01_​TRAIN

The monitor decides and the monitor calls. 02_MONITOR sees the drift and starts 01_TRAIN itself. The registry/ folder is the only thing the two of them share.

1 Data

Sixteen sensors, one number each. train_upto is the last batch this run may learn from.

Read it, cut it at train_upto

2 Label and split

One question: is this reading Acetone. Stratified 70 / 30 on it, seed 42.

70 % to learn on, 30 % to be judged on

3 Two fits, on purpose

Fit A on the 70 % gives the accepted accuracy; fit B on the whole range is what ships.

The number you report and the model you ship

4 Register

One model file and one registry row. Nothing else leaves this workflow.

Hand it to MLOps: a file and a row

One workflow, five versions. Four of the five are started by 02_MONITOR through Workflow Service Input, and nothing in here knows that. Send nothing and the declared workflow variables are used instead, which is why it still runs on its own.

It makes its own registry. Both writers sit on the LOCAL file system and work their absolute path out at run time, so the folder, the model file and registry.csv all appear on a first press with no registry/ beside them.

Why two fits, why rows_trained is not rows_measured, why the solver tolerance is 1e-8 and why created_at is the one column verify.py skips: README.md.

The reference profile — the other thing this workflow registers

Accuracy needs labels and labels are weeks late, so every version also leaves profile_v<N>.csv behind: one row per sensor — mean and sd — plus the share of "yes" the deployed model gave on its own training window. 02_MONITOR measures the next batch against it with no labels at all. Why Unpivoting: README.md.