Preparing content…

An analytics chart on the left and a linked data pipeline on the right
职业发展

Data Analyst vs. Data Engineer: Skills and Portfolio Projects

Compare metrics, reports, data pipelines and recovery checks with one dataset. Build an analysis or engineering sample before committing to a career move.

By: FermatMind Editorial

Published: Nov 17, 2026

Updated: Nov 17, 2026

25 min read

Content review completed · Nov 17, 2026

Subscribe to the English blog (Atom) · Add this link to an RSS reader that supports Atom for newly published articles and substantive updates.

A data analyst explains what changed and what a business team should check next. A data engineer builds the pipeline that keeps the underlying data reliable, traceable, and safe to rerun. Both may use SQL. To choose a direction, try the same orders dataset twice: deliver a metric explanation, then build a repeatable transformation.

Data analystData engineer
Typical outputMetrics report, possible explanations, next stepsProcessed tables, validation rules, alerts, rerun logs
Core skillsDefine metrics, check SQL results, explain them to business teamsTransform data with SQL or scripts, design tables, handle duplicates and errors
Sign of good workSomeone can reproduce the metric and act on itSomeone can rerun or recover the pipeline without double counting
First exerciseExplain the metrics in one set of ordersDeduplicate those orders and verify a second run
From an adjacent roleTurn an operations, sales, or finance report into a complete analysisTurn a development, testing, or automation task into a recoverable pipeline

Pick a first portfolio piece by reading job postings

Start with shared skills. Practice SQL grouping and joins, missing values, and duplicate records. For analysis, show how to calculate a business metric, what a change tells you, and what to check next. For engineering, document the inputs, expected output, and recovery steps so someone else can run it.

Compare two jobs you can actually apply for. Match target city, seniority, and hiring stage. Highlight required and preferred qualifications and the main output. If you already make reports, add a business explanation. If you develop or test software, add a rerun and recovery example. Duties can overlap. Ask: “What is a new hire's first deliverable? When numbers don't match, who explains them and who fixes the pipeline?”

A past mainland Chinese campus posting by Shanshu Technology has a data analysis entry covering business understanding, processing source data, analyzing results, recommending changes, and supporting engineering. An experienced-hire posting from ECCOM Network in Shanghai has a data development entry covering collection, processing, modeling, data services, and automation. Compare the responsibilities, without treating these different hiring stages as evidence about pay or ease of entry. For scope, see IBM's overview of data engineering.

Shared input: Duplicate orders, missing regions, and refunds across days

FermatMind Editorial created this fictional exercise. It contains no customer data. Times use Beijing time, and D1 and D2 are consecutive days. Each payment or refund event has a business key. Replayed inputs can repeat that key. The deduplicated output has one record per key. Use the arrival-time column to determine when each record became available and the event-time column to determine when payment or refund occurred.

RecordEvent timeArrival timeAmountRegionNote
Payment O101D1 10:00D1 10:01120East ChinaPaid
Payment O101 replayD1 10:00D1 10:03120East ChinaIdentical to original
Payment O102D1 11:00D1 11:01100North ChinaPaid
Payment O103D1 16:00D1 16:0180MissingPaid
Payment O104D2 09:00D2 09:01200East ChinaPaid
Refund R01 for O101D2 10:00D2 18:0030From original orderSuccessful, late arrival
Refund R02 for O104D2 11:00D2 11:0150From original orderSuccessful

The business asks: How much net payment value remains for orders grouped by their original payment date? Count a duplicate payment once, keep a missing region as “Unknown” in the sitewide total, and attribute a successful refund to the date the original order was paid. Net payments equal payments minus successful refunds that have arrived by the snapshot and been matched to their original paid order. The exercise assumes one valid payment per order. With multiple payments, confirm the transaction key and aggregation level first. Deduplicate refunds by event key too. Deduct the amount of a partial refund, never the full order value.

If finance asks how much money was returned today, group by when refunds succeeded and state the visible snapshot. Compensation, settlement, and other metrics need their own definitions. Both roles need to agree on the business question; a tool cannot make that choice for them.

The analyst's deliverable: A report and an explanation people can act on

Deduplicate and summarize successful refunds using the same D2 20:00 snapshot.

Original payment dayUnique paid ordersPaymentsSuccessful refunds receivedNet payments
D1330030270
D2120050150
Sitewide450080420

D1 payments are 120 + 100 + 80 = 300. Counting the O101 replay would give 420, overstating payments by 120. The D1 breakdown is 90 for East China, 100 for North China, and 80 for Unknown, still totaling 270. Filtering out Unknown leaves only 190. Use 190 as the denominator only if the report is explicitly about the known-region breakdown, rather than the entire site.

The analyst's note might say: “D1 net dropped from 300 in the noon snapshot to 270 in the evening. The original order's refund succeeded at 10:00 on D2, but only arrived at 18:00. This is a revision after backfill. The 30 yuan decrease doesn't establish a sudden drop in D1 sales. Check how often data arrives late and whether missing region values cluster at one entry point.” In a chart alone, a correction to history can look like real-time deterioration.

There was no random assignment or enough data to establish a cause. D1 and D2 could differ in order volume, channel mix, or how complete each observation window was. Check the anomaly with the orders or refunds team, and expand the sample and observation period before making larger decisions.

The engineer's deliverable: Run the same inputs twice without double counting

Save the raw snapshot and run timestamp first. Identify duplicates by the payment or refund key. If key and essential fields match, count one event. If the key matches but amount or state conflicts, use the upstream contract to distinguish a valid correction from a conflict. Flag unresolved records and alert the source owner. The last row to arrive is not automatically right.

A refund needs its original paid order. If R01 arrives before that order, hold it as unmatched, deduct nothing yet, and make the quality status visible downstream. Recompute the original payment day when you find the order. Stamp each output with its snapshot time and version so users can distinguish noon results from evening backfill.

TestExpected output
D2 12:00 initial runR02 visible, R01 absent; D1 300, D2 150, sitewide 450. Show the snapshot time and possible late refunds.
D2 20:00 updateR01 matches O101. Backfill D1 to 270. D2 stays 150, for a sitewide total of 420.
Same inputs rerunO101 and R01 count once even if replayed; the same snapshot stays at four orders and net 420.
Missing regionKeep O103 and its 80 in the sitewide total. Report Unknown = 80 and log the missing value.
Conflicting key or upstream not readyFlag the issue and notify the owner. Do not label a partial refresh as a completed run.

Use these as pre-release tests. When only R01 arrives, only D1 metrics change. Preserve distinct publication dates, data dates, and run IDs. The noon report should remain traceable so business users can still see why it said 450. Engineering makes the run, results, and recovery visible. Analysis explains what those states mean for the reader.

When the refund arrives first, what should each role tell downstream users?

The initial exercise used a late refund to revise an earlier day downward. Now add a separate fictional stress test: late on D2, a successful 20-yuan refund R03 arrives for O105 before the payment record. On D3 morning, O105's valid 70-yuan payment arrives, dated late on D2 and assigned to North China. This branch adds to the original example without changing its seven rows.

Data availableD1 netD2 netSitewide netHow to read this version
D2 at 20:00, original release270150420Neither R03 nor O105 is present.
Late D2, R03 arrives but O105 is missing270150420One unmatched refund of 20 yuan. No payment date can yet be assigned.
D3 morning, payment and refund match and are backfilled270200470Add 70 yuan in D2 payments and 20 in counted refunds, a net increase of 50.

The last row is reproducible: the original D2 net was 200 − 50 = 150 yuan; the new pair adds 70 − 20 = 50, giving 200. Sitewide payments become 570 and counted refunds 100, leaving 470. A successful refund status alone does not identify an order, region, or payment date. If engineering simply deducted R03 on D2, users would see 130 then 200, an unexplained dip unrelated to trading. A separate report grouped by refund date may count it under its own definition. That does not change the order-based metric here.

The analyst should give both what is verifiable and what remains missing: “The D2 20:00 report showed 150 yuan using matched data available then. A late payment and its matched refund now bring D2 net to 200 yuan on the original payment date. If reviewing yesterday's decision, keep the 150 yuan snapshot and explain the 50 yuan in the backfill note.” Readers can distinguish what informed their decision and what belongs in a retrospective total.

