Interview guides · Checkout

Checkout Machine Learning Engineer Interview Questions & Guide 2026

Checkout 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 Checkout, you will sit at the intersection of high-scale financial technology and cutting-edge artificial intelligence. Your work is critical to securing the global payments ecosystem, optimizing transaction authorization rates, and building sophisticated fraud detection systems that protect merchants and consumers alike.

The loop does not sample the job evenly, and arguing about that in the room costs you. Daily work is mostly incremental change inside code someone else wrote, while the loop samples narrow slices of it; prepare for the slices and save the realism argument for your questions at the end.

Focus areas

Bound every outbound call with a timeout
Choose indexes from the query's access path
Make every write idempotent under retry

Overview

This role is for engineers who thrive on complexity and are motivated by the tangible impact of their models. You will be responsible for the full lifecycle of machine learning solutions, from data ingestion and feature engineering to model training, deployment, and monitoring. At Checkout, you aren't just building models; you are building robust, production-grade systems that must operate with high reliability and low latency in a fast-paced environment.

The Checkout interview process

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

Take-Home Assignment

Nobody is in the room when this is read, so the submission stands in for what an ordinary pull request from you would cost to review. That makes navigability the first thing being judged. A reader should reach the code that does the actual work within a minute of cloning, without tracing three layers of indirection to get there. What most separates a strong submission from a weak one is separation itself: the logic that implements the prompt kept apart from file reading, argument parsing and printing, so each piece can be read, and tested, on its own.

What to demonstrate

Whether the entry point makes the shape of the program obvious: what is read, what is computed, what is written, in that order
Whether the core logic can be called without touching the filesystem, a socket or stdin, since that is what makes it testable at all
Whether names hold up on a second reading; a function called parse that also writes, or a total that means something different from the README, costs the reader more than a longer name would have
Whether the commit history reads as a sequence of steps someone could follow, rather than one commit containing everything

How to prepare

Clone your own finished submission into an empty directory and read it in the order a stranger would, README then entry point then downward, noting every place you rely on something the code never told you, and fix those places
Take an exercise you have already solved and restructure it so the module holding the rules imports no file, CLI or network library, leaving input parsing and output as a thin shell around it
Build the history while you work, committing at points where the tests pass, so the log shows the order you actually solved the problem in

Behavioral Discussion

This round is deciding whether a change you make without supervision can be allowed to reach production. It is scored on what you knew at the moment you decided, not on how it turned out, so a story that opens with the result and works backwards reads as luck retold as judgement. Say what the options were, what you did not know, what you did to shrink the unknown before committing, and what you accepted as the worst plausible case. The detail that separates answers is a bound: how many users, how much data, and for how long, if you had been wrong.

What to demonstrate

Whether the reasoning you give was available at the time you decided rather than after the result came in, since a story whose deciding evidence arrived later describes an outcome and not a judgement
Whether you can put units on the exposure (users, rows, minutes of degraded service) and whether the containment you chose actually bounded it: a canary bounds the request path it fronts, while a background job writing to a shared table reaches every user regardless of which version served their requests
Whether the reversal path existed before you shipped or was improvised during the incident, and whether it restores state or only stops further damage

How to prepare

For your three largest changes, write down the one thing you would have had to be wrong about for it to fail, and what your best estimate of it was on the day you shipped. If you never held an estimate, that is the gap the follow-up questions will find
Write the undo procedure for one of those changes as it existed at the time, then mark which steps restore data and which only stop new damage. Turning a flag off or reverting a deploy ends the new writes; rows already written come back only from a copy you kept, and a dropped column comes back empty unless something outside the schema holds the values
Rehearse one story from the decision point forward and stop before the outcome, then have someone ask what you would do next. If the story only works with the ending attached, it is an anecdote rather than a decision you can defend

Technical Case Studies

Designs are easy while everything is up. The half of this round that separates people who have operated systems from people who have only drawn them begins when a component is taken away: a consumer dies mid-batch, a replica falls behind, the same request arrives twice because a client retried. Decide early what the system is allowed to lose and what it may never lose, then let the failure behaviour follow from that rather than inventing it per component. The answer that lands is not the one with no failure modes; it is the one whose failure modes are stated and survivable.

