Design a feature store / fine-tuning data pipeline

Expected question

"Design a feature store and fine-tuning data pipeline for ML/LLM products. How do you version data, serve features online/offline, and feed training jobs reliably?"

Variant forms

Interviewers often ask the same design with different framing — recognize the archetype:

• "Design the data platform that feeds both RAG ingestion and periodic model fine-tunes."
• "How do you build training datasets from production logs without leaking PII?"
• "Design point-in-time correct features for ranking models used in production and in offline eval."
• "Our fine-tune jobs keep training on stale data — architect lineage from raw docs to model weights."
• "Design a pipeline for curated SFT examples with human review, versioning, and rollback."
• "How do you share embeddings/features across ten product teams without schema chaos?"
• "Design backfill and streaming feature updates for real-time personalization at scale."

Where this actually gets asked

This is the weakest-evidenced topic in this set for the six companies in scope — no company-specific attributed question was confirmed for OpenAI, Anthropic, Meta, Google, Microsoft, or Apple specifically. What exists publicly is mostly adjacent-company material (e.g., Scale AI's own interview guides describe RLHF/fine-tuning data pipeline stages) or generic MLOps system design content not pinned to a specific employer. Treat this as a well-established type of question in ML infrastructure interviews broadly — feature stores and training-data pipelines are a mature, widely-taught system design topic — rather than a confirmed big-tech-specific prompt. Answer it on its technical merits.

Executive summary

30-second thesis

I'd treat features as a versioned data product with point-in-time joins — train-serve skew and silent leakage are the bugs that make offline eval look like a lie.

2-minute answer

I'd start by locking entities, feature views, and the offline/online consistency contract. Point-in-time joins and label windows aren't optional if anything promotes to prod. Lineage IDs on every snapshot. Publish features separately from train/promote. Fine-tune data gets consent, PII redaction, and a holdout before it ever touches a job. When a model regresses, I'd slice freshness and train-serve skew before touching hyperparameters — that's usually where the scar is. Adjacent O proof for lineage discipline: enterprise_rag_platform ingest hashes — same idea, different store; not Lucid P.

Quantitative trade-offs

Numbers and thresholds in interview delivery should be labeled H unless the candidate can defend a measured baseline. Open repositories are O; researched public patterns are R. Do not upgrade O/R into employer P adoption.

ML fundamentals

ML fundamentals: leakage prevention, train-serve skew, delayed labels, feature lineage, and holdout/eval hygiene before any fine-tune promotion.

Migration and rollout

Feature stores lie at training/serving boundaries. Migrate the contract, not just the tables.

1. Freeze feature definitions + point-in-time join semantics in a registry before dual-writing.
2. Dual-compute training and online features for a holdout; diff distributions and null rates.
3. Only then cut online readers. Training jobs pin feature versions explicitly.
4. Drill: label join with as-of timestamps — if you can't prove no leakage, don't promote.
5. Rollback = previous feature view version. Kill switch = freeze publishers that break the contract.

Org ownership and operating model

• ML platform owns feature registry, point-in-time joins, and online/offline parity tests.
• Model owners own which features are allowed into a training set and the leakage review.
• Data producers own source freshness and schema change notices.
• SRE owns online store latency/availability.
• Governance owns PII classification on feature materialization.

Requirements

Functional

• Data scientists should be able to define a feature once and use it consistently for both model training (offline, batch) and real-time inference (online, low-latency).
• Support point-in-time correct joins — a training example must only see feature values that were actually available at the time the label was generated, not values computed later.
• Fine-tuning/training jobs need a versioned, reproducible snapshot of the data they were trained on.

Non-functional

• Online feature lookup: single-digit-millisecond latency at inference time.
• Offline feature computation: can tolerate minutes-to-hours latency, optimized for throughput over a huge historical dataset.
• Training-serving skew must be structurally prevented, not just tested for after the fact.

Core entities

• Feature definition: name, computation logic, data source, freshness SLA.
• Feature value: a computed value for one entity (user, document, session) at a point in time.
• Dataset snapshot: an immutable, versioned join of features + labels used for one training run.
• Training/fine-tuning job: references exactly one dataset snapshot, never a live, mutable query.

API / interface

Auth: service identity for writers; training jobs read via scoped dataset tokens.

