A customer sends a 2 GB CSV export with inconsistent dates, duplicate headers, and blank account IDs two days before a pilot. How would you inspect and clean it without hiding data problems from the customer?

I would preserve the raw export, profile it before pandas can normalize anything, and produce a traceable cleaned copy rather than edit the customer's file in place. - I would inspect the raw header row for repeated names and row-width mismatches, then read chunks as strings so automatic type inference does not hide malformed dates or account IDs. - I would calculate null, parse-error, duplicate, and distinct-ID counts across all chunks, normalize only agreed formats, and quarantine rows with blank IDs or impossible dates. - I would share a reconciliation such as 1.8 million received, 12,430 quarantined, and 99.3% loadable so the customer data lead can judge whether the pilot remains representative.

A customer's JSON API sometimes sends order_total as a string and sometimes omits currency. How would you use Pydantic so bad records do not quietly reach the deployment?

I would define an explicit Pydantic model at the adapter boundary and make validation failures visible as customer data issues. - I would use Decimal for order_total, an enum for supported currencies, and required versus optional fields based on the signed data contract. - I would allow only safe coercions, reject ambiguous values, and capture the record ID plus validation path without logging the full payload. - I would return a summary such as 47 of 10,000 orders rejected by field and give the customer sample IDs they can fix at the source.

Your Python adapter calls a customer inventory API that occasionally hangs for 90 seconds and returns 503 during nightly maintenance. What would you change before the first production run?

I would add bounded timeouts and selective retries so one unstable endpoint cannot stall the customer's whole workflow. - I would set separate connection and read timeouts, for example 3 and 20 seconds, based on measured normal latency rather than leaving them unlimited. - I would retry only transient failures such as 429, 502, 503, and 504 with exponential backoff, jitter, and a maximum attempt count. - I would expose retry and final-failure counts in the run summary so the customer knows whether inventory freshness met the agreed cutoff.

A hospital can only provide a 40 GB newline-delimited JSON file through SFTP, and the deployment container has 2 GB of memory. How would you process it?

I would stream the NDJSON as bytes and commit bounded batches so memory stays predictable and a transfer failure is recoverable. - I would read one newline-delimited record at a time from a buffered SFTP stream, decode the agreed encoding, validate the JSON object, and write batches of perhaps 5,000 records. - After each target commit, I would persist the exact byte position at a line boundary together with the source file size and fingerprint, and resume only if the remote file is unchanged. - I would measure throughput, peak memory, and rejected-row counts in staging, then give the hospital an expected completion window and a tested rerun procedure.

A customer SDK is synchronous, but your shared deployment service uses asyncio to call ten tenant endpoints concurrently. How would you integrate the SDK without freezing every tenant job?

I would isolate the blocking SDK behind a small adapter and run its calls off the event loop. - I would use asyncio.to_thread or a bounded thread pool for the synchronous I/O rather than pretending the SDK is asynchronous. - I would cap concurrency to protect both thread count and the customer's API, then measure whether parallelism actually improves the nightly window. - I would explain that rewriting the SDK is unnecessary for the pilot, while the adapter keeps that option open if it becomes the proven bottleneck.

During a 10,000-record customer import, record 4,201 fails because a required product code is missing. How should the Python job handle the exception and partial result?

I would handle the failure at record or batch scope, preserve successful work, and make the partial state explicit. - I would catch expected validation errors near the record, quarantine that item with a reason, and let independent records continue. - I would stop and roll back only for systemic failures such as an invalid schema, expired credentials, or a broken target transaction. - I would return a completion status like partial_success with loaded, rejected, and retriable counts so the customer never mistakes 9,999 records for a complete load.

A customer support engineer asks for logs to debug failed employee syncs, but payloads contain names, emails, and access tokens. What would you log?

I would emit structured operational context while redacting personal data and secrets by default. - I would log fields such as timestamp, deployment ID, customer-safe tenant ID, endpoint, status code, duration, attempt, and correlation ID. - I would mask emails, omit request bodies, and configure redaction for Authorization, cookies, tokens, and known sensitive JSON paths. - I would give support a correlation ID and aggregate failure reason so they can trace the sync without receiving employee records.

A customer's payroll API has no sandbox, and a field rename could break your adapter on payday. What pytest coverage would you add before deployment?

I would build contract tests around saved, sanitized examples and the exact request and response boundary we depend on. - I would test a normal response, missing optional fields, renamed or extra fields, authentication failure, pagination, and a representative error body. - I would use responses or respx to verify HTTP method, path, headers, timeout, and parsing without calling the live payroll system in CI. - I would ask the customer to approve the sanitized fixtures and run one read-only smoke call during their change window before payday.

