ByteDance ML System Design: Live Stream Violation Penalty

You must design the post-detection flow for a live streaming violation penalty system that receives real-time alerts from an ML detector (hate speech, nudity, copyright, etc.). The focus is on processing violation alerts, applying rule-based penalties, notifying stakeholders, and storing auditable records—not on how the ML model itself works.

Core tasks you’ll address:

• Ingest and validate violation alerts (stream ID, timestamp, violation type, confidence). Ensure deduplication and idempotency when alerts arrive from multiple detectors or retries.
• Decide penalty logic: map violation types + confidence + user history to actions (warning, temporary suspension, stream stop, permanent ban). Include escalation and decay of strikes.
• Real-time enforcement: stop or restrict a live stream within seconds, update user state, and ensure distributed enforcement consistency.
• Notifications & appeals: push in-app alerts and emails to streamers and moderators with appeal links and human review handoff.
• Data & audit: define schemas for violations, penalties, user state, and immutable audit logs for compliance and debugging.

High-level flow/stages you should walk through in an interview:

1. Ingestion & validation (queue, schema check)
2. Rule engine & decision (apply penalty rules, check history)
3. Enforcement & state update (execute action, write user/stream state)
4. Notification & human review (alerts, appeals pipeline)
5. Logging & monitoring (audit trails, metrics, retries)

Skill signals interviewers look for: knowledge of low-latency streaming pipelines, distributed consistency (idempotency, exactly-once considerations), schema design for auditability, scale strategies (partitioning, sharding, backpressure), and extensible rule engines. You should also discuss reliability, monitoring, and how to handle false positives and appeal workflows.