Atos Data Scientist Interview Questions & Guide 2026

The 60-second brief

At Atos, particularly within the Atos Amplify and Digital Transformation Consulting (DTC) divisions, a Data Scientist operates at the critical intersection of advanced technology and strategic business advisory. You are not just building models in isolation; you are a consultant-led technical expert responsible for translating complex, ambiguous business challenges into secure, scalable, and ethical AI-driven solutions. Your work directly impacts how public sector and enterprise clients navigate their digital transformation journeys, driving measurable operational efficiency and long-term return on investment.

Ask early whether the loop includes an asynchronous take-home or a timed live case, because the two are graded on different things. A take-home is read as an artifact: the question you decided to answer, what you did about missing or malformed records, and a conclusion stated plainly enough for someone to act on. A reviewer who cannot rerun your notebook discounts the result whatever score is printed in it. Hold to the stated time box and write down what you would have done with more of it, since the follow-up round is usually a live defence of the same work.

Focus areas

• Separate novelty effects from durable behaviour change
• Pick a randomisation unit that respects interference
• Decompose a metric move by segment and mix

Overview

As a Data Scientist, you will join the rapidly expanding Insight practice. This team is tasked with designing, prototyping, validating, and deploying end-to-end data science applications. Whether you are leveraging traditional machine learning frameworks or architecting cutting-edge Generative AI and Retrieval-Augmented Generation (RAG) systems, your contributions will directly influence client strategy, operational scaling, and responsible AI governance.

The role is highly collaborative and multidisciplinary. You will partner closely with client stakeholders, business analysts, data engineers, and project managers to deliver high-value use cases. This unique blend of deep technical execution and high-level client advisory makes the Data Scientist position at Atos exceptionally dynamic, offering you diverse project exposure across highly regulated industries and the public sector.

The Atos interview process

Atos candidates report 5 rounds · ≈ 4-6 weeks. The stages below are what candidates describe, not a published process.

Initial Touchpoint

An extra round usually exists because something is still open after the standard loop: a skill the earlier interviews did not sample, a level decision, or two interviewers who disagreed. It is rarely a rerun of what you already did well. Ask the recruiter who you are meeting, what function they sit in, and how long the session runs. That is an ordinary scheduling question, and the answer changes what you should prepare. What separates a strong candidate here is treating the round as a fresh evaluation with its own bar, rather than assuming earlier performance carries you through or sinks you.

What to demonstrate

• Whether you can answer well on ground the earlier rounds did not cover, without leaning on what you already said to someone else
• Consistency of the facts in your stories: the same sample size, timeframe, team size and scope of your own role as in earlier conversations
• How you handle an unfamiliar format live, including whether you ask what kind of answer is wanted before producing one

How to prepare

• Ask the recruiter for the interviewer's function, the length, and whether to expect a coding surface, a discussion, or a presentation. Preparing for a 30 minute conversation with a partner team is not the same work as preparing for a 60 minute technical block.
• Write out what each earlier round actually covered, then list the two or three areas nobody probed. That gap is the most likely subject of the extra round.
• Re-read the numbers in the project stories you have already told, so a second telling does not quietly contradict the first.

Discovery Call

Because the format is not fixed, prepare the reasoning rather than the ritual. Nearly every version of this round draws on the same underlying material: a design you can defend, a metric you can define exactly, an analysis whose assumptions you can state out loud. Only the wrapper changes, whether that is a take-home, a live case, a deep dive on past work, or a rough estimate on a whiteboard. Answers rehearsed to fit one shape stall the moment the shape differs. Practise naming the assumption behind a number, then saying how much the conclusion moves if that assumption is wrong.

What to demonstrate

• Whether your justification for a method survives the question 'why not the simpler thing', including when the simpler thing would have worked
• Precision under pressure: what exactly counts as an active user, a conversion or a success, over what window, with what exclusions
• Whether you carry an argument through to a recommendation instead of stopping at a list of tradeoffs

How to prepare

• For each project you plan to mention, write the metric definition in one sentence: numerator, denominator, time window, exclusions. Say it out loud once, because vagueness shows up in speech before it shows up on paper.
• Rehearse the same project at three lengths: two minutes, ten minutes, and a deep dive on one technical decision. Cutting live is harder than it sounds.
• For your headline result, write down what would have had to be true for it to be wrong, and how you ruled that out.

