Revision · Modern Frontier

This part is a snapshot labelled “as of 2025”: three shifts — LLM-backed features, edge/serverless compute, and a converging data layer — each run through the same trade-off lens as the timeless parts before it.

What this part covered

• Three new constraints — non-determinism as a feature, compute that moves to the user, and one dataset wanting three shapes; none repeal the fundamentals, they recombine them under new cost curves.
• Building with LLMs — an LLM is a slow, expensive, occasionally-wrong network dependency; it buys a feature you couldn’t hand-code and costs determinism, seconds of latency, a per-token bill, and a new abuse surface.
• Retrieval-augmented generation — RAG grounds the model in your data via a vector DB and ANN search, trading extra retrieval infrastructure for relevance and freshness the base model lacks.
• Streaming UX and guardrails — designing for tokens-over-time hides latency, while evals and guardrails bound the non-determinism instead of pretending it away.
• Edge & serverless — CDN workers and per-request functions buy single-digit-millisecond proximity to users but make statelessness mandatory and pull compute far from your data.
• Cold starts and data-locality — scaling to zero taxes you with cold starts (isolates vs microVMs vs containers), and the edge wins for proximity right up until data-locality bites.
• From batch to streaming — Kafka/Flink pipelines with event-time processing and the same exactly-once caveat replace batch for freshness, at the cost of streaming’s operational sharpness.
• Change data capture — log tailing is the right primitive for getting data out of the database reliably, echoing the outbox pattern from Part 10.
• The converging data layer — the lakehouse plus vector ANN indexes (HNSW/IVF) pull OLTP, OLAP, and semantic search toward one system, which buys one platform and costs the specialization each of the three had.

The takeaway

The thread matters most here, because frontier tools are sold loudly on what they buy and stay quiet about what they cost: the LLM, the edge, and the converged platform each hand you a headline capability and a bill in determinism, data-locality, or specialization. The durable parts — latency budgets, statelessness, the storage hierarchy, trade-off thinking — are exactly what let you read both columns before signing up. That accounting is where the book ends and your own designs begin.