andreahlert opened a new pull request, #326:
URL: https://github.com/apache/airflow-steward/pull/326
## Summary
Three small corrections to `docs/mode-economics.md`, all confined to the
existing **Reducing costs** and **Local and self-hosted inference** sections.
No new claims, no methodology changes, no new rows — only tightening claims
already on the page.
1. **Cache section referenced a pre-correction anchor.** Item 3 ("Cache
skill context") said _"The skill file (3 000–6 000 tokens) and stable project
configuration files are ideal cache candidates"_. That `3 000–6 000` figure is
the same anchor the PR description of #253 explicitly flagged as wrong and
corrected elsewhere on the page — the **What \"tokens\" means here** table now
reports measured ranges:
| Skill class | Range |
|---|---|
| small setup/utility | ~1 000–3 000 |
| typical workflow | ~3 500–9 000 (median ~5 300) |
| large multi-step security | ~11 000–36 000 |
The inline figure in the cache item was never updated, so the page
contradicted itself. Replace the figure with a pointer to the corrected anchor.
2. **Cache TTL caveat (same paragraph).** The pattern \"first invocation
pays; subsequent invocations cheap\" is real for bursty same-session workloads
but typically misses for the periodic / day-spaced workloads listed in the
per-mode tables above (Anthropic prompt cache TTL: 5 min default, 1 h extended
at higher write cost). One sentence flagging the constraint avoids the footgun.
3. **Local-inference table: amortisation needed a denominator.** The
Consumer-GPU row listed `~\$0.10–0.50/hr amortised` with no basis.
\"Amortised\" without capex / lifespan / utilisation is uninterpretable. Inline
the assumption (`capex amortised over ~3 yr lifespan × moderate utilisation`)
so a reader can sanity-check the range against their own hardware.
## Out of scope (intentionally)
The page still has broader gaps that I think deserve discussion, not a
unilateral edit: a measurement-date / coverage / tokenizer banner, splitting
measured-vs-planned rows in the per-mode tables, p50 anchors on the wide
ranges, a date stamp on the cross-class multipliers. I'll open a separate issue
for those so #281's eval harness can inform the answers rather than my picking
them in isolation.
## Test plan
- [x] `markdownlint` / `doctoc` / `typos` clean locally (only edits are
in-paragraph text; no new anchors, ToC unchanged).
- [x] Internal anchor `#what-tokens-means-here` resolves to the existing
section heading on the same page.
- [x] Diff is 9 lines added / 4 removed, scoped to a single file.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
