlap output regardless of its lease.
The reported peak timestamp for a job_type lies inside at least one of that type's intervals.

Follow-up

A handler is not idempotent and you have found 400 overlapping jobs. Which of them actually caused damage, and what would you query to find out?
Peak concurrency for one job_type is 4 against a configured cap of 4. Is the cap working, or is the data hiding attempts that never started?
How would you compute both answers incrementally as records arrive rather than in a daily batch?

Identify the heaviest tenants in a five-minute window under memory pressure

Medium · top-k · heavy hitters · streaming

The edge service handles about 3,000 requests per second across roughly 50,000 tenants, peaking near 9,000. Expose the 50 heaviest tenants by request count over the trailing five minutes so limits can be tightened before one tenant's backfill starves the fleet. You may not retain five minutes of raw records. Give the exact solution and its memory, then the bounded-memory approximation with its error stated as a formula, and say which you would ship and at what tenant cardinality that choice changes.

Approach

Do the exact version first, because it is affordable at this cardinality: a ring of 300 one-second counters per tenant, advanced lazily, is 1,200 bytes of counters per tenant and roughly 60 to 90 MB for 50,000 tenants with overhead. Carry a running total and subtract the bucket you overwrite so a window read is O(1) rather than 300 adds.
Extract the top 50 with a size-k min-heap over the tenant sums: O(d log k) for d tenants, against O(d log d) to sort them all. Maintaining the heap continuously instead of on query requires a tenant-to-heap-index map, because incrementing a count already inside the heap means sifting from a known position, and without that map you rebuild the heap on every request.
State the approximation precisely rather than gesturing at sketches. Misra-Gries with m counters retains every item whose true count exceeds N/(m+1), and each retained count underestimates the truth by at most N/(m+1). With m = 1,000 and N = 900,000 requests in the window the error is roughly 900 requests, which is fine for spotting a tenant sending 50,000 and useless for ranking two tenants 200 apart.
Say what breaks when the window slides: Misra-Gries and Space-Saving are insert-only and cannot be decremented as records age out. The workable construction is one summary per sub-window, say ten seconds, with 30 summaries merged at query time, and the merged error is the sum of the per-summary errors, so the bound degrades linearly in the number of sub-windows.
Choose and defend it: at 50,000 tenants the exact rings cost under 100 MB in a process that already holds more, so ship exact. Keep the sketch for the case that actually motivates it, a per-principal or per-IP key where cardinality runs to millions and is not bounded by anything you control.
Raise the fleet problem before it is asked: each of 20 to 40 instances sees only its share, and the top 50 of one shard is not the top 50 of the fleet. Either aggregate counts centrally or accept that a per-instance threshold multiplied by instance count is the limit you are really enforcing.

Follow-up

The heaviest tenant is heavy because of one export job rather than user traffic. Should the limiter treat those as the same tenant?
Two tenants sit tied at the boundary of the top 50. Does your answer flap, and does the flapping matter?
You switch to per-principal keys and cardinality goes to 10 million. Walk through what changes.

When would you choose a NoSQL database over a relational database for an ML feature store?

Medium · data, sql and pipelines

Approach

Handle the rows that do not match: that is usually the actual question.
Name the grain you start from and join outward from it.
Say which index the query would use, and what makes it unusable.

Follow-up

How does the query change if that join becomes one-to-many?
How would you run this migration without downtime?

Keep soft-deleted accounts from blocking re-registration

Easy · soft delete · unique constraints · engine differences

app_user holds user_id, tenant_id, email CITEXT, password_hash (NULL for SSO principals), email_verified_at, auth_version, status ('invited','active','suspended','deactivated'), created_at, updated_at, deleted_at. Two live accounts for one address inside a tenant must be impossible, but an address freed by a soft delete must be reusable, and the same tenant may delete and re-register it repeatedly. Write the uniqueness DDL for PostgreSQL 16, then the equivalent for MySQL 8 where partial indexes do not exist, and say what each permits once three deleted rows already hold that address.

