Interview guides · Nscale

Nscale Software Engineer Interview Questions & Guide 2026

Timed coding practice for Nscale software engineer candidates: full attempts under a clock, then scored against a rubric.

The 60-second brief

At Nscale, a Software Engineer is at the forefront of building the next generation of AI-native infrastructure. You are not just writing code; you are architecting the platforms that power high-performance computing, managing complex fleet automation, and optimizing developer experiences for critical AI workloads.

Ask how many rounds there are and what each one is before you plan your weeks, because the composition is what your hours should follow. A loop with three coding rounds and one short design conversation deserves a different split from the reverse, and whoever schedules it will usually say plainly which it is.

Focus areas

Evolve APIs without breaking pinned SDK clients
Scope every query and cache key by tenant
Keep money in integer minor units

Overview

This role is inherently cross-functional and strategic. Whether you are working on Infrastructure Software Engineering to manage massive GPU clusters or focusing on Developer Experience to improve internal velocity, your work directly impacts the scalability and reliability of the products that support Nscale customers. You will solve problems that are often at the bleeding edge of cloud-native technology, requiring a blend of deep technical rigor and a product-focused mindset.

The Nscale interview process

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

Initial Screening

Half of this call is the part candidates treat as small talk: start date, notice period, work authorisation and its timing, location and time zone, on-call, and the number. Those are what kill offers late, after several engineers have each spent a day. Surfacing a hard constraint now costs you nothing and occasionally buys you something, since a loop compressed to fit a competing deadline can usually only be arranged if it is asked for early. The common failure is deflecting the compensation question twice, then discovering at offer stage that the band never reached your number.

What to demonstrate

Whether your hard constraints are compatible with the role before a loop gets booked: earliest start, notice period, what authorisation you hold and when it needs action, days on site, willingness to carry a pager
Whether you give a compensation range with something behind it, such as current total compensation or a competing timeline, rather than leaving the band untested
Whether your stated timeline is real, since a competing deadline raised now is something scheduling can sometimes work around and the same deadline raised at offer stage usually is not

How to prepare

Write each constraint down in one line before the call and state them as facts rather than negotiating them live under a question you were not expecting
Set your range from two or three current data points for that level and location, and name the structure you are quoting in, so the number is comparable to the one they are holding
If another process is running, say where it stands and by when, and ask directly whether this loop can be scheduled inside that window

Technical Interviews

The same problem is scored by two different mechanisms depending on the format, and preparing for one does not cover the other. With a person watching, partial progress is visible and a hint is a correction you can absorb; silence is the expensive failure, because nobody can read a half-written function. With an automated grader there is no partial credit for what you were about to do, nobody to ask, and the worked examples in the prompt are the entire specification. Read them as a contract, down to whether an empty result should be an empty list or no output at all.

What to demonstrate

In a live session, whether your commentary tracks what your hands are doing, and whether a hint redirects you or gets defended against
In an automated one, whether you cover the cases the examples do not show, since the hidden cases are where the score moves
Whether you manage the clock on purpose: abandoning an approach that is not converging while there is still time to write something simpler that finishes

How to prepare

Have someone hand you a problem and feed you one deliberately wrong hint. Practise testing it against a concrete case instead of accepting or rejecting it on authority.
Do one timed run a week in a plain browser editor with autocomplete, linting and your own snippets switched off, which is closer to what these environments give you
For the automated format, write the harness before the solution: a main that feeds the worked examples plus an empty and a single-element case and prints expected against actual, so a wrong submission is caught by you first

Behavioral Discussions

Many of these questions are about something that went wrong, and the grading sits mostly in the hours after you knew. Who found out first, whether that was you or an alert or a user, how long it took you to say it out loud, and whether the people who needed the news got it while they could still act on it. Engineers under-tell this part because it feels like confessing. The pattern it is looking for is the opposite: the quiet fix, an incident absorbed without telling anyone, after which nothing changed and the same failure is still available.