Technical Evaluation

A handful of shapes account for most of what gets asked in this format: a ranking or deduplication inside groups, a running or rolling total, a period-over-period comparison, and a cohort tracked forward over time. Recognising the shape quickly is most of the speed here; deriving it from scratch while a clock runs is where the time goes. Know that a window function keeps every row while a GROUP BY collapses them, and know which one the question needs. If the exercise is in Python instead of SQL, the same shapes arrive as groupby with transform, shift and merge, and the same grain mistakes are available.

What to demonstrate

• Whether you reach the right construct without a detour, such as ROW_NUMBER over a partition to deduplicate instead of a self-join against a MAX subquery
• Whether you know what your window frame actually is, since adding ORDER BY inside OVER changes the default frame and silently changes a running total
• Whether the thing runs. A near-miss that throws an error scores below a plainer query that returns the right rows.

How to prepare

• Write each of the four shapes once from memory against a small schema and keep the working version somewhere you will reread it: dedupe with ROW_NUMBER, a running total, a month-over-month change with LAG, and a retention table
• Compute one running total twice on data with tied timestamps, once on the default frame and once with ROWS BETWEEN UNBOUNDED PRECEDING AND CURRENT ROW, and look at where the two disagree
• If Python is on the table, rebuild the dedupe and the running total with groupby and cumsum, then assert the two implementations return identical rows

Live Technical Test

This round decides whether someone can hand you a schema and a question and trust the number that comes back. Correctness under a clock is the bar, not clever syntax. The habit that separates strong from weak answers is checking the grain: after every join, know how many rows you expect and whether the count moved. Most wrong answers in this format are not wrong logic, they are a fan-out from a key that turned out not to be unique, or a filter applied before an aggregate when it belonged after. Say what you expect before you run it.

What to demonstrate

• Whether your row counts survive each join, and whether you notice on your own when they do not
• Deliberate handling of rows that fail to match, including whether the question needs an inner join or a left join with the non-matches kept and counted
• Whether NULLs are treated on purpose, given that a NULL compares equal to nothing and that COUNT of a column skips it
• Reaching a defensible answer inside the window instead of a refined one after it

How to prepare

• Take a two-table schema, write a join that fans out on purpose, then fix it by collapsing the many-side to one row per key before joining. Repeat until the fix is reflex rather than recall.
• Write a funnel as one query and print the distinct user count at each stage, then confirm each stage is a subset of the one above it rather than assuming it
• Do a few timed runs in a plain text box with no autocomplete and no formatter, since assessment editors often have neither

Final Interviews

Where a loop ends with a senior leader, that conversation is rarely another skills test. The technical signal already exists by then, so the questions tend to open up: what you would look at first, where a metric you have heard about could mislead, what you would push back on. The decision being made is scope, which in practice means level and how much you would be trusted to own unsupervised. Treating it as a formality is the usual mistake. An open question late in the day is still being scored, and a vague answer reads as someone who has not run anything themselves.

What to demonstrate

• Whether your view of the business has anything specific behind it, given that you are working only from what is public and are expected to say so
• Whether the scope of work you describe owning matches the scope of the role, instead of sitting a level below it
• Whether you can disagree with something concrete and stay useful about it, rather than agreeing with everything said in the room
• Whether your questions are ones only this person could answer, as opposed to ones the recruiter already covered

How to prepare

• Build one view you could defend for two minutes using only public information: what the funnel probably looks like, which metric likely drives decisions, and where that metric could mislead. Being wrong for a stated reason survives this round; having no view does not
• Write down the largest piece of work you have owned from question to decision, who else touched it, and what you decided alone, then check that it reads at the level you are interviewing for
• Prepare one thing you would want changed if you joined and phrase it as a question rather than a verdict, so it opens a conversation instead of closing one

Where candidates lose points

Watching an experiment daily and stopping when it crosses significance

A fixed-sample test controls type I error at one pre-declared look. Checking repeatedly and stopping at the first p < 0.05 inflates the false positive rate to roughly 0.15 to 0.20 for ten looks, and it rises further with more frequent checks, because the p-value takes a random walk that will eventually dip below the threshold under the null. The usual defences are a fixed horizon declared before launch, group-sequential boundaries such as O'Brien-Fleming that spend alpha across a planned number of looks, or always-valid confidence sequences that are correct under continuous monitoring. Compounding it, the effect size reported conditional on having crossed the threshold is biased away from zero, and the bias is larger the lower the power was, so an underpowered test that 'won' typically overstates the lift it found.

