Recommendation System — System Design Academy

Complete System Design Tutorial

The design separates the latency-sensitive online path, the correctness-critical durable transition, and asynchronous projection or repair work. Each chapter below explains one decision shown in the diagrams.

1. Requirements and scope

Functional requirements

Execute the primary recommendation system workflow and return an observable result.
Expose the concrete operations shown below with stable retry identities.
Allow operators to inspect, disable, repair, or replay ambiguous state.
Keep derived analytics and notifications off the correctness-critical path.

Non-functional requirements

Design for 1B events/day · 10K candidates/request · p99 < 200 ms.
Protect the named business invariant under concurrency and partial failure.
Scale by an explicit aggregate/partition key without one global mutex.
Degrade intentionally when cache, queue, store, or regional ownership is unavailable.

Primary invariant: each accepted recommendation system operation has one stable result and retries do not repeat its business effect.

↑ top

2. Capacity assumptions

State assumptions before choosing technology. The supplied workload is 1B events/day · 10K candidates/request · p99 < 200 ms. Convert it into peak throughput, retained records, bandwidth, and failure budget; then connect each number to a component.

Dimension	Working estimate	Design implication
Traffic	1B events/day · 10K candidates/request · p99 < 200 ms	Separate online compute, durable state, and asynchronous consumers.
Peak factor	5–10× average	Admission control, autoscaling headroom, and hot-key protection.
Stored state	write rate × retention × record bytes	Partition count, indexes, compaction, backup, and archival tier.
Network	peak operations × payload size	Batching, compression, edge placement, and egress budget.
Availability	99.9–99.99%	Multi-zone replicas, tested failover, and dependency error budgets.
Recovery	RPO/RTO by invariant	Ack boundary, replay retention, reconciliation, and operator runbooks.

Do not cargo-cult the numbers: show the arithmetic, name the uncertain variable, and say which production metric will replace the interview estimate.

↑ top

3. API and data model

GET /v1/recommendations?surface=&cursor=

POST /v1/eventsimpression/click/conversion

PUT /internal/models/{version}signed promotion

GET /internal/features/{entity}point-in-time vector

Authoritative records and indexes

feature(entity_id, feature_name, event_time, value, version)

model(model_id, version, artifact_uri, feature_schema, state)

exposure(request_id, user_id_hash, candidates, scores, model_version)

embedding(entity_id, vector_version, shard, updated_at)

Command and event envelopes

# Command envelope
operation: GET /v1/recommendations?surface=&cursor=
request_id: req_01J...
idempotency_key: idem_7f2...
aggregate_key: recommendation-system_123
expected_version: 17
payload: {...}

# Durable event
event_id: evt_01J...
aggregate_key: recommendation-system_123
event_type: recommendation_system_committed
schema_version: 3
trace_id: tr_8ac...

The authoritative aggregate and idempotency record share a transactional or conditional-write boundary. Event streams, search indexes, caches, analytics, and dashboards are rebuildable projections.

↑ top

4. Why this architecture and what it costs

The boxes are not a shopping list. Each boundary exists because it owns a different invariant, scaling dimension, data lifecycle, security policy, or failure domain. The table makes those choices explicit and shows when the design should remain simpler.