Approach

Start from what is actually unique: not (tenant_id, email), but (tenant_id, email) among live rows. PostgreSQL says that directly — CREATE UNIQUE INDEX app_user_live_email ON app_user (tenant_id, email) WHERE deleted_at IS NULL. A full constraint over the same two columns burns the address permanently the first time someone deletes an account.
Keep case-insensitivity in the type or the index, never in the application: CITEXT as given, or UNIQUE (tenant_id, lower(email)) as an expression index where the extension is unavailable. A case-sensitive unique column is exactly how two accounts for one human appear.
For MySQL 8 the predicate has to move inside the key: add a discriminator column that is a constant 0 while the row is live and is set to user_id on delete, with UNIQUE (tenant_id, email, deleted_marker). Live rows share the constant and still collide; deleted rows differ from each other and stop colliding.
State the NULL variant and its dependency: leaving the marker NULL for deleted rows also works, because a unique index treats NULLs as distinct — true in MySQL, and true in PostgreSQL only under the default NULLS DISTINCT, which PostgreSQL 15 lets you reverse. Check the polarity against the three existing deleted rows: constant-on-live is what preserves the collision you want, and reversing it silently admits duplicate live accounts.
Say what a soft delete must do besides setting deleted_at: increment auth_version so existing tokens stop validating, leave resource.owner_user_id and resource_revision.actor_user_id intact, and accept that the address is retained — erasure is a different requirement answered by scrubbing the column, not by a DELETE that would break those references.

Worked solution (20 min)

Create the PostgreSQL partial unique index, insert a live row, soft delete it, and insert the same address again.
Repeat the delete-and-reinsert cycle three times and confirm three deleted rows coexist with exactly one live row.
Write the MySQL form with the discriminator, then deliberately reverse the polarity so live rows carry NULL, and show two live duplicates commit.
Attempt a second live insert on both engines and map the resulting 23505 / ER_DUP_ENTRY to the 409 the handler should return.

Expected result: PostgreSQL rejects only the second live row and admits unlimited deleted ones. MySQL behaves identically with UNIQUE (tenant_id, email, deleted_marker) where live = 0; with the polarity reversed, two live rows commit and nothing afterwards can tell them apart.

Verify it yourself

Group live rows by (tenant_id, lower(email)) after a randomised insert/delete run — no group exceeds one.
Inserting 'User@x.test' after 'user@x.test' is rejected; if it is accepted, the case handling is in application code rather than in the index.
EXPLAIN a login lookup and confirm the partial index is chosen — it is only usable by queries that carry the deleted_at IS NULL predicate, so a query without it falls back to a scan.

Follow-up

A deleted account re-registers with the same address the next day. Do the old resource rows follow the new user_id, and how does the API keep the two principals apart?
How do you honour an erasure request while resource_revision.actor_user_id still references this table?
What changes if a user may hold membership in two tenants?

Can you explain the trade-offs between batch inference and real-time inference?

Medium · ml system design and production

Approach

Name what you would monitor after launch and what triggers a retrain.
Separate the offline training path from the online serving path.
Fix the product goal and the online metric before choosing any model.

Follow-up

How would you detect drift before the metric drops?
What happens when a feature is missing at serving time?

What are the key considerations when designing a high-availability machine learning API for inference?

Medium · ml system design and production

Approach

Name what you would monitor after launch and what triggers a retrain.
Fix the product goal and the online metric before choosing any model.
Separate the offline training path from the online serving path.

Follow-up

How would you roll the new model out safely?
How would you detect drift before the metric drops?

What is your strategy for handling large-scale data ingestion for training pipelines?

Medium · ml system design and production

Approach

Fix the product goal and the online metric before choosing any model.
Say where features come from at serving time and how they match training.
Separate the offline training path from the online serving path.