The same adapter must run for three customers with different endpoints and credentials. A teammate proposes copying the file and editing constants for each deployment. What would you do instead?

I would keep one adapter and move customer-specific values into validated runtime configuration. - I would define required settings such as base URL, tenant ID, timeout, and feature flags, then fail fast at startup if any are invalid. - I would store credentials in the customer's secret manager or deployment platform, never in source control, images, or sample config. - I would provide a non-secret example file and a startup configuration summary with secret values redacted, making each deployment reproducible.

Your Python adapter works on your laptop, but the customer requires it to run in their Kubernetes cluster with no root user and no internet access. How would you containerize it?

I would build a small, pinned container that runs as a non-root user and contains everything needed before it reaches the customer network. - I would use a slim Python base, lock dependency versions, copy only required files, and use a multi-stage build if native packages need compilation. - I would set a non-root UID, a clear entrypoint, a health or job exit signal, and write temporary data only to an approved writable path. - I would test the image without network access, provide its digest and vulnerability scan, and document the CPU, memory, and configuration the customer must supply.

On day one of a retail deployment, you receive read-only access to a warehouse with 300 unfamiliar tables and little documentation. How would you find the data needed for a stockout workflow?

I would map the smallest schema slice needed for stockouts before writing production queries. - I would inspect table and column metadata, row counts, keys, recent timestamps, and a few safe samples for likely inventory, product, store, and order tables. - I would trace candidate identifiers and update cadence with the customer's data engineer instead of guessing from column names alone. - I would produce a one-page source map showing tables, joins, owners, freshness, and open questions so the customer can validate my understanding.

The customer wants to join CRM accounts to billing records, but CRM uses strings like ACCT-0042 while billing stores integer 42. How would you make the join reliable?

I would define and validate an explicit cross-system ID mapping rather than burying string cleanup inside the final join. - I would confirm with both system owners that the prefix and leading zeros are formatting only, then normalize into a staging key. - I would test one-to-one assumptions and report unmatched, duplicated, and malformed IDs before calculating customer revenue. - I would retain both source IDs in the output so finance can trace any account back to CRM and billing.

A support director says the pilot should reduce average ticket resolution time by 20%. How would you write the SQL aggregation so the result supports that customer KPI?

I would agree on one KPI definition and apply the same SQL aggregation to a comparable baseline and pilot cohort. - I would confirm whether reopened tickets and nonbusiness hours are excluded; because the sponsor said average, I would use AVG unless we explicitly choose median to reduce the effect of outliers. - I would compare equivalent pre-pilot and pilot windows, segment by severity, and return the ticket count with the metric so a changed case mix is visible. - I would report the calculation directly, for example mean resolution time fell from 10.0 to 7.8 hours, a 22% reduction across 1,240 tickets, which exceeds the 20% target.

A daily customer export contains repeated order IDs with different updated_at values, and finance needs exactly one current row per order. How would you deduplicate it deterministically?

I would define a deterministic winner for each order and preserve the duplicate evidence for finance. - I would use ROW_NUMBER partitioned by order_id and ordered by updated_at descending, followed by a stable source sequence or ingestion ID as the tie-breaker. - I would collapse byte-for-byte duplicate payloads safely, but quarantine rows that share the latest timestamp and disagree on business fields when no trusted tie-breaker exists. - I would reconcile order counts and revenue before and after deduplication and document the winning rule so rerunning the export produces the same result.

In a customer's opportunities table, closed_at is NULL for open deals, but some old closed deals also have NULL because of a migration bug. How would you avoid misleading the sales forecast?

I would not treat NULL as one business meaning until the customer confirms how deal state should be derived. - I would use the authoritative status field with closed_at, inspect contradictory combinations, and quantify the legacy gap by period. - I would write predicates with IS NULL and explicit CASE branches rather than comparisons that produce unknown SQL results. - I would show the sales owner how many records are excluded or classified as unknown and request a source correction before calling the forecast complete.

A global customer stores event timestamps without offsets, while teams in New York and Berlin expect a daily dashboard in local time. What would you do before building it?

I would identify the source time zone, convert events to UTC for storage, and apply the requested business zone only at reporting boundaries. - I would ask the system owner whether timestamps represent server local time, user local time, or UTC and test records around daylight-saving changes. - I would use timezone-aware types and named IANA zones such as America/New_York rather than fixed offsets. - I would document each dashboard's cutoff and show sample events near midnight so both customer teams agree which day receives them.