Treating last-touch attribution as the causal value of a channel

The attribution label on dim_user is the output of a rule that assigns full credit to whichever touch happened to be recorded last inside a lookback window, and that rule systematically rewards channels that sit close to the conversion, especially branded search and retargeting, which largely intercept demand that already existed. Reallocating spend on those labels moves budget toward the channels that are best at being last, which is why attributed return on ad spend often improves while total signups do not. Nothing in the touchpoint data can settle this, because the counterfactual of not running the channel was never observed. The credible reads are a geo holdout or a scheduled pause, sized in advance on the total-signups metric rather than on the attributed one, and the honest framing in the meantime is that the label describes correlation with conversion and not incremental contribution.

Explaining an aggregate move without decomposing the mix shift

Split the change in the aggregate into within-segment movement and movement in segment weights before you explain it. Every segment's rate can fall while the overall rate rises, purely because volume shifted toward segments that already had higher rates.

Generalising beyond the population the sample actually supports

State the frame the sample was drawn from and where it diverges from the population you want to talk about: time window, platform, geography, opt-in. If a group is excluded from the frame, either weight to known margins or narrow the claim rather than quietly extending it.

Questions & practice

Permutation test for a difference in conversion rates

Medium · permutation test · resampling · inference

Write a two-sided permutation test from scratch for a difference in conversion rates, using no scipy hypothesis function. Input: a DataFrame with unit_id, variant in {control, treatment} and converted in {0,1}, one row per randomisation unit. Compute the observed difference in proportions, then build the null distribution by reshuffling the variant labels while holding each arm's size fixed. Report the p-value as (1 + the count of permuted statistics at least as extreme in absolute value) / (B + 1) with B at least 10,000, and return the permutation distribution.

Approach

1. Name the null being tested: the sharp null that each unit's outcome is the same under either label. That is what licenses permuting labels, and it is stronger than the null of equal means, which matters when someone asks whether the test is valid under unequal variances.
2. Extract converted to a single numpy array of 0s and 1s and record n_treatment. Every permutation is then just a reshuffle of one array, and the treatment mean is the mean of the first n_treatment entries of the shuffled array.
3. Vectorise the B permutations with rng.permuted on a tiled 2-D array, or with argsort of a (B, n) random matrix. A Python loop calling np.random.shuffle B times is correct but roughly an order of magnitude slower and often runs past the time limit.
4. Use the +1 correction in both numerator and denominator. Without it a p-value of exactly 0 is reportable, which is false: the observed labelling is itself one of the permutations, so the smallest attainable p-value is 1/(B+1).
5. Compare the resulting p-value against a two-proportion z-test as a sanity check. At these sample sizes they should agree closely; a large divergence means the statistic or the shuffle is wrong, not that the permutation test found something subtle.

Follow-up

• The arms are 200 and 20,000 units. Does the permutation test stay valid, and what happens to its resolution at B = 10,000?
• Give a 95 percent confidence interval for the difference. Can you get it from this permutation distribution, and if not, what would you run instead?
• The randomisation unit is user_id but the outcome is per session. What breaks, and what is the fix?

Sessionise an event stream with gap and midnight rules

Hard · sessionisation · vectorisation · event streams

Sessionise a raw event stream. Input: a DataFrame with visitor_id, user_id (often NULL), occurred_at_utc and event_name, unsorted, up to 5 million rows. A session breaks when the gap from that visitor's previous event exceeds 30 minutes, and is force-closed at UTC midnight so no session spans two calendar dates. A gap of exactly 30 minutes does not break. Emit one row per session with session_id, visitor_id, the user_id as of the last event in the session, started_at_utc, ended_at_utc, session_date, duration_seconds and event_count. Vectorise; do not loop per visitor.

Approach

1. Sort by ['visitor_id', 'occurred_at_utc', 'event_id'] once, then express the whole problem as one boolean vector: a row starts a new session when the visitor changed, or the gap exceeds 30 minutes, or the UTC date differs from the previous row's UTC date. Cumsum that vector and you have the session key.
2. Get the comparison direction right on the gap: the rule is strictly greater than 1800 seconds, so an event at exactly 1800 seconds continues the session. Write it as gap > pd.Timedelta(minutes=30), and make the tie a test case rather than an assumption.
3. Derive the midnight break from the date change, not from inserting synthetic boundary rows. A date change implies a break even when the gap is two seconds, which is precisely the force-close rule and is why the two conditions are ORed rather than one subsuming the other.
4. Aggregate with a single groupby on the session key: min and max of occurred_at_utc, size for event_count, and last for user_id, which is correct because the frame is already sorted so 'last' is the final event in the session. That is the identity-as-of-session-end rule.
5. Compute duration_seconds as (max - min).dt.total_seconds(), which makes a single-event session 0 seconds. Say so explicitly, because a downstream mean session duration is sensitive to whether single-event sessions are 0 or excluded.