What to demonstrate

Whether retries are safe in your design: at-least-once delivery requires either a consumer that dedupes against a key it stores, or an operation that is naturally idempotent. Saying exactly-once without one of those present is the tell
What you gave up by serving reads from asynchronous replicas, including whether you noticed a user can fail to see their own write unless you pin them to the primary for a window or wait on a known replication position
Whether you can state the recovery path and not only the redundancy: what detects the failure, what happens to work already in flight, and how long the system stays degraded
Whether the durability choice tracks the data: dropping an analytics event under load and dropping a payment record are not the same decision, and a design that treats them identically has not made one

How to prepare

Take a design you have already drawn and remove one component at a time, writing what the user sees and what the stored data looks like afterwards. The instructive cases are partial: a write committed to storage while the downstream notification never went out
Find every point where duplicate work becomes visible to a user and write down the dedupe key and where it lives. A client-generated request id retained for a bounded window covers most of it, so state what happens once that window closes
Set a staleness budget in seconds for each read path before picking a replication strategy, and note beside each what actually breaks when the budget is exceeded
Practise the cross-shard question: pick a key, then write which operations now span two shards and whether you are paying for a distributed transaction, accepting that the operation is not atomic, or changing the key

Where candidates lose points

Assuming an isolation level prevents the anomaly you actually have

Isolation levels are named by the SQL standard but implemented differently, so any claim about one is only true of a named engine. PostgreSQL defaults to READ COMMITTED, where every statement takes a fresh snapshot, so two statements inside one transaction can legitimately disagree about the same row. Its REPEATABLE READ is snapshot isolation: it removes non-repeatable and phantom reads but permits write skew, where two transactions each read a set, each conclude their own write is safe, both commit, and the combined result violates a constraint that no single row expresses. Only SERIALIZABLE closes that, and it closes it by aborting a transaction with a serialization failure (SQLSTATE 40001), which means the guarantee is theoretical unless the application has a retry loop. InnoDB's REPEATABLE READ is a different mechanism again - plain SELECTs read a consistent snapshot while locking reads and writes see the latest committed row - so a read-modify-write inside one transaction can act on a value that the transaction's own earlier SELECT never returned.

Paginating with LIMIT/OFFSET over a set that changes while the client is reading it

OFFSET n makes the database produce and discard n rows before returning anything, so the cost of a page grows with its depth rather than with its size and page 500 costs five hundred pages of work. The correctness problem is worse than the cost: if a row is inserted or reordered between two page fetches, rows shift across the offset boundary and are either skipped entirely or returned twice, and neither outcome leaves any trace in the response for the client to detect. Keyset pagination - WHERE (sort_key, id) < ($last_sort_key, $last_id) ORDER BY sort_key DESC, id DESC LIMIT n, backed by an index in exactly that order - reads only the rows it returns and is stable against concurrent inserts. It requires the tie-break column: a timestamp is not unique, and duplicate sort keys straddling a page boundary reintroduce the skip it was adopted to remove.

Hardcoding to the sample inputs

Solve the stated problem rather than the two examples; special-casing a literal to make a sample pass is obvious immediately and reads as either a misunderstanding or an attempt to fake progress. If you genuinely cannot generalise yet, say which part is a stub and what would replace it.

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.

Questions & practice

Explain your approach to debugging a model that is performing well in training but poorly in production.

Medium · machine learning fundamentals

Approach

Pick the metric from the cost of each error type, not from habit.
State the learning problem: the label, the unit of prediction and how the model is used.
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?

How do you determine the success metrics for a new model before implementation?

