Skip to content

foundation, stage-contract, decisions, index: milestones 4 to 6 on SMDC and Batch - #89

Merged
rusholme merged 1 commit into
mainfrom
foundation/milestones-4-6
Oct 21, 2026
Merged

rusholme merged 1 commit into
mainfrom
foundation/milestones-4-6

Conversation

@rusholme

Copy link
Copy Markdown
Collaborator

What and why

Task 6 of the unattended run foundation-m4to6-2026-10-21: the documentation of what tasks 1 to 5 built and measured on rebuild of Caltech-IPAC/rapid. Task 4's ledger line is TASK DONE 4, so milestones 4 to 6 are documented in full.

The pull requests documented: Caltech-IPAC/rapid#284 (s3:// prefixes, job submit), #286 (the SMDC schema, inputs and baseline), #287 (reference on run(job)), #288 (difference on run(job), refreshed from dev to d6aac515), #289 (the chain on Batch, the change group over Batch products and its undo: milestone 4), #290 (job run records the image digest and reports peak memory), #292 (the field uncapped on Batch with milestones 5 and 6's measurements).

Files

  • system/foundation.md:
    • the opening names the five foundation stages and says the chain has run on SMDC and Batch;
    • new "Where outputs live": the outputs root, prefix.get and prefix.put, attempts under s3://;
    • new "A job on Batch": job submit, the container's job run, retry by resubmission, the image digest and peak memory, one image and job definition revision per commit, a code fix as a new run;
    • the thin chain's table gains reference and difference rows in place of the "port comes first" row, with new reference and difference sections (zero point, the drawn rfid and pid, the two PSFs told apart by PSFKIND, one differencer published, the cap);
    • the settings paragraph covers [reference] and [difference];
    • new "On SMDC": the five parts of the procedure as system behaviour (schema replaced from schema.sql with grants, write-once inputs bucket, the live baseline on the scratch root, the non-live Batch run, the change group including the reference, its undo, the uncapped run);
    • new "Measurements", dated 2026-10-21: per stage per epoch capped and uncapped, peak memory, sources before the cap, objects and packets, the matcher, the synthetic member chain, change-group time, the comparison with dev, the paragraph against the hour per image, and what was not measured for want of data;
    • "Not decided here": the reference and difference port item leaves (decided by the new sections); chain cost and change-group cost narrowed by the measurements; two items added, the PSF kind in the PSF key (#288) and the read role's SELECT (#284, #286).
  • system/stage-contract.md: five stages on run(job); the toolkit() seam; job submit's queue and container command; run_job records the image digest and returns peak_rss_kib; admit and finalize the only stages left on the older contract (they cannot run on Batch); the port pattern; the leaf module seams gone (#288); selftest's new holdings.
  • system/decisions.md: one ruling, decision-smdc-prototype (Ben, 2026-10-21), which the foundation page's SMDC section cites.
  • system/index.md: the foundation.md, stage-contract.md and decisions.md scope cells.
  • system/reference.md, system/difference.md: the first sentence of the run-model note only (see divergences).

Divergences from the brief, with the reason

  • The run-model notes on reference.md and difference.md changed by one sentence. The brief leaves the pages carrying the note untouched. Their notes said "This stage still runs on the older contract", which became false with #287 and #288. Only that sentence changed: it now says the page describes the older-contract port and points at the foundation page's section. The bodies are untouched; their rewrite stays with the next documentation PR.
  • A ruling added to decisions.md. The brief names three pages. AGENTS.md requires a ruling the pages state to go on decisions.md in the same pull request; the SMDC section states "no backup, no fence" by Ben's 2026-10-21 ruling.
  • No Batch job UUIDs, run ids or bucket names on the site. The runbook in the pipeline repository carries them; the page carries the measurements and points at the runbook.
  • The forced-photometry cost line moved from "about two seconds" to "two to four seconds per window image", from #292's Batch split (3.8 s for one image, 7.1 s for two).

Build

Not built on the host: there is no Sphinx environment there and provisioning one is an install. docs.yml is the check. scripts/check-public-safety.sh passes locally; the changed pages have no em dashes and none of the AGENTS.md excess-vocabulary words.

Left

🤖 Generated with Claude Code

…reference and difference stages, Batch and the SMDC measurements

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@rusholme
rusholme merged commit 430ac2b into main Oct 21, 2026
2 checks passed
@rusholme
rusholme deleted the foundation/milestones-4-6 branch October 21, 2026 18:40
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant