Data engineering interview prep

Walk into your
data engineering
interview ready.

Fundamentals they grill. System design for the platform round. Live incident sims for the senior judgment they probe.

7 layers of depth · system-design method · new reps weekly

System design Try a sim

No credit card · sign in with Google or GitHub

The platformdata flows up
L7∿Semantic & Metrics Layermetric drift · KPI delta -34%1322
13 courses→2 sims→2 design→
L6◈Query Engines & OLAPTrino · ClickHouse · Druid708
7 courses→8 design→
L5⚙Orchestration & PipelinesAirflow · SLAs · lineage1125
11 courses→2 sims→5 design→
L4⚡Compute EnginesSpark · Flink · DuckDB3511
3 courses→5 sims→11 design→
L3⬡Open Table FormatsIceberg · Delta · Hudi427
4 courses→2 sims→7 design→
L2⇄Ingestion & TransportKafka · CDC · backfills11113
11 courses→1 sim→13 design→
L1◼Storage & File FormatsParquet · S3 · GCS235
2 courses→3 sims→5 design→
ParquetKafkaDebeziumIcebergDelta LakeHudiSparkFlinkDuckDBAirflowTrinoClickHouseDruidPinotSnowflakedbt
Three surfaces · every round covered

Learn it, design it, then defend it under pressure.

A data engineering loop tests three things. Know the fundamentals cold, design the platform on a whiteboard, and reason through a live incident the way a senior does - one surface for each.

The internals interviewers probe, down to the bytes.

Annotated source dives, failure-mode walkthroughs, and Docker labs you run locally - from storage up to the semantic layer. When they ask why, you answer from the mechanics, not a tutorial you half-remember.

  • Answer 'why', not just 'which tool'
  • Docker labs on your machine
  • Storage → semantic, every layer
Browse 59 courses
parquet_footer.py · annotated source
# the footer holds the stats that let a reader skip row groups
for rg in footer.row_groups:
    stats = rg.column("event_ts").statisticsmin/max → predicate pushdown
    if stats.max < lo or stats.min > hi:
        continue  # whole row group skipped, zero I/O
This is why sorted data scans 60-80% cheaper - and why an unsorted write quietly kills it.
docker compose up · run the same query locally, watch the bytes read
See where you stand

Answer these five phone-screen questions.

Anyone can pick the right option. The onsite is the three “why”s that come after it - the failure mode, the tradeoff, the line you'd defend to a principal. Clear the screen here. The real bar sits a lot deeper.

Two answers · one question

When they dig in, which answer do you give?

Interviewers push on four moments every data engineer hits — a page, a bill, a bad number, a design review. One answer guesses and hopes. The other reads the mechanics and knows exactly why. Pick one.

PageFlink throughput just dropped 60%
P2 · on-call: you
The answer that guesses
  1. Scale up the task managers
  2. Restart the job
  3. Post "looking into it"
  4. 2h in — still red
The answer that knows why
  1. Read the network-buffer metrics
  2. Backpressure isolates to one operator
  3. Patch the serialization hot path
  4. Resolved — 19 min

You don't scale blindly. You read the buffers and fix the one operator that's choking.

Faster resolution

Every moment above is drawn from a real production incident in the curriculum — the depth that turns a shaky answer into a confident one comes from building the systems yourself in Petascale Labs.

Free to start · new reps every week

Walk in ready. Prep every round.

One subscription unlocks all three surfaces. Know the fundamentals, design the platform, reason through the incident. No credit card to start; cancel anytime.

See full pricing →