roadmap
- 01
- 02
- 03
- 04
- 05
- 06
- 07
- 08
Answer pandas interview questions by thinking in whole columns, not row loops. Covers DataFrames and selection, missing data, groupby and merges, reshaping, time series and memory.
416 of 416 questions in your plan: 0 known, 0 in progress, 0 due, 0 skipped, 416 new. 0 cut, outside your plan.
Questions on pandas usually arrive with a notebook attached: the hands-on round of data analyst, data scientist and data engineer loops, or a take-home assignment that hands over a DataFrame and a business question such as revenue per customer per month. Interviewers notice the reflex before any follow-up. A loop over iterrows, or an apply wrapping plain Python where a column expression would do, says more than the answer it produces.
The follow-ups are built from results that look right and are not: a merge that returned more rows than the left table had, an ID column read as integers that lost its leading zeros, an assignment through a chained selection that never reached the original frame, a pivot that refused duplicate entries where pivot_table would have averaged them silently.
Levels separate on suspicion more than on vocabulary. An early-career candidate reaches the number; a stronger one names what they checked before believing it: key uniqueness before the join, the row count after it, the dtypes after the read. Size enters at senior level: why a frame can occupy far more memory than the file it came from, and which aggregates can be computed one chunk at a time.
The plan opens below the library, on what a table is: typed columns held together by row labels that operations align on, and what each column type costs. Many later surprises are type or alignment surprises in disguise, such as an integer column that turned to floats when one value went missing. NumPy comes next, the array layer pandas is built on, whose fixed-width integers have no slot for a missing value. Then comes pandas itself, the calls a hands-on round expects you to type without looking up. The modules after it are phrased in terms of tables rather than one library's methods, and return to each area the API introduced to ask why it behaves as it does.
Selection and vectorization pair up because one produces the column the other operates on. The trouble spots are the seam between label and position, whether a selection shares memory with its source, and where whole-column operations stop paying. Absent values get a module before aggregation, since the default that skips them moves the denominator of every mean computed next.
Grouping and combining tables carry most hands-on rounds, and the combining topics are organised around failures, from the join that multiplied rows to keys that almost matched. Reshaping follows because a pivot is a grouping laid out differently, which is how one can aggregate without being asked; time-indexed data extends the idea, resampling being grouping into time bins.
File I/O sits late on purpose: the reader's guesses about types and missing markers make sense only once those ideas are in place, and its chunked pass leads into scale. The plan closes on trusting the result, with notebooks that cannot be rerun and checks that find where records were dropped.
For someone who writes SQL daily, most of the plan translates on sight, so a short runway belongs where the translation fails. A `merge` matches null keys to each other where a SQL join would not, `groupby` leaves null-key rows out of its result by default, and arithmetic between two Series lines up on labels, not row position. A week out, the one change worth making is to raise importance to must know, leaving study order on by frequency, where it opens; level can keep its full span, since one hands-on round mixes a selection question with a cardinality follow-up. A longer runway is better spent reproducing than reading: work the modules in printed sequence, study order on by topics, with a notebook open on a public dataset, and make each silent failure happen before reading its answer. Data engineers and senior candidates should not skim the API module: its reading, memory and speed topics hold the calls their follow-ups turn on, and the I/O, scale and correctness modules then explain why those calls behave as they do. The plan stops at the table: pandas' own plotting calls are in it, chart design is not, and neither are statistical inference or model fitting. Past the point where one machine is not enough, the [Apache Spark roadmap](/roadmaps/spark) takes over. For the full analyst loop, where pandas shares the hands-on round with SQL and the case discussion leans on statistics, the [Data Analyst roadmap](/roadmaps/data-analyst) places this plan among them.
For someone who writes SQL daily, most of the plan translates on sight, so a short runway belongs where the translation fails. A merge matches null keys to each other where a SQL join would not, groupby leaves null-key rows out of its result by default, and arithmetic between two Series lines up on labels, not row position. A week out, the one change worth making is to raise importance to must know, leaving study order on by frequency, where it opens; level can keep its full span, since one hands-on round mixes a selection question with a cardinality follow-up.
A longer runway is better spent reproducing than reading: work the modules in printed sequence, study order on by topics, with a notebook open on a public dataset, and make each silent failure happen before reading its answer. Data engineers and senior candidates should not skim the API module: its reading, memory and speed topics hold the calls their follow-ups turn on, and the I/O, scale and correctness modules then explain why those calls behave as they do.
The plan stops at the table: pandas' own plotting calls are in it, chart design is not, and neither are statistical inference or model fitting. Past the point where one machine is not enough, the Apache Spark roadmap takes over. For the full analyst loop, where pandas shares the hands-on round with SQL and the case discussion leans on statistics, the Data Analyst roadmap places this plan among them.