technologybriefs
10:00in productionCh. 1 · Failure-first architecture/ 10:00 · ceiling 15 min
Software · Systems

Apache Hadoop

Hadoop doesn’t make big data easy—it makes it possible on broken hardware.

Hadoop is not a database or a query engine. It is a fault-tolerant batch-processing substrate built for scale, not speed or simplicity.

Chapters & takeaways4
  1. 1:04
    Failure-first architecture

    Hadoop is built for failure—not as a contingency, but as its core design assumption.

  2. 2:34
    Data locality by default

    It stores data in large blocks across nodes, then ships code—not data—to where the data lives.

  3. 4:22
    What 'automatic' really means

    Automatic failure handling works only for node crashes—not network partitions, silent corruption, or misconfigured shuffles.

  4. 5:59
    Storage layers keep splitting

    HDFS remains central, but Ozone (2020) shows Hadoop’s storage layer is still evolving—not stabilising.

Worth your time?

Yes. Study the whole thing.

3.5/ 5
What works
  • fault-tolerant storage via HDFS replication
  • parallel computation via MapReduce's split-process-merge flow
  • automatic task rescheduling after node failure
  • rack-aware data placement to reduce cross-rack traffic
What does not
  • real-time processing
  • interactive querying
  • developer ergonomics
  • cloud-native integration (as of 2006)
Study it if
  • data engineers managing offline ETL pipelines
  • organisations with petabytes of append-only log data
  • teams willing to trade latency for resilience on cheap hardware
Skip it if
  • application developers building low-latency services
  • analysts needing ad-hoc SQL access
  • organisations without dedicated cluster operations staff
The written brief1 min read

What it is and the problem it solves

Hadoop is an open-source framework for reliable, scalable, distributed computing on commodity hardware. It solves the problem of storing and processing massive datasets when hardware failure is expected, not exceptional.

How it works

Hadoop splits files into large blocks and distributes them across cluster nodes. It transfers packaged code to those nodes to process data in parallel. This exploits data locality: nodes process data already stored locally.

What works

HDFS reliably stores large files across many machines with replication and rack awareness. MapReduce executes parallel computations while automatically rerouting work around failed nodes. Heartbeat monitoring and failover support keep jobs running despite frequent hardware faults.

What does not

Hadoop does not handle real-time or interactive queries well. It does not simplify data engineering—users must manage partitioning, serialization, shuffle bottlenecks, and failure recovery manually. It does not abstract away distributed systems complexity.

What it changes

It changes how organisations store and process petabyte-scale logs, web crawls, and offline analytics by replacing expensive proprietary data warehouses with open-source, rack-aware clusters that self-heal from node failures.

Is it worth your time

Yes—if you need fault-tolerant, scalable batch processing on commodity hardware and can absorb the operational overhead of cluster coordination, manual tuning, and MapReduce’s rigid two-stage model.

Same field · Software4 of 102
10:03
Go (programming language)Robert Griesemer, Rob Pike, Ken Thompson · 2009Go is a statically typed, compiled programming language designed in 2007 by Robert Griesemer, Rob Pike, and Ken Thompson at Google to improve programming productivity for multicore, networked machines and large codebases. It combines static typing and runtime efficiency like C, readability and usability like Python, and high-performance networking and multiprocessing. Publicly announced in November 2009 and released as version 1.0 in March 2012, Go emphasizes simplicity, safety, and developer tooling — especially around concurrency via goroutines and channels, packaging, testing, and deployment — while deliberately omitting features like inheritance, assertions, pointer arithmetic, and implicit type conversions. Its design reflects a focus on the software development process itself rather than just language syntax or semantics.
9:44
Apache CassandraFacebook · 2008Cassandra is a distributed, wide-column database built for write-heavy, multi-datacentre workloads. It prioritises availability and scalability over consistency. It requires upfront schema design aligned to access patterns. It was developed at Facebook for inbox search and released open-source in July 2008.
10:34
Amiga 10001985The Amiga 1000 is not defined by what it promised, but by what it delivered on day one: a working, shipped, ROM-locked implementation of preemptive multitasking — alongside deliberate physical design choices that treated the case as functional infrastructure. Its limitations are architectural, not aspirational.
9:14
Apache HBase2006HBase is a wide-column, non-relational, distributed database modelled after Google's Bigtable. It runs on HDFS or Alluxio. It provides fault-tolerant storage for large sparse datasets with fast read/write operations, high throughput, low I/O latency, and CP consistency per the CAP Theorem. It features compression, in-memory operation, and Bloom filters on a per-column basis. Tables can serve as input and output for MapReduce jobs and are accessible via Java API, REST, Avro, or Thrift.
Up next in Technology

FM broadcasting

Edwin Howard Armstrong · 11:43

FM didn’t replace AM—it exposed AM’s noise problem as solvable, not inevitable.

11:43