A reconciliation query takes 18 minutes on the customer's production warehouse, and you are not allowed to create indexes yourself. How would you work with their DBA?

I would bring the DBA evidence from the query plan and a specific, reversible request instead of asking for a generic performance fix. - I would capture EXPLAIN output, table sizes, filters, join keys, estimated versus actual rows if permitted, and the required runtime target. - I would reduce unnecessary columns or scans first, then propose an index on the proven filter or join pattern with its expected write and storage cost. - I would ask the DBA to test it in a safe environment and agree on rollback, then report the measured improvement to the customer workflow owner.

Your deployment loads an invoice header and its line items into the customer's database. A line insert can fail after the header succeeds. How would you protect the customer's records?

I would write each invoice in one bounded transaction and make retries idempotent by the source invoice ID. - I would validate the invoice, begin the transaction, insert the header and all lines, and commit only after every constraint passes. - On failure I would roll back, record a safe reason, and rely on a unique source invoice ID or upsert rule so an uncertain retry cannot create a second header. - I would keep the transaction scoped to one invoice or small batch rather than the entire nightly file so a bad record cannot hold customer locks for hours.

A customer wants hourly updates from a 200-million-row events table, but full reloads miss the one-hour SLA. How would you introduce an incremental watermark?

I would load only rows beyond a durable watermark, with an overlap that catches late-arriving events. - I would choose a stable source field such as updated_at plus event_id, because timestamps alone can tie or arrive out of order. - I would persist the watermark only after the target batch succeeds and reread a small overlap window with idempotent upserts. - I would show the customer the measured lag, overlap size, and late-arrival count so the one-hour freshness claim is auditable.

After the first production load, the source has 1,002,481 customers but the target has 998,920. How would you reconcile the deployment before telling the customer it succeeded?

I would reconcile counts and business totals by stage, then account for every difference before declaring success. - I would compare source, extracted, validated, rejected, and loaded counts using the same snapshot and filters. - I would add checks for distinct customer IDs, key nulls, duplicates, and control totals such as active accounts by region. - I would give the customer a signed-off report that explains the 3,561-record gap by reason and names any unresolved exceptions.

Forward Deployed Engineer interview questions

100 real questions with model answers and explanations for Junior candidates.

See a Forward Deployed Engineer resume example →

Practice with flashcards

Spaced repetition · Hunter Pass

Questions

I would preserve the raw export, profile it before pandas can normalize anything, and produce a traceable cleaned copy rather than edit the customer's file in place.

I would inspect the raw header row for repeated names and row-width mismatches, then read chunks as strings so automatic type inference does not hide malformed dates or account IDs.
I would calculate null, parse-error, duplicate, and distinct-ID counts across all chunks, normalize only agreed formats, and quarantine rows with blank IDs or impossible dates.
I would share a reconciliation such as 1.8 million received, 12,430 quarantined, and 99.3% loadable so the customer data lead can judge whether the pilot remains representative.

Why interviewers ask this: The interviewer is evaluating whether you combine practical data handling with traceability and honest customer communication.

deploymentapivalidation

I would define an explicit Pydantic model at the adapter boundary and make validation failures visible as customer data issues.

I would use Decimal for order_total, an enum for supported currencies, and required versus optional fields based on the signed data contract.
I would allow only safe coercions, reject ambiguous values, and capture the record ID plus validation path without logging the full payload.
I would return a summary such as 47 of 10,000 orders rejected by field and give the customer sample IDs they can fix at the source.

Why interviewers ask this: A strong answer uses typed validation to protect production while giving the customer actionable evidence.

python

I would add bounded timeouts and selective retries so one unstable endpoint cannot stall the customer's whole workflow.

I would set separate connection and read timeouts, for example 3 and 20 seconds, based on measured normal latency rather than leaving them unlimited.
I would retry only transient failures such as 429, 502, 503, and 504 with exponential backoff, jitter, and a maximum attempt count.
I would expose retry and final-failure counts in the run summary so the customer knows whether inventory freshness met the agreed cutoff.

Why interviewers ask this: The interviewer wants to see safe network defaults tied to the customer's freshness and reliability expectations.

containersdeploymentmemory

I would stream the NDJSON as bytes and commit bounded batches so memory stays predictable and a transfer failure is recoverable.

I would read one newline-delimited record at a time from a buffered SFTP stream, decode the agreed encoding, validate the JSON object, and write batches of perhaps 5,000 records.
After each target commit, I would persist the exact byte position at a line boundary together with the source file size and fingerprint, and resume only if the remote file is unchanged.
I would measure throughput, peak memory, and rejected-row counts in staging, then give the hospital an expected completion window and a tested rerun procedure.

