Skip to content

fix(acceptance): correct TI03, TI04 and RG07, follow titiler-eopf to v0.12.4 - #14

Merged
lhoupert merged 13 commits into
mainfrom
fix--suite-bugs
Nov 5, 2026
Merged

lhoupert merged 13 commits into
mainfrom
fix--suite-bugs

Conversation

@lhoupert

@lhoupert lhoupert commented Nov 4, 2026 •

Copy link
Copy Markdown
Collaborator

Fixes three wrong results from the 3 Nov runs on EODC's rc stores, makes TI03 test the stretches consumers use, and renews an expired known issue.

  • TI03 on the 0.12 item route sampled the tilejson's zooms 0–24, so z0 and z24 tested nothing. It now renders one tile per store level, as the asset route does. A flat tile is compared with up to four others in the footprint, and fails unless one has both variety and enough valid pixels. The report says what was compared, and the summary names any level that passed this way.
  • L2A's stretch: TI03 and TI04 tested at 0,0.3, which no item link uses, so a fully clouded scene clipped to one colour. They now test at 0,1, the stretch the items link. min_distinct moves from 16 to 5 with it, because each PNG level now covers more reflectance: 5 levels at 0,1 span 0.0196, 16 at 0,0.3 span 0.0188. A level where no tile varies still FAILs, and the message now says where the pixels sit.
  • TI10 (new) renders the same levels at the Explorer's true colour stretch, 0,0.5, with min_distinct 10. A level flat on the tiles tried is a WARN, because a clouded scene clipping isn't a store defect, and a tile that can't be read is a FAIL. It runs only where a config sets render.second_stretch, which today is L2A, and on one API, 0.12, because whether data clips doesn't depend on the titiler version.
  • TI04 gets the same comparison for a grey tile, and takes its zoom from the store's levels.
  • Unreadable tiles: a tile TI03, TI04 or TI10 can't read or decode now fails the check, including one fetched for the comparison. A decode error used to stop the run and lose the rest of the battery.
  • RG07 passed without checking anything on stores that send no Last-Modified (data.eodc.eu): obstore reports the epoch, so every item looked newer than its store. It now SKIPs.
  • titiler-eopf v0.12.4: the configs, the reader pin and the constraints follow the version /rstaging reported on 3 Nov. Nothing under titiler/ changed since v0.12.2; the regenerated lock brings its dependency updates (rio-tiler 9.4, obstore 0.12).
  • OLCI: the /raster TI02 known issue's review date passed on 31 Oct and the >0.12 flip isn't ready, so the next review is 30 Nov.

For reviewers:

  • This restores four pins changed in build(deps): bump the minor-and-patch group in /acceptance with 5 updates #13. morecantile, obstore, rasterio and xarray go back to titiler-eopf v0.12.4's versions (7.0.3, 0.12.0, 1.5.0, 2026.4.0), as the rule in pyproject.toml says: the suite runs what the image runs. Dependabot will propose them again unless its /acceptance config changes.
  • TI03 is relaxed, so please read that part closely. Against a real titiler at the new settings, these all still FAIL TI03: a stretch above the data, a stretch far too narrow, and simulated constant, undecoded and double-decoded data.
  • What this gives up: simulated scale errors of x3 to x20, and x0.1 on one scene, can now pass where they failed before by accident. TI03 isn't a calibration check, and one is still needed.
  • Known limits:
    • A wholly calm-ocean scene is still expected to FAIL TI03 falsely. T27SVB passes narrowly: its centre tiles at z10 to z14 have 2–5 distinct values, and z10, z11 and z13 pass only on other tiles.
    • TI03 sends 0,1 without the links' color_formula, so it's closer to what users get but not identical.
    • L1C keeps 0,1 with min_distinct 16, so it may have the same dark-scene problem. No L1C scene has been rendered yet.
  • A level that passes through the comparison stays PASS, named in the summary. A review argued for WARN, since the centre tile stays unverified. It's PASS because a cloud isn't a defect, and it would otherwise WARN on healthy scenes. Easy to change if you prefer WARN.
  • TI10 warned on 2 of the 6 scenes tried: T25WFP (cloud, z13) and T27SVB (dark ocean). The Explorer's stretch comes from a landing-page config read from a June clone.
  • Worst case, the registered L2A plan with both endpoints is 286 requests of the default 300. The 30 Nov date is a proposal.

How this was made

Claude Code wrote the fixes and tests from the bug list in the 4 Nov handoff. Measuring real tiles showed the TI03 fix listed there wasn't enough, and an agent panel found that TI03 was testing a stretch no item link uses. @lhoupert chose to test both the links' and the Explorer's stretches and accepted the scale-error trade-off; @lhoupert's review of the code is pending. Independent /code-review passes found most of what the later commits fix. Tested locally with uv sync --extra reader and uv run pytest (174 passed, including the real titiler-eopf app at v0.12.4) and pip-audit on the lock (no known vulnerabilities); this repo's CI runs security checks only. Each new rule's test was checked against a mutant of that rule. Store and titiler checks with a local v0.12.4, reading EODC only, on the committed config: T33WWP, T33WXQ, T27SVC, T27SVB, T35WNU and T25WFP all pass overall. Not run against /rstaging, titiler 0.11 (/raster) or real faulty stores, and TR01 skips on https stores.

…tells a flat patch from a constant render

The item route's tilejson says 0-24, so z0 and z24 tested nothing, and a saturated cloud at the centre of a scene failed as a constant render.
obstore reports the epoch for a missing header, so every item looked newer than its store.
/rstaging runs it. The reader code is unchanged since v0.12.2; the regenerated constraints bring its dependency updates.
…30 Nov

The review date passed on 31 Oct and the >0.12 flip isn't ready yet, so the config's own rule says to extend it.
…eport what was compared

A tile only clears another if it has enough valid pixels, and a tile that can't be read now fails. TI04 also takes its zoom from the store's levels.
… crashing the run

The decode error stopped the run and lost the rest of the battery.
@lhoupert lhoupert changed the title fix(acceptance): correct TI03 and RG07, follow titiler-eopf to v0.12.4 fix(acceptance): correct TI03, TI04 and RG07, follow titiler-eopf to v0.12.4 Nov 4, 2026
…say why store levels are missing

Some bad bodies still stopped the run, and unusable store metadata brought back z0 and z24 on the item route.
@lhoupert
lhoupert marked this pull request as ready for review November 4, 2026 17:34
Keeps the image's versions in the constraints, as pyproject.toml's rule says, over the bumps in main.
… the Explorer's as TI10

A cloudy scene clipped to one colour at 0,0.3, which no item link uses. The threshold moves with the stretch so dark scenes keep the same bar.
…TI10's report honest

TI10 runs on one API, says which levels it couldn't judge, and always sends its stretch. Only decoding errors now read as a bad tile.
@lhoupert
lhoupert merged commit 55edf5e into main Nov 5, 2026
5 checks passed
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