TL;DR: The artifacts live in S3 either way; a registry is the metadata and workflow layer on top that buys you lineage, atomic promotion semantics, an approval and audit trail, and a discovery and CI/CD hook point. S3 answers "where is the file"; the registry answers "which model is in production, how it got there, and what trained it."
How to approach it
A deceptively simple senior screen, especially at banks. The interviewer is baiting a specific weak answer, "S3 has versioning, so you don't need one", to see whether you understand that the registry's job is lifecycle and lineage, not storage. Concede the storage point immediately, then explain what sits on top.
A strong answer
Start with the concession that frames everything: the artifacts still live in S3 either way. A registry isn't an alternative storage system; it's a metadata and workflow layer pointing at that storage. The question is really "what does the pointer layer buy you," and the answer is four things:
Lineage. A registry version links to the training run that produced it: code commit, hyperparameters, dataset snapshot, evaluation metrics. An S3 key called fraud_model_v3_final_FIXED.pkl links to nothing. When production misbehaves, "which data trained the thing currently serving" must be a thirty-second lookup, not an archaeology project through Slack threads.
Promotion semantics. Production needs a notion of "the current model" that's separate from any artifact. With MLflow aliases, serving code loads models:/fraud@champion, and promotion or rollback is an alias reassignment, atomic, instant, no redeploy. With raw S3 you end up encoding state in path conventions (/prod/latest/), and now a copy command is your deployment mechanism, with no record of who ran it or why.
Approval workflow and audit. The registry records every transition: who promoted version 12, when, with what attached evaluation results. In regulated industries this isn't nice-to-have, model risk management frameworks like the Fed's SR 11-7 effectively require demonstrable controls on what enters production. "Can you show me the approval trail for the model that declined this loan?" is a real auditor question, and "we check CloudTrail for S3 writes" is a failing answer.
Discovery and a CI/CD hook point. One place listing every model, version, owner, and status, and an event source: registration triggers validation, approval triggers deployment. Without it, every team wires bespoke glue between training and serving.
The audit question is best rehearsed as the two conversations it produces, because the contrast is the answer. The auditor asks: "show me the approval trail for the model that declined this loan on March 12." The S3-convention team's honest reply is a reconstruction: "the prod path held fraud_v3_final_FIXED.pkl then... we think; here are the CloudTrail write events, here is a Slack thread where someone says 'shipping the new one,' and the person who ran the copy left in May." Every sentence is true and none of it is a control. The registry team's reply is a lookup: "version 12, promoted to @champion on March 3 by the model-risk approver, with these attached evaluation results, trained by this run on this data snapshot; here is the export." Same models, same S3 bytes underneath, ten minutes versus three weeks, and under a framework like SR 11-7 the first team has a finding while the second has a screenshot. That contrast, not any feature list, is why regulated shops treat the registry as mandatory.
The one-liner worth landing: S3 answers "where is the file"; a registry answers "which model is in production, how it got there, and what it was trained on." Companies that skip the registry rediscover this the first time two teams disagree about which artifact is actually live.
The registry never stores the bytes; it points at them and wraps them in the semantics S3 alone cannot express.
What interviewers probe next
- "S3 has object versioning, why isn't that enough?", versioning gives you immutable blobs, not semantics. It can't tell you which version is production, can't gate promotion on approval, and carries zero training metadata.
- "When is a registry overkill?", one model, one team, batch-only scoring: a documented S3 convention plus a config file is honestly fine for a while. Knowing the threshold (multiple models or your first compliance question, whichever comes first) reads better than registry absolutism.
- "Git LFS or DVC for models instead?", they version artifacts alongside code, which helps reproducibility, but still lack promotion states, approvals, and a serving-facing "current champion" pointer. Complementary, not substitutes.
- "How does the registry trigger deployment?", webhooks or event polling on alias change kick the CD pipeline; the registry is the contract boundary between training-side and serving-side automation.
Common mistakes
Saying the registry "stores models better", it doesn't store them at all, and this phrasing reveals the misunderstanding being screened for. Skipping audit and approval, which at JPMorgan or Capital One is most of the actual motivation. And dismissing the question as trivial: candidates who answer in one sentence miss that this is where the interviewer decides whether to ask you the governance follow-ups that distinguish senior hires.