Why interviewers ask this: This checks whether the candidate can adapt basic file processing to real customer infrastructure constraints.

asyncconcurrencyendpoints

I would isolate the blocking SDK behind a small adapter and run its calls off the event loop.

I would use asyncio.to_thread or a bounded thread pool for the synchronous I/O rather than pretending the SDK is asynchronous.
I would cap concurrency to protect both thread count and the customer's API, then measure whether parallelism actually improves the nightly window.
I would explain that rewriting the SDK is unnecessary for the pilot, while the adapter keeps that option open if it becomes the proven bottleneck.

Why interviewers ask this: The interviewer is testing concurrency fundamentals, scope discipline, and the ability to explain a pragmatic choice.

error-handlingpython

I would handle the failure at record or batch scope, preserve successful work, and make the partial state explicit.

I would catch expected validation errors near the record, quarantine that item with a reason, and let independent records continue.
I would stop and roll back only for systemic failures such as an invalid schema, expired credentials, or a broken target transaction.
I would return a completion status like partial_success with loaded, rejected, and retriable counts so the customer never mistakes 9,999 records for a complete load.

Why interviewers ask this: A strong answer distinguishes isolated bad data from systemic failure and communicates partial outcomes clearly.

tokens

I would emit structured operational context while redacting personal data and secrets by default.

I would log fields such as timestamp, deployment ID, customer-safe tenant ID, endpoint, status code, duration, attempt, and correlation ID.
I would mask emails, omit request bodies, and configure redaction for Authorization, cookies, tokens, and known sensitive JSON paths.
I would give support a correlation ID and aggregate failure reason so they can trace the sync without receiving employee records.

Why interviewers ask this: The interviewer checks whether observability helps the customer without creating a second data exposure.

coveragepytestapi

I would build contract tests around saved, sanitized examples and the exact request and response boundary we depend on.

I would test a normal response, missing optional fields, renamed or extra fields, authentication failure, pagination, and a representative error body.
I would use responses or respx to verify HTTP method, path, headers, timeout, and parsing without calling the live payroll system in CI.
I would ask the customer to approve the sanitized fixtures and run one read-only smoke call during their change window before payday.

Why interviewers ask this: This evaluates whether tests protect a customer integration contract rather than only isolated helper functions.

deploymentendpoints

I would keep one adapter and move customer-specific values into validated runtime configuration.

I would define required settings such as base URL, tenant ID, timeout, and feature flags, then fail fast at startup if any are invalid.
I would store credentials in the customer's secret manager or deployment platform, never in source control, images, or sample config.
I would provide a non-secret example file and a startup configuration summary with secret values redacted, making each deployment reproducible.

Why interviewers ask this: A strong answer prevents configuration drift while respecting each customer's secret-handling process.

kubernetescontainerspython

I would build a small, pinned container that runs as a non-root user and contains everything needed before it reaches the customer network.

I would use a slim Python base, lock dependency versions, copy only required files, and use a multi-stage build if native packages need compilation.
I would set a non-root UID, a clear entrypoint, a health or job exit signal, and write temporary data only to an approved writable path.
I would test the image without network access, provide its digest and vulnerability scan, and document the CPU, memory, and configuration the customer must supply.

Why interviewers ask this: The interviewer is looking for reproducible packaging that fits customer security and runtime constraints.

documentationdeploymentwarehouse

I would map the smallest schema slice needed for stockouts before writing production queries.

I would inspect table and column metadata, row counts, keys, recent timestamps, and a few safe samples for likely inventory, product, store, and order tables.
I would trace candidate identifiers and update cadence with the customer's data engineer instead of guessing from column names alone.
I would produce a one-page source map showing tables, joins, owners, freshness, and open questions so the customer can validate my understanding.

Why interviewers ask this: This tests disciplined schema discovery and early collaboration in an unfamiliar customer environment.

joins

I would define and validate an explicit cross-system ID mapping rather than burying string cleanup inside the final join.

I would confirm with both system owners that the prefix and leading zeros are formatting only, then normalize into a staging key.
I would test one-to-one assumptions and report unmatched, duplicated, and malformed IDs before calculating customer revenue.
I would retain both source IDs in the output so finance can trace any account back to CRM and billing.

Why interviewers ask this: The interviewer evaluates whether the candidate treats customer identifiers as governed data, not convenient strings.

sqlaggregation

I would agree on one KPI definition and apply the same SQL aggregation to a comparable baseline and pilot cohort.