Medium · machine learning fundamentals

Approach

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

Follow-up

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

Collapse a redelivered event batch into per-aggregate high-water marks

Easy · hashing · at-least-once · aggregation

You drain a batch of up to 5,000,000 events, each (aggregate_id BIGINT, aggregate_version INT, event_type, payload). The log guarantees order within one aggregate only; the batch merges 64 partitions, and a relay failover has redelivered a range, so an older version for an aggregate can appear after a newer one. Given a map of last_applied_version per aggregate, produce the events worth applying, at most one per (aggregate_id, version), plus the count discarded. Target O(n) time. State the memory for 2,000,000 distinct aggregates and what you do when it does not fit.

Approach

One pass, one hash map from aggregate_id to the highest version kept, and a discard counter. An event whose version is at or below last_applied_version for its aggregate is dropped without further work, which is the whole reason the event carries its version rather than a delta. O(n) expected time, O(d) space in distinct aggregates.
Keep the maximum, never the last occurrence. The redelivered range means the final appearance of an aggregate in the batch can be an older version than one seen earlier in the same batch, so last-wins applies stale state over newer state and the projection regresses with no error anywhere.
Cost the memory instead of calling it large: an 8-byte key plus a 4-byte version is 12 bytes of payload, and an open-addressed table held at a 0.7 load factor costs roughly 17 bytes per entry before per-slot metadata, so 2,000,000 aggregates is tens of megabytes in a native layout and several times that in a runtime that boxes both key and value.
If the distinct set exceeds memory, partition on hash(aggregate_id) mod P and reduce each partition independently. Every event for one aggregate hashes to the same partition, so the per-partition result is exact and the merge is concatenation rather than a second reduction.
Reject sorting the batch by (aggregate_id, version) as the default. It is O(n log n) and buys nothing, because max is associative and commutative and needs no ordering; sorting earns its cost only when the downstream consumer must receive the events in order rather than a per-aggregate winner.
Separate the two mechanisms out loud: in-batch deduplication does not make the consumer idempotent, because the same event redelivered tomorrow arrives in a different batch entirely. The projection write itself still has to be keyed on (aggregate_id, version).

Worked solution (20 min)

Write the pass: look up last_applied_version, skip if the event's version is not greater, otherwise upsert into the keep-map only when the incoming version exceeds the version already held, incrementing the discard counter on every skip.
Hand-trace one aggregate whose events arrive as v5, v3, v4, v5 with last_applied_version = 2, and confirm the output holds v5 once while the counter reads 3.
Compute the table footprint for 2,000,000 entries at 12 bytes of payload and a 0.7 load factor, then state the multiplier for a runtime that boxes keys and values.
Add the hash-partitioning fallback and say in one sentence why the per-partition results need no cross-partition merge logic.

Expected result: A single-pass O(n) reduction keyed on aggregate_id that keeps the maximum version rather than the last occurrence, O(d) space with the byte cost stated for 2,000,000 aggregates, a hash-partitioning fallback keyed on aggregate_id, and an explicit statement that batch-local deduplication does not replace a projection write keyed on (aggregate_id, version).

Verify it yourself

Shuffling the batch order changes nothing in the output, which is the test that separates max-wins from last-wins.
Feeding the same batch twice produces an identical keep-set and only doubles the discard counter.
Every emitted event has a version strictly greater than the supplied last_applied_version for its aggregate.

Follow-up

The payload is a patch rather than a snapshot, so applying only the highest version loses the intermediate changes. What changes in your reduction?
How do you detect that version 7 arrived while version 6 was never delivered, and what should the consumer do about the gap?
Two events for one aggregate carry the same version with different payloads. Which one is wrong, and how would you find out?