POST /v1/feature-sets
{"name":"user_engagement_v3","entity_keys":["user_id"],"ttl_hours":24}
→ 201 {"feature_set_id":"fs_...","version":1}

POST /v1/feature-sets/{id}/materialize
{"mode":"offline","as_of":"2026-07-01T00:00:00Z"} → 202 {"job_id":"mat_..."}

GET /v1/online-features
{"feature_set_id":"fs_...","entities":[{"user_id":"u_1"}],"features":["click_7d"]}
→ 200 {"entities":[{"user_id":"u_1","values":{"click_7d":0.12},"event_time":"..."}]}

POST /v1/training-datasets
{"feature_set_id":"fs_...","label_source":"s3://labels/...","point_in_time":true,
 "train_window":{"start":"...","end":"..."}}
→ 202 {"dataset_id":"ds_...","lineage_id":"lin_..."}

GET /v1/training-datasets/{dataset_id}
→ {"status":"ready","uri":"s3://...","row_count":12000000,"lineage":{...}}

Staff+ callout: point-in-time correctness belongs in the training-dataset API contract, not a notebook convention.

Data Flow

Offline materialization feeds training datasets; online path serves low-latency feature vectors at inference.

Rendering architecture diagram…

High-level design

Maps to functional requirements from step 1 — the component architecture that makes the API and data flow real.

Rendering architecture diagram…

The core architectural decision is the dual-store pattern: the same feature computation logic writes to both an offline store (optimized for large historical range scans, used to build training datasets) and an online store (optimized for point lookups by entity ID, used at inference time). The alternative — compute features differently for training vs. serving — is the single most common root cause of training-serving skew in real ML systems, and a strong answer should name this explicitly as the failure mode being designed against.

Deep dives below target non-functional requirements (latency, scale, failure, cost, security).

Deep dive 1: point-in-time correctness

Hard constraint: a training row may only see feature values that existed at label time — joining "latest" values into historical labels is future leakage, and offline metrics will cheerfully lie about it.

The naive dataset-building approach joins the latest feature values to historical labels. This leaks future information into training — a feature that reflects "user's total purchases," computed today, joined against a label from six months ago, tells the model something it couldn't have known at prediction time. The correct join is as_of_time-aware: for each label, fetch the feature value as it existed at that label's timestamp, not the current value.

This is the deep-dive question a Staff+ candidate should raise unprompted, because it's the kind of bug that produces a model that looks great in offline eval and silently underperforms in production — exactly the failure mode a rigorous eval practice (see system-design/07) is supposed to catch, but only if the training data itself wasn't already contaminated.

Deep dive 2: data lineage and reproducibility

Hard constraint: if you can't say exactly which snapshot trained a checkpoint, you can't debug or comply — "we track lineage" as a slide is not enough.

A fine-tuning job that can't say exactly what data it was trained on is a compliance and debugging liability — if a fine-tuned model produces a bad output, you need to trace it back to what training data might be responsible, and re-run that exact snapshot to reproduce a bug.

Adjacent open system (O), not a full feature store: enterprise_rag_platform's ingestion pipeline stamps every indexed chunk with a real content hash and ingestion timestamp that survive every downstream transformation (ADR-0005/ADR-016) — the same "can you actually answer when was this data captured and has it changed since" discipline a training-data pipeline needs, applied to a retrieval index instead of a training set. The generalizable principle: lineage metadata that's computed but not enforced (or not propagated through every transformation) is not real lineage — a real audit found exactly this gap (three places in the codebase reconstructed the data object explicitly and silently dropped the lineage fields to their defaults) and it's the same class of bug that would let a training pipeline silently lose track of what a dataset snapshot actually contains.

Deep dive 3: freshness vs. compute cost trade-off

Hard constraint: you can't recompute every feature for every entity on every write without lighting money on fire — so name which features are event-driven vs batch, with an example.

Recomputing every feature for every entity on every write is wasteful; recomputing only on a schedule risks staleness. The workable default: event-driven incremental computation for features that change frequently (recent activity counts), scheduled batch recomputation for features that are expensive and change slowly (aggregate historical statistics) — naming which features belong in which bucket, with a concrete example, is what separates a Staff-level answer from "we'll just recompute everything hourly."