I would confirm whether reopened tickets and nonbusiness hours are excluded; because the sponsor said average, I would use AVG unless we explicitly choose median to reduce the effect of outliers.
I would compare equivalent pre-pilot and pilot windows, segment by severity, and return the ticket count with the metric so a changed case mix is visible.
I would report the calculation directly, for example mean resolution time fell from 10.0 to 7.8 hours, a 22% reduction across 1,240 tickets, which exceeds the 20% target.

Why interviewers ask this: A strong answer connects SQL details to a defensible customer success metric.

queries

I would define a deterministic winner for each order and preserve the duplicate evidence for finance.

I would use ROW_NUMBER partitioned by order_id and ordered by updated_at descending, followed by a stable source sequence or ingestion ID as the tie-breaker.
I would collapse byte-for-byte duplicate payloads safely, but quarantine rows that share the latest timestamp and disagree on business fields when no trusted tie-breaker exists.
I would reconcile order counts and revenue before and after deduplication and document the winning rule so rerunning the export produces the same result.

Why interviewers ask this: The interviewer checks for deterministic SQL and awareness that deduplication rules affect customer reporting.

migrationsfundamentals

I would not treat NULL as one business meaning until the customer confirms how deal state should be derived.

I would use the authoritative status field with closed_at, inspect contradictory combinations, and quantify the legacy gap by period.
I would write predicates with IS NULL and explicit CASE branches rather than comparisons that produce unknown SQL results.
I would show the sales owner how many records are excluded or classified as unknown and request a source correction before calling the forecast complete.

Why interviewers ask this: This evaluates SQL NULL semantics together with judgment about ambiguous customer data.

I would identify the source time zone, convert events to UTC for storage, and apply the requested business zone only at reporting boundaries.

I would ask the system owner whether timestamps represent server local time, user local time, or UTC and test records around daylight-saving changes.
I would use timezone-aware types and named IANA zones such as America/New_York rather than fixed offsets.
I would document each dashboard's cutoff and show sample events near midnight so both customer teams agree which day receives them.

Why interviewers ask this: The interviewer wants practical time-zone handling that prevents customer KPI shifts at day boundaries.

indexesquerieswarehouse

I would bring the DBA evidence from the query plan and a specific, reversible request instead of asking for a generic performance fix.

I would capture EXPLAIN output, table sizes, filters, join keys, estimated versus actual rows if permitted, and the required runtime target.
I would reduce unnecessary columns or scans first, then propose an index on the proven filter or join pattern with its expected write and storage cost.
I would ask the DBA to test it in a safe environment and agree on rollback, then report the measured improvement to the customer workflow owner.

Why interviewers ask this: A strong answer combines basic query diagnosis with respect for customer ownership and change control.

databasedeployment

I would write each invoice in one bounded transaction and make retries idempotent by the source invoice ID.

I would validate the invoice, begin the transaction, insert the header and all lines, and commit only after every constraint passes.
On failure I would roll back, record a safe reason, and rely on a unique source invoice ID or upsert rule so an uncertain retry cannot create a second header.
I would keep the transaction scoped to one invoice or small batch rather than the entire nightly file so a bad record cannot hold customer locks for hours.

Why interviewers ask this: The interviewer checks atomicity plus sensible transaction scope in a customer data load.

I would load only rows beyond a durable watermark, with an overlap that catches late-arriving events.

I would choose a stable source field such as updated_at plus event_id, because timestamps alone can tie or arrive out of order.
I would persist the watermark only after the target batch succeeds and reread a small overlap window with idempotent upserts.
I would show the customer the measured lag, overlap size, and late-arrival count so the one-hour freshness claim is auditable.

Why interviewers ask this: This evaluates a junior candidate's understanding of incremental loads, recovery, and freshness commitments.

deployment

I would reconcile counts and business totals by stage, then account for every difference before declaring success.

I would compare source, extracted, validated, rejected, and loaded counts using the same snapshot and filters.
I would add checks for distinct customer IDs, key nulls, duplicates, and control totals such as active accounts by region.
I would give the customer a signed-off report that explains the 3,561-record gap by reason and names any unresolved exceptions.

Why interviewers ask this: A strong answer proves completeness with traceable controls rather than relying on a successful job exit.

