Skip to content
Tech Interview Prep
Technical interview guide

Model Versioning & Registries

Tracking every trained model artifact alongside the data and code that produced it, so results are reproducible.

Read
45 min
Practice MCQs
25
Interview Q&A
25
Edition
v4
Editorial status
Reviewed
Relevant for
MLOps Engineer

Scope: MLflow, SageMaker AI, Vertex AI, Azure Machine Learning, and NIST AI RMF guidance current 2026-09-01.

Curated: · Written: · Reviewed:

25 questions

MCQ-4A serving job resolves the registry alias `champion` at start-up and writes the resolved version number into its deployment record. Overnight a release engineer reassigns `champion` from version 5 to version 7; the running job keeps serving version 5 until it restarts. Which statement about `champion` is correct?
MCQ-5A fraud-detector service loads its model as `models:/fraud-detector@champion` once at process start and keeps that version for the life of the process. It logs the alias name on every request but not the resolved version. Three replicas started at 08:00, 12:00 and 16:00 on Tuesday. The `champion` alias moved from v8 to v9 at 10:00 and from v9 to v10 at 16:00. An incident is traced to requests served between 13:00 and 15:00. Which versions served those requests?
MCQ-20A team moved the `production` alias on `fraud-scorer` from version 7 to version 8 at 14:02 and has not redeployed the serving endpoint. At 14:30 the endpoint's metadata reports version 7's artifact digest, and requests still return version 7's predictions. What does this finding indicate?
MCQ-1Registry `reg.internal` keeps one repository per team namespace. Team `risk` pushed `fraud-scorer:v3` on 3 March (digest `sha256:4b1e…`), re-pushed `fraud-scorer:v3` on 17 March after a retrain (digest `sha256:9d07…`), and moved the `production` alias to that re-push. Serving pods run under a service account in namespace `platform`, and unqualified names resolve against the caller's namespace. On 20 March a rollback to `fraud-scorer:v3` served the 17 March model, and its evaluation numbers did not match the 3 March run. Which deployment reference would have served the 3 March bytes?
MCQ-19A platform team copies the `fraud-detect` model from a staging registry to a production registry. The copy job reports success for every artifact. For this model: staging holds version 7 with artifact digest `sha256:4b1e`; the production registry now holds version 12 with the same digest `sha256:4b1e`; the production `latest` tag still resolves to version 11, digest `sha256:77aa`. The serving system loads whatever the `latest` tag resolves to. Which version will serve after this promotion?
MCQ-3A model registry for `fraud-scorer` lists five versions. v41 has a passing fairness evaluation and an approval record, and it currently serves 100% of production traffic. v42 was registered after v41 and has no evaluation run and no approval record. v43 was registered after v42, has a passing fairness evaluation, but no approval record. v44 was registered after v43, has a failed fairness evaluation, but an approval record. v45 was registered after v44, has no evaluation run, but an approval record. A new compliance rule requires that any version promoted to production must have both a passing fairness evaluation and an approval record. Which version satisfies the rule?
MCQ-17Your model registry stores immutable versions, and a version receives live traffic only once the `production` alias points at it. A promotion job moves `production` to any version whose status is `Approved`; a review service is the only writer of that status. Version 12 was registered at 09:00 with status `PendingManualApproval`. The promotion job ran at 09:05, exited 0, and logged `eligible versions: 0`. The registry API reports `production -> version 11`, and the serving layer resolves that same alias. The canary dashboard shows zero requests to version 12. What is the correct next step?
MCQ-24A registry's nightly governance job runs at 02:00 UTC and enforces approval expiry with these rules. If an approval's `expires_at` has passed and no revalidation ticket is open, the job moves the `champion` alias to the most recent version with a live approval and rejects promotion requests for the expired version. If an open revalidation ticket exists and was opened fewer than 14 days ago, the job leaves the alias alone and marks the version `expired-pending-review`. If the open ticket is 14 days old or older, the job treats it as stale and applies the same block as the no-ticket case. `fraud-scorer` v12 holds the `champion` alias. Its approval expired at 2025-03-01T00:00Z. v11 is the most recent version with a live approval. A revalidation ticket for v12 was opened at 2025-02-17T09:00Z and is still open. The job runs at 2025-03-05T02:00Z. What does the job do?