Deep dive 4: exact reproducibility as a bit-level contract, not a policy statement

Hard constraint: reproducing a training set means re-running the same transformation code on the same input bytes — hashing outputs alone misses silent library upgrades.

"We track lineage" is a Staff-level claim; a Principal-level answer specifies what reproducing a training set actually requires: a content-hash of the feature computation logic itself (not just its output) alongside the input snapshot's own hash, so that "rerun this exact training set six months later" means re-executing the identical transformation code against the identical input data, byte-for-byte — not just "the same feature names with roughly the same values." Without hashing the transformation logic, a silent library upgrade or a refactored feature-computation function can change output values without changing anything the lineage system was tracking, breaking reproducibility invisibly. This matters concretely for compliance and incident response: if a fine-tuned model is later found to have learned something harmful from its training data, "what exact data and exact transformation produced this checkpoint" needs to be an answerable, bit-exact question, not an approximate one.

The backfill cost this implies is also a real, numbers-driven decision: recomputing a feature over N months of historical data at a non-trivial per-entity compute cost (e.g., an aggregate statistic requiring a full scan of historical events) can be a multi-hour-to-multi-day batch job at real data volumes — a Principal-level answer sizes this cost before committing to "we'll just backfill it," and considers whether a cheaper, approximate backfill (computed only for entities actually needed by the current training run, not the full historical population) is acceptable given the reproducibility contract above.

Deep dive 5: schema evolution, dual-write sync, and RTBF (interview-critical, keep short)

Schema / backfill: never mutate a feature definition in place for breaking changes — bump a SemVer (user_engagement@2.0.0) and materialize into a new offline partition. Legacy training snapshots keep pointing at v1; new jobs opt into v2. That isolates backfill blast radius without rewriting history.

Streaming dual-write: one computation path fans out to online + offline. Online writes must be idempotent (entity_id + event_time); offline commits via checkpointed batches. On crash, replay the log — Redis/Dynamo overwrites safely; uncommitted lake files are abandoned. Do not invent a full XA/2PC story unless asked — idempotent online + checkpointed offline is the Staff+ answer.

GDPR vs immutability: "immutable snapshot" conflicts with right-to-be-forgotten. Prefer pseudonymization at ingest, or rebuild a scrubbed snapshot with a new content hash — never silently rewrite a published lineage id. Say this in one breath when compliance comes up.

Staff+/Principal signal rubric

• Mid-level: proposes a single feature table used for both training and serving; may not spontaneously identify training-serving skew as a risk.
• Senior: proposes separate online/offline stores; identifies training-serving skew as a named risk to design against.
• Staff+: insists on point-in-time-correct joins unprompted, and can explain precisely why naive "join latest value" leaks future information.
• Principal: additionally specifies reproducibility as a content-hash contract over both the transformation logic and the input snapshot (not just "we track lineage" as a policy statement), and sizes the real compute cost of a historical backfill before committing to it — treating exact reproducibility as an incident-response and compliance requirement, not a nice-to-have.

Follow-up questions to expect

• "How do you backfill a new feature into historical training data?" (Answer: this requires the feature's computation logic to be replayable against historical raw data, which is why raw source retention — not just computed feature retention — matters.)
• "How do you detect feature drift between training and production?" (Answer: monitor the statistical distribution of online-served feature values against the distribution in the training snapshot, alert on divergence.)
• "What's different about this for a fine-tuning pipeline specifically, versus a classic tabular ML feature store?" (Answer: the "features" become prompt/completion pairs or preference comparisons, but point-in-time correctness and lineage requirements are identical in spirit — you still need to know exactly what data a given fine-tuned checkpoint saw.)
• "How do prompt templates / tokenization fit a fine-tuning feature store?" (Answer: version the prompt template as a first-class artifact; prefer tokenizing in the training job unless you need bit-exact token arrays in the store — changing a tokenizer is a major feature version.)
• "In 45 minutes, what do you refuse to design?" (Answer: a full Feast/Tecton product catalog. Anchor on dual-store + PIT join + immutable snapshot + one of: schema versioning, dual-write sync, or RTBF.)

Related

• enterprise_rag_platform ADR-0005/ADR-016: Ingestion data contract + lineage
• system-design/08: Fine-tuning/RLHF training pipeline at scale