Follow-up

How would you roll the new model out safely?
How would you detect drift before the metric drops?

How do you ensure data consistency across distributed systems?

Medium · system design

Approach

Name the failure you are designing for, then the recovery path.
Fix the scope first: who calls this, how often, and what they do when it fails.
Choose a partition key and say what query it makes expensive.

Follow-up

How does this behave when that dependency is down for an hour?
What breaks first when traffic grows ten times?

Keep one unresponsive destination from stalling all webhook delivery

Medium · fan-out · circuit breaker · timeouts · bulkheads

Egress delivery sends about 1.5k webhooks/second to 40k destinations, with a per-destination concurrency cap of 4 and a 10-second connect-plus-read timeout. One destination begins accepting connections and never responding; within the hour 150 destinations behave the same way. Design the delivery path so unrelated destinations are unaffected: the pool structure, the timeouts, the retry policy, the per-destination circuit, and what is recorded so a retry is not a second effect at the receiver. State how many in-flight slots the degraded destinations hold and why that number decides the design.

Approach

Start with the number, and with the law that produces it. In-flight work is arrival rate times time in service, so 1.5k/second against a healthy 200 ms response needs about 300 concurrent slots. Per destination the same product applies, ceilinged by the concurrency cap: at the fleet average of 0.0375 deliveries/second per destination (1.5k spread over 40k) a 10-second timeout is 0.375 slots. A destination that has queued retries behind it is a different regime - every slot refills the instant an attempt expires, so it sits pinned at its cap of 4 - and 150 of those hold 600 slots, more than a pool sized for healthy traffic, entirely consumed by endpoints that will never answer. The per-destination cap bounds one endpoint and says nothing about the aggregate, which is exactly why it alone is not containment.
Contain with bulkheads and an admission bound rather than a larger pool. Cap total in-flight per pool and shard destinations across pools by a hash of destination id, so a correlated group - one provider, one region - cannot exceed its pool's share. A delivery refused admission and re-queued with backoff is strictly better than one holding a slot on behalf of a receiver that is not listening.
Treat the timeout as two timeouts, and be exact about what shortening one buys. Connect and read are separate failures and both must be shorter than the budget of whatever is waiting. Occupancy is min(cap, arrival rate x timeout), so dropping the read ceiling from 10 seconds to 3 cuts a merely slow destination's occupancy proportionally, 0.375 slots to 0.11 at the fleet average. It does not cut the 4 slots held by one of the 150: a destination with a retry backlog arrives far above cap/timeout - 0.4/second at a 10-second timeout, 1.33/second at 3 - so it stays pinned at the cap either way and only the slot-seconds per attempt fall. What that does buy is detection rate: 3.3x more failures observed per second on the same four slots, which is how fast the circuit reaches its threshold. Pick the value from the measured latency distribution of successful deliveries, with their high percentile as the floor, not from a round number.
Add a circuit per destination, counting a timeout as a failure. Once open, fail fast without taking a slot - that is the whole point, converting 4 held slots into zero. Half-open on a schedule with exactly one probe and close only if the probe succeeds, so a permanently dead endpoint costs one request per interval instead of a growing retry queue.
Make retries safe and non-synchronising. Back off with full jitter, sleeping a random value in [0, min(cap, base x 2^attempt)], because a fixed delay re-synchronises every failed delivery to one destination into a simultaneous burst. Delivery is at-least-once, so the payload carries the event id under the signature and the receiver deduplicates on it; record the attempt against (destination, event id) rather than a bare success flag, so a lost response does not become a second business effect on the other side.

Worked solution (25 min)

Compute healthy in-flight from rate times latency, then slots held by 150 destinations at the cap and the full timeout, and compare both against one pool size.
Write the pool sharding rule and the admission bound, and state what a refused delivery does next.
Pick connect and read timeouts from the success-latency distribution, then compute min(cap, arrival rate x timeout) for an average destination and for one with a retry backlog, and say which of the two the shorter timeout actually moves.
Write the circuit's state machine with its open threshold, probe interval and close condition, and the backoff formula with full jitter.