Locked questions

21
A customer operations lead asks why your integration uses POST to create a case, PATCH to update it, and returns 202 during long processing. How would you explain the choices in customer terms?
concurrency
22
At 06:00, a customer's OAuth-based connector starts returning 401 even though it worked yesterday, and their daily report is due at 07:00. What do you check and communicate?
communicationoauth
23
A customer API says there are 12,400 assets, but your first adapter run imports only the first 100. How would you fix pagination and prove completeness?
pagination
24
During a customer backfill, their API returns 429 after 60 requests per minute and includes Retry-After. The sponsor wants you to add more workers. What would you recommend?
resiliencebackfillapi
25
A customer sends payment webhooks to your deployment, and a security reviewer asks how you know an attacker did not forge the payload. What would you implement and explain?
deploymentwebhooks
26
A customer's webhook provider retries the same shipment event five times when your endpoint responds slowly. How would you stop five shipment records from appearing?
endpointswebhooks
27
A customer plans to rename employeeNumber to workerId in their API next month, while two regions cannot upgrade together. How would you keep the deployment running?
deploymentapi
28
Customer developers receive only 500 integration_failed when an employee record is rejected. What error contract would you propose so they can resolve issues without seeing internals?
29
A customer asks for real-time inventory updates, but their source system can export only every 30 minutes and the business acts on stockouts twice a day. Would you build streaming?
system-designstreaming
30
A customer says your API is broken, but you suspect their proxy removes the Authorization header. How would you reproduce the issue and communicate the result?
communicationapiauth
31
A logistics VP says, 'Use our data to make deliveries better,' and expects a pilot in four weeks. How would you turn that into a first scoped workflow?
32
A customer claims the new triage workflow should save analysts time, but nobody knows the current effort. What baseline and success criteria would you establish?
fundamentals
33
In a customer whiteboard session, you need to explain a proposed flow from Salesforce through your adapter to Snowflake and an operations dashboard. What would you draw?
snowflakesessions
34
Your pilot depends on customer security approving OAuth, data engineering creating a view, and operations naming test users. How would you prevent these dependencies from quietly blocking week three?
oauthdependencies
35
A customer wants a production deployment in six weeks. How would you create milestones that reveal risk earlier than the final cutover?
milestonesdeployment
36
A customer asks you to test a mapping change directly in production because staging data is stale. How would you preserve development, staging, and production boundaries?
37
A customer's deployment pipeline currently copies a Python adapter to a server manually. What CI/CD checks would you add before automating production release?
ci-cddeploymentpython
38
The first production sync will run every hour, and the customer asks how they will know it is healthy without reading raw logs. What observability would you add?
observability
39
You are preparing to switch 500 customer users from a legacy workflow on Friday evening. What belongs on your cutover checklist and rollback trigger?
rollback
40
After a successful pilot, the customer's operations team must own the integration without calling you for every failure. What would you put in the runbook and handoff?
runbooks
41
During kickoff, a customer asks whether you are their Solutions Engineer, project manager, or the person who will build the integration. How would you clarify your FDE role?
initiation
42
A customer opens discovery by demanding a chatbot, but their real complaint is that agents spend 15 minutes searching policy documents. How would you respond without simply accepting the proposed solution?
43
After a discovery call, customer engineering expects a daily file while operations expects live updates. How would you restate requirements before anyone starts building?
44
You need sample claims data and API access to start a two-week healthcare pilot, but the customer says, 'Tell us what you need.' What exactly would you request?
api
45
A customer asks for a unique Salesforce adapter even though the standard template covers 90% of their fields. How would you decide between configuration and one-off code?
config
46
A nontechnical sponsor wants the dashboard in one week, but row-level access controls will push delivery to two weeks. How would you explain the tradeoff?
sponsor
47
A customer executive insists on a Friday production launch, but reconciliation still shows 8% of invoices missing. How would you push back respectfully?
react
48
Customer security has not approved a service account, so your pilot has been blocked for three days. What update would you send to the sponsor?
sponsor
49
In tomorrow's customer demo, search works for English documents but not scanned PDFs, which represent 25% of their archive. How would you present the demo?
50
Your first plan used customer webhooks, but after a week you learn their legacy system cannot send them. How would you recover without losing customer confidence?
system-designwebhooks
51
A 4,000-employee customer asks for SSO and SCIM before a pilot starts in three weeks. How would you separate the two requirements?
52
A customer identity architect can offer either SAML 2.0 or OIDC for a web application, and kickoff is Friday. How would you discuss the choice?
initiation
53
A retailer wants SCIM to provision 12,000 users nightly and disable leavers within 30 minutes. How would you test the integration?
54
A bank has six Okta groups but your product has three roles, and access review is due in ten days. How would you build the RBAC mapping?
rbac
55
A nightly Salesforce sync needs a service account before production cutover next week. What would you ask the customer to create?
56
A hospital offers an API token with administrator access for a read-only patient-directory import of 200,000 records. What would you do?
tokensapi
57
A customer's production API secret expires in 48 hours, but the integration processes 30,000 orders per day. How would you rotate it?
secretsapiconcurrency
58
A manufacturer will connect only through a VPN, IP allowlist, or private endpoint, and the PoC has five days left. How would you choose a path?
endpoints
59
A customer's Java agent reports a TLS certificate error when calling your API, and cutover is tomorrow morning. What would you inspect and communicate?
communicationapitls
60
A financial customer sends a 90-question security questionnaire due in four business days. How would you provide evidence without inventing answers?
