tests/test_eum.py::TestEUM::test_wrapper fails on master today:
ValueError: 110321 is not a valid eumItem
It fails on the first assertion, at tests/test_eum.py:8. eumWrapper.CreateItemHashtable() asks the native EUM library for every item it knows and converts each id through eumItem(...). The native library reports item 110321, "Carbon Emmision Factor". The Python eumItem enum in mikecore/eum.py ends at eumIThickness = 110320, so the conversion raises and the whole test dies before reaching anything it meant to check.
Why CI does not see it
The workflows pin different versions of the same native package:
.github/workflows/test_linux.yml:31 — nuget install DHI.MikeCore.Linux.rhel7 -Version 20.0.0.github/workflows/build_linux.yml:31 — nuget install DHI.MikeCore.Linux.ubuntu -Version 22.1.0
So the tests run against a native library old enough not to know 110321, while the wheel that ships is built against one that does. Green CI, red suite for anyone installing the released package. Aligning the two versions would make CI reproduce this immediately.
The underlying shape of the problem
mikecore/eum.py carries a hand-transcribed copy of the EUM item and unit tables — 607 eumItem members — with the maintenance burden stated in the source itself, twice (lines 9 and 611):
# Must be updated with every new release, or if the EUM.xml is updated
Every binding to this library repeats that transcription and then drifts on its own schedule, so this failure is not a one-off missing constant — it is the mode this design fails in. Any new item added on the native side breaks CreateItemHashtable() in any binding that has not been hand-updated, and the break is total rather than per-item, because one unmappable id takes down the whole table.
Two fixes, and they are independent
Immediate: add the missing member so the current native library round-trips:
eumICarbonEmmisionFactor=110321
(spelling copied from the native key, which has the typo)
Structural: stop the next one from being a test failure at all.
The root need is a canonical, persistent, machine-readable source for EUM codes — a stable public URL listing every item and unit id with its key — that bindings can generate from and that users can look a code up in. Today there is no such page: MIKE Core SDK / EUM describes what EUM is, and mikeio's EUM guide shows how to search types from Python, but neither enumerates the codes. The only complete list is the one each binding transcribes by hand, which is why they drift.
The model to copy is the CF standard name table: a numbered version, a persistent URL per version, a machine-readable XML artifact alongside the human-readable page, and a documented process for adding a name. Every client generates from it instead of transcribing it, and a citation like "CF standard names v88" is unambiguous years later. EUM has the same shape of problem — a controlled vocabulary of quantities and units consumed by many independent clients — and none of that infrastructure.
With that in place, eum.py becomes generated from a versioned artifact rather than maintained by hand, and a native version bump regenerates instead of breaking.
Until then, a smaller mitigation is worth having regardless: make id-to-enum conversion tolerant, so an id the enum does not know yields an unknown/passthrough item instead of raising and taking the whole hashtable down with it. One unrecognised code should not cost every other code in the table.
The immediate fix unblocks the suite. The canonical source is what stops this recurring on every release.
tests/test_eum.py::TestEUM::test_wrapperfails onmastertoday:It fails on the first assertion, at
tests/test_eum.py:8.eumWrapper.CreateItemHashtable()asks the native EUM library for every item it knows and converts each id througheumItem(...). The native library reports item110321,"Carbon Emmision Factor". The PythoneumItemenum inmikecore/eum.pyends ateumIThickness = 110320, so the conversion raises and the whole test dies before reaching anything it meant to check.Why CI does not see it
The workflows pin different versions of the same native package:
.github/workflows/test_linux.yml:31—nuget install DHI.MikeCore.Linux.rhel7 -Version 20.0.0.github/workflows/build_linux.yml:31—nuget install DHI.MikeCore.Linux.ubuntu -Version 22.1.0So the tests run against a native library old enough not to know
110321, while the wheel that ships is built against one that does. Green CI, red suite for anyone installing the released package. Aligning the two versions would make CI reproduce this immediately.The underlying shape of the problem
mikecore/eum.pycarries a hand-transcribed copy of the EUM item and unit tables — 607eumItemmembers — with the maintenance burden stated in the source itself, twice (lines 9 and 611):# Must be updated with every new release, or if the EUM.xml is updatedEvery binding to this library repeats that transcription and then drifts on its own schedule, so this failure is not a one-off missing constant — it is the mode this design fails in. Any new item added on the native side breaks
CreateItemHashtable()in any binding that has not been hand-updated, and the break is total rather than per-item, because one unmappable id takes down the whole table.Two fixes, and they are independent
Immediate: add the missing member so the current native library round-trips:
(spelling copied from the native key, which has the typo)
Structural: stop the next one from being a test failure at all.
The root need is a canonical, persistent, machine-readable source for EUM codes — a stable public URL listing every item and unit id with its key — that bindings can generate from and that users can look a code up in. Today there is no such page: MIKE Core SDK / EUM describes what EUM is, and mikeio's EUM guide shows how to search types from Python, but neither enumerates the codes. The only complete list is the one each binding transcribes by hand, which is why they drift.
The model to copy is the CF standard name table: a numbered version, a persistent URL per version, a machine-readable XML artifact alongside the human-readable page, and a documented process for adding a name. Every client generates from it instead of transcribing it, and a citation like "CF standard names v88" is unambiguous years later. EUM has the same shape of problem — a controlled vocabulary of quantities and units consumed by many independent clients — and none of that infrastructure.
With that in place,
eum.pybecomes generated from a versioned artifact rather than maintained by hand, and a native version bump regenerates instead of breaking.Until then, a smaller mitigation is worth having regardless: make id-to-enum conversion tolerant, so an id the enum does not know yields an unknown/passthrough item instead of raising and taking the whole hashtable down with it. One unrecognised code should not cost every other code in the table.
The immediate fix unblocks the suite. The canonical source is what stops this recurring on every release.