LinkedIn#

Q: How is LinkedIn's connection graph different from Twitter's?

LinkedIn connections are symmetric: both sides must accept. Follows can be asymmetric. The graph also exposes 1st, 2nd, and 3rd degree distance to power search and PYMK.

Q: How does People You May Know work?

PYMK is a recommendation pipeline that mines shared connections, schools, employers, and behavioral signals. Triangle closing on the graph is the dominant signal.

Q: How does LinkedIn job search scale?

Elasticsearch clusters index jobs by title, skills, location, and seniority. Queries fan out across shards and rank with a learned model that uses user profile features.

Q: How is LinkedIn's feed different from Facebook's?

It blends posts, articles, jobs, and recommendations with stronger emphasis on professional intent. Ranking favors employer pages, mutual connections, and dwell time.

Q: How is privacy enforced in LinkedIn search?

Profile visibility settings are checked at query time. Anonymous viewers, blocked users, and 3rd-degree limits are applied as filters before ranking returns results.

Problem statement (interviewer prompt)

Design LinkedIn: professional profiles, an asymmetric connection graph (1st/2nd/3rd degree), feed of posts + articles, people + job search, recommendations (PYMK, jobs you may like), and messaging. Handle 1B users with strong privacy controls.

flowchart LR
  U([User])
  P[Profile Service]
  G[Connection Graph]
  F[Feed]
  S[People Search]
  J[Jobs Service]
  MSG[Messaging]
  U --> P
  U --> G
  U --> F
  U --> S
  U --> J
  U --> MSG

    classDef client fill:#dbeafe,stroke:#1e40af,stroke-width:1px,color:#0f172a;
    classDef edge fill:#cffafe,stroke:#0e7490,stroke-width:1px,color:#0f172a;
    classDef service fill:#fef3c7,stroke:#92400e,stroke-width:1px,color:#0f172a;
    classDef datastore fill:#fee2e2,stroke:#991b1b,stroke-width:1px,color:#0f172a;
    classDef cache fill:#fed7aa,stroke:#9a3412,stroke-width:1px,color:#0f172a;
    classDef queue fill:#ede9fe,stroke:#5b21b6,stroke-width:1px,color:#0f172a;
    classDef compute fill:#d1fae5,stroke:#065f46,stroke-width:1px,color:#0f172a;
    classDef storage fill:#e5e7eb,stroke:#374151,stroke-width:1px,color:#0f172a;
    classDef external fill:#fce7f3,stroke:#9d174d,stroke-width:1px,color:#0f172a;
    classDef obs fill:#f3e8ff,stroke:#6b21a8,stroke-width:1px,color:#0f172a;
    class U client;
    class P,G,F,S,J,MSG service;

flowchart TB
  subgraph Clients
    Web
    Mobile
  end

  subgraph Edge
    CDN
    LB
    GW[REST + GraphQL]
  end

  subgraph Profile
    PSVC[Profile Service]
    PDB[(Profiles - Espresso<br/>document store)]
    SKILL[Skills / Endorsements]
  end

  subgraph Graph[Economic Graph]
    CONN[Connection Service]
    GDB[(Distributed Graph<br/>LiquidGraph)]
    DEG[Degree / shortest path]
  end

  subgraph Search
    PEO[People Search]
    JOB[Jobs Search]
    GAL[Galene / Lucene search]
    QU[Query understanding]
  end

  subgraph Feed
    HOMEFEED[Home Feed]
    AGG([Aggregator])
    RANK([ML Ranker])
    CANDS([Candidate Gen<br/>followed + suggested])
  end

  subgraph Jobs
    JOBS[Jobs Service]
    JDB[(Jobs DB)]
    REC([Job Recommendations])
    ALERT[Job Alerts]
    APPLY[Easy Apply / ATS]
  end

  subgraph Msg
    INMAIL[InMail]
    DM[Messaging Service]
    WS[WebSocket gateway]
  end

  subgraph Async
    KAFKA[[Kafka pipelines]]
    SAMZA([Samza / Flink])
    EMB([Embeddings - Pinot])
    LAKE[(Data lake)]
  end

  subgraph ML
    PYM[People You May Know - PYMK]
    PYMW[People You May Work With]
    NOTIF([Notifications Ranker])
    ADS[Ads / Sponsored]
  end

  Clients --> CDN --> LB --> GW
  GW --> Profile --> PDB
  GW --> Graph --> GDB
  GW --> Search --> GAL
  GW --> Feed --> AGG --> CANDS
  CANDS --> Graph
  Feed --> RANK
  GW --> Jobs --> JDB
  GW --> Msg --> WS
  Profile --> KAFKA
  Graph --> KAFKA
  Jobs --> KAFKA
  KAFKA --> SAMZA --> EMB
  EMB --> PYM
  EMB --> JOBS
  PYM --> CANDS
  ADS --> RANK

