System Interview Framework

Ask Interviewer These Questions First

Interactive probing questions with typable fields, automated conversions, mathematical proofs, and engine tradeoffs.

Step 1 of 8
Traffic Scenario Testing (Probing Presets)
Stress-test probing values under surge events:
PROBING8 Questions
1
Processing Paradigm & Ingestion Mode

Is the data to be processed real-time, event-based, or ETL? Can it be hybrid?

2
User Scale (DAU & MAU Conversion)

What is your Daily Active Users (DAU) & Monthly Active Users (MAU)?

3
Inbound + Outbound Requests (Total Traffic Hits)

How many inbound requests does this application process, and what is the outbound fan-out to get total request hits?

4
Payload Sizing & Data Storage Volume

What is the request & response payload size (KB/MB) and monthly storage accumulation?

5
Reliability, Failures & Dual Channels

What if requests fail? Should we use dual communication channels (API + Kafka Outbox) and retry jobs?

6
Workload Characteristics (Read/Write Ratio)

What is the read-to-write ratio of incoming operations?

7
Latency SLAs & Tail Budget (p90, p95, p99)

What is your expected response time across p90, p95, and p99 latency percentiles?

8
Cache Sizing, TTL & Retention

How many records can be stored in the cache server for a day and for how many hours?

SIZINGStep 1 of 8

Is the data to be processed real-time, event-based, or ETL? Can it be hybrid?

Interviewer Dialogue & Probing Target:

"Should checkout confirm inventory synchronously with ACID guarantees, or can order fulfillment publish events to an asynchronous event bus?"

Library & Engine Fit:Debezium CDC + Kafka + PostgreSQL

Why Hybrid is standard: Direct dual-writes to DB + Kafka risk partial failures (database commits but network drops before Kafka message sends). We solve this with the Transactional Outbox Pattern using Debezium CDC to tail the PostgreSQL WAL log into Kafka topics.

1 / 8 Probing Topics
Interactive Pipeline

Connected System Flow Topology

Click any node to view exact applied math & library tradeoffs

System Architecture Subsystem Sizing & Tradeoffs

Two-column view: Left shows real-time derived capacity; Right shows mathematical proofs, probing questions, and library comparisons.

Tier 1 — User Ingress & Concurrency

Anchor Metric
users
%
%
acts
sec
x
Derived DAU
Peak PCU
Peak API RPS
How We Derived the Anchor RPS:
1. DAU = 10,000,000 * 20% = 2,000,000 users
2. PCU = 2,000,000 * 20% = 400,000 concurrent
3. Peak RPS = (400,000 * 30 / 3600) * 2x = 6,667 req/s
Library Tradeoff: Cloudflare vs CloudFront

Cloudflare Workers execute V8 isolates in 0ms cold starts for edge rate-limiting, whereas AWS CloudFront integrates natively with AWS WAF and Private Origin VPC links.

Tier 2 — Kafka Event Bus & Partitions

Pub/Sub Streaming
KB
x
MB/s
MB/s
parts
tries
Ingress MB/s
Partitions
Brokers
Kafka Partition & Broker Derivations:
1. Ingress MB/s = (6667 RPS * 1 KB) / 1024 = 6.51 MB/s
2. Partitions = ceil(6.51 / 10 MB/s) = 6
3. Brokers = max(3, ceil(19.53 / 500)) = 3 nodes
Library Tradeoff: Apache Kafka vs RabbitMQ

RabbitMQ provides complex routing keys and per-message ACKs but degrades under millions of queued items; Kafka uses append-only disk segments with zero-copy DMA (sendfile) for 10x higher throughput.

Tier 3 — Database & Redis Cache

ACID & Caching
%
%
qps
items
KB
days
DB Replicas
Cache RAM
Monthly Storage
Database Replica & Cache Derivations:
1. Read QPS = 20000 * 80% = 16000 QPS
2. DB Replicas = ceil(16000 / 5000) = 4 nodes
3. Cache RAM = (100,000 * 3 KB * 1.3x) / 1000 = 5.7 GB
Library Tradeoff: Redis vs Memcached

Memcached is multi-threaded for pure key-value caching, but Redis Cluster provides persistence snapshots (AOF/RDB), atomic increment counters for flash-sale inventory, and pub/sub cache invalidation.