What to demonstrate

How the problem was found, and whether that route was one you had built or one that happened to you, since a user reporting it first means your instrumentation did not cover that failure
Whether time-to-detect and time-to-tell are separate numbers in your account and whether you know both, because a fast fix that nobody heard about until the retro is a different answer from a slow one that was announced immediately
Whether the resolution left something durable behind, a check that fires or a default that changed, rather than depending on people remembering to be careful
Whether you can say what the failure cost without either inflating it or waving it away

How to prepare

Reconstruct one incident you were part of as a timeline with clock times: first bad request, first signal, first person who knew, first message outside the team, mitigation, permanent fix. The gaps between those entries are what gets asked about
Look up the configuration of the signal that caught it, including its evaluation window and threshold. An alert defined on a five-minute aggregate cannot fire until the condition holds across that window, which puts a floor under time-to-detect that has nothing to do with how severe the failure was. Be able to say what that floor was and whether anyone had chosen it deliberately
Prepare one story where you escalated early and the severity turned out to be smaller than you thought, including what it cost the people you pulled in. Without it, every answer you give about raising alarms is unfalsifiable

Final Decision-Making

Where the day includes a partner from product, design or data, that conversation is weighted like the technical ones and prepared for least. They are deciding one thing: whether having you in the room makes their decisions cheaper. That means options with costs attached, not implementation detail and not "it depends". An estimate someone can plan against — a range, the assumption that would push it to the high end, and what you would drop to hit the low one — is worth more than a confident single number, which everyone present already knows is wrong.

What to demonstrate

Whether an estimate comes as a range with the assumption most likely to break it, and states what a specific scope cut would actually buy
Whether a technical constraint is handed over as a choice with consequences on their side, rather than as a verdict they have no standing to argue with
Whether you establish what decision is on the table before proposing anything
Whether risk is raised while it can still change the plan, with the trigger that would confirm it, instead of reported afterwards as a slip

How to prepare

Take a project that shipped late and write the two-sentence warning you could have given three weeks earlier, naming what you would have needed decided at that point
Rehearse one estimate out loud until it arrives in three parts: the range, the single assumption that would blow it, and the smallest thing you would cut to protect the date
Rewrite an objection you have actually made — the "we can't do that" version — as two options with their costs, so the choice ends up with the person who owns it

Where candidates lose points

Holding money in a floating-point type, or rounding it more than once

Binary floating point cannot represent 0.01 or 0.1 exactly, so sums drift and two code paths that should agree disagree by cents nobody can trace back. The fix is integer minor units or an exact decimal type end to end, with sub-cent rates expressed as scaled integers such as micro-units, because a per-request price genuinely is smaller than a cent. The second half of the trap is rounding position: rounding each line and then summing gives a different total from summing and rounding once, and half-up and half-even diverge systematically across many lines, so rounding must happen at one named place and every downstream reader must carry the rounded value rather than recompute it from quantity and rate.

Checking a quota with a select and then writing

Under read-committed isolation, two concurrent transactions both observe a count below the limit and both insert, so the limit is exceeded by exactly the concurrency. Repeatable read does not rescue it either: it provides a stable snapshot, and this is write skew, which snapshot isolation permits by design. The options are serialisable isolation, which detects the conflict and aborts one transaction with a serialisation failure and therefore obliges the caller to retry; a single statement with the predicate inside the write; or a constraint that makes the surplus insert fail outright. The reason this pattern survives review is that it is correct in every test that runs one request at a time.

Quoting amortised or average cost as if it were a worst-case guarantee

Appending to a dynamic array is amortised O(1), but the append that triggers a resize copies every element, and hash lookup is constant only while the hash spreads the actual keys. Say which guarantee you are offering when the caller cares about the latency of one call rather than the total over many.

A queue or buffer with no bound

Every producer-consumer boundary needs a capacity and a policy for reaching it: block the producer, shed load, or drop the oldest entry. Unbounded buffering converts a temporary slowdown into memory exhaustion and hides the backpressure signal that would have revealed the consumer was falling behind.

Questions & practice

How would you approach identifying a bottleneck in a distributed system with limited monitoring?