61
A customer needs 2 million Salesforce Account and Opportunity records copied nightly, but its REST API allocation is shared with other apps. How would you design the first version?
restdesign
62
A Workday customer cannot grant API access before a 14-day pilot but can schedule a custom report export each morning. What would you propose?
api
63
A customer gives your loader a Snowflake ACCOUNTADMIN role, and a 500 GB backfill must finish this weekend. What would you change?
snowflakebackfill
64
An insurer will deliver a 20 GB claims file over SFTP every night by 1 AM, and results are due by 5 AM. How would you make the batch reliable?
batch
65
A factory runs SAP ECC and needs purchase orders in your platform within 30 minutes, but it has no approved REST path. What integration options would you assess?
rest
66
A French customer sends a 300,000-row CSV that opens correctly in Excel but your parser shifts columns and corrupts accented names. What would you inspect?
schemaexcel
67
A customer's daily orders feed adds a new column and changes total_amount from a number to text two days before launch. How would you handle the schema drift?
schemaiac
68
A retailer has customer records in Salesforce, Snowflake, and an ERP, but email addresses change and duplicates exist. Which source keys would you map?
snowflake
69
A support organization wants an LLM assistant in a six-week pilot for 40 agents and 12 possible workflows. How would you select the first workflow?
70
A legal customer says its contract-review assistant must be 'accurate' before a 30-day PoC. How would you create a baseline and evaluation set?
71
A 2,000-employee customer wants answers about policies that change monthly. When would you use RAG instead of a prompt-only solution?
72
A bank requires every policy answer to show evidence and refuses confident answers when no source is found. How would you design that behavior?
design
73
A retailer updates return policies every Friday, but its assistant sometimes quotes last month's rule. What would you inspect and change?
74
A SaaS customer will pilot the assistant with three subsidiaries, and documents from one subsidiary must never appear in another's answers. What would you implement?
cloud
75
During a 10-day RAG pilot, customer security finds instruction-like text on an untrusted wiki page, but the content owner resists quarantining it because 300 support articles link to the page. How do you handle the deployment?
soft-skillsdeployment
76
A healthcare customer wants a pilot in 7 days using 20,000 support transcripts, but its DPA with the external model vendor is not signed and the files contain health details. What temporary deployment path would you propose?
procurementdeployment
77
A call-center assistant must respond within two seconds and stay below $0.03 per request at 100,000 requests per day. How would you evaluate the model path?
decision-making
78
A claims team wants the LLM to draft decisions for 25 reviewers during a four-week pilot. How would you collect feedback and keep humans in control?
feedback
79
During a live executive demo, the assistant invents a refund policy for a $3 million expansion opportunity. What do you do in the room and afterward?
80
A customer-facing RAG service is degraded during Monday peak traffic, and answers may omit citations. What safe fallback would you ship?
81
In a 60-minute interview exercise, you must connect a synthetic customer REST API to a deployment and explain the result to the customer's data engineer. What would you build first?
restdeployment
82
An Okta user completes SAML sign-in but your application denies access, while the customer's test user works. How would you debug this without weakening access control?
83
A customer user edits a case while your integration also PATCHes it, and the integration silently overwrites the new owner. How would you change the REST flow?
rest
84
A pilot saves 6 minutes on each of 30 daily cases for 20 analysts over 220 workdays. Loaded analyst cost is $45 per hour and annual deployment cost is $180,000. How would you calculate and present the value?
deployment
85
An operations dashboard must show all 120 stores, including stores with zero failed syncs, but your SQL returns only 83 after joining the error table. What would you change and verify?
sql
86
An integration works in your staging tenant but fails in the customer's production tenant two hours before acceptance testing. What would you compare?
acceptance
87
A customer agent cannot resolve your private API hostname from its data center, and cutover is in six hours. How would you separate DNS from firewall issues?
dnsnetworkingapi
88
You shadowed a senior FDE on a deployment but personally built the customer adapter and ran cutover checks. How would you walk an interviewer through what you shipped without overstating ownership?
discoverydeploymentownership
89
Error rate rises from 0.2% to 8% ten minutes after a production cutover serving 15,000 users. When and how would you roll back?
rollback
90
A nightly SFTP batch contains corrupted rows halfway through, and the customer asks you to replay the entire 8 million-row file before noon. What would you do?
batch
91
You have a 30-day PoC for a logistics customer, but its legacy AS/400 cannot produce the required shipment events until day 20. How would you reset the plan?
92
The customer's data engineer misses three weekly mapping sessions, and a six-week deployment is now blocked on 18 undefined fields. What do you do?
sessionsdeploymentfundamentals
93
Halfway through a four-week pilot, the customer adds multilingual search, mobile support, and a second data source. How would you respond?
94
A customer asks for an untested Saturday cutover because its executive sponsor announced Monday launch to 8,000 users. What would you do?
sponsor
95
Your custom integration fails during a live demo to 20 customer stakeholders, and only 12 minutes remain. How do you recover?
stakeholder-managementcommunication
96
You misconfigured a field mapping and delayed a customer's user-acceptance test by one day. What would you tell them and change?
acceptance
97
A customer needs a custom audit-log export for a $400,000 renewal, but your product team says it will not build it this quarter. How would you proceed?
98
The customer security lead demands private connectivity, while the analytics lead wants a public API pilot live in ten days. How would you resolve the conflict?
api
99
After an adapter demo using 500 sanitized records, a skeptical customer head of data asks, 'What has this pilot actually proved?' How would you answer?
validation
100
You join a healthcare deployment and have five days to understand prior authorization well enough to map 40 fields and meet the customer director. How would you learn the domain?
joinsauthdeployment

