Skip to content

DE Mastery GitHub

Start

Curriculum
Roadmap
How to study with this course

Tracks

FND Foundations of distributed data

Overview
01 Why distributed?
02 Partitioning & shuffles
03 Replication & consensus
04 Time
05 Delivery semantics
06 Storage formats
07 Back-of-envelope estimation

SPK Apache Spark

Overview
01 Spark architecture
02 Lazy evaluation and the DAG
03 The DataFrame API
04 Spark SQL and data sources
05 Catalyst and the physical plan
06 Joins and aggregations at scale
07 Partitioning, bucketing, and the small-file problem
08 Memory, caching, and Adaptive Query Execution
09 UDFs and the Pandas API
10 Structured Streaming I — the micro-batch engine and exactly-once
11 Structured Streaming II — watermarks, state, and streaming SCD2
12 Testing and production Spark
BYO Build your own mini columnar query engine

LKH Lakehouse & open table formats

Overview
01 The Iceberg metadata tree
02 Commits & concurrency
03 Catalogs, branching & write-audit-publish
04 Row-level changes — SCD2, CoW & MoR
05 Table layout & maintenance
06 Delta Lake
07 Hudi & the format landscape
08 Engines on the lake
BYO Build your own mini table format

JVM Java for data engineers

Overview
01 Java for Python developers
02 Collections and streams
03 Build and dependencies
04 Serialization
05 Concurrency and the JVM
06 Testing and debugging

KFK Apache Kafka

Overview
01 The log — events, state, retention and replay
02 Topics, partitions, replication — leaders, ISR, acks, durability
03 Producers — batching, idempotence, ordering under retries
04 Consumers & groups — manual commits, at-least-once, no loss across crashes
05 Transactions & exactly-once — read-process-write, read_committed, fencing
06 Schema Registry & contracts
07 Kafka Connect & CDC
08 Kafka Streams
09 Event & topic design
10 Testing & observability (apps)
11 Cluster configuration & partition reassignment
12 Security — SASL/SCRAM authentication & ACL authorization
13 Deployment architecture — multi-region DR
14 Operating & troubleshooting a cluster
BYO Build your own mini Kafka

FLK Apache Flink

Overview
01 Philosophy & architecture
02 DataStream API
03 Time & windows
04 Flink SQL & dynamic tables
05 Joins — regular, interval, temporal (point-in-time) & lookup
06 State, checkpoints & exactly-once — surviving a TaskManager crash
07 Process functions — timers, keyed state & broadcast rules
08 Table API & PyFlink — one pipeline, two languages
09 Streaming patterns — dedup, top-N, split/merge, MATCH_RECOGNIZE, OVER
10 CDC & lake sinks — Debezium changelog to Iceberg, commit on checkpoint, schema evolution
11 Performance & state at scale — backpressure, parallelism, redistribution, skew
12 Deploy & operate — upgrade a stateful job via savepoint without losing state
BYO Build your own mini stream processor

STH Streamhouse

Overview
01 Architecture history — Lambda → Kappa → lakehouse → streamhouse
02 Apache Paimon — primary-key tables, merge engines, partial-update
03 Apache Fluss — PK tables, lake tiering, streaming vs batch consistency
04 Iceberg as streaming storage — commit interval vs small files
05 Streaming dimensional modeling — SCD2 dim + fct with temporal (as-of) join
06 Streaming data quality — in-flight DQ, quarantine / DLQ, completeness
07 Serving layer — OLAP on the streamhouse
08 Latency tiers & cost — seconds, minutes, hours

Foundations of distributed data · FND-01

Why distributed?

2 h

From your world. Why Redshift has a leader node and compute nodes; when one big box + DuckDB beats a cluster

Lesson Exercises 1 Quiz Interview drills Explain it back

Practice quiz with instant feedback. For a scored run that counts toward your local progress, use uv run de quiz FND-01.

1A job is 95% parallelizable and 5% inherently serial. By Amdahl's law, what is the maximum possible speedup, however many machines you add? A20x B95x CUnbounded — more machines always help D5x

The maximum speedup as n→∞ is 1/(1-p) = 1/0.05 = 20x. The serial 5% is the wall.

Source: https://dl.acm.org/doi/10.1145/1465482.1465560

2In a Redshift cluster, which component is most similar to Amdahl's serial fraction (1 - p)? AThe leader node that does the final merge/sort BA compute node that scans its slice CThe S3 spectrum layer DThe network switch

The final serial merge on the leader node cannot run in parallel across compute nodes. It limits scaling, in the same way as (1 - p).

Source: https://docs.aws.amazon.com/redshift/latest/dg/c_high_level_system_architecture.html

3For an aggregation over Parquet that scans a lot of data, which single factor most often has the largest effect on runtime and cost? ABytes read (the columns and row groups that the query actually scans) BNumber of CPU cores CThe programming language of the client DRAM clock speed

Columnar pruning cuts the bytes read by 10-100x. The bytes scanned usually dominate. This is why Athena bills by scanned bytes.

Source: https://duckdb.org/why_duckdb

4When is scaling OUT (more machines) likely the WRONG choice? (select all) (select all that apply) AThe dataset fits easily in the RAM of one machine BAn inherently serial step dominates the job CThe latency of a single small query matters more than throughput DThe working set is 50 TB and grows weekly

A, B and C all favor one box. D truly exceeds a single machine and justifies scale-out.

Source: https://dl.acm.org/doi/10.1145/42411.42415

5What is the fundamental reason that a shuffle is expensive in a scaled-out system? AData must cross the network, which is orders of magnitude slower than RAM BShuffles always re-read the source files from disk CShuffles require a GPU DShuffles disable columnar encoding

A shuffle moves records between nodes over the network. Network latency and bandwidth are far worse than in-memory access. So the movement of data between nodes is the cost.

Source: https://research.google/pubs/pub62/

6Gustafson's law differs from Amdahl's law mainly because it assumes that: AThe problem size grows to use the available machines BThe serial fraction is always zero CNetworks are infinitely fast DAll work is CPU-bound

Gustafson fixes the time and grows the workload with n, so the parallel fraction grows. Amdahl fixes the workload. The two laws answer different questions. Both are valid.

Source: https://dl.acm.org/doi/10.1145/42411.42415

NextFND-02 · Partitioning & shuffles

DE Mastery · source on GitHub · Found an error? Report it on GitHub