Component	Owned responsibility	Why separate it	Rejected alternative	Cost and evolution trigger
User Request	Owns one bounded domain decision, transition, or operational capability. In this design it participates in: Build request context and fetch online features with strict per-source deadlines.	Separates a distinct scale, ownership, policy, or failure boundary. It prevents the external caller from also owning the contract and failure behavior of Recommendation API.	Keep this behavior inside an adjacent service. Prefer that simpler option until independent scale, ownership, policy, or failure isolation is measurable. Here that simpler choice would couple “Feature store returns stale vector” to the external caller.	Costs an interface, deployment, telemetry, ownership, and failure mode. Keep it in-process first; extract when independent scale or reliability pays for the boundary. Measure the latency, saturation, lag, or mismatch at the User Request boundary before extracting or scaling it.
Recommendation API	Owns authentication, validation, quotas, deadlines, and routing. In this design it participates in: Retrieve bounded candidates in parallel from collaborative, vector, and trending sources.	Separates a distinct scale, ownership, policy, or failure boundary. It prevents User Request from also owning the contract and failure behavior of Context Builder.	Put protocol, authentication, quotas, and business transitions in one process. Acceptable at low scale, but it couples external traffic risk to domain releases. Here that simpler choice would couple “Candidate source exceeds deadline” to User Request.	Costs another hop, policy distribution, and a high-blast-radius fleet. Introduce it when many clients or services need one protocol and protection boundary. Measure the latency, saturation, lag, or mismatch at the Recommendation API boundary before extracting or scaling it.
Context Builder	Owns one bounded domain decision, transition, or operational capability. In this design it participates in: Union/dedupe, apply availability/safety policy, then score with a pinned model.	Separates a distinct scale, ownership, policy, or failure boundary. It prevents Recommendation API from also owning the contract and failure behavior of Online Feature Store.	Keep this behavior inside an adjacent service. Prefer that simpler option until independent scale, ownership, policy, or failure isolation is measurable. Here that simpler choice would couple “Model artifact and feature schema mismatch” to Recommendation API.	Costs an interface, deployment, telemetry, ownership, and failure mode. Keep it in-process first; extract when independent scale or reliability pays for the boundary. Measure the latency, saturation, lag, or mismatch at the Context Builder boundary before extracting or scaling it.
Online Feature Store	Owns authoritative versioned state or immutable evidence. In this design it participates in: Rerank for diversity, freshness, exploration, and business constraints.	Defines the source of truth and commit boundary. It prevents Context Builder from also owning the contract and failure behavior of Candidate Router.	Let each service persist its own interpretation. That removes one clear commit boundary and makes retries, support, and reconciliation ambiguous. Here that simpler choice would couple “Feedback loop narrows diversity” to Context Builder.	Costs schema migration, replication, backup, and shard ownership. Split storage only when access patterns, retention, or compliance boundaries materially differ. Measure the latency, saturation, lag, or mismatch at the Online Feature Store boundary before extracting or scaling it.
Candidate Router	Owns one bounded domain decision, transition, or operational capability. In this design it participates in: Return results and log candidates/scores/exposures so training and evaluation are reproducible.	Separates a distinct scale, ownership, policy, or failure boundary. It prevents Online Feature Store from also owning the contract and failure behavior of Collaborative Index.	Keep this behavior inside an adjacent service. Prefer that simpler option until independent scale, ownership, policy, or failure isolation is measurable. Here that simpler choice would couple “Exposure logs drop during traffic spike” to Online Feature Store.	Costs an interface, deployment, telemetry, ownership, and failure mode. Keep it in-process first; extract when independent scale or reliability pays for the boundary. Measure the latency, saturation, lag, or mismatch at the Candidate Router boundary before extracting or scaling it.
Collaborative Index	Serves a low-latency, disposable projection with authoritative fallback. In this design it participates in: Build request context and fetch online features with strict per-source deadlines.	Separates a distinct scale, ownership, policy, or failure boundary. It prevents Candidate Router from also owning the contract and failure behavior of Content / Vector Index.	Read authoritative storage for every request. Simpler initially, but latency and source-of-truth load scale together and a hot key becomes a database incident. Here that simpler choice would couple “Feature store returns stale vector” to Candidate Router.	Costs invalidation, staleness, stampede protection, and memory. Add or split it only after the workload (1B events/day · 10K candidates/request · p99 < 200 ms) proves repeated reads dominate. Measure the latency, saturation, lag, or mismatch at the Collaborative Index boundary before extracting or scaling it.
Content / Vector Index	Serves a low-latency, disposable projection with authoritative fallback. In this design it participates in: Retrieve bounded candidates in parallel from collaborative, vector, and trending sources.	Separates a distinct scale, ownership, policy, or failure boundary. It prevents Collaborative Index from also owning the contract and failure behavior of Trending Store.	Read authoritative storage for every request. Simpler initially, but latency and source-of-truth load scale together and a hot key becomes a database incident. Here that simpler choice would couple “Candidate source exceeds deadline” to Collaborative Index.	Costs invalidation, staleness, stampede protection, and memory. Add or split it only after the workload (1B events/day · 10K candidates/request · p99 < 200 ms) proves repeated reads dominate. Measure the latency, saturation, lag, or mismatch at the Content / Vector Index boundary before extracting or scaling it.