Medium · data structures and algorithms

Approach

Choose the data structure from the access pattern, not from familiarity.
State the target complexity and say which constraint rules the naive version out.
Walk one small example through your approach before writing the whole thing.

Follow-up

What is the worst case, and how likely is it on real data?
How does this change if the input no longer fits in memory?

3 keys and 3 doors: What is the minimum number of tries needed to know exactly which key fits in each door?

Medium · data structures and algorithms

Approach

Walk one small example through your approach before writing the whole thing.
Restate the input: its shape, its size, and what is guaranteed about it.
Choose the data structure from the access pattern, not from familiarity.

Follow-up

What is the worst case, and how likely is it on real data?
Which test case would catch an off-by-one here?

Order a job dependency graph and find its critical path

Medium · topological-sort · dag-longest-path · cycle-detection · critical-path

A workspace defines up to 50,000 jobs with up to 200,000 dependency edges and an estimated duration_seconds per job. Given the edge list, reject the graph if it contains a cycle and name one cycle's nodes; otherwise return a valid execution order, the earliest possible completion time with unlimited workers, and the set of jobs whose slack is zero. Then say which single job to shorten in order to cut the completion time, and by exactly how much. State the complexity of each part.

Approach

Kahn's algorithm for the order: compute indegrees, seed a queue with zero-indegree nodes, emit and decrement. O(V + E), which at 50,000 and 200,000 is milliseconds. If fewer than V nodes are emitted, the graph contains a cycle.
Kahn detects a cycle but cannot name one. The nodes left with indegree above zero contain every cycle, so run one DFS restricted to that residual subgraph with three-colour marking and report the stack slice from the grey node the back edge points at. That is the difference between a usable error message and 'dependency cycle detected'.
Earliest completion with unlimited workers is the longest path, which is NP-hard on a general graph and linear on a DAG. State the precondition, then relax in topological order: earliest_finish[v] = duration[v] + max(earliest_finish[u] for u in preds(v)), taking the max over an empty predecessor set as zero. The makespan T is the maximum over all nodes. O(V + E).
Second pass in reverse topological order for latest_finish, then slack[v] = latest_finish[v] - earliest_finish[v]. Zero-slack nodes form the critical path, and there can be several disjoint critical paths, so return the set rather than one chain. slack[v] = 0 is exactly the statement that some longest path runs through v; equivalently, the longest path through v has length T - slack[v].
The speed-up bound is the point of the question, and the obvious form of it is wrong. Shortening a zero-slack job v by d, with 0 <= d <= duration[v], cuts the makespan by min(d, T - L_avoid(v)), where L_avoid(v) is the longest path in the graph with v deleted: the longest path that avoids v, not the second-longest path overall. The two coincide only when the runner-up path misses v. Counterexample: A of 10 s feeds both B of 5 s and C of 4 s, so T = 15 s and the second-longest path is 14 s, yet shortening A by 10 s leaves a makespan of 5 s. The realised gain is the full 10 s, because both paths ran through A and shrank together, while min(10, 15 - 14) predicts 1 s. The reason is structural: shortening v reduces every path through v by d and leaves every other path alone, so the new makespan is max(T - d, L_avoid(v)).
Compute L_avoid(v) the direct way: delete v and re-run the same forward relaxation, O(V + E) per candidate. The cheaper equivalent skips the deletion, since L_avoid(v) only ever matters through that max: set duration[v] := 0, recompute the makespan as T0(v) = max(T - duration[v], L_avoid(v)), and the gain is min(d, T - T0(v)), which is identical for every d <= duration[v]. Only zero-slack jobs are candidates, because shortening a job with positive slack changes the completion time not at all. One relaxation is milliseconds at this size, so ranking a critical set in the hundreds costs O(k(V + E)) and is worth doing exactly; a critical set in the tens of thousands is not, and there you evaluate a shortlist, longest jobs first, and say that the answer is the best of that shortlist rather than the optimum.