Expected result: Destinations sharded across bounded pools with a total in-flight cap per pool, separate connect and read timeouts chosen from the success-latency distribution, a per-destination circuit that fails fast while open and probes once per interval, and full-jitter backoff on retry. A shorter read timeout lowers occupancy only for destinations arriving below cap/timeout; a dead destination with a retry backlog stays pinned at its cap of 4 whatever the timeout, and the circuit opening is what takes it to zero. Deliveries carry a signed event id the receiver deduplicates on, and attempts are recorded per (destination, event id). Degraded destinations hold at most their pool's share rather than 600 slots of a shared pool.

Verify it yourself

Slots held by degraded destinations are bounded by the pool shard, not by how many destinations degrade.
An open circuit reduces slots held by that destination to zero between probes.
Lowering the read timeout does not reduce successful deliveries, verified against the measured success-latency distribution, and does not by itself reduce the slots a backlogged dead destination holds.
A redelivered event is discarded by the receiver on its event id rather than applied twice.

Follow-up

The destination is not dead - it answers in 9.5 seconds with a 200. Does a failure-rate circuit open? Should anything shed that traffic, and on what signal?
One destination requires deliveries in order. What does a per-destination concurrency of 4 do to that guarantee, and what would you change to offer it?
A destination has been parked six hours with 900k undelivered events. What does resuming look like, and is delivering the whole backlog the right call?

Read latency spikes on a sixty-second sawtooth

Medium · cache stampede · single-flight · ttl · tail latency

The cached listing read path serves about 14k reads/second at an 85% hit rate. p99 sits at 35 ms for 57 seconds, jumps to 900 ms for 3, and repeats. During each spike the primary shows several hundred identical listing queries starting within the same millisecond, all carrying one large tenant's id. Cache entries use a 60-second TTL. Give the mechanism, the ordered checks, the fix, and the correctness hazard your fix must not introduce.

Approach

Match the period to a configured number before theorising about load. A spike every 60 seconds against a 60-second TTL is an entry expiring, and you confirm it by correlating spike timestamps with the entry's write time rather than with the traffic curve. If the period had matched a cron or a GC interval instead, this is a different investigation.
Establish the concurrency of the miss. Several hundred identical queries in one millisecond means the miss path has no coalescing: every request that arrives between expiry and repopulation recomputes. The herd size is that key's arrival rate times its recompute time, so at 1.2k reads/second for the hot key and a 250 ms recompute you expect about 300 concurrent misses, which matches what is observed.
Add single-flight on the miss path so one caller per key recomputes under a short-lived lock while the rest wait for its result. Prefer stale-while-revalidate where the read tolerates it: return the expired value immediately and refresh asynchronously, which removes the latency spike rather than serialising it into a queue of waiters.
De-synchronise the keys. Write TTLs with jitter, for example 60 seconds plus or minus 10%, so a deploy or a mass invalidation does not align every key on the same second and turn a per-key herd into a fleet-wide one.
Name the hazard the fix must not introduce. Serving a stale listing is acceptable only because the API reports the projection watermark, and a reader that loaded the old value before a write can repopulate the entry after the invalidation, so the bounded TTL is what actually caps staleness rather than the delete. Keep read-after-write pinned to the primary for the writing session regardless.
Verify on miss concurrency, not hit rate. The hit rate barely moves, because the herd is one miss multiplied; the number that must change is distinct origin queries per key per minute.

Follow-up

The same sawtooth appears on a key that is invalidated on write rather than expired. Is that the same bug?
How does your answer change if the recompute takes 4 seconds instead of 250 ms?
What exactly does a client see during a stale-while-revalidate window, and how does the watermark let them tell?