Write the update path that detects a concurrent edit

Medium · optimistic concurrency · transactional outbox · audit history

resource carries version INT NOT NULL DEFAULT 1. resource_revision holds revision_id, resource_id, version, actor_user_id, change_kind, patch JSONB, request_id, created_at with UNIQUE (resource_id, version). outbox_event holds aggregate_type, aggregate_id, aggregate_version, event_type, payload, status. A PUT carries the version the client read. Write the exact statements for the single transaction that applies the edit, records the revision and enqueues 'resource.updated', and give the handler's branch on zero affected rows. Then say what PostgreSQL 16 does under READ COMMITTED when two of these updates hit one row at once.

Approach

One transaction, three writes, no network call inside it: UPDATE resource SET title = $3, version = version + 1, updated_at = now() WHERE resource_id = $1 AND tenant_id = $4 AND version = $2; then INSERT the resource_revision row at version $2 + 1; then INSERT the outbox_event row at the same aggregate_version. The event goes to a table rather than a broker because no transaction spans both.
Branch on the affected-row count before doing anything else. Zero has three causes — stale version, wrong tenant, row gone — so re-read once and map to 409 carrying the current version, or 404 for an id outside the caller's tenant, which also stops the endpoint confirming that another tenant's id exists.
State the engine behaviour instead of assuming it. Under READ COMMITTED the second UPDATE blocks on the row lock, and when the first commits PostgreSQL re-evaluates the WHERE clause against the newly committed row, so the version predicate now fails and the statement reports zero rows. Under REPEATABLE READ the identical collision raises SQLSTATE 40001 instead, so the handler must fold both shapes into one conflict response.
Keep UNIQUE (resource_id, version) even though the predicate already serialises writers. It is what makes a lost update unwritable if any other path ever reaches the revision table, and it converts a logic bug into 23505 rather than into a silently missing history row.
Refuse to auto-retry the whole PUT. A retry re-reads the winner's state and reapplies an intent formed against data that no longer exists — the silent overwrite the version token was added to detect. Return the conflict; merge field-wise only if the patches are provably disjoint.
Note that now() is the transaction timestamp in PostgreSQL, so resource.updated_at, the revision's created_at and the outbox row share one instant, which is what later makes reconciliation between the three tables unambiguous.

Worked solution (25 min)

Write the three statements plus the rowcount branch and confirm they sit in one BEGIN/COMMIT with no outbound call between them.
Run two clients that both read version 7 and apply their updates 5 ms apart; assert one 200 and one 409.
Assert resource.version = 8, exactly one resource_revision row at version 8, and one outbox_event row at aggregate_version 8.
Repeat at REPEATABLE READ and record the different failure shape (SQLSTATE 40001) the handler must also map to 409.
Delete the version predicate and re-run: both writes commit and the first edit disappears with no error raised anywhere.

Expected result: Exactly one of the two commits. version = 8 with one revision row and one outbox row at version 8, the loser gets 409 carrying version 8, and no state exists in which the resource moved but the revision or the outbox row is missing.

Verify it yourself

COUNT(*) of resource_revision for that resource equals resource.version, with no gap in the version sequence.
Kill the process between the UPDATE and the outbox INSERT — the transaction rolls back and no event describes a change that never happened.
Under a 50-way race, resource.version equals the number of 2xx responses rather than the number of requests.

Follow-up

A client sends the version it read ten minutes ago and the resource has moved three versions. What is in your 409 so it can resolve the conflict without a full re-fetch?
Two editors, two disjoint fields, no overlap. Does your answer still refuse the second write, and should it?
Every write now touches a second hot table. How do you keep the outbox insert and its partial index from becoming the write bottleneck at 1.2k writes/second?

Find version gaps and relay lag with window functions

Medium · window functions · event tables · frame semantics

outbox_event holds event_id, aggregate_type, aggregate_id, aggregate_version, event_type, payload, status ('pending','published','dead'), attempts, created_at, published_at. A projection is missing rows and you must decide whether the relay skipped events or the consumer dropped them. Write three queries over the last seven days: one listing every aggregate_id whose published aggregate_version sequence has a hole, one giving per-day counts with a running total, and one returning the newest published event per aggregate. For each, say where the window function is evaluated relative to WHERE and LIMIT. PostgreSQL 16.

