DataX common mistakes
DataX · Understand the traps and how to avoid them.
When asked to classify a data asset, the candidate assigns a label based on the department that owns the system rather than on the asset's content, sensitivity, and regulatory category.
Candidates conflate organizational ownership with data classification and rely on the source system name (e.g., 'HR system') instead of inspecting the actual fields and their sensitivity.
What to do instead
Practice classifying sample datasets by examining column names, sample values, and business context; use a checklist of classification criteria (PII, PHI, financial, public) and apply it to the data itself, not to the system owner.
When drawing an ER diagram or normalizing a schema, the candidate stops at third normal form without considering query patterns, and later cannot justify denormalization for performance.
Candidates memorize normalization rules as an absolute goal and overlook that modeling decisions depend on access patterns, latency requirements, and storage trade-offs.
What to do instead
For each modeling scenario, explicitly state the expected query patterns and justify each normalization or denormalization step with a performance or integrity rationale; practice comparing 3NF and star-schema designs for the same business case.
When designing an ETL/ELT pipeline, the candidate chooses between ETL and ELT based on tool popularity rather than on data volume, transformation complexity, target platform capabilities, and governance needs.
Candidates learn ETL and ELT as competing buzzwords and fail to map the architectural trade-offs (e.g., ELT leverages warehouse compute but requires strong governance) to concrete requirements.
What to do instead
Build a decision matrix with criteria such as data volume, transformation complexity, latency, cost, and governance; for each practice scenario, score both approaches and defend the choice in writing.
When asked to handle missing values, the candidate immediately deletes rows with nulls or imputes the mean without first investigating the missingness mechanism (MCAR, MAR, MNAR).
Candidates treat missing data as a purely mechanical cleanup step and skip the diagnostic step of determining whether missingness is random or related to other variables.
What to do instead
Before choosing a treatment, compute missingness rates per column, test for relationships between missingness and other variables, and then select deletion, imputation, or flagging with a documented rationale.
When interpreting a p-value or confidence interval, the candidate states that there is a 95% probability the true parameter lies in the computed interval, or that a p-value is the probability the null hypothesis is true.
Frequentist interpretations are counterintuitive, and candidates rely on colloquial phrasing from popular media rather than the formal definition of repeated-sampling coverage.
What to do instead
Memorize and practice writing correct frequentist statements: a 95% CI means that 95% of intervals constructed this way would contain the true parameter; a p-value is the probability of observing data at least as extreme assuming the null is true. Drill with varied scenarios.
When selecting a chart type for a dashboard, the candidate uses a pie chart for more than five categories or a 3D chart for part-to-whole comparisons, and does not order categories by value.
Candidates prioritize visual appeal over perceptual accuracy and are unaware that angle and area judgments are less precise than position and length judgments.
What to do instead
Apply a chart-selection heuristic: use bar charts for comparisons, line charts for trends over time, and pie charts only for a few parts of a whole; always sort categories by value and avoid 3D effects. Practice critiquing sample dashboards against these rules.
When asked about data privacy, the candidate treats pseudonymization and anonymization as interchangeable and assumes that hashing direct identifiers alone makes a dataset non-personal.
Candidates learn the terms superficially and do not consider re-identification risks from quasi-identifiers (e.g., age, ZIP code, gender) or from linking attacks.
What to do instead
Study the distinction: pseudonymized data remains personal data because re-identification is possible with additional information; anonymization requires that re-identification be irreversibly prevented. Practice assessing datasets for quasi-identifier combinations and applying techniques such as generalization, suppression, and k-anonymity.
When comparing batch and streaming ingestion, the candidate recommends streaming for every use case because it is 'more modern', without evaluating latency requirements, cost, and complexity.
Candidates associate streaming with being advanced and overlook that most analytical workloads tolerate batch latency and that streaming adds operational overhead.
What to do instead
For each scenario, derive the required latency from the business question (e.g., real-time fraud detection vs. monthly reporting), then compare batch, micro-batch, and streaming on cost, complexity, and freshness; justify the choice explicitly.
When asked to define a data quality dimension, the candidate lists dimensions such as accuracy, completeness, and timeliness but cannot give a measurable operational definition or a threshold for each.
Candidates memorize dimension names without linking them to concrete metrics, so they cannot translate a quality requirement into a testable rule.
What to do instead
For each dimension, write an operational definition and a measurement formula (e.g., completeness = non-null records / total records, threshold 99.5%); practice mapping business requirements to measurable quality rules and alert thresholds.
When designing a data warehouse, the candidate mixes grain levels in a single fact table (e.g., combining daily and monthly sales rows) and then cannot reconcile aggregate totals.
Candidates focus on including all desired metrics and lose track of the declaration of grain, which is the most important design decision in dimensional modeling.
What to do instead
Always declare the grain of the fact table in one sentence before adding dimensions or measures, and enforce it by ensuring every row represents exactly one instance of that grain; practice by reviewing sample fact tables and identifying grain violations.