Data cleaning skills
employers test in 2027

Real data is messy, and decisions built on messy data are wrong. That is why almost every analyst, BI, data engineering and data science interview includes data cleaning — sometimes as explicit questions, more often hidden inside a take-home.

The work is also a large share of the job. In Anaconda's 2022 State of Data Science survey, respondents reported spending about 38% of their time on data preparation and cleansing;¹ in the 2020 edition, loading and cleaning together took around 45%.² Being fast and careful at it is one of the most practical ways to stand out.

Time spent on data preparation and cleansing (Anaconda, 2022)

~38%

Loading plus cleaning, Anaconda 2020 survey

~45%

Typical data analyst pay in India (AmbitionBox, Oct 2026)

₹6.5–7.2 lakh

Typical data engineer pay — where cleaning becomes pipelines

₹11.5–12.7 lakh

Table: Data cleaning by role (editorial summary)

Table: How data cleaning is tested in hiring (editorial summary)

"How do you approach a new, messy dataset?"

Profile it, list issues, agree definitions with the question in mind, clean in reproducible steps, validate counts and totals against a known number, and document every decision.

"A column has 30% missing values. What do you do?"

Find out why (system change, optional field, data not collected). If it's crucial and not random, flag and analyse separately; if not needed, exclude the column; if filling, justify the method and show sensitivity.

"How do you find duplicate customers with slightly different names?"

Standardise (case, spaces, punctuation), compare on stable keys like phone or email, use fuzzy matching for names with a review step, and decide rules for which record wins.

"After joining two tables, revenue doubled. Why?"

Duplicate keys causing row multiplication. Check key uniqueness and row counts before and after; aggregate before joining.

"How do you handle dates from different time zones?"

Store in UTC with time zone information, convert consistently for reporting, and check day boundaries.

"When would you remove an outlier?"

When it's a confirmed error. Otherwise keep it, analyse separately or use robust statistics, and say how it affects results.

"How do you make cleaning repeatable?"

Scripts or queries instead of manual edits, version control, tests on key assumptions and a README of rules.

"How would you stop bad data reaching a dashboard?"

Validation rules at ingestion, automated tests (not null, unique, accepted values, freshness), alerts and a quarantine for failing records.

Table: How data preparation has changed (editorial summary)

• A cleaning notebook or script on a genuinely messy public dataset.
• A data-issues log — each problem, decision and impact.
• Before-and-after checks — row counts, totals, distinct values.
• One automated test that would catch the problem next time.
• A verified data cleaning score to link.

1. Week 1: Profiling — Profile three messy datasets in SQL and pandas.
2. Week 2: Missing values and duplicates — Practise strategies and their trade-offs.
3. Week 3: Formats, text and dates — Standardise categories, units, time zones.
4. Week 4: Joins and validation — Key checks, row counts, reconciliation; take the Data Cleaning and Preparation Certification.
5. Week 5: Repeatability — Turn your cleaning into a script with tests; publish it with a data-issues log.

Related tests: Python for Data Analysis, Data Quality and Governance and SQL Fundamentals. For tools, OpenRefine is a useful free option for one-off messy datasets.⁵

Here is what the top of a strong take-home looks like for the food delivery dataset described above.

Table: An example data-issues log for a take-home (illustrative)

A table like this takes ten minutes and tells reviewers you noticed every planted problem, handled each deliberately and know how it changed the result.

1. Profile three datasets — For each, list row counts, nulls per column, distinct values and suspicious ranges before touching anything.
2. Name standardisation — Take a list of company or restaurant names with variants and standardise them with rules plus a fuzzy-match review.
3. Time-zone fix — Convert mixed UTC and IST timestamps and see how daily totals change.
4. Join audit — Join two tables, check key uniqueness and row counts, and write down any rows lost or duplicated.
5. Automate one check — Turn your most important cleaning rule into a test that fails loudly.

Careful data preparation underpins every data role, and it becomes a career of its own in engineering and governance. Typical pay on AmbitionBox, October 2026:

Table: Roles built on data preparation and their typical pay in India (AmbitionBox, October 2026; indicative)

• Answering before profiling.
• Silent drops of rows with no record.
• Fixing in Excel by hand with no trail.
• Outliers removed because they look odd.
• No reconciliation to a known total.

Time-allocation figures come from Anaconda's State of Data Science surveys as reported by Predictive Analytics World (2022) and HPCwire (2020). Tool references are pandas, dbt and OpenRefine documentation. Pay figures come from AmbitionBox (read October 2026). Test formats and the history table are editorial summaries. Science of Apps publishes the free certification linked here. Last reviewed October 2026.

• Data profiling — summarising a dataset's structure, values and quality before using it.
• Imputation — filling missing values with estimates.
• Deduplication — finding and resolving repeated records.
• Outlier — a value far from the rest of the data.
• Validation rule — a check data must pass, such as "order total is never negative".
• Data observability — monitoring data freshness, volume and quality continuously.