Worked solution (30 min)

Build four fixtures. A: 12 jobs, two branches of 100 s and 95 s that share no job. B: fixture A plus one back edge. C: two disjoint paths tied at 100 s. D: the shared-prefix case, one job of 10 s feeding a 5 s job and a 4 s job, so the longest path is 15 s and the runner-up is 14 s.
Run Kahn; on fixture B confirm it emits fewer than V nodes, then run the residual-subgraph DFS and print the actual cycle.
Compute earliest_finish forward and latest_finish backward, and list the zero-slack set for each fixture.
For each zero-slack job v, recompute the makespan with duration[v] := 0 to get T0(v), and record both the correct bound T - T0(v) and the wrong one, T - second_longest_path, side by side.
Apply the shortening for real (20 s off the critical branch of A, 10 s off the shared prefix of D) and diff the recomputed makespan against each prediction.

Expected result: Fixture A: makespan 100 s, and shortening by 20 s leaves 95 s, a gain of 5 s. Both formulas agree here, because the 95 s branch avoids the shortened job. Fixture D: makespan 15 s, and shortening by 10 s leaves 5 s, a gain of the full 10 s, which T - T0(v) = 15 - 5 = 10 predicts and T - second_longest = 1 does not. Fixture C: the zero-slack set covers both tied paths, and shortening a job on one of them alone gains nothing, since the other path still runs 100 s.

Verify it yourself

The cycle printed by the DFS is a real cycle: every consecutive pair is an edge present in the input.
Summing durations along any zero-slack path equals the makespan exactly.
For every candidate, the predicted gain equals the makespan recomputed from scratch with the shortened duration. Assert the equality; do not eyeball it.
Fixture D is the regression test for the bound, and a fixture whose branches are disjoint is not: it passes under either formula and proves nothing.

Follow-up

Only m workers are available. What happens to your answer, and what can you still promise about the schedule you produce?
Edges arrive incrementally as the customer edits the pipeline. How do you detect a cycle at insert time without re-running Kahn over 250,000 elements?
Durations are estimates. How would you express completion time as a distribution, and what breaks about the critical path once you do?

Migrate a live partitioned event table without blocking ingest

Hard · online migration · ddl locks · expand and contract · partitioning

usage_event is range-partitioned daily on ingested_at, holds roughly 250M rows per day across 400 live partitions, and is written at 10-40k rows/second. Two changes are required: quantity must move from double precision to numeric(20,6), and a new environment column must become NOT NULL with a default of 'production'. Ingest cannot stop. Give the ordered plan, naming for each step the lock it takes, what that lock blocks, and roughly how long it is held. Identify the one step that cannot be rolled back cleanly once traffic depends on it.

Approach

Classify the two changes before planning anything. Adding a column with a non-volatile default has been metadata-only since PostgreSQL 11, so it is cheap. Changing double precision to numeric is not binary-coercible, so alter column ... type rewrites every partition under ACCESS EXCLUSIVE and rebuilds its indexes; on this volume that is hours of blocked ingest and is simply not an option, which is why the plan is expand-and-contract rather than one statement.
Expand: add quantity_numeric numeric(20,6) and environment with its default on the parent. Both are catalogue-only but both take a brief ACCESS EXCLUSIVE that cascades to partitions, so run each with lock_timeout set to a second or two and retry on failure. A queued ACCESS EXCLUSIVE request blocks every reader behind it, which is how a metadata-only change turns into an outage.
Dual-write: deploy producer code that populates both columns on every insert, and leave it running before anything reads the new column. This is the step that cannot be reverted cleanly. Once readers depend on quantity_numeric, reverting the writer leaves rows with a null there, and the gap is only discoverable by re-reading the old column, which the readers have stopped doing.
Backfill older partitions in batches keyed on the primary key, oldest first, committing every few thousand rows with a pause between batches, and skipping the partition still receiving writes until it rotates. Each batch is an ordinary UPDATE taking row locks only. The cost is bloat and WAL rather than blocking, so watch dead tuples and let autovacuum keep pace instead of wrapping 400 partitions in one transaction.
Make NOT NULL cheap with the three-step form: add constraint ... check (environment is not null) not valid (brief ACCESS EXCLUSIVE, no scan), then validate constraint (SHARE UPDATE EXCLUSIVE, scans while reads and writes continue), then set not null, which from PostgreSQL 12 uses the validated check and skips its own full scan. Do this per partition, then on the parent.
Switch and contract: move reads to the new column behind a flag, verify over a full period that both columns agree on freshly written rows, drop the old column (metadata-only), and only then remove the dual-write. Any index on the new column goes on with CREATE INDEX CONCURRENTLY per partition, since CIC is not supported on a partitioned parent: create the parent index with ONLY, build each child concurrently, then ALTER INDEX ... ATTACH PARTITION until the parent index becomes valid.

