Skip to content
Tech Interview Prep
Technical interview guide

Probability Distributions

The handful of named distributions (normal, binomial, Poisson) that show up repeatedly, and what each models.

Read
44 min
Practice MCQs
25
Interview Q&A
25
Edition
v6
Editorial status
Reviewed

Scope: Standards and library references current as of 2026-09: NIST/SEMATECH e-Handbook, SciPy, NumPy, statsmodels, and scikit-learn.

Curated: · Written: · Reviewed:

25 questions

MCQ-2The gap between two retries is modeled as exponential with a mean of 0.1 s, so the density is f(t) = 10 * exp(-10t). The ops dashboard plots f(t) and labels the peak 'probability of a 0 s gap: 10'; a reviewer rejects the model on that line. Which pair of numbers should the dashboard carry instead: the probability that the next gap is exactly 0 s and the probability that it exceeds 0.2 s?
MCQ-5A worker service's job arrivals are modelled as a Poisson process at 5 jobs per second, and the team is deciding whether it can sleep for 200 ms between polls without falling behind. What does the model give for the probability that a 200 ms window contains no arrival at all?
MCQ-3A nightly reconciliation job fans out over 10 shards. Each shard is independently locked by a vendor batch window with probability 0.10, and any single locked shard aborts the run. Locks are independent across shards, and the job makes one pass over them. What is the chance a given night's run aborts?
MCQ-11In reliability and queueing models based on independent Bernoulli trials with constant success probability p, which statement precisely defines the relationship between the geometric and negative binomial distributions?
MCQ-12A team ports a Monte Carlo queueing simulation from R to SciPy. The R code drew service times with `rexp(n, rate = 0.25)`; the SciPy port calls `scipy.stats.expon.rvs(scale = 0.25, size = n)`. The port runs without error, but the simulated mean service time is 0.25 s where the model specifies 4 s. Which parameterisation difference between the two ecosystems explains the inversion?
MCQ-6A storage team models disk time-to-failure as exponential with a mean of 1000 hours. One disk in the fleet has run for 3000 hours without failing. The team is deciding whether to replace it preemptively and asks for the probability that it survives the next 1000 hours.
MCQ-8A team fits a log-normal to request latency and reports its two parameters. Which statement about the log-normal is correct?
MCQ-10An engineer estimates the capacity parameter N of an integer ID sequence drawn uniformly from {1, ..., N} based on n observed requests, and also fits a continuous uniform model U(0, B) to task runtimes. Which statement correctly captures the estimation properties and mechanics of uniform bounds?
MCQ-13A 200-request latency sample is sorted ascending. The 197th value is 290 ms and the 198th is 310 ms. The SLO requires 99% of requests under 300 ms. Which statement about the empirical CDF at 300 ms is correct?
MCQ-7A page render issues 100 downstream calls one after another, so the render's latency is the sum of those calls' latencies. The calls are independent, each has mean 8 ms, standard deviation 6 ms, and mild right skew (skewness about 2). The team must publish a p99 for the render before load testing. Which estimate is right?
MCQ-23A gateway fans out each request to 100 independent backends in parallel and returns only when all have replied. Each backend's latency has p99 = 100 ms (that is, 1% of its calls exceed 100 ms) and the backends' slow events are independent. Roughly what fraction of requests have end-to-end latency above 100 ms?
MCQ-25A fraud model assumes each account's daily transaction count is Poisson with rate λ. The team validates the assumption weekly by checking that the sample mean stays near its historical value of 3.1. After a new client began batching retries, the mean over the latest 50,000-account window is still 3.1, but the sample variance has risen from 3.1 to 9.4. What does this window tell the team?