Follow-up

• Sessions are used as the denominator of a conversion rate. How does moving the inactivity gap from 30 to 45 minutes move that rate, and in which direction?
• A visitor's clock is 40 minutes ahead, so their events arrive with future occurred_at values. What does your sessioniser do, and what would you rather it did?
• The same person signs up mid-session on mobile and continues on desktop. How many sessions and how many users does your output show, and is that the right answer?

Cluster bootstrap for a per-session rate randomised on users

Hard · cluster bootstrap · randomisation unit · variance

An experiment randomised on user_id reports a per-session conversion rate, so sessions inside a user are correlated. Input: one row per session with user_id, variant in {control, treatment} and converted in {0,1}. Write a cluster bootstrap from scratch: resample users with replacement within each arm, keep every session of a drawn user, recompute each arm's ratio of converted sessions to sessions, and take the difference. Return the point estimate, a 95 percent percentile interval from at least 2,000 resamples, the naive session-level interval that ignores clustering, and the ratio of their widths.

Approach

1. Name the estimand precisely: it is a ratio of sums, sum(converted) over sum(sessions) within an arm, not the mean of per-user rates. Those differ whenever session counts vary across users, and the ratio is what the reported metric is.
2. Resample the cluster, not the row. Draw n_users user ids with replacement inside each arm and take every session belonging to each draw, including duplicate draws of the same user. Keeping the user count fixed per arm rather than the session count is what preserves the sampling design.
3. Precompute per-user (converted_sum, session_count) once, so each resample is two vector lookups and a division rather than a repeated filter over the session frame. That turns 2,000 resamples from minutes into under a second.
4. Take the 2.5th and 97.5th percentiles of the 2,000 differences for the interval, and report the point estimate from the full data rather than from the bootstrap mean, since the bootstrap mean carries the resampling bias.
5. Compute the naive interval from the session-level binomial standard error and compare widths. The expected inflation is roughly sqrt(1 + (m-1)*rho), with m the mean sessions per user and rho the intraclass correlation of converted within users, so a computed ratio far from that value points at a bug in one of the two intervals.

Worked solution (35 min)

1. per_user = df.groupby(['variant','user_id'])['converted'].agg(['sum','size']); split into two arrays per arm.
2. point = (t_sum.sum() / t_n.sum()) - (c_sum.sum() / c_n.sum()).
3. For b in range(B): idx = rng.integers(0, len(t_sum), len(t_sum)); ratio_t = t_sum[idx].sum() / t_n[idx].sum(); same for control; store the difference. Vectorise by drawing a (B, n) index matrix if memory allows.
4. ci = np.percentile(diffs, [2.5, 97.5]); naive_se = sqrt(p_t*(1-p_t)/n_sessions_t + p_c*(1-p_c)/n_sessions_c); naive_ci = point +/- 1.96*naive_se.
5. width_ratio = (ci[1]-ci[0]) / (naive_ci[1]-naive_ci[0]).

Expected result: A point estimate identical to the full-data ratio difference, a percentile interval containing that point estimate, and a width ratio greater than 1 whose value approximates sqrt(1 + (m-1)*rho) for the data's mean cluster size m and intraclass correlation rho. A ratio of approximately 1.0 means sessions were resampled instead of users.

Verify it yourself

• Construct data where every user has exactly one session: the cluster and naive intervals then agree to within Monte Carlo error, since m = 1 makes the design effect 1.
• Construct data where converted is constant within each user: rho is 1, and the width ratio should approach sqrt(m).
• The bootstrap distribution's mean sits close to the point estimate; a large offset means the resampled arms have drifted in size.
• Doubling B narrows the Monte Carlo noise in the percentile endpoints without shifting the interval's centre.

Follow-up

• Users average 3.4 sessions and the intraclass correlation is 0.12. What width ratio do you predict before running it, and does your bootstrap land there?
• Give the delta-method standard error for this ratio and say when you would prefer it to the bootstrap.
• Half the users in the treatment arm have exactly one session. What does that do to the cluster bootstrap's coverage, and how would you check it?