Follow-up

A CREATE INDEX CONCURRENTLY fails halfway through the partition list. What state is the table in, how do you detect it, and what do you run?
The producer computes quantity itself. What happens to a request already in flight when the dual-write deploy lands, and does it matter?
Give two queries that prove the backfill is complete: one cheap enough to run every minute, one authoritative.

Model credential revocation so history survives the delete

Easy · soft delete · partial indexes · check constraints · audit history

tenant_api_key stores key_id, tenant_id, workspace_id, name, key_prefix, secret_hash, scopes text[], status (active, revoked, expired, compromised), auth_version, created_at, expires_at, last_used_at, revoked_at, revoked_reason. Rotation inserts a new row and revocation never deletes, because an incident review asks which credential served a request last quarter. Write the constraints that enforce: a label is unique only among a tenant's live keys, revoked_at and status can never disagree, and scopes is never empty. Then write the authentication lookup predicate, and name one column in this table that must stay out of it.

Approach

Reach for a partial unique index rather than a plain UNIQUE: create unique index on tenant_api_key (tenant_id, name) where revoked_at is null. Any number of revoked rows may share a label, the live namespace stays unique per tenant, and the revoked majority is not in the index at all, so it stays small on a table that only grows.
Tie the nullable timestamp to the enum so the two cannot drift: check ((revoked_at is not null) = (status in ('revoked','compromised'))) and check ((revoked_at is null) = (revoked_reason is null)). A revocation that records no reason is the one an incident review cannot use.
Write the emptiness check as check (cardinality(scopes) > 0), not array_length(scopes, 1) > 0. array_length returns NULL for an empty array, a CHECK constraint passes when its expression is NULL, so the array_length version accepts exactly the value it was written to reject.
Make the lookup a single index probe with every liveness condition inside it: where secret_hash = $1 and revoked_at is null and (expires_at is null or expires_at > now()) and auth_version = $2, backed by a unique index on secret_hash. Nothing is filtered in application code, so there is no path that forgets a clause.
Keep last_used_at out of that predicate. It is written asynchronously and is allowed to lag by a minute, so it is a usage signal; feeding it into an authorisation decision makes the decision depend on a write that may be late, batched away or lost.
Flag the modelling smell while you are here: expired is derivable from expires_at < now(), so storing it as a status obliges a job to keep it true and guarantees the column is wrong between the expiry instant and that job's next run. Derive it in the predicate; keep the stored status for states that are decisions rather than clock readings.

Worked solution (20 min)

Create the table with all three constraints on a scratch database and insert two revoked rows sharing (tenant_id, name); the partial index should accept both.
Insert a second live row with that same name and confirm the violation names the partial index.
Run update tenant_api_key set revoked_at = now() leaving status = 'active' and confirm the CHECK rejects it; then try insert ... scopes = '{}' against both the cardinality and the array_length forms and note that only one rejects it.
Run explain (analyze, buffers) on the lookup predicate for a live key and confirm an index scan on secret_hash with rows removed by filter equal to zero.

Expected result: A partial unique index on (tenant_id, name) where revoked_at is null; check ((revoked_at is not null) = (status in ('revoked','compromised'))); check ((revoked_at is null) = (revoked_reason is null)); check (cardinality(scopes) > 0); a single-probe lookup on secret_hash plus revoked_at, expires_at and auth_version; last_used_at excluded, with a note that 'expired' should be derived rather than stored.

