System-Design Drills

The interview canvas

Use this order so your answer remains navigable:

1. requirements, scale, and success/failure semantics;
2. events/data model and time/identity;
3. high-level flow and component responsibilities;
4. state, partitioning, and storage;
5. ordering, backpressure, idempotency, and recovery;
6. latency/capacity calculation;
7. observability, deployment, security, and trade-offs.

Drill A — Event-driven backtesting platform

Support 100 researchers, equities/futures, daily and tick data, reproducible runs, parameter sweeps, cost/fill models, and promotion toward live strategies.

Must cover: immutable/versioned data, point-in-time universe, run manifest, deterministic event order, compute scheduling and quotas, artifact/results store, experiment metadata, failure isolation, cache correctness, and research/live parity.

Pressure questions:

• How do you prevent look-ahead centrally rather than by convention?
• What exactly makes a run reproducible six months later?
• How do you avoid one parameter sweep starving the cluster?
• Which results can safely be cached, and what is the cache key?
• How is a strategy promoted without copying its logic?

Drill B — Real-time market-data distribution

Consume multiple feeds at 100k events/sec average and 1m peak. Thousands of symbols; several strategies need sub-50-ms freshness; raw data must be replayable.

Must cover: adapters, source sequences, gap recovery, normalized schema, raw journal, partitioning, bounded queues, per-data-type overload policy, snapshots, consumer isolation, freshness and p99 telemetry, schema evolution.

Pressure questions:

• Can top-of-book updates be dropped? Can depth deltas?
• What happens when one consumer is slow?
• How does a restarted consumer rebuild state?
• How do you compare source event time to platform latency safely?

Drill C — Order and pre-trade risk service

Accept orders from multiple strategies, enforce limits, route to brokers/venues, track state, and remain safe through duplicate requests and ambiguous timeouts.

Must cover: client order IDs/idempotency, order state machine, independent risk boundary, durable intent, venue adapters, cancel/fill races, reconciliation, kill switch, audit, permissions, and controlled degradation.

Pressure questions:

• The send timed out. Do you retry?
• Risk state is stale. Fail open or closed?
• How do limits update without inconsistent partial rollout?
• What is the source of truth for position after restart?

Lightweight capacity math

State assumptions and calculate orders of magnitude:

events/day = events/sec × active seconds
raw bytes/day = events/day × encoded bytes/event
required workers ≈ arrival rate × service time / target utilization

At 1,000,000 events/sec and 100 bytes/event, raw payload alone is about 100 MB/sec before replication, indexes, metadata, and compression. Numbers expose design constraints; precision theatre does not.

Self-score (0–2 each)

• clarified functional and non-functional requirements;
• defined event identity, time, ordering, and state;
• addressed overload and recovery before prompting;
• connected storage/partitioning to access pattern;
• quantified throughput/latency/storage;
• covered observability and business invariants;
• explained alternatives and why one was chosen;
• gave a concise final summary.

Target: 12/16 before the interview.