Trending Store	Owns authoritative versioned state or immutable evidence. In this design it participates in: Union/dedupe, apply availability/safety policy, then score with a pinned model.	Defines the source of truth and commit boundary. It prevents Content / Vector Index from also owning the contract and failure behavior of Candidate Union.	Let each service persist its own interpretation. That removes one clear commit boundary and makes retries, support, and reconciliation ambiguous. Here that simpler choice would couple “Model artifact and feature schema mismatch” to Content / Vector Index.	Costs schema migration, replication, backup, and shard ownership. Split storage only when access patterns, retention, or compliance boundaries materially differ. Measure the latency, saturation, lag, or mismatch at the Trending Store boundary before extracting or scaling it.
Candidate Union	Owns one bounded domain decision, transition, or operational capability. In this design it participates in: Rerank for diversity, freshness, exploration, and business constraints.	Separates a distinct scale, ownership, policy, or failure boundary. It prevents Trending Store from also owning the contract and failure behavior of Policy / Safety.	Keep this behavior inside an adjacent service. Prefer that simpler option until independent scale, ownership, policy, or failure isolation is measurable. Here that simpler choice would couple “Feedback loop narrows diversity” to Trending Store.	Costs an interface, deployment, telemetry, ownership, and failure mode. Keep it in-process first; extract when independent scale or reliability pays for the boundary. Measure the latency, saturation, lag, or mismatch at the Candidate Union boundary before extracting or scaling it.
Policy / Safety	Owns one bounded domain decision, transition, or operational capability. In this design it participates in: Return results and log candidates/scores/exposures so training and evaluation are reproducible.	Separates a distinct scale, ownership, policy, or failure boundary. It prevents Candidate Union from also owning the contract and failure behavior of Online Ranker.	Embed policy in every caller. Rules then drift by deployment version and emergency changes require fleet-wide redeployment. Here that simpler choice would couple “Exposure logs drop during traffic spike” to Candidate Union.	Costs an interface, deployment, telemetry, ownership, and failure mode. Keep it in-process first; extract when independent scale or reliability pays for the boundary. Measure the latency, saturation, lag, or mismatch at the Policy / Safety boundary before extracting or scaling it.
Online Ranker	Owns one bounded domain decision, transition, or operational capability. In this design it participates in: Build request context and fetch online features with strict per-source deadlines.	Separates a distinct scale, ownership, policy, or failure boundary. It prevents Policy / Safety from also owning the contract and failure behavior of Diversity Reranker.	Keep this behavior inside an adjacent service. Prefer that simpler option until independent scale, ownership, policy, or failure isolation is measurable. Here that simpler choice would couple “Feature store returns stale vector” to Policy / Safety.	Costs an interface, deployment, telemetry, ownership, and failure mode. Keep it in-process first; extract when independent scale or reliability pays for the boundary. Measure the latency, saturation, lag, or mismatch at the Online Ranker boundary before extracting or scaling it.
Diversity Reranker	Owns one bounded domain decision, transition, or operational capability. In this design it participates in: Retrieve bounded candidates in parallel from collaborative, vector, and trending sources.	Separates a distinct scale, ownership, policy, or failure boundary. It prevents Online Ranker from also owning the contract and failure behavior of Model Registry.	Keep this behavior inside an adjacent service. Prefer that simpler option until independent scale, ownership, policy, or failure isolation is measurable. Here that simpler choice would couple “Candidate source exceeds deadline” to Online Ranker.	Costs an interface, deployment, telemetry, ownership, and failure mode. Keep it in-process first; extract when independent scale or reliability pays for the boundary. Measure the latency, saturation, lag, or mismatch at the Diversity Reranker boundary before extracting or scaling it.
Model Registry	Owns one bounded domain decision, transition, or operational capability. In this design it participates in: Union/dedupe, apply availability/safety policy, then score with a pinned model.	Separates a distinct scale, ownership, policy, or failure boundary. It prevents Diversity Reranker from also owning the contract and failure behavior of Exposure Log.	Keep this behavior inside an adjacent service. Prefer that simpler option until independent scale, ownership, policy, or failure isolation is measurable. Here that simpler choice would couple “Model artifact and feature schema mismatch” to Diversity Reranker.	Costs an interface, deployment, telemetry, ownership, and failure mode. Keep it in-process first; extract when independent scale or reliability pays for the boundary. Measure the latency, saturation, lag, or mismatch at the Model Registry boundary before extracting or scaling it.
Exposure Log	Preserves durable ordering, replay, and consumer checkpoints. In this design it participates in: Rerank for diversity, freshness, exploration, and business constraints.	Separates a distinct scale, ownership, policy, or failure boundary. It prevents Model Registry from also owning the contract and failure behavior of Offline Lake / Training.	Call every downstream consumer synchronously. That couples user latency and availability to analytics, notification, indexing, and audit dependencies. Here that simpler choice would couple “Feedback loop narrows diversity” to Model Registry.	Costs eventual consistency, schema evolution, replay tooling, idempotent consumers, and lag operations. Earn it when bursts or downstream failure must not block acknowledgment. Measure the latency, saturation, lag, or mismatch at the Exposure Log boundary before extracting or scaling it.
Offline Lake / Training	Owns one bounded domain decision, transition, or operational capability. In this design it participates in: Return results and log candidates/scores/exposures so training and evaluation are reproducible.	Separates a distinct scale, ownership, policy, or failure boundary. It prevents Exposure Log from also owning the contract and failure behavior of operators and repair tooling.	Keep this behavior inside an adjacent service. Prefer that simpler option until independent scale, ownership, policy, or failure isolation is measurable. Here that simpler choice would couple “Exposure logs drop during traffic spike” to Exposure Log.	Costs an interface, deployment, telemetry, ownership, and failure mode. Keep it in-process first; extract when independent scale or reliability pays for the boundary. Measure the latency, saturation, lag, or mismatch at the Offline Lake / Training boundary before extracting or scaling it.