Verify it yourself

Two revoked keys with the same name insert cleanly; a second live one does not.
The array_length form of the emptiness check accepts an empty array and the cardinality form rejects it.
No statement anywhere in the design deletes or updates a revoked row.
Revoking a key touches exactly one row and takes effect on the next lookup with no application-side filtering.

Follow-up

Rotation issues a replacement while the old key stays live for a 30-day overlap. What does the uniqueness rule become, and what does the UI show to tell two same-named keys apart?
A password reset bumps the principal's auth_version. No row in this table changed. How does the next request fail, and what query counts how many keys that bump just killed?
A key turns up in a public repository. Which columns let you find it, and what do you write to the row?

Per-tenant rate limiting that holds across one hundred pods

Medium · rate limiting · hot keys · atomicity · write skew

The same gateway enforces each tenant's plan rate limit in requests/second and its monthly quota. Traffic is heavily skewed: a handful of tenants exceed all others combined, while a small tenant's requests may all land on one of the 120 pods. Both limits must hold in aggregate rather than per pod, per credential or per region, and the check may add at most 2 ms at p99. Specify the algorithm, where the state lives, the exact operation performed per request, and the behaviour when the counter store is unreachable - separately for the rate limit and the quota.

Approach

Do the arithmetic on per-pod buckets before discarding them, and be precise about which half fails. Giving each pod rate/N keeps the sustained aggregate correct - 120 buckets of 5 rps sum to the 600 rps limit - but a tenant whose traffic all lands on one pod is throttled at 5 rps, a hundred and twentieth of what it pays for. The over-admission comes from burst depth, not sustained rate: if each pod carries a full bucket depth B, the aggregate burst is 120B against a limit that intended to allow B.
Move the rate limit to a shared counter accessed in exactly one atomic round trip - a token bucket or GCRA as a single server-side script returning allow/deny plus a retry-after - never a read followed by a write. Keep the key tenant-scoped and region-local; a globally shared counter costs a cross-region round trip that alone exceeds the 2 ms budget, so shard the global limit per region in proportion to observed regional share and accept that a tenant whose traffic shifts regions is briefly throttled until the shards rebalance.
Cut round trips for the largest tenants with leased tokens: a pod leases 50 at a time and spends them locally, reducing counter round trips by 50x at the cost of bounded over-admission of pods x lease_size, which is 6,000 requests in the worst case here. Shrink the lease for tenants with small limits, where that bound would swamp the limit itself.
Treat the quota as a different problem, not a longer window. It is a count that other in-flight requests are changing, so select-the-usage-then-insert is write skew: read-committed permits it and repeatable read permits it too, because snapshot isolation is exactly what allows two transactions to read a consistent count and both write. Enforce it inside one statement - an atomic increment whose returned value is compared, or UPDATE ... SET used = used + $1 WHERE used + $1 <= limit RETURNING - or with a constraint that makes the surplus write fail.
Write the unavailability policy per limit rather than one policy for both. The rate limit fails open to a conservative local bucket, because it protects capacity and failing closed turns a counter blip into a total outage. The quota fails open for tenants whose last known usage was well below the cap and closed for those at or above it, since that is the only decision with revenue attached, and reconciles from the usage ledger on recovery - which is why the monthly counter must be re-seedable from that ledger rather than being the only copy.

Follow-up

One tenant is 40% of all traffic and its counter key lands on a single shard. What do you change, and what does that cost in accuracy?
Clients receive 429s and retry. What stops the well-behaved ones from synchronising into a herd?
A tenant reports being throttled below its limit. Which evidence do you produce, and does your design emit it today?

Leasing sandboxed job runs without double execution

Medium · leases · fencing · state machines · metering

job-runner leases work from a queue and executes customer workloads in sandboxes with CPU, memory, wall-clock and egress limits. Run durations span 200 ms to 30 minutes, thousands are concurrent, and job_run.status must reach exactly one terminal state among succeeded, failed, timed_out, cancelled and lost. A worker can pause 45 seconds for garbage collection, or be briefly partitioned, after its sandbox has exited but before it writes the outcome. Design the lease duration, the fencing, and timeout ownership, and state how billable_seconds is computed for a run whose worker never returns.

