[CodyCBakerPhD]

Problem

For an NWB file in a BIDS dataset, dandi.metadata.nwb.get_metadata() took the BIDS part of the metadata from df.get_metadata(). df is an NWBBIDSAsset, and its get_metadata() also extracts the NWB metadata, via NWBAsset.get_metadata() → nwb2asset() → get_metadata(). That calls back into the same BIDS branch, so the cycle repeats until RecursionError.

The result was still correct, because NWBAsset.get_metadata() falls back to default metadata on errors at each level. But the BIDS validator ran at every one of roughly 100 levels. On ieeg_epilepsyNWB from bids-examples, one get_metadata() call made 98 nested nwb2asset() calls and took about 60 s. That is what made test_bids_nwb_metadata_integration time out on Windows CI.

Fix

Take only the BIDS metadata, via bids.BIDSAsset.get_metadata(df, ...), which was the intent of that branch. The NWB metadata is extracted by the rest of get_metadata() itself.

Measured on the same file:

	nwb2asset() calls	time	subject_id / sex
before	98 (nested)	60.9 s	01 / U
after	0	0.7 s	01 / U

---

🤖 Generated with Claude Code

https://claude.ai/code/session_01Qn5WBSiQgoZoL4fytF6nEr

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 90.82%. Comparing base (ad8f64d) to head (3ee7aba).

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #1958      +/-   ##
==========================================
+ Coverage   90.77%   90.82%   +0.05%     
==========================================
  Files          92       92              
  Lines       14130    14159      +29     
==========================================
+ Hits        12826    12860      +34     
+ Misses       1304     1299       -5     

Flag	Coverage Δ
unittests	90.82% <100.00%> (+0.05%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[CodyCBakerPhD]

@yarikoptic this one was discovered back in the bids-dandisets days but fixing was delayed until closer to 'eventual' BEP032 integration

Might as well do it now though, with all the other validation improvements going on

[yarikoptic]

I traced how this became a recursion and checked the fix locally. LGTM; the notes below are non-blocking.

How the loop formed

• #1076 (Aug 2022): NWBBIDSAsset.get_metadata() merges BIDSAsset.get_metadata() with NWBAsset.get_metadata() → nwb2asset() → metadata.get_metadata(). That function knew nothing about BIDS yet, so there was no loop.
• #1164 (Dec 2022, for dandi ls) added a BIDS branch to metadata.get_metadata() that calls the overridable df.get_metadata(). At first it sat in the else: for non-NWB files, which was safe. 11c18f5 then put it before the NWB branch (if bids: … elif nwb:). For an NWB file in BIDS, df is an NWBBIDSAsset, which calls back into metadata.get_metadata(): the loop.
• #1183 (Jan 2023) changed elif to if and added ieeg_epilepsyNWB and test_bids_nwb_metadata_integration. The test passed because NWBAsset.get_metadata() catches the RecursionError at the innermost level (except Exception with ignore_errors=True), after which every outer level completes normally.
• Every level builds a new BIDSDatasetDescriptionAsset and validates the whole dataset (bidsschematools for the metadata, deno for the version). That is where the ~60 s and the Windows timeouts come from.

On the "band-aid" question

It isn't one. A band-aid would guard against recursion depth; this PR instead removes the overridable call that closed the loop. get_bids_metadata() is defined once on BIDSAsset and never reads file content, so this call can no longer recurse, whatever subclass df is.

There is a more general option, possibly simpler: for NWB files, the BIDS branch in nwb.get_metadata() contributes nothing. Every key it can set (age, genotype, identifier, sex, species, strain) is then overwritten by _get_pynwb_metadata(), with None if the NWB file lacks it. The only key that survives is bids_version, which is always None (#1961). Skipping the branch for NWB files would make metadata.nwb stop depending on files.bids for NWB files, and would also remove two full-dataset validations per NWB file. Filed as #1962; fine as a follow-up.

Verified locally (Python 3.13)

• The PR's new tests against master's code fail with 68 nested nwb2asset() calls (53 s and 57 s). test_bids_nwb_metadata_integration still passes there, but takes 77 s and logs RecursionError warnings from bases.py.
• On this branch all 3 pass in about 1 s. The spy does catch the calls, because bases.py imports nwb2asset inside the function.
• NWB values still take priority over BIDS values, both in the flat dict and in the NWBBIDSAsset merge.

Nits (optional)

• The get_bids_metadata() docstring mentions "files such as participants.tsv". Here bidsschematools only provides path entities (subject_id, session_id); participants.tsv is not read.
• get_bids_metadata() returns the object cached in _asset_metadata and modifies it in place (digest, path; ZarrBIDSAsset also sets encodingFormat). This is pre-existing and harmless today. Now that it is a public method meant to be reused, copying it first (.model_copy(deep=True)) would be safer.
• A short comment on why the test uses get_metadata.__wrapped__ would help: it bypasses the cache, so a cached result can't make call_count == 0 pass trivially. The two new tests could also share the copytree/path setup.

---

Generated by Claude Code

[CodyCBakerPhD]

#1183 (Jan 2023) changed elif to if and added ieeg_epilepsyNWB and test_bids_nwb_metadata_integration. The test passed because NWBAsset.get_metadata() catches the RecursionError at the innermost level (except Exception with ignore_errors=True), after which every outer level completes normally.

Hah, nice to see the story laid out