System Design Interview: Designing a Job Board Like Indeed
14 views
"Design a job search platform" is a realistic system design question because it combines search, ranking, write-heavy ingestion and user notifications. Here is a structured answer for a 45-minute round.
1. Requirements
Functional: employers post and edit jobs; job seekers search by keywords, location, salary and remote options; seekers save jobs, set alerts, and apply with a stored resume; employers review applicants. Non-functional: search results in under 300 ms, new postings searchable within a minute or two, high availability for search, strong consistency for applications (an application must never be lost). Scale assumptions: 30 million active jobs, 100 million monthly users, 5,000 search queries per second at peak, a few hundred postings per second including scraped and partner feeds.
2. High-level components
- Posting service: validates and stores jobs in a relational database (jobs, companies, locations).
- Ingestion pipeline: pulls jobs from employer ATS feeds and partner XML/JSON feeds, normalizes titles and locations, and deduplicates.
- Search index: an inverted index cluster (Elasticsearch/OpenSearch) holding a denormalized job document.
- Search service: parses queries, applies filters, calls ranking.
- Application service: stores applications and resume references.
- Alerts service: matches new jobs to saved searches and sends emails or push notifications.
3. Indexing flow
The posting service writes to the database and emits an event (via an outbox table to a log such as Kafka). An indexer consumes events and upserts the search document: title, normalized title, company, description, skills extracted by NLP, geo point, salary range, remote flag, posted date. This gives near-real-time freshness without coupling writes to the search cluster. Expired or filled jobs produce delete events.
4. Deduplication
The same job often arrives from the company site, a staffing agency and an aggregator. Compute a fingerprint from normalized company, title, location and a shingle hash of the description; cluster near-duplicates with MinHash and keep the canonical source (preferably the employer's own). Interviewers like hearing this because duplicate listings hurt user trust.
5. Search and ranking
Stage one retrieves candidates with filters (geo distance, remote, salary) and keyword relevance (BM25 plus synonym expansion: "software engineer" also matches "developer"). Stage two re-ranks the top few hundred with a learned model using freshness, click-through and apply rates, title match, and user signals such as past searches. Shard the index by job ID with replicas for read throughput; geo queries use geohash fields.
6. Saved-search alerts
Running every saved search against the index every hour does not scale. Instead, use a reverse approach: when a new job is indexed, match it against stored queries (a percolator-style index of saved searches), batch matches per user, and send a daily or instant digest depending on preferences.
7. Applications
Applications go to a relational store with an idempotency key (user_id, job_id) so double-clicks do not create duplicates. Resumes live in object storage with encryption at rest and access limited to the employer for that job. Forward applications to the employer's ATS through an async queue with retries.
8. Abuse and quality
Detect scam postings (requests for payment, mismatched domains), rate-limit employer accounts, and verify companies. Mention privacy: let seekers hide their profile from their current employer.
9. Wrap-up
Summarize: relational source of truth, event-driven indexing, dedupe, two-stage ranking, percolator alerts, and durable applications. Candidates often know these pieces but lose the thread when an interviewer asks a deep follow-up. Linkjob AI (https://linkjob.ai) provides real-time help during live job interviews and technical interviews, helping you recall the next layer of a design while you keep talking.