Suggested practice checklist

For a candidate senior enough that the loop turns on design and judgement rather than on whether the coding round gets finished. Five days build one system properly and then stress it; coding gets a single maintenance day, on the assumption that the risk at this level is an unexamined tradeoff rather than a missed algorithm.

Day 1 — Numbers before diagrams

Build your own reference card of the figures you will re-derive all week: bytes for a realistic record, requests per second implied by a given daily active count, and the storage that a year at a given write rate produces. Derive each one rather than copying it, because the derivation is what survives a follow-up.
Turn one product statement into capacity requirements. From ten million daily users at four writes and forty reads each, state the peak-to-average factor you are assuming and why, then produce peak write QPS, peak read QPS and a year of storage.
Write the two numbers whose order of magnitude changes the design, the read-to-write ratio and the working-set size against memory per node, and state the threshold at which each one flips your answer.

Deliverable: A one-page numbers card and one worked capacity estimate with every assumption written down.

Day 2 — One system, from requirements to schema

Spend the first ten minutes producing only functional requirements, non-functional targets with numbers attached, a p99 latency, a durability expectation, a consistency requirement, and an explicit out-of-scope list.
Define the interface before the boxes: the three or four endpoints, their parameters, what each returns, and which of them are idempotent.
Write the data model, then write the single access pattern that justifies it, and state what the schema would have to become if the dominant access pattern were the other one.

Deliverable: One design carried to endpoint-and-schema depth, with non-functional targets expressed as numbers and a written out-of-scope list.

Day 3 — The consistency you are actually buying

Write out what a client sees under asynchronous replication when its write commits on the leader and its next read is served by a lagging follower, then write the two fixes, pinning that session's reads to the leader for a bounded window or carrying a version token the replica must reach, and the cost of each.
Work the quorum arithmetic on paper for N of three with W and R of two, and separate what R + W > N does guarantee, that any read set intersects any write set, from what it does not: on its own it is not linearizability, and a sloppy quorum that accepts writes on nodes outside the preference list breaks even the intersection.
Take two storage choices with different defaults, a single-leader relational store committing synchronously and a quorum-replicated store that converges eventually, and write the specific product behaviour that would be wrong under each, rather than a general statement about which is stronger.

Deliverable: A page separating what quorum overlap guarantees from what it does not, with one concrete product misbehaviour attached to each gap.

Day 4 — Failure is the design

For one write path, work through the case where the client times out after the server has already committed, then design the idempotency key: who generates it, how long it is retained, and what the duplicate request returns.
Express the retry policy as parameters rather than as a word: maximum attempts, base delay, backoff factor, jitter, and which error classes are retried at all. Then state why retrying a non-idempotent write without a key is a correctness bug and not merely waste.
Compute the fan-out effect on tail latency. If a request waits on ten backends and each independently exceeds its p99 one percent of the time, the chance at least one is slow is 1 - 0.99^10, about ten percent. Then write why independence is the optimistic assumption and what correlates them in practice.
Name the backpressure mechanism for one queue or one dependency in the design, a bounded queue with shedding or a concurrency limit, and write what the caller is told when it engages.

Deliverable: One write path with an idempotency design, a parameterised retry policy, and a written tail-latency calculation with its assumption named.

Day 5 — Scaling the hot path

Choose cache-aside or write-through for one read path and write the staleness window each produces, then name the invalidation event and what the system does when that event is lost.
Design against the stampede: either coalesce requests so only one recomputes a missing key, or refresh early with jittered expiry, and write why identical TTLs on keys populated in the same moment produce a synchronised expiry and a thundering herd.
Shard one table by a key you choose, then answer the two questions that break the choice: which queries now require a scatter-gather, and what happens to the distribution when one tenant is a hundred times larger than the median.
Write the cost of adding a node under plain modulo placement, where nearly every key moves, against consistent hashing, where roughly one key in n+1 moves, and state what virtual nodes are for.