MCQ-1A support team counts tickets per customer per week. Across customers the mean is 2 and the variance is 6; tickets within a customer are independent, but each customer has a persistent rate. The histogram is right-skewed, though a normal curve fits the central mass. The team wants a p99 threshold for alerting. Which model and threshold should they use?
MCQ-14You fit a lognormal distribution to 5,000 request latencies. A Q-Q plot against the fitted lognormal is straight through the middle, but the largest 1% of points bend above the reference line. What should you conclude for a p99 SLO?
MCQ-15A team fits a normal distribution to 2,000,000 latency residuals. A Kolmogorov–Smirnov test returns D = 0.0021 with p ≈ 4e-8. The fit feeds one decision: whether the normal model's p99 of 412 ms is close enough to the empirical p99 of 412.4 ms to use in a capacity plan. What should the team do?
MCQ-16A read service serves 90% of requests from a warm cache in 5 ms and the remaining 10% from the origin in 300 ms. The dashboard reports a mean latency of 34.5 ms, and the team's SLO is 'mean latency < 50 ms', currently green. Which reading of the dashboard is correct?
MCQ-22Over the last 30 days a service's per-minute request rate was at or below 1,500 rps for 90% of minutes, between 1,500 and 4,000 rps for 9% of minutes, between 4,000 and 6,000 rps for 0.9% of minutes, and above 6,000 rps for 0.1% of minutes; the mean was 900 rps. The SLO allows at most 1% of minutes to be overloaded. Which capacity target meets the SLO with the least headroom?
MCQ-24A checkout service takes 1,000,000 requests a day. Latency across the day is reported as mean 120 ms, median 40 ms, p99 1,400 ms. The capacity plan needs the average number of requests in flight, and the on-call dashboard needs the daily count of requests above 1.4 s. Which reading of these numbers is correct?
MCQ-21A service records 1,000 request latencies. Thirty requests hit the 5,000 ms client timeout and are stored as 5,000 ms. The dashboard computes p99 with the nearest-rank definition from those 1,000 stored values and reports 5,000 ms. Which conclusion about that reported p99 is correct?
MCQ-9A team models request sizes as a Pareto distribution with tail index α = 1.5. From 10,000 samples they compute a sample mean of 120 KB and a sample standard deviation of 2,000 KB, then estimate p99 as mean + 2.33×std = 4,780 KB. The largest observed request is 50,000 KB. They plan to increase the sample to 100,000 to stabilize the p99 estimate. Which statement is correct?
MCQ-4A checkout experiment has a 2% baseline conversion rate. The team wants to estimate that rate to within ±10% relative at 95% confidence. A previous experiment on a 20% baseline needed 1,540 visitors for the same relative precision. How many visitors should they plan for the 2% baseline?
MCQ-17A service's response times are sampled 8 times: mean 120 ms, sample standard deviation 20 ms. An engineer reports a 95% confidence interval for the mean of 120 ± 1.96 × 20/√8 = [106.1, 133.9] ms. Which assessment of that interval is correct?
MCQ-18A checkout test has variant B at 3 conversions from 4 sessions and variant A at 40 from 100. The team puts a Beta(1, 1) prior on each rate and reports posterior means. What posterior mean does variant B get?
MCQ-19A service's latency SLO is stated at p99.9. You have 200 sampled requests; the largest is 900 ms and the second largest is 610 ms. You draw 10,000 bootstrap resamples of the 200 samples, compute the 99.9th percentile of each resample using the nearest-rank definition (the ⌈np⌉-th smallest value), and take the 97.5th percentile of those 10,000 values as the upper confidence bound. What is that upper bound, and why?
MCQ-20A four-process Monte Carlo job is meant to pool 4,000,000 draws of a mean estimate. Each worker's entry point begins with `rng = np.random.default_rng(7)` and draws 1,000,000 samples; the driver concatenates the four arrays and reports the sample mean with a standard error of s / sqrt(4,000,000). Compared with the single-process run (same code, 1,000,000 draws, half-width known to be correct), the pooled mean matches to within floating-point rounding and the reported 95% half-width is half as large. Which conclusion is right?