Approach

Resolve the lease-duration conflict rather than picking a compromise. A lease shorter than the longest legitimate run re-dispatches work that is still executing; a lease long enough for a 30-minute run leaves a crashed 200 ms run undetected for half an hour. Use a short lease - about 30 seconds - renewed by a heartbeat every 10 seconds while the supervisor is alive, each renewal pushing leased_until to now() + 30 s. Detection latency is then one lease, not one heartbeat: a dead worker's last successful heartbeat landed at most 10 s before it died, so leased_until expires 20 to 30 s after the death, plus whatever interval the reaper scans on. What the 3:1 ratio of lease to heartbeat buys is headroom - one or two missed heartbeats change nothing, and a stall only becomes a re-dispatch once it outlives the lease remaining when it began, between 20 and 30 s here. Both costs follow from those two numbers: heartbeat write load proportional to concurrent runs (3,000 runs at a 10-second interval is about 300 updates/second), and any pause longer than the lease - the 45-second GC - re-dispatching a run that is perfectly healthy.
Fence the outcome write so the store, not the worker's memory, arbitrates: UPDATE job_run SET status = $1, finished_at = $2, exit_code = $3 WHERE run_id = $4 AND status = 'running' AND lease_token = $5, and require exactly one row affected. Zero rows means this worker was fenced while it was paused, and its correct behaviour is to discard the result, not to retry - the retry is how a resumed worker overwrites a newer attempt.
Make the status machine monotone with a check constraint or a trigger so no terminal row can move, and model a retry as a new row with parent_run_id set and attempt incremented rather than a reset of the old one. That is what keeps how many times did this actually execute answerable afterwards and keeps each attempt's resource usage attributable to itself.
Put the wall-clock timeout in the supervisor, never in the workload: a customer-supplied program asked to time itself out will not. Distinguish timed_out, which the supervisor observed and caused, from lost, which nobody observed at all, and leave exit_code null unless the supervisor actually saw the process exit. Recording lost as failed asserts an outcome no one witnessed and then bills and retries on that assertion.
Decide billable_seconds for the unobserved case explicitly, because it cannot be derived. It is not now() - started_at, which bills queue and pause time the customer never consumed. The two defensible policies are leaving it null while status = 'lost' and billing nothing, or flooring it at the last heartbeat's observed running time as a stated lower bound; pick one and write it down. Add a per-tenant concurrency cap on sandbox slots so one tenant cannot occupy the whole fleet while these questions are being answered.

Worked solution (30 min)

Draw the timeline: last successful heartbeat at t, sandbox exits at t, worker pauses from t to t+45 s, so leased_until = t+30 s and the lease expires there. Mark where the second worker starts and where the first worker's write arrives, then say how short the pause would have had to be to change nothing.
Write the fenced terminal update and state the expected rows-affected in both the healthy case and the fenced case.
Size the heartbeat: 3,000 concurrent runs at a 10-second interval is about 300 row updates/second on job_run - decide whether that lands on the same table and what it contends with.
Write the billable_seconds rule for each terminal status and say which status leaves it null.

Expected result: A short lease of roughly 30 seconds renewed by a 10-second heartbeat, so detection latency is one lease - 20 to 30 seconds after a worker dies, plus the reaper's scan interval - independent of the 200 ms to 30 minute run spread, with one or two missed heartbeats tolerated rather than fatal; a terminal write fenced by AND status = 'running' AND lease_token = $x where zero rows affected means discard; retries as new rows with parent_run_id; a supervisor-owned wall-clock kill that separates timed_out from lost; billable_seconds left null while status = 'lost', or floored at the last heartbeat as a stated policy, never inferred from wall-clock.

Verify it yourself

In your timeline the fenced write must affect zero rows. If it affects one, the lease alone was doing the work and the token adds nothing.
At 3,000 concurrent runs and a 10-second heartbeat, check 300 updates/second against the other write load on that table before accepting the interval; if it is the dominant writer, the heartbeat belongs somewhere else. Check the ratio too: at lease = 3 x interval a lost heartbeat is survivable, whereas at lease = interval every scheduling hiccup is a double execution.
State whether a duplicate execution is billed, then confirm that summing billable_seconds across a parent and its retry matches that stated policy rather than matching it by accident.