Deliverable: A caching and sharding decision for one path, each with its failure mode and its rebalancing cost written beside it.

Day 6 — Keep the coding hand in, at the bar that applies to you

Solve one medium problem in thirty minutes, then spend twenty more making it production-shaped: named invariants, validation at the boundary, and errors that distinguish a caller mistake from an internal fault.
Write the tests you would require of a colleague's version of that function: one for empty input, one for the boundary, and one for the case the implementation is most likely to get wrong.
Read a piece of your own code from six months ago and write the change you would ask for, phrased as you would actually phrase it in review.

Deliverable: One problem hardened to review standard, with its test list and one written review comment.

Day 7 — Defend it while being interrupted

Run a forty-five-minute design mock with an interviewer briefed to change a requirement halfway, a tenfold traffic increase or a new strict consistency requirement, and to push on one number you estimated.
Rehearse the two sentences a senior loop is listening for: naming the tradeoff you are choosing against and why, and saying what you would measure to learn that the choice was wrong.
Prepare the design you regret: a real decision, the constraint that produced it, what it cost, and what you changed afterwards.

Deliverable: Mock notes recording how the design changed under the new requirement, plus a written account of one regretted decision.

The bug you decided not to fix

Engineers over-index on what they repaired. A stronger answer covers something you knowingly left broken: the alert you tuned down, the data inconsistency you documented instead of chasing, the cleanup you deferred past two quarters. Give the reasoning and the condition that would have reopened it, so it reads as a decision and not as neglect.

How do you handle disagreements with stakeholders regarding model performance versus business requirements?

Medium · behavioural and collaboration

Approach

State the situation in two sentences and spend the rest on the reasoning.
Close with what you would do differently, concretely.
Pick a story where you made the decision, not one where you watched it.

Follow-up

What did you decide not to do, and why?
What would you do differently if you ran that again?

How do you handle model versioning and tracking in a production pipeline?

Medium · behavioural and collaboration

Approach

Close with what you would do differently, concretely.
Name the disagreement and how you resolved it with evidence.
Give the blast radius: what could have broken, and what you measured.

Follow-up

What did you decide not to do, and why?
What would you do differently if you ran that again?

Unblock an engineer without taking the keyboard

Easy · mentoring · leases · at-least-once

A teammate has spent two days on a job handler that occasionally writes duplicate rows. They are certain the queue is delivering twice by mistake. You suspect a lease expiring under a slow handler, so the job is running concurrently with itself. Describe how you have unblocked someone in this position: what you asked before offering a hypothesis, what you showed them rather than told them, and what you left them owning. Then say what you would do if their theory turned out to be the right one.

Approach

Ask before diagnosing, and ask for things answerable from data they already have: the attempt count on the job rows that produced duplicates, the handler's observed duration against its lease expiry, and whether the duplicate rows share a natural key that a unique constraint could have caught.
Teach the shape rather than the answer. A lease cannot distinguish a dead worker from a slow one, so a handler that outruns its lease is running twice by design, and deploys deliver the other half by killing handlers mid-run on every rollout. Both of their candidate theories produce identical duplicate rows, which is why the evidence has to come from timings rather than from argument.
Hand over a checklist they execute: a natural key on every write the handler performs so the second copy collides rather than appends, the record of intent written before any external effect, a lease heartbeat while running, and the metric that shows it working.
Keep ownership with them deliberately. Pair on the first write, then step back; if you finish it yourself you have closed one ticket and left the same person stuck on the next redelivery.
Close on the systemic gap that let two days pass, which is usually a missing dashboard for attempt counts or an undocumented at-least-once contract, and fix that rather than only the bug.

Follow-up

How would you distinguish a genuine double-delivery from a lease expiry using only the data already stored?
Their handler calls an external endpoint before recording that it did. What do you tell them to change first?
What do you do the third time the same person brings you the same class of bug?