Read an experiment from first exposure, not assignment

Hard · experiments · window-functions · srm

fct_experiment_exposure holds experiment_id, unit_type, unit_id, variant, user_id, assigned_at_utc, first_exposed_at_utc, is_in_analysis_population and planned_end_utc. fct_event holds user_id, occurred_at_utc, is_core_action, and carries events up to a known data cut, :data_cut_utc. For one experiment randomised on unit_type = 'user', return per variant: exposed units, units with at least one core action in the seven days after that unit's own first exposure, the rate, and the variant share of exposed units. Only units whose seven-day window has fully elapsed as of the data cut belong in the readout. Units appearing under more than one variant are excluded from both arms and counted separately.

Approach

1. Run the contamination pass as an aggregate, not a window: SELECT unit_id FROM fct_experiment_exposure WHERE experiment_id = :exp GROUP BY unit_id HAVING COUNT(DISTINCT variant) > 1, then anti-join it away. PostgreSQL rejects COUNT(DISTINCT variant) OVER (PARTITION BY unit_id) outright, since DISTINCT is not implemented for window functions; if you want the test inline, MIN(variant) OVER (PARTITION BY unit_id) <> MAX(variant) OVER (PARTITION BY unit_id) is the equivalent that does run.
2. Do not resolve contamination by keeping the earliest variant. A unit that saw both arms carries treatment from both, so assigning it to either one biases that arm.
3. Define the population as is_in_analysis_population = TRUE AND unit_type = 'user' AND first_exposed_at_utc < planned_end_utc AND first_exposed_at_utc + interval '7 days' <= :data_cut_utc. The horizon filter is what makes the readout reproducible next week instead of drifting with every re-run; the data-cut filter is the one that actually buys seven days of follow-up, since a unit exposed an hour before the horizon otherwise contributes an hour of observation to a seven-day rate.
4. Measure the outcome on a per-unit relative window: LEFT JOIN fct_event on user_id with is_core_action = TRUE and occurred_at_utc in [first_exposed_at_utc, first_exposed_at_utc + interval '7 days'). LEFT JOIN so units with no outcome stay in the denominator at zero rather than being deleted by an inner join.
5. Check the sample ratio before reading the effect: variant share of exposed units against the intended split, tested as a binomial. Run it on the truncated population as well as on the full exposed set, because if one arm exposes later on average the data-cut filter removes more of that arm and can manufacture a ratio mismatch the randomisation did not have. A mismatch on the full set means the exposure data is not a valid randomisation and invalidates the readout rather than being a footnote under it.
6. Report the per-variant rate, the absolute difference, and the fact that the variance unit is unit_id. That is straightforward here only because the grain is already one row per user; a per-session outcome under user randomisation would need a delta-method or bootstrap standard error instead.

Worked solution (35 min)

1. Count units per variant and units with more than one variant; report the contaminated count before anything else.
2. Build the exposed population CTE with both the horizon filter and the seven-days-before-data-cut filter, assert one row per unit_id, and record per variant how many units the second filter removed.
3. LEFT JOIN the bounded outcome events and flag conversion with a boolean aggregate.
4. Aggregate per variant to exposed, converted, rate and share, and run the binomial check on the share, both before and after the truncation.
5. Re-run with a fixed calendar window from experiment_start_utc and record the difference as the size of the dilution you avoided.

Expected result: One row per variant with exposed, converted, rate and share, plus a scalar count of excluded contaminated units. Every retained unit was first exposed at least seven days before the data cut, so each one has a full seven days of observation and the rates are directly comparable.

Verify it yourself

• Exposed counts across variants sum to the population count minus the contaminated units.
• MAX(first_exposed_at_utc) among retained units is at least seven days before :data_cut_utc.
• Variant shares sum to 1.
• No unit_id contributes to two variant rows.
• The truncation drops a similar share of each arm; if it does not, note the imbalance alongside the SRM result rather than reading the effect.
• Swapping the relative window for a calendar one changes the rates; if it does not, exposure timing is balanced and you can say so.

Follow-up

• Some units were assigned days before they were exposed. What does analysing the assigned set instead do to the estimated effect, and in which direction?
• The split is 51/49 on 400,000 exposed units. Do you read the result?
• The treatment arm exposes on average two days later than control. What does that do to a fixed calendar outcome window, and which arm does it favour?

Find reactivation gaps in account paid-period history

Hard · gaps-and-islands · window-functions · subscriptions

fct_subscription_period holds account_id, subscription_id, period_start_utc, period_end_utc, period_status and change_reason. A mid-period plan or seat change closes one row and opens another, so a single continuous paid tenure is often many rows, and an account may hold two overlapping subscriptions. Collapse rows with period_status in ('active','past_due') into continuous tenures per account, treating gaps of three days or less as continuous. Return account_id, tenure_start, tenure_end, and for every tenure after the first, the gap in days that preceded it.

