Design ML A/B Testing

Hard45 min
1 / 30
understanding9 min read

Problem Statement: ML Experimentation at Netflix Scale

Problem Statement: ML Experimentation at Netflix Scale — ML A/B testing interview depth

Problem Statement: ML Experimentation at Netflix Scale

Design production ML A/B testing and experimentation platform that assigns traffic to model arms, logs exposures and outcomes, and computes trustworthy lift for product and ML teams. This section covers problem statement ml experimentation at netflix scale in the understanding phase.

Why interviewers probe here

Netflix, Google, and Uber run thousands of concurrent model experiments; interviews test whether you treat experimentation as a distributed systems problem—not a spreadsheet.

Operational detail

Frame the system as assigning each request to a treatment arm (model version or policy), logging exposures and outcomes, and computing statistically valid lift without harming user experience.

Failure and edge cases

Simpson's paradox from naive aggregation, peeking at p-values, and SRM (sample ratio mismatch) when bucketing breaks.

Interview checkpoints

  • Checkpoint (understanding): Separate experiment design from inference serving
  • Checkpoint (understanding): Name guardrail metrics vs primary KPI
javaOne Dark Pro
1public record ExperimentArm(String id, String modelVersion, int weightBps) {}
pythonOne Dark Pro
1@dataclass(frozen=True)
2class ExperimentArm:
3 id: str
4 model_version: str
5 weight_bps: int # basis points sum to 10_000
typescriptOne Dark Pro
1export interface ExperimentArm { id: string; modelVersion: string; weightBps: number; }

Why interviewers care

ML A/B Testing interviews reward crisp scope, explicit trade-offs, and failure stories—not generic microservice diagrams.

Interview checkpoint

Name one failure story for Problem Statement: ML Experimentation at Netflix Scale that proves you understand real outages, not happy-path diagrams.

Key Highlights

  • Experiment control plane: hypotheses, arms, allocation policies
  • Data plane: traffic splitter, model endpoints, metric collectors
  • Stakeholders: ML scientists, product analysts, platform SRE
  • North-star metrics: watch time, conversion, latency guardrails
Key insight
Google, Amazon, and Meta run thousands of models in production; interviews test whether you treat inference as a distributed systems problem—not a notebook export.
Avoid
Cold starts after scale-to-zero, OOM on oversized batches, version skew when router cache lags registry, and thundering herd when a viral model spikes.

Section Rescue Kit

Buzzwords to use:

Randomization UnitGuardrail Metric

Safe statements:

  • "Let me separate experiment control plane (registry, analysis) from data plane (assignment + ML A/B testing platform)."
  • "I will confirm randomization unit and guardrail metrics before discussing hash bucketing."
Design ML A/B Testing - System Design | WinJob | WinJob