Questions

1A customer sends a 2 GB CSV export with inconsistent dates, duplicate headers, and blank account IDs two days before a pilot. How would you inspect and clean it without hiding data problems from the customer?

2A customer's JSON API sometimes sends order_total as a string and sometimes omits currency. How would you use Pydantic so bad records do not quietly reach the deployment?

3Your Python adapter calls a customer inventory API that occasionally hangs for 90 seconds and returns 503 during nightly maintenance. What would you change before the first production run?

4A hospital can only provide a 40 GB newline-delimited JSON file through SFTP, and the deployment container has 2 GB of memory. How would you process it?

5A customer SDK is synchronous, but your shared deployment service uses asyncio to call ten tenant endpoints concurrently. How would you integrate the SDK without freezing every tenant job?

6During a 10,000-record customer import, record 4,201 fails because a required product code is missing. How should the Python job handle the exception and partial result?

7A customer support engineer asks for logs to debug failed employee syncs, but payloads contain names, emails, and access tokens. What would you log?

8A customer's payroll API has no sandbox, and a field rename could break your adapter on payday. What pytest coverage would you add before deployment?

9The same adapter must run for three customers with different endpoints and credentials. A teammate proposes copying the file and editing constants for each deployment. What would you do instead?

10Your Python adapter works on your laptop, but the customer requires it to run in their Kubernetes cluster with no root user and no internet access. How would you containerize it?

11On day one of a retail deployment, you receive read-only access to a warehouse with 300 unfamiliar tables and little documentation. How would you find the data needed for a stockout workflow?

12The customer wants to join CRM accounts to billing records, but CRM uses strings like ACCT-0042 while billing stores integer 42. How would you make the join reliable?

13A support director says the pilot should reduce average ticket resolution time by 20%. How would you write the SQL aggregation so the result supports that customer KPI?

14A daily customer export contains repeated order IDs with different updated_at values, and finance needs exactly one current row per order. How would you deduplicate it deterministically?

15In a customer's opportunities table, closed_at is NULL for open deals, but some old closed deals also have NULL because of a migration bug. How would you avoid misleading the sales forecast?

16A global customer stores event timestamps without offsets, while teams in New York and Berlin expect a daily dashboard in local time. What would you do before building it?

17A reconciliation query takes 18 minutes on the customer's production warehouse, and you are not allowed to create indexes yourself. How would you work with their DBA?

18Your deployment loads an invoice header and its line items into the customer's database. A line insert can fail after the header succeeds. How would you protect the customer's records?

19A customer wants hourly updates from a 200-million-row events table, but full reloads miss the one-hour SLA. How would you introduce an incremental watermark?

20After the first production load, the source has 1,002,481 customers but the target has 998,920. How would you reconcile the deployment before telling the customer it succeeded?

Locked questions