Approach

1. Filter to paid rows only: period_status IN ('active','past_due'). Trialing periods are not tenure, and including them turns every trial that never converted into a one-period tenure followed by a fake churn.
2. Order by period_start_utc and take a running maximum of all prior ends: MAX(period_end_utc) OVER (PARTITION BY account_id ORDER BY period_start_utc, period_end_utc, subscription_id ROWS BETWEEN UNBOUNDED PRECEDING AND 1 PRECEDING). Those three columns are the only stable ordering this schema exposes, so check first that they are unique within an account; if rows tie on all three, the island numbering is order-dependent between runs and you need a real row key before the result is reproducible.
3. LAG on its own is wrong here because with overlapping or nested periods the immediately preceding row by start date is not the one that ends latest, so the running maximum is the part that cannot be shortcut.
4. Flag a new island when prior_max_end IS NULL OR period_start_utc > prior_max_end + interval '3 days', then number islands with a running SUM of the flag over the same ordering and an explicit ROWS frame.
5. Group to (account_id, island) taking MIN(period_start_utc) and MAX(period_end_utc), then LAG(tenure_end) OVER (PARTITION BY account_id ORDER BY tenure_start) to compute the preceding gap in days for every tenure after the first.
6. Sanity-check with change_reason, which is the only lineage this schema carries: list its distinct values first, then confirm that rows recording a plan or seat change sit inside a tenure rather than opening one, and that every tenure after the first opens on a row whose reason records a restart rather than an ordinary renewal. Do not reconcile against a churn timestamp on dim_account, which this schema does not define; and where such a column does exist, a cancellation timestamp records when the request was made and routinely sits weeks before the period it ends.

Follow-up

• Why three days of grace? What do 0 and 30 days each do to the count of accounts classed as reactivated?
• An account runs two concurrent subscriptions for different teams. One tenure or two, and what does the revenue reader expect?
• How would you turn these tenures into a monthly gross logo churn series without double-counting an account that churned and returned in the same month?

How do you manage your time and prioritize deliverables when working across multiple client projects simultaneously?

Medium · product sense and metrics

Approach

1. Fix the population and the time window before naming any metric.
2. Restate the decision this analysis has to support, and who acts on the answer.
3. State what result would change your recommendation, so the answer is falsifiable.

Follow-up

• What would you do if the primary metric and the guardrail moved in opposite directions?
• How would you detect that the metric is being gamed rather than genuinely improving?

Tell a novelty effect apart from a durable one

Medium · novelty effect · exposure cohorts · composition

A redesigned navigation shows a 6% lift in weekly core actions in calendar week one, 3% in week two and 1% in week three, on a continuously enrolling user-randomised experiment. Enrolment ran throughout, so later exposure cohorts are younger accounts. Using fct_experiment_exposure (unit_id, variant, first_exposed_at_utc), dim_user.account_created_at_utc and fct_event, determine whether this is novelty decay, a composition artefact, or a real effect that shrank, and state what evidence would settle it.

Approach

1. Separate calendar time from exposure time first. With continuous enrolment, calendar week two mixes units in their second week of exposure with units in their first, and the tenure mix changes daily, so a calendar trend is not evidence of decay.
2. Rebuild the readout as a triangle indexed on days since first_exposed_at_utc. For each exposure cohort compute the treatment effect at exposure-day k, then average across cohorts at fixed k using only cohorts that have reached k. Real decay shows as a falling curve in k; a composition artefact flattens out.
3. Use the one clean discriminator novelty offers: it requires a prior version to be novel against. Split on whether the unit existed before experiment start. Users who signed up afterwards never saw the old navigation, so a stable effect among them is not novelty, and an effect confined to pre-existing users is.
4. Check the primacy direction as well. An early negative effect that recovers is the mirror image and argues for shipping rather than against it, and it is invisible without the same exposure-indexed curve.
5. Say what would settle it. A long-run holdback, a few percent of traffic kept on control for eight weeks or more and read from exposure-week six onward. No three-week test can distinguish a decaying effect from one that has reached a lower plateau.

Worked solution (30 min)

1. Assign each exposed unit an exposure cohort from date(first_exposed_at_utc) and compute treated and control means per cohort per exposure-day k.
2. Average the per-cohort effects at each k only across cohorts that have observed day k, leaving later k blank rather than averaging over whoever has reached it.
3. Split the same triangle on pre-launch versus post-launch account_created_at_utc from dim_user.
4. Compare the curves: decay confined to the pre-launch split is novelty, decay in both is a real effect shrinking, and a flat exposure-time curve under a falling calendar curve is composition.
5. Recommend the holdback and state the earliest exposure-week at which a plateau claim becomes supportable.