The engineer should show why two reruns cannot turn a 50-yuan increase into 100. Record input snapshots, unmatched records, affected partitions, reconciliation results, and release version. On D3 recompute only D2; D1 remains 270 yuan. If the same payment or refund key returns, check for a duplicate rather than adding it again. Both snapshots must point to what was visible at the time. If the new amount conflicts with an existing record, verify whether it is an authorized correction before replacing it; arrival order is not enough.

Four questions before handing off your portfolio piece

  1. What if the user asks “How much did we refund yesterday?” Keep the original payment-date definition in the report title and specify the fields and other table needed for refund-date reporting. Identify whether they want to review refund timing or revenue on original orders.
  2. Could another person recover the last row if you hid it? Have them find the late payment, refund match, and old snapshot, and calculate the 50-yuan change and 470-yuan total. Put any missing explanation in the notes or acceptance criteria.
  3. What should people see if the backfill fails? Keep the last complete release, and show refresh failure, data timestamp, and unmatched status. Agree on who investigates missing data and when to update users about the next refresh.
  4. How would you show your part in a portfolio? Analysis should show the business question, definition, two snapshots, and recommendation. Engineering should show inputs, rules, expected outputs, failure paths, and proof of a rerun. Both may include SQL. Show how someone can accept and take over your particular output.

Compare real job descriptions and try both tasks for 20 minutes

  • Analysis. Use the input to produce a D1/D2 metrics table and a short memo (80–120 Chinese characters in the original exercise) explaining the definition, revision, and first thing to investigate. Ask someone what number they might misread, then revise the title or notes.
  • Engineering. Split payments and refunds. Write five processing rules and four assertions on paper, in a spreadsheet, or in SQL. Ask someone to rerun identical inputs, supply a refund a day late, or introduce a key conflict, then check whether your plan explains every outcome.

Notice which rework you would do again. If you like figuring out why the business changed and explaining what the data warrants, explore more analysis roles. If you enjoy fixing occasional failures, revising old data, and delivering a repeatable process, explore engineering. If both appeal to you, ask what new hires maintain, who takes care of upstream gaps, and whether report owners also explain results.

Coming from another major? Use the job-description and skills-evidence checklist to turn exercises into portfolio work. If you enjoy troubleshooting and verification, try the troubleshooting and handoff exercises to see what you would repeat. The Business Intelligence Analysts profile gives an adjacent analysis example for defining metrics and checking results. The Data Warehousing Specialists profile is one engineering reference for modeling, maintenance, and handoff. If you are still deciding, use the 94-question FermatMind career assessment to organize preferred activities, work styles, and job conditions. Compare two real postings, then try the 20-minute exercises above.

Frequently asked questions

Do both data analysts and data engineers use SQL? Yes. An analyst might join orders and refunds to answer a question and check the denominator against source records. An engineer might use SQL in a transformation and add checks for repeat records, missing fields, and late events. Check the posting for the actual tools and the outputs you would own.

Can I move from data analyst to data engineer? Start with a report you can already explain. Build the upstream transformation, run it twice on the same inputs, and then add a late record and a duplicate. Document how the result changes and how report users find out. Compare the exercise with engineering job postings to see what orchestration, testing, or operational skills you still need.

What is a good portfolio project for a first data role? Choose one small dataset you can share. For analysis, submit the metric definition, query, and a concise recommendation. For engineering, submit the transformation, data-quality checks, rerun instructions, and a record of what happened when a record arrived late.

Related tags

职业发展

Go to topic hubs

Continue from the article into a more structured topic entry surface.

Take the test

If you want to turn reading into self-measurement, continue into an assessment.

About FermatMind

FermatMind helps young people everywhere understand themselves and explore careers. We started in the Chinese-speaking world. Explore your preferences with MBTI and your interests with the Holland assessment, then use more than 1,000 occupation profiles to compare tasks, skills, and entry paths and find a next step worth trying.

Data Analyst vs. Data Engineer: Skills and Projects | FermatMind