Follow-up

A worker returns from a 45-second GC pause and writes succeeded. Walk through what the database does statement by statement.
Heartbeats are now a few hundred writes a second against job_run. How do you keep that off the path that matters, and what do you lose by moving it?
The re-dispatched workload has an external side effect the first attempt already performed. What does your design owe the customer, and what can it not fix?

Regional error rate explodes after a dependency merely slows

Hard · retry storm · circuit breaker · partial outage

A control-plane read replica in one region degrades from 4 ms to 120 ms. Within ninety seconds that region's gateway error rate rises from 0.01% to 40% and its p99 becomes bimodal, one mode near the old p99 and one at the client timeout. The other two regions are unaffected. The gateway retries control-plane reads three times with exponential backoff and no jitter. Give an ordered checklist that separates trigger from amplifier, the offered-load arithmetic, and the controls that break the loop.

Approach

Split the incident into three questions before touching a control: what started it, what amplified it, and what would make recovery slow. Here they are the replica slowdown, the retry policy interacting with queueing, and a synchronised unjittered herd at recovery. They are different mechanisms and each needs its own control.
Read the distribution rather than the mean. A bimodal p99 with one mode pinned at the client timeout is two populations, not one degraded path; split latency by cache hit and miss and confirm the fast mode is hits and the timeout mode is misses that reached the replica.
Do the load arithmetic. Three retries turn one client request into up to four upstream requests, so offered load reaches roughly 4x on a dependency that is already slower, and it arrives at the worst moment. With utilisation approaching one, queueing delay grows superlinearly, which is why a 30x latency increase upstream does not produce a 30x increase downstream, it produces timeouts.
Break the loop with controls that bound offered load rather than with more attempts: a concurrency limit on the control-plane client so at most N calls are in flight and the remainder fail fast, a circuit breaker scoped per dependency and region, and a retry budget capping retries at a small fraction of base traffic so amplification has a ceiling that does not depend on how many clients are retrying.
Add full jitter to whatever retries survive, sleeping uniformly in [0, min(cap, base x 2^attempt)], so attempts de-correlate instead of arriving in waves aligned to the moment of failure.
Decide the unreachable-dependency behaviour in advance, because it is the actual product decision underneath: serving from an expired credential cache keeps the product available while extending a revoked key's life past the stated bound, and failing closed converts a dependency degradation into a total outage. State the mode and the staleness number rather than letting the timeout choose.

Follow-up

The replica recovers. Describe what happens in the first ten seconds with your controls in place versus without them.
Which single metric would have paged before the error rate moved, and why is upstream latency by itself not it?
Requests that fail fast under the concurrency limit still need an answer. What does the gateway return, and what does it do to the usage event it would otherwise have emitted?

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.

What the behavioural round is actually measuring

Every story you tell gets read for blast radius and judgement: what could have broken, who else it touched, what you knew at the moment you decided. Nobody can audit your code in an hour, so they audit your reasoning instead. Pick work where the call was genuinely yours and the consequences were real enough to remember.

Why do you want to work at Nscale?

Medium · behavioural and engineering judgement

Approach

State the situation in two sentences and spend the rest on the reasoning.
Give the blast radius: what could have broken, and what you measured.
Name the disagreement and how you resolved it with evidence.

Follow-up

How did you know your change caused the improvement?
What did you decide not to do, and why?

Tell me about a time you had to advocate for a technical decision that was unpopular with your team.

Medium · behavioural and engineering judgement

Approach

State the situation in two sentences and spend the rest on the reasoning.
Name the disagreement and how you resolved it with evidence.
Close with what you would do differently, concretely.

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 ambiguity when requirements are not clearly defined?

Medium · behavioural and engineering judgement

Approach

Close with what you would do differently, concretely.
State the situation in two sentences and spend the rest on the reasoning.
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?
What did you decide not to do, and why?