Approach

Gaps: compute lead(aggregate_version) OVER (PARTITION BY aggregate_id ORDER BY aggregate_version) in a subquery, then filter next_version <> aggregate_version + 1 in the outer query. Window functions are evaluated after WHERE, GROUP BY and HAVING and before the outer ORDER BY and LIMIT, so the predicate cannot sit in the same WHERE clause and PostgreSQL 16 has no QUALIFY.
Say what the seven-day filter does to the answer: it truncates every partition, so the first row per aggregate has no predecessor inside the window and a hole spanning the boundary is invisible. Widen the window, or join to resource.version as the authority for the true maximum.
Running total: SELECT date_trunc('day', created_at) AS d, count() AS n, sum(count()) OVER (ORDER BY date_trunc('day', created_at) ROWS UNBOUNDED PRECEDING). An aggregate inside a window call is legal because grouping runs before windowing. The grouping key is unique per row here so ROWS and RANGE agree, but write the frame anyway — over ungrouped rows with tied timestamps the default RANGE frame pulls in every peer row and the total jumps.
Newest per aggregate: DISTINCT ON (aggregate_id) ... ORDER BY aggregate_id, aggregate_version DESC is the cheap PostgreSQL-only form when an index matches that order; row_number() OVER (PARTITION BY aggregate_id ORDER BY aggregate_version DESC) = 1 is the portable form and needs a subquery for the same evaluation-order reason as the gap query.
Interpret rather than report: no gaps plus a normal p95 of published_at - created_at points at the consumer; gaps or a fat lag tail point at the relay; rows still 'pending' with attempts > 0 point at neither, because they never left the database.
Be explicit that the partial index on (created_at, event_id) WHERE status = 'pending' does not serve any of these — they read published rows. Name the index a recurring monitor would need, and say why a query run twice a year may not deserve one.

Follow-up

Relay failover redelivers events. Does a duplicate break the gap query, and how would you detect one from this table alone?
Turn the gap check into a continuous monitor rather than a query someone runs after an incident. What does it watch?
The consumer claims it never received event 4,812,006. What do you look at, in what order?

How do you balance model complexity with the need for low-latency production environments?

Medium · ml system design and production

Approach

Say where features come from at serving time and how they match training.
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?
What happens when a feature is missing at serving time?

How do you ensure your models are scalable and maintainable?

Medium · system design

Approach

Name the read and write paths separately; they rarely have the same bottleneck.
Name the failure you are designing for, then the recovery path.
Choose a partition key and say what query it makes expensive.

Follow-up

What would you drop to keep the system up under load?
How does this behave when that dependency is down for an hour?

What strategies do you use for feature selection in high-dimensional datasets?

Medium

Approach

Say what you would check first and why it is the highest-information step.
Work from the requirement backwards to the design.
Clarify what is being asked and what a complete answer contains.

Follow-up

What assumption would you test first?
How would you know your answer was wrong?

Shard by tenant when one tenant outgrows a single shard

Hard · sharding · hot partition · routing · online migration

One primary holds resource, resource_revision, outbox_event and idempotency_key for every tenant and is at its write ceiling at 1.2k writes/second. tenant_id leads every index. Shard across eight primaries. One tenant holds 22% of all rows and by itself exceeds a single shard's write capacity. Design the routing, the split of that tenant, and the online move of a tenant between shards with writes continuing. State what breaks for queries that are tenant-scoped today, and exactly what a write must do when it arrives at the old shard after the cutover.