MCQ-11`set_alias(model, alias, version)` overwrites the alias's current target; the registry has no conditional form. A team guards promotions with a read-then-write check: the pipeline reads the alias, and if the read returns version 11 (the version it started from), it calls `set_alias`. The read and the write are separate calls, and nothing re-checks the alias at write time. Alias `champion` on model `fraud` points to version 11 at 02:00:00. Two pipelines run: - 02:00:02. Pipeline A reads `champion`: version 11. Check passes. - 02:00:03. Pipeline B reads `champion`: version 11. Check passes. - 02:00:04. Pipeline B calls `set_alias('fraud', 'champion', 13)`. - 02:00:05. Pipeline A calls `set_alias('fraud', 'champion', 12)`. Both jobs log success. Version 13 scored higher than version 12 on the holdout set. What does `champion` point to at 02:01?
MCQ-25A nightly batch scoring job splits 60 partitions evenly across 12 workers, 5 partitions each. Each worker resolves the `champion` alias once, when it picks up its first partition, and reuses that resolved version for all of its partitions. Workers 1–5 start at 02:00; workers 6–12 start at 02:10. At 02:05 an engineer promotes version 42 to `champion`, replacing version 41. How many partitions were scored with version 41?
MCQ-2A fraud-scoring service runs 40 replicas that each pull the model artifact from the registry by digest and cache it in memory for the life of the process. The registry treats every version as an immutable, self-contained inference bundle and returns HTTP 409 on any write to an existing version. Version 7 of `fraud-scorer` bundles the weights and the feature-preprocessing module. The weights are unchanged, but a preprocessing change shifts a categorical feature's bucket boundaries so about 3% of requests now map to different feature values. The team must ship this and be able to roll back by moving one alias. What must they do?
MCQ-6A registry stores one lineage record per model version. Version 12's record holds the training code commit, the URI and SHA-256 of the raw training snapshot archive, the hyperparameter file, and the container image digest. The feature pipeline that turns that raw snapshot into training features is a pinned library; it reads its column definitions and rolling-window sizes from a config service that is edited in place, and that config is not part of the lineage record. A year later the team re-runs training from the record: the snapshot checksum matches, the code commit checks out, and the image digest resolves, but validation AUC comes back at 0.79 against the recorded 0.86. What is the correct diagnosis?
MCQ-22A team ships `support-bot-v3`, a LoRA adapter, on a base model pulled by the mutable tag `mistral-7b-instruct`. The registry entry records the adapter's SHA-256, the base tag, and an 84% score on a frozen 500-item eval. Six months later, re-running the same eval with the same adapter file and unchanged serving code yields 71%; the adapter SHA matches, and the new outputs are fluent but formatted differently from the recorded run. Which addition to the registry entry would have made the cause identifiable?
MCQ-10Endpoint /score runs 12 pods that each loaded model `fraud:37` (digest sha256:4d1a...) once at startup on 2025-03-11 12:00. The pods have not restarted since, the model is never reloaded, all configuration is read at startup, and the endpoint returns the raw model output before any calibration; a downstream service applies a 0.80 threshold. The deployment record for the pod set stores the model digest and the image digest. Request R carries only a customer id, and each pod fetches that customer's features from an online store at request time. R returned 0.812 from all 12 pods at 23:00 on 2025-03-11 and 0.786 from all 12 pods at 08:00 on 2025-03-12; both calls sent byte-identical request bytes. Which mechanism accounts for the two readings?
MCQ-13Live fraud precision fell from 0.71 to 0.62 in the hour after last night's v4 deploy. v3 is still registered with its immutable artifact and scored 0.71 in an evaluation 14 months ago. The one upstream change since then: six months ago the client SDK stopped emitting device_age_days, so the field is absent from 38% of live requests, while only 0.8% of v3's training rows lacked it. Which expectation for a v3 rollback is best supported?
MCQ-15A fraud-scoring service loads a registered model whose signature declares `amount` as double and `channel` as string. A pipeline change makes the upstream job emit `amount` in cents instead of dollars. The signature check passes, the batch scores, and the mean predicted fraud probability falls from 0.31 to 0.04. Which change would have caught the unit change before the batch scored?