Expected result: The decision rests on the exposure-indexed curve, not the calendar one. Novelty is indicated only when decay is present among units that existed before launch and absent among units created after it. Identical decay in both splits means a real effect shrinking; no decay in exposure time means the calendar trend was composition. A three-week test cannot separate decay from a plateau, and a long-run holdback can.

Verify it yourself

• Within a cohort, denominators must not fall as k grows. If they do, the population is being conditioned on survival rather than held fixed at exposure.
• Re-aggregating the exposure-indexed effects with the observed enrolment weights must reproduce the reported calendar-week effects, which proves the two views are the same data.
• Verify with a query that no unit in the post-launch split has a pre-launch session, otherwise that split is not clean and the discriminator is worthless.

Follow-up

• The holdback costs a few percent of traffic on the winning arm. How do you justify that, and when would you retire it?
• Weekly core actions per user is a ratio whose numerator and denominator live at different grains. Which standard error are you computing, and why is the naive one wrong?
• What would you conclude if the effect were 6%, 3%, 1% in exposure time as well, among post-launch signups only?

Attribute a flow-completion drop to one client build

Medium · release-regression · instrumentation · staged-rollout

Core-flow completion rate on surface = 'ios', defined as distinct flow_instance_id reaching 'flow_completed' within 30 minutes of 'flow_started' with no 'error_shown' on the same id in between, fell from 78% to 71% over four days. A new iOS build began a staged rollout on day one. Using fct_event columns flow_id, flow_instance_id, app_version, event_name, user_id and occurred_at_utc, produce the completion rate per build per day, size the loss in absolute completions, and say whether the fix is a rollback. Treat the possibility that the instrumentation changed rather than the flow.

Approach

1. Cut the rate by app_version and day, each build against its own denominator. A pooled series during a staged rollout is a weighted average whose weights move daily, so it declines in proportion to the rollout share even when the old build is perfectly flat; confirming that proportionality is itself the evidence the regression is build-specific rather than environmental.
2. Test the denominator before believing the rate. Compute flow_started events per distinct user_id on each build: if the new build re-mints flow_instance_id on retry, or fails to carry the id from start to completion, the denominator inflates and the numerator falls with nothing changing for the user.
3. Normalise to a user-level outcome that is immune to the id question, namely completions per distinct user_id per day on each build. This is the number that says whether anyone actually failed to finish.
4. Check the error path: pull error_shown counts and their properties for flow_instance_id values on the new build. A genuine regression produces errors; an instrumentation break produces missing completions with no corresponding error volume, and the two prescribe different fixes.
5. Size the real loss as the per-user completion gap multiplied by daily users on the new build, and give the rollback recommendation conditional on the error evidence rather than on the rate.

Follow-up

• The rollout is at 45% and product wants to go to 100% tomorrow. What do you say, and what would you need by when?
• If the flow_instance_id is genuinely being re-minted, what is the correct historical treatment of the four affected days?
• How would you have caught this on day one instead of day four?

Suggested practice checklist

For someone who has spent the last year in notebooks, dashboards or modelling work and has not written raw SQL under time pressure. The first four days rebuild query fluency against a fixture you control and can verify by hand; the last three attach that fluency to the rest of the loop.

Day 1 — Build a fixture you can check answers against

• Create a local Postgres or SQLite database with four tables (users, sessions, events, orders) holding roughly 200 rows you generated yourself, so you know the contents well enough to predict every result.
• Deliberately seed the cases that break queries: a user with no sessions, a session with no events, two orders sharing a timestamp, a NULL in one join key, and one duplicated user row.
• Before writing any SQL, hand-compute five answers on paper (how many users placed at least one order, median orders per ordering user, and three others) and save them as the ground truth for the week.

Deliverable: A one-command seed script plus a text file of five hand-computed answers to grade every later query against.

Day 2 — Joins, filters and NULL semantics

• Answer "which users have no orders" three ways (LEFT JOIN with IS NULL, NOT EXISTS, NOT IN) and confirm that the NOT IN version returns zero rows once the subquery contains a NULL, because the comparison is never TRUE.
• Reproduce the LEFT JOIN that silently collapses to an inner join by putting a right-table predicate in WHERE, then fix it by moving the predicate into the ON clause, and record both row counts.
• Create a fan-out bug on purpose by joining orders to order_items and summing the order total, then correct it with a pre-aggregated subquery and explain in one line which table changed the grain.

