Why NoSQL systems converged on relational features

  • Question: DynamoDB now looks relational — why? (1:13:08)
  • DynamoDB's trajectory (Amazon, one of the first NoSQL databases): key/value store → JSON support → table-like structures, queried with PartiQL, Amazon's SQL dialect (1:13:33, 1:14:00).
  • Reason for convergence: customers actually want the guarantees discussed in the lecture — a relational model that stops you shooting yourself in the foot (1:14:00).
  • The stated position: like 1 + 1 = 2, the relational model is simply the correct way to represent data; vector indexes and JSON are accoutrements added on top (1:14:24).

What database should you pick?

  • Question: in system-design interviews, interviewers sometimes prefer NoSQL answers (1:18:03).
  • Answer: always pick Postgres — "I'm dead serious" (1:18:19).
    • Rock solid, with enterprise-grade features.
    • Applies beyond interviews: for startups and any new project, Postgres should be your first choice.
    • Rationale: spend your time building the application, not choosing a DBMS. Nobody knows or cares whether you run Postgres or MongoDB because users don't interact at that level (1:18:19).
    • If you eventually outgrow it, that is the point at which you pay someone a lot of money to help it scale (1:19:04).

Career in databases

  • The job market argument: unlike operating systems (IBM, Red Hat, and not much else), there is an enormous number of organizations building databases — Microsoft, Amazon, Google, Facebook, plus many startups (1:19:36).
  • Even though relational databases are ~50 years old, people are still building new systems, and there is a lot of money in it (1:20:09).
  • Disclosure: the lecturer's own database startup failed that summer (1:20:09).
  • Red flags when evaluating a company (1:20:09, 1:20:34):
    • Check the startup's finances by talking to people.
    • Companies "chasing fads" — a database that suddenly rebrands as a vector database or an AI database "looks a little dicey".
    • EnterpriseDB is named as one to avoid.
  • Another option: companies building an in-house system from scratch — they are their own customer, so they never worry about acquiring customers (1:20:59).
  • Why companies recruit from this class: they don't want "random JavaScript programmers" touching DBMS internals (1:21:26).
  • The lecturer has been told this course is used as the onboarding process at some database companies, and that some companies' interviews are "basically entire lectures from this class" (1:21:26, 1:21:55).

Miscellaneous

  • OpenAI acquired Rockset (a relational database startup founded by the people behind RocksDB at Facebook) for a large sum that summer; what they will do with it is unknown (1:14:47, 1:15:14).
  • No DJ in the course this year due to legal issues involving the Warhol ruling and a morality clause (1:22:18, 1:22:33).