MCQ-7A fraud team registers model v3 on 2025-03-01, recording only the training table name prod.txns and the query text SELECT * FROM prod.txns WHERE event_time < '2025-03-01'; that run read 4.2 billion rows. prod.txns is append-only (rows are never updated or deleted in place), and a nightly job drops every daily partition whose date is older than 90 days. On 2025-09-15 an auditor re-runs the recorded query to reproduce the training set. What does the re-run return?
MCQ-12A fraud-scoring service runs model v7 behind a mutable alias. v8 ships alongside a store migration that replaces `txn_velocity_1h` with `txn_velocity_5m` in the online feature store. After v8's precision drops, the on-call repoints the alias to v7; within a minute the error rate climbs from 0.1% to 9%, with HTTP 500s reading `missing feature txn_velocity_1h`. Which change would have made the rollback restore correct serving?
MCQ-14A nightly job loads 40 partner-supplied `.pt` checkpoints with `torch.load(path)` on PyTorch 2.6. After the upgrade, 12 of them fail with `UnpicklingError: GLOBAL mylib.CustomLayer was not an allowed global`. A teammate proposes switching the call to `torch.load(path, weights_only=False)` so all 40 load again. What is the right call?
MCQ-9A model registry enforces two separate permissions: `registry.write` to upload a version, and `registry.promote` to move a version to the `production` alias. The training job runs under a service account holding both, and the CI bot that starts the job authenticates with that same account's key. In a postmortem, a poisoned dependency in the training image called the promote API itself 90 seconds into the run, so a version was serving production traffic before the pipeline's evaluation step had even started. The team wants to keep the pipeline fully automated, with a human approving each promotion. Which change stops a compromised training job from putting a version into production?
MCQ-18A platform team runs one registry with three repositories: app-dev, app-staging, app-prod. The CI service account ci-dev has push and pull on all three because permissions were granted at the registry level. A developer pipeline accidentally ran against app-prod and overwrote the v1.2.3 tag with a dev build; production deployed it. The team wants to stop this class of failure while keeping dev iteration fast. Which change best addresses the root cause?
MCQ-16A model registry holds 1,200 model versions; tags are free-form strings. A deployment controller polls every 30 seconds and deploys any version carrying the tag `stage=prod` to the production serving cluster. The `ml-platform` group (40 engineers) has push access to every repository, and the release bot's service account also has push. Last quarter, 9 versions reached production without review: the registry audit log attributes 6 of the tag writes to engineers in `ml-platform` and 3 to the release bot's service account, from CI jobs running on feature branches. Engineers still need push access to publish their own artifacts for dev and staging. Which change stops unreviewed versions from reaching production?
MCQ-8A team is deciding whether to promote ranking model v7 over the incumbent v6. Both are scored on the same frozen 50,000-example holdout. Overall AUC: v7 = 0.812, v6 = 0.804. On the 3,000-example new-accounts slice: v7 = 0.71, v6 = 0.79. The 95% CI on the overall AUC difference is ±0.011. The promotion gate requires no slice regression larger than 0.02. What should the team do?
MCQ-23An audit standard counts a production deployment as governed only when it references an immutable model version whose approval record is current. The registry expires approvals after 90 days. The registry holds 3,200 versions across 410 models, and the platform team reports 97% governance coverage, computed as the share of versions carrying a description and a lineage record. An auditor samples the 40 deployments currently serving production traffic. Of those, 24 are pinned to an immutable version; 5 of the 24 point to a version whose approval record is 120 days old. The other 16 resolve the mutable alias `champion` at request time; the alias currently points to an approved version for 10 of them and to a version with no approval record for 6. How many of the 40 deployments does the standard count as governed?
MCQ-21Your team runs a model registry. Today is 2024-06-15. The `fraud` model has five versions: | version | alias | other references | alias removed | | --- | --- | --- | --- | | v6 | — | none | 2023-11-01 | | v7 | — | nightly batch job `fraud-batch` pins `models:/fraud/7`; last run 2024-06-14 | 2023-12-15 | | v8 | — | none | 2024-05-01 | | v9 | — | named in open incident INC-4412, opened 2024-05-20 | 2024-03-01 | | v10 | challenger | A/B test running through 2024-06-30 | — | Artifacts are retained for 90 days after last use, and a version named in an open incident is held until the incident closes. Which version can be deleted today?