What each run costs in GCP compute, and where it goes.
Unit prices are real billed rates for this GCP project (derived once from its own BigQuery billing export), not public list pricing - applied to each run's own measured machine types, boot-disk sizes, and VM lifetimes. The shared deploy node's cost runs from launch to when this data was generated (a few minutes before the VM is actually deleted - negligible, since it's the smaller of the two totals). Each GCP Batch job gets its own dedicated VM; its lifetime splits into boot (provisioning, measured from the job's own status events), run (actual tool execution, also measured), and teardown - a small fixed estimate, since GCP Batch exposes no event for when the VM is actually deleted after the job finishes. See anvil_generate_cost_data.py for the full methodology and pricing constants.
Cost per run
Each bar is a run, split between the shared deploy node and GCP Batch jobs.
No cost data yet.
Per-job cost (GCP Batch)
One row per GCP Batch job, split into its boot / run / teardown phases.
Rows are ordered top-to-bottom by when the job's VM was scheduled, and positioned left-to-right by elapsed time since the first job in the run started provisioning. Boot and run are measured from the job's own GCP Batch status events; teardown is a small fixed estimate applied to every job equally (see the note above) - useful for showing the shape of the overhead, not a precise per-job figure. Only jobs whose machine type and status events were both captured are priced; a run with mostly-failed jobs may show fewer rows than its total job count.
No GCP Batch job cost data yet.