Yelp moved a petabyte without moving the data
Yelp migrated about a thousand Cassandra pods to Amazon EKS by repointing the disks rather than streaming the data, and the week its engineers spent doing it by hand is what made the automation cheap. Thoughtworks, separately, names cost attribution as a job of the serving platform rather than of the model.
How we Organise
A model cannot say which team spent the money
Thoughtworks argues that the per-token price a leadership team approves is one line item. The serving infrastructure decides the rest. An agent loop re-reads its context on every pass, tool output returns to the context window, and a long system prompt is billed on every call, so volume and repetition set the bill rather than the unit price. The essay names four jobs for an inference platform. One is attribution. That means knowing which team, workflow or customer spent the money.
We tell a client that a cost nobody can attribute is a cost nobody can govern. It is attributed in the observability layer, or nowhere at all. A contract will not do it.
A team whose cost cannot be stated on its own can be neither funded nor stopped.
The essay carries no cost figures. So the distance between an approved price and a first invoice goes unmeasured here, and for a chat workload it may be small.
How we Build
A week of doing it by hand was what made the automation cheap
Yelp moved about a thousand Cassandra pods and nearly a petabyte of data onto Amazon EKS without streaming a byte. It created persistent volumes in the new cluster pointing at the same EBS volume IDs. Its engineers first ran the procedure by hand for a week in development, which produced a specification carrying every failure path they hit. The automation was then generated from that runbook rather than from a plan. It needed minor adjustment after two days of testing. The rollout took three weeks.
Our position is that an increment earns its place by changing the scope of the one after it. The sequencing here is the argument. Nobody asked a model to design the migration.
Work moves in increments thin enough to be shipped and evaluated before the next is scoped, and what an increment proves changes what the next one is.
This is not the first source to bear on that statement. A September preprint strained it. Martin Fowler’s site carried a confirming case in October. It stands contested rather than settled.