aiagent.org Official Logo aiagent.org

System Design Architectures

Interactive technical blueprints, capacity estimates, database sharding strategies, and visual flowcharts for the top 30 system design interview questions.

Target Architectures
Architecture Blueprint

Design Auto-Complete

Real-time typeahead query suggestion system generating dynamic matches under high write/read traffic.

1. System Requirements

Functional
  • Provide top 5 search query suggestions based on historical popularity matching the typed prefix.
  • Suggestions must update dynamically as the user types.
  • Must support custom query blacklisting.
Non-Functional / SLA
  • Extremely low-latency updates: suggestions must appear in < 100ms.
  • High availability and scalability to process 100,000 queries/sec.

2. Scale & Resource Estimates

Throughput: 100 Million daily active users. Assuming 10 search queries/day and 5 keystrokes per query = 5 Billion typeahead requests/day = ~58,000 requests/sec average.

Storage & RAM: 10M unique queries * 15 bytes = 150 MB database storage. Extremely lightweight, allowing complete in-memory storage.

3. System API Endpoints

GET /api/v1/search/suggest?q=sys
Response: { suggestions: ["system design", "system metrics", "system software", "sys log", "sysctl"] }

4. Database Schema Design

Table: query_frequencies (MySQL - for MapReduce batch updates)
- query (VARCHAR - Primary Key)
- frequency (BIGINT)

Trie Cache (Redis Cache - Key-Value)
- Key: prefix:{char_sequence}
- Value: Sorted Set of suggestions (Score: frequency)

5. Interactive High-Level Architecture

graph TD Client[User Client] -->|Type Keystroke| LB[Load Balancer] LB -->|HTTPS| Gateway[API Gateway] Gateway -->|Get Suggestions| Suggest[Suggest Service] Suggest -->|Prefix Query| Cache[(Redis Trie Cache)] Suggest -->|Blacklist filter| Blacklist[(Redis Bloom Filter)] Client -->|Complete Search| SearchService[Search Service] SearchService -->|Log Search| LogBroker[Kafka Log Queue] LogBroker -->|Batch Freq Count| MR[MapReduce Engine] MR -->|Rebuild Trie| Cache

6. Scalability Deep Dive & Bottlenecks

Stores prefix suggestions inside an in-memory Trie (Prefix Tree) cache. To optimize lookup speed, each Trie node contains a list of its top 5 hot queries, reducing retrieval time from $O(L + K \log K)$ to $O(L)$ where $L$ is query length. A nightly MapReduce offline pipeline aggregates raw search log queries, calculates frequencies, and refreshes the Trie structure in Redis.

7. Production Scaling Strategies

Trie Sharding by Prefix: Partition the Trie across 26+ Redis nodes by first character of the query. 'a-c' on node 1, 'd-f' on node 2, etc. Each node holds ~50M prefix entries. Hot prefixes (e.g., 'how to') are replicated to 3+ nodes for read scaling.

MapReduce Refresh Pipeline: Nightly Hadoop job processes ~10B daily search logs. Aggregate query frequencies, apply time-decay weighting (recent queries weighted 2×). Build a fresh Trie snapshot and atomically swap it into Redis via keyspace rename. Zero-downtime refresh.

Regional Edge Caching: Cache the top 10K most popular prefixes at CDN edge PoPs. Queries for 'weather', 'news', 'amazon' resolve in <5ms from edge without hitting origin. Reduces origin Trie lookups by 60%.

Query Log Sampling: At 100K queries/sec, storing every query is expensive. Sample 10% of queries for the MapReduce pipeline. Apply statistical estimation to extrapolate true frequencies. Keeps log storage under 1 TB/day while maintaining 95%+ accuracy.

Personalized Suggestions: Blend global Trie results with user-specific recent queries (stored in Redis per user_id, last 100 queries). Weight personal results 1.5× higher than global. Falls back to global-only for new users.

Latency Budget: Total autocomplete response must be <50ms (P99). Budget: 5ms network, 10ms Trie lookup, 5ms ranking, 10ms serialization. If Trie lookup exceeds 15ms, fall back to cached top-10 results for that prefix.

8. Runnable Production Starter Code

Scaffold