Approach

Route on a unit smaller than a tenant from the start. Make the routing key (tenant_id, bucket) with a fixed bucket count - 64 over eight shards - and keep a directory mapping each (tenant_id, bucket) to a shard, carrying a version and cached in every service. An ordinary tenant has all 64 buckets pointing at one shard and behaves exactly as it does today; only the hot tenant has its buckets spread. Hashing tenant_id alone spreads tenants evenly, gives you no way to move one, and has no answer at all for a tenant larger than a node. The bucket count is the part you cannot change later without rehashing rows, so pick it well above the shard count and rebalance by moving buckets, not by re-bucketing.
For the tenant that exceeds one node, its buckets must land on different shards - that is the whole point of bucketing it, and buckets confined to its own shard would rename the rows while leaving every write on the node whose ceiling it already exceeds. Size it from measured numbers rather than from its row share: 22% of rows says nothing about write rate. The current primary tops out near 1.2k writes/second on this hardware and workload, so a tenant peaking at W writes/second needs its buckets spread over at least ceil(W / headroom-per-shard) shards, where the divisor is the share of each shard's ceiling you are willing to give it while that shard still serves other tenants - not the full 1.2k. Size on its peak, not its mean.
Fix the co-location invariant at the right grain. What must commit in one transaction is a resource, its resource_revision row and its outbox_event row, so the bucket is a property of the resource: derive it once at creation and stamp it into resource_id, and every later revision and event routes with its parent for free. Per-tenant co-location was never the requirement, and mistaking it for one is what makes a tenant look unsplittable. What genuinely breaks is an invariant spanning two resources of one tenant - a per-tenant counter, uniqueness across its resources - which now needs either a home-shard table or two-phase commit, and 2PC at this write rate is not a serious option.
Keep the idempotency constraint arbitrating, because it is now enforced per shard. PRIMARY KEY (tenant_id, idempotency_key) only continues to reject a retry if the same key always lands on the same shard, so derive the create-path bucket from hash(tenant_id, idempotency_key) and mint the new resource_id inside that bucket, which also puts the key row and the resource it guards in one transaction. A bucket chosen from anything that differs between a request and its retry - a timestamp, the worker id, a client-supplied resource id - splits one key across two shards, both inserts succeed, and the write endpoint's retry safety is silently gone.
State what the split costs the hot tenant's reads. Its listing, one 21-entry index scan today, becomes a scatter-gather: every bucket-shard returns 21 rows, a coordinator merges and discards the surplus, latency becomes the slowest shard's rather than the median's, and the keyset cursor has to carry a position per bucket instead of one (updated_at, resource_id) pair. Counts over that tenant fan out the same way. The relay becomes one leader per shard; consumers are unaffected because their ordering guarantee was always per aggregate and a resource's events never leave its bucket.
Move one bucket at a time, reversibly, and fence the straggler at the shard rather than at the caller. Copy from a snapshot while the bucket stays read-write, tail changes until the remaining delta is a few seconds of writes, fence writes for that (tenant_id, bucket) alone with a retryable status, apply the final delta, bump the routing version. Scoping the fence to a bucket is what makes a seconds-long freeze affordable. Then have each shard store the routing epoch it believes it holds for each (tenant_id, bucket) and reject any write carrying an older one: without that token, a service on a stale map commits successfully to a database nothing will ever read again, and the loss stays invisible for days. Outside the data path, anything that aggregated across tenants in one query - admin reporting, the A-Z index, global counters - becomes a fan-out across eight shards with a merge, and per-tenant uniqueness survives only on tables that stay whole on the tenant's home shard.

Worked solution (40 min)

Write the routing lookup keyed by (tenant_id, bucket), its version field, where it is cached and invalidated, and the request-path cost.
From the tenant's measured peak write rate and the per-shard headroom you will grant it, compute how many shards its buckets must span, assign them, and show no single shard carries its whole write rate.
Trace one create end to end: which value picks the bucket, where resource_id gets it stamped, and why the revision, outbox and idempotency rows land on the same shard.
Write the move steps for one bucket, then the epoch check the shard performs on every write, and trace a stale-map write through it.

