Paths
Five routes through the same material. Path 0 is every case study, one page each,
and it is the one to scan if you are only here once. The rest are aimed at a goal:
an exam, a SQL screen, a systems interview, or what to read next.
Path 0 · 27 pages
Case Studies for Modern Data Systems
How real systems work, one page each. Most-read first, then everything in course order.
Path 1 · ≈ 10–14 hrs
Minimum Viable Exam Prep
Quizzes to find your gaps, the P-sets for mechanics, CA walkthroughs, then the outcomes.
Path 2 · ≈ 3–5 hrs
Data / SQL Interview Prep
The patterns interviewers test, the SQL playground, and how Postgres scales in the wild.
Path 3 · ≈ 6–10 hrs
Systems Design Interview Prep
IO cost model, index selection, a real transaction disaster, and distributed primitives.
Path 4 · post-class
Going Deeper · Reading List
The papers the field was built on, and how teams running these systems write it up.
Free to read. Opening a page asks you to sign in with Google or an email link, which is what keeps your progress.
Path 0
Case Studies for Modern Data Systems
27 pages
How real systems actually work.
Most read
The rest, in course order
M0
Pandas: small analytics
A laptop holds 16 GB of RAM. Spotify holds billions of listens. Everything after this page is that gap.
Open →
M0
Claude & OpenAI: agent memory
The context window is short-term memory, wiped the moment the session ends. Anything that must outlast it lives in a database.
Open →
M0
UberEats: a big service
A burrito order has three sides, and all three have to agree the instant it is placed. That is the job a transaction does.
Open →
M2
Uber and Maps: geo-hashing
One grid key instead of two coordinates, so “drivers within 2 km” is a few cell lookups instead of distance math over millions of rows.
Open →
M2
Chrome Safe Browsing: Bloom filters
Chrome checks every URL you open against millions of bad ones, without shipping the list or asking a server.
Open →
M3
BigQuery scaling
“Scan a petabyte” becomes “scan three columns on a thousand workers.” Storage and compute bought separately.
Open →
M3B
Spotify search
100 million tracks, and a search reads one short list per word instead of scanning any of them.
Open →
M3B
Spotify activity
A billion plays a day. Writes never overwrite anything; they append, and the sorting happens later.
Open →
M3B
Spotify Wrapped
December 1 is the year’s biggest read spike, and it costs one row per user, because November already did the work.
Open →
M3B
The $550k query disaster
One missing index, one query, a hundred times a day. $550,000 a year.
Open →
M4
ACID at Stripe
One Stripe charge is four writes across four tables. All four land or none do, and there is no third outcome.
Open →
M5
GFS: the internet on dying disks
Build storage out of drives you expect to fail, and buy reliability with replication instead of hardware.
Open →
M6
OpenClaw’s memory
Files are enough for one agent. The moment several share a memory you need a database, and locking comes back with it.
Open →
M6
DB design for a startup
Most apps never outgrow one machine. Start on one Postgres, add read replicas, and reach for sharding far later than you think.
Open →
M6
SQL beyond tables
One engine stores documents, arrays, geo points and vectors, all through the same SELECT. Most “we need NoSQL” arguments end here.
Open →
M6
DB design at big-tech scale
The stack splits in three: OLTP for writes users see, OLAP for analytics over billions of rows, and a pipeline moving data between them.
Open →
M6
Data quality & observability
A fare field vanished from one in ten Uber sessions. Every nightly job ran green. Someone noticed forty-five days later.
Open →
M6
SQL vs NoSQL
One Postgres row now holds a rigid ID, a JSON blob and a vector. Most of the old reasons to leave SQL have expired.
Open →
M6
Key-value stores
Song lyrics do not want to be columns. One key, one blob, one hop, and no joins ever.
Open →
M6
Claude’s KV cache
An agent resends the same long prefix every turn. Cache it and the cost per turn collapses.
Open →
M6
Privacy
Deleting names does not anonymise anything. One sensitive place on a known day is enough to put the name back.
Open →
Path 1
Minimum Viable Exam Prep
≈ 10–14 hrs
The straight line from "haven't studied" to "ready to sit the exam." Quizzes first to find your gaps, then P-sets for the mechanics, then CA section walkthroughs to see solved examples, then re-read the concept summaries.
Path 2
Data / SQL Interview Prep
≈ 3–5 hrs
For SQL screens and analytics-engineer interviews. Start with the patterns interviewers actually test (Round 1 = single pattern, Round 2 = composition), then drill on the playground, then read how Postgres scales in the wild for the inevitable "how would you scale this?" follow-up.
Path 3
Systems Design Interview Prep
≈ 6–10 hrs
For data-infra, backend, and platform-engineering interviews. Start with the IO cost model (the substrate every systems answer rests on), then index selection, then a real-world transaction disaster, then distributed primitives. Each step is a concrete pattern you can name in an interview.
Path 4
Going Deeper — Reading List
post-class
For when class is over and you want to keep reading. Everything here is
outside the course: the papers the field was built on, and engineering
write-ups from teams running these systems now.