CSUE301 · University Elective · Semester 5

Big Data
Analytics

Seven chapters, one running example. CityCourier — a food-delivery platform running three neighbourhood hubs out of one kitchen — places sixteen orders in a fifteen-minute lunch rush in Chapter 1, and every chapter afterward returns to those same sixteen rows, asking a different question of them: where they're stored, who processes them, how they're cached, streamed, learned from, and finally summarized onto a dashboard.

Small enough to check by hand, on purpose. Everything a real system does to a billion rows, this course does to sixteen — so the arithmetic is always verifiable, and the mechanics never change with scale, only the size of the numbers.
01
Unit 1 · 4 hrs · 7 sections

Big Data Fundamentals

The five V's, the data life cycle, MapReduce and BSP, Lambda/Kappa architectures, performance metrics, the CAP theorem, and responsible-AI governance — all worked on CityCourier's founding sixteen orders, including the veracity bug that echoes through every later chapter.

Open →
02
Unit 2 · 5 hrs · 6 sections

Distributed Storage Systems

HDFS blocks and replication, the NameNode/DataNode split, the four NoSQL data models, consistent hashing, quorum replication, B-Tree vs LSM-Tree indexing, and scatter-gather query execution.

Open →
03
Unit 3 · 5 hrs · 6 sections

Parallel & Distributed Data Processing

YARN's cluster resource management, a MapReduce job traced task by task, combiners, data skew and salting, Spark's DAG execution model, and task-retry versus lineage-based fault tolerance.

Open →
04
Unit 4 · 3 hrs · 4 sections

In-Memory Data Processing

Why memory beats disk by roughly 80×, RDD caching and persistence, LRU eviction and spill-to-disk, and how Spark's Catalyst optimizer pushes filters down to storage that a raw RDD never could.

Open →
05
Unit 5 · 5 hrs · 6 sections

Streaming Analytics & Real-Time Processing

Tumbling, sliding, and session windows; event time versus processing time; watermarks and allowed lateness; Kafka's topic/partition/consumer-group model; micro-batch versus true streaming; and exactly-once semantics.

Open →
06
Unit 6 · 4 hrs · 5 sections

Scalable Machine Learning

Data versus model parallelism, distributed gradient descent (and the size-weighting bug that quietly wrecks it), distributed K-means, embarrassingly-parallel Naive Bayes, and fair train/test splitting at scale.

Open →
07
Unit 7 · 4 hrs · 5 sections

Scalable Visualization & Analytics

Why exact answers stop scaling; Bloom filters, HyperLogLog, and Count-Min Sketch as three answers to membership, cardinality, and frequency; OLAP cubes with free roll-up and drill-down; and sampling for legible dashboards.

Open →
Colour contract, every chapter Hub N's server Hub S's server Hub E's server the number worth noticing