    classDef client fill:#dbeafe,stroke:#1e40af,stroke-width:1px,color:#0f172a;
    classDef edge fill:#cffafe,stroke:#0e7490,stroke-width:1px,color:#0f172a;
    classDef service fill:#fef3c7,stroke:#92400e,stroke-width:1px,color:#0f172a;
    classDef datastore fill:#fee2e2,stroke:#991b1b,stroke-width:1px,color:#0f172a;
    classDef cache fill:#fed7aa,stroke:#9a3412,stroke-width:1px,color:#0f172a;
    classDef queue fill:#ede9fe,stroke:#5b21b6,stroke-width:1px,color:#0f172a;
    classDef compute fill:#d1fae5,stroke:#065f46,stroke-width:1px,color:#0f172a;
    classDef storage fill:#e5e7eb,stroke:#374151,stroke-width:1px,color:#0f172a;
    classDef external fill:#fce7f3,stroke:#9d174d,stroke-width:1px,color:#0f172a;
    classDef obs fill:#f3e8ff,stroke:#6b21a8,stroke-width:1px,color:#0f172a;
    class WS edge;
    class GW,PSVC,SKILL,CONN,DEG,PEO,JOB,GAL,QU,HOMEFEED,JOBS,APPLY,INMAIL,DM,PYM,PYMW,ADS service;
    class PDB,GDB,JDB,LAKE datastore;
    class KAFKA queue;
    class AGG,RANK,CANDS,REC,SAMZA,EMB,NOTIF compute;
    class ALERT obs;

Distinctive parts#

Economic Graph: members, companies, skills, jobs as nodes. Used for PYMK, job match, talent insights.
People You May Know (PYMK): ML-driven candidate gen using 2nd/3rd-degree contacts, school/company overlap, embeddings.
Galene = LinkedIn's Lucene-based search platform (now superseded internally) with custom relevance for "name + headline + company" queries.
Espresso = LinkedIn's distributed document store (Helix-managed, Avro schemas).
Voldemort historically for KV; replaced by Espresso + Venice.

Messaging#

InMail (one-off to non-connections, paid) vs Direct Messaging (connections).
Same kind of WS + inbox queue as WhatsApp but simpler (no E2E).

Jobs#

Index = job postings + filters (location, exp, remote).
Recommendations combine collaborative + content (skills, profile embedding).

Trade-offs#

Strongly typed graph unlocks rich features but ops complexity.
Multi-DC writes for profile but mostly single-region for graph hot path.
Privacy & compliance (GDPR, recruiter visibility) drive complex ACL.

Glossary & fundamentals#

Concepts referenced in this design. Each row links to its canonical page; the tag column shows whether it is a high-level (HLD) or low-level (LLD) concept.

Tag	Concept	What it is	Page
`HLD`	CDN	edge caching for static assets	cdn
`HLD`	Pub/Sub & message brokers	topics, consumer groups, delivery semantics	pub-sub-pattern
`HLD`	Realtime protocols	WS / SSE / polling / gRPC streaming	realtime-protocols
`HLD`	Batch & stream processing	Lambda vs Kappa, watermarks, windows	batch-stream-processing
`LLD`	REST API design	verbs, statuses, pagination, errors	rest-api-design

Quick reference#

Functional#

Profile (work history, education, skills).
Connection graph (1st/2nd/3rd degree).
Feed of activity (posts, articles, reactions).
People search, jobs search.
Messaging + InMail.
Notifications.

Non-functional#

1B users; ~300M MAU.
p99 profile page < 300 ms.
Multi-region active for reads.

Capacity#

~50B notifications/day at peak.
Feed traffic: 100M+ DAU × 30 feed loads.

Schema highlights#

member(id, headline, location, skills[], experience[], education[])
connections(a_id, b_id, status, since)
posts(id, author, ts, body, media[], visibility)
jobs(id, company_id, title, location, skills[], desc)
Bi-directional shards on connections (by a and by b) for both-direction queries.

Trade-offs#

Graph queries (2nd-degree) = expensive joins; use precomputed degree counts + traversal limits.
Recommendation quality = key product differentiator; large investment in ML.
Recruiter Search has different latency vs free search; separate index.

Refs#

LinkedIn engineering blog (Espresso, Voldemort, Pinot, Helix, Kafka, Samza).
Galene/Search papers.
"LinkedIn's Economic Graph" talks.
ByteByteGo "Design LinkedIn".

FAQ#

How is LinkedIn's connection graph different from Twitter's?#

LinkedIn connections are symmetric: both sides must accept. Follows can be asymmetric. The graph also exposes 1st, 2nd, and 3rd degree distance to power search and PYMK.

How does People You May Know work?#

PYMK is a recommendation pipeline that mines shared connections, schools, employers, and behavioral signals. Triangle closing on the graph is the dominant signal.

How does LinkedIn job search scale?#

Elasticsearch clusters index jobs by title, skills, location, and seniority. Queries fan out across shards and rank with a learned model that uses user profile features.

How is LinkedIn's feed different from Facebook's?#

It blends posts, articles, jobs, and recommendations with stronger emphasis on professional intent. Ranking favors employer pages, mutual connections, and dwell time.

How is privacy enforced in LinkedIn search?#

Profile visibility settings are checked at query time. Anonymous viewers, blocked users, and 3rd-degree limits are applied as filters before ranking returns results.

Video walkthrough

Design LinkedIn: System Design Mock with Senior SWE @ Amazon : via Exponent