Was reading Netflix’s post on their write-ahead log service. The part about multi-partition writes stuck with me.
The problem:
Their KV stores can’t update rows across partitions atomically. If a write fails partway, some rows change and some don’t.
What they do instead:
Write every mutation to a separate durable KV table, each with a sequence number
After the last one, write a completion marker
Send a Kafka message saying the batch is ready
Consumer reads the full batch from the durable table
Checks the marker exists, then applies the rows in order
How failures play out:
Consumer crashes: batch is still in the durable table, so it gets replayed
Database is down: retry later, nothing is lost
Client dies mid-batch: no marker, so the partial batch is never applied
The detail I liked:
Kafka is only the trigger. The durable table is the source of truth for what was supposed to happen.
It’s the same idea as a database WAL on a single node, moved into a shared service so any team can use it.
blog: https://t.co/RieT7RKGVm
I think the declining interest in computer science exposes a deeper problem with how we understand the discipline.
For decades, CS education has been optimized around producing software developers teach programming languages, algorithms, frameworks, databases, and eventually how to implement applications. This made sense when human implementation was the primary bottleneck.
AI is changing that bottleneck.
We should be asking whether universities are teaching students to understand computation or simply preparing them to participate in an existing software production model.
I would fundamentally change the curriculum. More operational semantics, type theory, compiler construction, distributed systems, memory models, formal verification, and computational complexity. Less emphasis on reproducing known implementations.
Students should derive algorithms, establish invariants, construct counterexamples, prove transformations preserve semantics, and investigate where existing abstractions break.
Instead of another CRUD application, give an undergraduate a database engine, compiler, or incremental computation problem. Ask them to identify a fundamental limitation, derive an alternative, establish its correctness, and measure its actual computational cost.
There is also an economic contradiction nobody discusses enough companies might be eliminating the very entry-level work through which future senior engineers acquire expertise. AI can reduce implementation costs, but engineering judgment still requires developing an understanding of real systems, failure modes, and architectural trade-offs.
Universities should respond by becoming more research-oriented. Students should reproduce papers, challenge published results, build substantial technical artifacts, and graduate with evidence of their ability to investigate unfamiliar problems.
i see deeper opportunity is producing people who can invent better computational abstractions new execution models, programming languages, runtimes, storage engines, and verification systems.
Programming is an interface to computation. Computer science is the discipline that allows us to understand and redesign computation itself.
We shouldn't persuade students to study CS because software engineering jobs might survive AI. We should persuade them because cheaper implementation makes the ability to reason about what should exist, how it should execute, and whether it is correct even more valuable
CS is goated. Idk, I've never had a single problem with it. I genuinely love it. Most of what universities teach is BS anyway. You can self-learn computer science if you have the curiosity