Expected result: A versioned (tenant_id, bucket) to shard directory cached in every service, with ordinary tenants pointing all their buckets at one shard. The hot tenant's buckets are deliberately assigned to different shards, enough of them that its peak write rate divided across them fits the headroom granted on each, because no single node can carry it. The bucket is derived once and stamped into resource_id, so a resource, its revisions and its outbox rows commit together on one shard while the tenant spans several; only cross-resource invariants within the tenant lose their single-shard transaction. The create-path bucket comes from hash(tenant_id, idempotency_key) so the per-shard unique constraint still arbitrates retries. The tenant's listing becomes a scatter-gather with a per-bucket cursor. Migration moves one bucket at a time - snapshot copy, tail, seconds-long fence for that bucket, final delta, routing-version bump - and a routing epoch per (tenant_id, bucket) is checked by the shard so a stale caller's write is rejected rather than committed where nothing will read it.

Verify it yourself

The hot tenant's buckets resolve to more than one shard, and its peak write rate split across them is inside the headroom each shard has after its other tenants.
A resource's rows, its revisions, its outbox rows and its idempotency key row always resolve to a single shard, so no one write spans two.
The same (tenant_id, idempotency_key) routes to the same bucket on every retry, so the per-shard unique constraint still rejects the duplicate.
A write carrying an old routing epoch is rejected by the shard rather than merely misrouted by the caller.

Follow-up

The fence lasts 90 seconds instead of 4 because the final delta keeps growing. What is happening, and what do you do while the tenant is fenced?
Two tenants must merge into one account. What does that cost under this scheme, and which step is not reversible?
A shard is lost entirely. Which tenants are affected, and what is the source of truth for rebuilding them?

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

Four days sample coding, design, fundamentals and the practical rounds at deliberately shallow depth, which is enough to surface the topics you did not know were in scope. That map, rather than a guess made on day one, decides where the last three days go.

Day 1 — Coding, one pass at shallow depth

Solve one problem from each of six families, an array with two pointers, hash counting, binary search, a tree traversal, a graph traversal and one dynamic program, under a hard twenty-minute cap with no extensions, marking each finished, late, or stalled.
For every stall, write the exact move you could not make rather than the subject, so the note reads could not turn the recurrence into a loop rather than bad at dynamic programming.
Fix nothing today. The value of the pass is the unfixed record.

Deliverable: Six timed attempts marked finished, late or stalled, each stall carrying a named blocking move.

Day 2 — Design, one pass at shallow depth

Spend twenty minutes each on three different shapes, a read-heavy feed, a write-heavy ingest path, and something needing a transaction across two entities, stopping each at requirements, interface and data model.
After each, write the first question you could not answer, which is usually a number you could not estimate or a failure mode you had no vocabulary for.
Mark which of the three you would be most relieved not to be asked, and treat that as data rather than as a preference.

Deliverable: Three shallow designs, each with the first unanswerable question written at the bottom.

Day 3 — Fundamentals and the practical rounds

Answer eight short questions in writing at four minutes each, covering the material that fills the gaps between the big rounds: what happens between a URL and a rendered page, what an index costs on write, when a process is preferable to a thread, and what conditions a deadlock requires.
Do one thirty-minute practical task of the kind a take-home compresses: read an unfamiliar two-hundred-line file and write what it does, what you would change, and the one thing you remain unsure of.
Score every answer fluent, correct but slow, or absent, and keep the absent ones visible.

Deliverable: Eight scored short answers and one written reading of unfamiliar code.

Day 4 — The rounds that are about you, and the map

Deliver three behavioural answers aloud against a timer, a conflict, a failure you owned, and a decision made without enough information, marking any that ran past three minutes or contained no number.
Assemble the map: every marked item from days one to three on a single page, sorted by how likely it is to appear in your loop rather than by how uncomfortable it felt.
Choose exactly two areas for the remaining three days and write down what you are deliberately abandoning.

Deliverable: A one-page scored map of the whole surface area with two areas chosen and the rest explicitly abandoned.