Decision 01

Chosen boundary vs simpler monolith

Separate candidate generation from ranking: broad retrieval first, expensive scoring second.

A single deployable with one transactional store is the right starting point for a small recommendation system workload. The proposed boundary becomes justified when separate candidate generation from ranking: broad retrieval first, expensive scoring second. Separating it contains the named failure “Feature store returns stale vector” without placing unrelated work on the critical path.

Decision 02

Correctness vs availability

Use point-in-time-correct features to prevent training/serving skew and leakage.

The design coordinates only where each accepted recommendation system operation has one stable result and retries do not repeat its business effect. Everywhere else it accepts bounded staleness or asynchronous propagation. The alternative—strong coordination for every projection—simplifies the mental model but increases tail latency and turns partitions into broad write unavailability.

Decision 03

Operational cost and evolution

Blend freshness, diversity, safety, and exploration with model score; log exposures for unbiased learning.

This choice adds dashboards, ownership, schema compatibility, capacity planning, and recovery runbooks. Start with the minimum topology that supports “Union/dedupe, apply availability/safety policy, then score with a pinned model.” and extract another service only when measured throughput, blast radius, data lifecycle, or team ownership changes independently.

Why this design: the synchronous path contains only the decisions required to protect each accepted recommendation system operation has one stable result and retries do not repeat its business effect. Durable projections, notifications, indexing, analytics, and repair are isolated so their load or failure cannot corrupt the acknowledged operation.

Correctness boundary

Separate candidate generation from ranking: broad retrieval first, expensive scoring second.

Scale boundary

Use point-in-time-correct features to prevent training/serving skew and leakage.

Operational boundary

Blend freshness, diversity, safety, and exploration with model score; log exposures for unbiased learning.

↑ top

5. Critical flow and algorithm

Build request context and fetch online features with strict per-source deadlines.

Retrieve bounded candidates in parallel from collaborative, vector, and trending sources.

Union/dedupe, apply availability/safety policy, then score with a pinned model.

Rerank for diversity, freshness, exploration, and business constraints.

Return results and log candidates/scores/exposures so training and evaluation are reproducible.

Retry-safe mutation skeleton

identity = authenticate(request)
prior = idempotency.get(identity, request.key)
if prior: return prior.result

current = authoritative_store.read(partition_key)
next_state, events = domain.transition(current, request)
authoritative_store.compare_and_set(current.version, next_state)
idempotency.commit(request.key, next_state.result)
outbox.append(events)
return next_state.result

Exactly-once business effect: the network can duplicate requests. Stable identities, conditional state transitions, immutable event IDs, and reconciliation make the user-visible effect happen once.

↑ top

6. Caching, hot keys, and backpressure

Cache policy

Cache immutable or slowly changing reads with versioned keys, TTL jitter, short negative entries, and request coalescing.

Hot partition or tenant

Measure per-key and per-tenant skew. Replicate safe reads, isolate the offender, salt only when ordering permits, and enforce budgets.

Backpressure

Bound queues, buffers, connection pools, and concurrency. Shed low-priority work and scale from lag/oldest-item age rather than CPU alone.

Stampede and retry storms

Use single-flight/coalescing, jittered expiry, bounded retries, exponential backoff, and circuit breakers so one dependency failure does not multiply traffic.

Cache outage rule: never allow every miss to fall through simultaneously. Use local protection, request coalescing, circuit breakers, bounded pools, and controlled fallback.

↑ top

7. Partitioning and consistency

partition = stable_hash(primary_or_aggregate_key) → shard owner
write = conditional_update(expected_version)
read_projection = versioned and rebuildable

Strong boundary

each accepted recommendation system operation has one stable result and retries do not repeat its business effect. Use a uniqueness constraint, serializable/conditional write, expected version, or fencing token at this boundary.

Eventual boundary

Analytics, notifications, search, read models, and operational dashboards may lag when freshness is observable and replay exists.

Operation	Consistency	Reason
Authoritative mutation	Conditional / serializable	Protect uniqueness, ownership, money, inventory, order, or legal state transition.
Read after own write	Primary, session token, or cache warm	The caller must observe the result it just committed.
Ordinary read projection	Bounded stale where safe	Replicas and caches improve latency when version/freshness is visible.
Analytics and operations	Eventual + replay	Freshness is less important than isolating the online path.

Route every mutation for one aggregate to one owner/partition epoch.
Use cursors rather than offsets for changing ordered collections.
Retain tombstones until replication and repair windows close.
Move shard ownership through versioned maps and dual-read verification.

↑ top

8. Failure drills and recovery

Failure	Runtime behavior	Recovery proof
Feature store returns stale vector	Stop unsafe progress, serve only explicitly safe degraded behavior, and recover from durable state.	Reconciliation returns zero invariant gaps; lag/replica health and user-visible SLO return to baseline.
Candidate source exceeds deadline	Stop unsafe progress, serve only explicitly safe degraded behavior, and recover from durable state.	Reconciliation returns zero invariant gaps; lag/replica health and user-visible SLO return to baseline.
Model artifact and feature schema mismatch	Stop unsafe progress, serve only explicitly safe degraded behavior, and recover from durable state.	Reconciliation returns zero invariant gaps; lag/replica health and user-visible SLO return to baseline.
Feedback loop narrows diversity	Stop unsafe progress, serve only explicitly safe degraded behavior, and recover from durable state.	Reconciliation returns zero invariant gaps; lag/replica health and user-visible SLO return to baseline.
Exposure logs drop during traffic spike	Isolate the key/tenant/priority, apply admission control, and preserve durable work.	Reconciliation returns zero invariant gaps; lag/replica health and user-visible SLO return to baseline.

Recovery checkpoint

# Recovery checkpoint
scenario: Feature store returns stale vector
operation_id: op_01J...
aggregate_key: recommendation-system_123
last_durable_version: 18
projection_checkpoint: partition-07@offset-482901
retry_policy: same_identity / bounded_backoff
reconciliation_status: pending
owner: Offline Lake / Training

Every dependency needs a timeout, retry budget, circuit policy, and explicit fail-open, fail-closed, stale, or read-only mode. “Highly available” is incomplete until degraded behavior is named.

↑ top

9. Security, regions, and observability

Security and abuse

Authenticate at ingress and authorize the exact resource/tenant on every operation.
Encrypt transport and durable data; tokenize or isolate sensitive material.
Rate-limit by tenant, actor, resource, and network; detect enumeration and anomalous fan-out.
Redact secrets and personal data before logs or analytical storage.
Audit privileged mutations, support access, policy changes, and forced recovery.

Signals that prove correctness

User-visible success rate and p50/p95/p99 latency.
Timeouts, circuit state, saturation, and replica health by dependency.
Queue lag, oldest-item age, retries, and dead-letter growth.
Conflict, duplicate-suppression, and reconciliation mismatch counts.
Business drill: Feature store returns stale vector — alert before the user-visible invariant is violated.