Deliverable: One annotated .sql file holding the three join traps, each with the wrong result and the corrected result side by side.

Day 3 — Window functions and frames

• Write three window queries against the fixture: a running order total per user, the rank of each order within its user by value, and the day gap to that user's previous order, then check each against the day-one ground truth.
• Run ROW_NUMBER, RANK and DENSE_RANK over a column containing ties, print all three side by side, and write one sentence on when each is the correct choice.
• Switch one query from the default frame (RANGE BETWEEN UNBOUNDED PRECEDING AND CURRENT ROW, which is what you get when ORDER BY is present and no frame is written) to ROWS BETWEEN UNBOUNDED PRECEDING AND CURRENT ROW, and explain why the output differs only when the ORDER BY column has duplicates.

Deliverable: Three verified window queries plus a short note explaining the RANGE versus ROWS difference in your own words.

Day 4 — The four analytical query patterns

• Write a monthly retention grid: first order month per user, then months-since-first as the column, and verify that month zero equals the cohort size exactly.
• Sessionize the events table under a 30-minute inactivity rule using LAG plus a cumulative sum over a new-session flag.
• Build a four-step funnel that counts distinct users rather than events at each step, and state the rule you applied to a user who reaches step three without ever logging step two.

Deliverable: One file with the retention, sessionization and funnel patterns, each carrying a one-line note on the assumption it bakes in.

Day 5 — Write SQL the way you will have to write it live

• Set a 12-minute timer and solve three medium prompts in a plain editor with no execution and no autocomplete, then run them and tally syntax errors separately from logic errors.
• Narrate one solution aloud while writing it, stating the grain of each intermediate result (one row per user, one row per user-day) before you type its body.
• Rewrite your slowest solution as a CTE chain where every CTE name states its grain, and time yourself re-solving it from blank.

Deliverable: A recording of one narrated solution plus an error tally that separates syntax from logic.

Day 6 — One day for everything that is not SQL

• Write the preconditions of the two-sample t-test from memory, then check them: independent observations, and a difference in means whose sampling distribution is approximately normal, which at large sample sizes follows from the central limit theorem rather than from normality of the raw values.
• Write the difference between an odds ratio from logistic regression and a relative risk, and state the condition under which the two are close (low outcome prevalence).
• Prepare a 90-second answer to "how would you know this model is any good" that names the metric, the baseline you would beat, and the cost of the errors you care about.

Deliverable: One page of notes covering test preconditions, the odds-ratio caveat and the model-quality answer.

Day 7 — Full loop rehearsal

• Run a 45-minute mock with someone willing to interrupt: 20 minutes of SQL, 15 minutes defining a metric, 10 minutes on a past project.
• Re-solve from blank the two queries you were slowest on this week and compare the times against day five.
• Write a five-line answer to "walk me through a project" that puts a number in the first sentence and names the decision the work changed.

Deliverable: Mock feedback notes plus a timed project narrative you can deliver without reading it.

Bring numbers, including the ones you are unsure of

An answer without a quantity is hard to interrogate, so interviewers keep probing until they find one. Come with the baseline, the change, the window it was measured over, and how confident you were. If the effect never got measured, say so and say what you would have measured. Fabricated precision is worse than an honest gap.

Why are you applying for the Data Scientist position at Atos, and how does it align with your long-term career goals?

Medium · behavioural and stakeholder questions

Approach

1. Close with what you would do differently, concretely.
2. State the situation in two sentences and spend the rest on your reasoning.
3. Name the disagreement or constraint, and how you resolved it with evidence.

Follow-up

• What would you do differently if you ran that project again?
• What did you decide not to do, and why?

How do you handle a situation where a client or stakeholder disagrees with your analytical approach or model results?

Medium · behavioural and stakeholder questions

Approach

1. Name the disagreement or constraint, and how you resolved it with evidence.
2. State the situation in two sentences and spend the rest on your reasoning.
3. Close with what you would do differently, concretely.

Follow-up

• What did you decide not to do, and why?
• How did you know the outcome was caused by your change?

Tell me about a recent professional success and a notable failure. What did you learn from each?

Medium · behavioural and stakeholder questions

Approach

1. Close with what you would do differently, concretely.
2. Name the disagreement or constraint, and how you resolved it with evidence.
3. Pick a story where you drove the decision, not one where you observed it.

Follow-up

• How did you know the outcome was caused by your change?
• What would you do differently if you ran that project again?