← Waymo Interview Insights

Waymo·Software Engineer·Technical Phone Screen·Senior

SeniorRejected
May 2026

Summary

Did an additional round at Waymo for a software engineering role. The interviewer was friendly enough but spent a weirdly long time explaining the autonomous driving stack to someone who already had relevant experience, and by the time we got to the actual question I could already feel it wasn't going well.

Questions Asked (1)

Q1

You have two sets of experiments with different latency profiles in the planning module. How do you decide which configuration is better?

A/B Testing & ExperimentationTechnical Trade-offsSystem Design
Author's notes

This is a deceptively tricky question because it's not just 'pick the lower latency one.' You have to think about what latency even means in a planning context, tail latencies versus medians, how latency interacts with safety margins, whether the two experiment setups are even comparable.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying the goal: what does 'better' mean for the planning module? Then compare the configurations using both latency metrics (e.g., p50, p95, p99) and downstream system-level outcomes (e.g., safety, comfort, route completion). Finally, recommend a decision based on trade-offs, risk tolerance, and statistical significance.

Pro tip: In safety-critical systems like autonomous driving, latency tails (p99, p99.9) often matter more than averages because a single slow response can cause a catastrophic failure. Always discuss tail latency and worst-case behavior, not just mean or median.

1. Clarify the objective and constraints

Define what 'better' means: is the priority minimizing average latency, reducing tail latency, improving safety, or meeting a hard real-time deadline? Also identify constraints like compute budget, memory, and regulatory requirements.

2. Analyze latency profiles statistically

Compare distributions using percentiles (p50, p90, p95, p99, p99.9), not just means. Check for statistical significance with appropriate tests (e.g., bootstrap, t-test) and consider variance and outliers.

3. Evaluate system-level impact

Assess how each configuration affects downstream metrics: safety (collisions, near-misses), comfort (jerk, acceleration), and efficiency (route completion time). Use A/B testing or simulation to measure these outcomes.

4. Weigh trade-offs and risks

Consider the cost of worst-case latency (e.g., missed deadlines) versus average performance. If one config has lower average but higher tail, it may be riskier. Quantify the trade-off in terms of acceptable risk.

5. Make a recommendation with justification

Choose the configuration that best aligns with the primary objective and risk tolerance. If inconclusive, propose further experiments or a hybrid approach. Clearly state assumptions and limitations.

Key Points to Mention

  • Tail latency (p99, p99.9) is critical in safety-critical systems; a single slow response can be catastrophic.
  • Statistical significance and confidence intervals to ensure differences are not due to noise.
  • Downstream metrics: safety, comfort, and route completion, not just latency.
  • Trade-offs between average and worst-case performance, and how they map to risk.
  • Real-time deadlines and hard constraints in the planning module.
  • A/B testing methodology: randomization, sample size, and avoiding confounding factors.

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.