Regional evolution

Begin multi-zone with tested failover, backups, and reconciliation.
Add global stateless ingress and regional read projections.
Keep one write owner per aggregate or partition where coordination is expensive.
Move ownership with epochs/fencing; avoid silent simultaneous writers.
Use expand/contract schemas, canaries, shadow reads, and reversible cutovers.

↑ top

10. Interviewer follow-ups

How do you solve cold start?

Separate candidate generation from ranking: broad retrieval first, expensive scoring second. Walk the concrete lifecycle step “Build request context and fetch online features with strict per-source deadlines.” Then test it against “Feature store returns stale vector”. State the retry identity, durable checkpoint, recovery signal, and cost of the alternative.

How do you evaluate offline and online?

Use point-in-time-correct features to prevent training/serving skew and leakage. Walk the concrete lifecycle step “Retrieve bounded candidates in parallel from collaborative, vector, and trending sources.” Then test it against “Candidate source exceeds deadline”. State the retry identity, durable checkpoint, recovery signal, and cost of the alternative.

How do embeddings update?

Blend freshness, diversity, safety, and exploration with model score; log exposures for unbiased learning. Walk the concrete lifecycle step “Union/dedupe, apply availability/safety policy, then score with a pinned model.” Then test it against “Model artifact and feature schema mismatch”. State the retry identity, durable checkpoint, recovery signal, and cost of the alternative.

How do you add diversity?

Separate candidate generation from ranking: broad retrieval first, expensive scoring second. Walk the concrete lifecycle step “Rerank for diversity, freshness, exploration, and business constraints.” Then test it against “Feedback loop narrows diversity”. State the retry identity, durable checkpoint, recovery signal, and cost of the alternative.

How do you prevent feedback loops?

Use point-in-time-correct features to prevent training/serving skew and leakage. Walk the concrete lifecycle step “Return results and log candidates/scores/exposures so training and evaluation are reproducible.” Then test it against “Exposure logs drop during traffic spike”. State the retry identity, durable checkpoint, recovery signal, and cost of the alternative.

Answer pattern: state the invariant, choose a mechanism, walk one failure, name the recovery signal, and explain the cost of the alternative.

↑ top

11. How to present this in 45 minutes

Time	What to cover
0–5 min	Clarify scope, workload, SLOs, and the correctness invariant.
5–10 min	Estimate throughput/storage and define concrete APIs and records.
10–20 min	Draw online, durable-write, and asynchronous/repair paths.
20–30 min	Explain the exact lifecycle, partition key, and concurrency control.
30–38 min	Discuss caching, hot keys, replication, and backpressure.
38–43 min	Run failure drills, security, observability, and regional evolution.
43–45 min	Summarize the five most important choices and trade-offs.

Final decision summary

Invariant	each accepted recommendation system operation has one stable result and retries do not repeat its business effect.
Partitioning	Stable aggregate/key hash with versioned ownership and measured hot-key isolation.
Commit	Expected version or uniqueness at the authoritative state boundary.
Async work	Durable outbox/log, idempotent consumers, checkpoints, and replay.
Recovery	Same operation identity, durable truth lookup, and business reconciliation.

Closing summary

CLOSING “The design protects each accepted recommendation system operation has one stable result and retries do not repeat its business effect. The online path stays bounded, authoritative state is versioned and partitioned, and asynchronous work is replayable and observable.”

↑ top

User Request

Recommendation API

Context Builder

Online Feature Store

Complete System Design Tutorial

1. Requirements and scope

Functional requirements

Non-functional requirements

2. Capacity assumptions

3. API and data model

Authoritative records and indexes

Command and event envelopes

4. Why this architecture and what it costs

Chosen boundary vs simpler monolith

Correctness vs availability

Operational cost and evolution

5. Critical flow and algorithm

Retry-safe mutation skeleton

6. Caching, hot keys, and backpressure

7. Partitioning and consistency

Strong boundary

Eventual boundary

8. Failure drills and recovery

Recovery checkpoint

9. Security, regions, and observability

Security and abuse

Signals that prove correctness

Regional evolution

10. Interviewer follow-ups

11. How to present this in 45 minutes

Final decision summary

Closing summary