Day 5 — First chosen area, to the depth you skipped

Work the higher-ranked area in four focused blocks, choosing items one level above where you stalled rather than repeating what already works.
After each block write the rule you extracted in one sentence with its precondition attached, since a rule carrying no precondition is exactly what fails under a variation.
Re-attempt the day-one or day-two item that exposed this area and compare against the original timing.

Deliverable: Four worked blocks, a timed re-attempt against the original, and three one-sentence rules with preconditions.

Day 6 — Second chosen area, where the gap is coverage rather than speed

Treat the second area differently from the first. Day five drilled something you could already half-do; this one is usually a topic you had simply never met, so build one worked reference example end to end and keep it, rather than attempting six problems badly.
Write down the vocabulary you were missing on day two or three, five terms at most, each with the one sentence that makes it usable in an answer rather than the textbook definition.
Redo the shallow attempt that exposed this area and note whether you now fail later in the problem, because moving the failure point is the realistic gain from a single day and is worth more than a score that did not change.

Deliverable: One worked reference example for the newly covered area, a five-term vocabulary list, and a note on where the failure point moved.

Day 7 — Reassemble the loop

Sit two rounds back to back with no gap, ordering them so the area you chose second comes last, because the map was built from rested, isolated attempts and the loop will reach your weaker area when you are already spent.
Write where the second round suffered from the first, which is normally the point at which structure collapses into narration.
Reduce the week to one page holding only the rules you can state without reading them.

Deliverable: Mock notes on cross-round carryover plus a one-page card of rules you can recite from memory.

Disagreement needs an outcome you did not choose

Conflict answers where you were right and everyone came round are the weakest ones. Stronger: the evidence you went and collected, what would have changed your mind, and what you did in the weeks after the call went against you. Implementing a design you argued against, properly, is a specific and checkable behaviour.

How do you handle data drift and model degradation in a production environment?

Medium · behavioural and collaboration

Approach

Give the blast radius: what could have broken, and what you measured.
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 would you do differently if you ran that again?
How did you know your change caused the improvement?

Tell me about how you plan ML projects end-to-end.

Medium · behavioural and collaboration

Approach

Pick a story where you made the decision, not one where you watched it.
Name the disagreement and how you resolved it with evidence.
Close with what you would do differently, concretely.

Follow-up

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

Ship under a deadline and bound the debt you chose

Medium · pagination · technical debt · tradeoffs

You have four days to ship a tenant-facing listing endpoint. The version you would defend uses keyset pagination over (tenant_id, status, updated_at DESC, resource_id DESC); the version you can finish uses LIMIT/OFFSET with no matching index. Describe a deadline call you actually made of this shape: what you shipped, what you knowingly deferred, how you bounded the damage with a mechanism rather than an intention, and the specific numeric condition that would force the follow-up. Name who you told and where you wrote it down.

Approach

Name the deferred failure precisely instead of calling it slow. OFFSET n makes the database produce and discard n rows, so cost grows with page depth; without an index matching the sort, every matching row is read and sorted before the limit applies; and rows inserted between two page fetches shift across the boundary so items are skipped or repeated with nothing in the response to signal it.
Bound the blast radius with something mechanical rather than a promise: cap maximum page depth, cap page size, restrict the endpoint to one internal caller, or keep it behind a flag. State which failure each cap removes and which it leaves standing.
Attach a number to the trigger and wire it to an alarm: the first tenant crossing N resources, or the endpoint's p99 crossing its share of the 400 ms budget, so the debt announces itself instead of waiting to be remembered.
Write it where the next engineer looks, which is the code and the ticket, not a chat message: what was deferred, why, the cap, and the trigger.
Report what actually happened in your real example, including the case where the trigger never fired and the debt was correctly never repaid.

Follow-up

At what page depth does the offset version breach your latency budget, given your page size and row counts?
What breaks first when you switch to keyset pagination later, and what does a client holding an old page token see?
Who would have overruled you if you had asked for two more days, and did you ask?