Dataset Viewer: Xet permission error in recording_physical_references/archive

#96
by maxwellinked - opened

The hosted Dataset Viewer for maxwellinked/time-lapse-artifacts, configuration recording_physical_references, split archive, returns HTTP 500 while reading image cells. It reproduces with a request for just one row.

Observed and rechecked on October 18, 2026. The repository is public and ungated.

Reproduce

Viewer

Unauthenticated one-row API request

Expected: HTTP 200 with the requested row and its image cell.

Actual: HTTP 500, X-Error-Code: TransformRowsProcessingError. The initial response reported:

RuntimeError: Previous task error: I/O error: I/O error: Permission denied (os error 13)

A later retry reported the same underlying error as:

OSError: I/O error: I/O error: Permission denied (os error 13)

Comparison checks

The tested image references are pinned to commit 042469fe4bff3c7ffad36b47be25503bf73f0da6:

Referenced JPEG Bytes Decoded dimensions Verification
0d5d812ffbe76b2e5e860a9f.jpg 182,004 836 x 627 HTTP 200; full decode; SHA-256 match
b7101a726643f0bccd90f547.jpg 9,787,585 8064 x 6048 HTTP 200; full decode; SHA-256 match

The affected API response's X-Revision was ee38d767e31d6d6f784a6b2f71d2e7dda3e52b88. Hub main at inspection was cf628dc15559fe743a9f18d478b3602924958fa8. These revisions are included as diagnostic context; the difference has not been established as the cause.

Traceback excerpt

The failure occurs while Pillow requests the first 16 image bytes through HfFileSystem/Xet:

File "/src/libs/libapi/src/libapi/rows_utils.py", line 37, in _transform_row
  transformed_row[featureName] = get_cell_value(...)
File "/src/libs/libcommon/src/libcommon/viewer_utils/features.py", line 426, in get_cell_value
  return image(...)
File "/src/libs/libcommon/src/libcommon/viewer_utils/features.py", line 115, in image
  image = PILImage.open(f)
File "/usr/local/lib/python3.14/site-packages/PIL/Image.py", line 3650, in open
  prefix = fp.read(16)
File "/usr/local/lib/python3.14/site-packages/huggingface_hub/hf_file_system.py", line 1307, in read
  return super().read(length)
File "/usr/local/lib/python3.14/site-packages/fsspec/spec.py", line 1897, in read
  out = self.cache._fetch(self.loc, self.loc + length)
File "/usr/local/lib/python3.14/site-packages/fsspec/caching.py", line 234, in _fetch
  self.cache = self.fetcher(start, end)
File "/usr/local/lib/python3.14/site-packages/huggingface_hub/hf_file_system.py", line 1253, in _fetch_range
  return b"".join(self.fs._read_xet(self._xet_file_data, self.size, start=start, end=end))
File "/usr/local/lib/python3.14/site-packages/huggingface_hub/hf_file_system.py", line 1126, in _read_xet
  group = session.new_download_stream_group(**xet_download_group_kwargs(xet_file_data, headers))
RuntimeError: Previous task error: I/O error: I/O error: Permission denied (os error 13)

An unrelated dataset's Viewer page also displays the same RuntimeError and stack.

The evidence suggests a shared Viewer/Xet runtime issue. The denied filesystem path and underlying cause remain unconfirmed. Could the Viewer team inspect the first underlying Xet I/O error and the service's filesystem permissions, restore the affected image-row requests, and refresh cached results if necessary?

@lhoestq — following up with a changed failure mode and a narrower reproduction. Could the Viewer team inspect the main dataset's pending/stalled Parquet metadata and cache dependencies, and perform a targeted refresh or requeue if appropriate? The exact blocking job is not visible from the public API.

Fresh anonymous requests on October 21, 2026 at 12:12 UTC, with Cache-Control: no-cache:

Request Result
Main paired view, offset=100, length=100 HTTP 500; X-Error-Code: ResponseNotReady
Main video-only canonical control, offset=0, length=100 HTTP 500; X-Error-Code: ResponseNotReady
Companion photograph dataset, offset=100, length=100 HTTP 200; 100 rows returned; 431 total
Main Hub Parquet listing HTTP 200, but empty {}

Main remains at 94ba224581bc337877b6972cc8add0789210c725; companion is fd6543d7c559f23ca32858526cd3b67f8e153453, matching the successful companion response's X-Revision. The failed main requests did not supply X-Revision.

The main generated Parquet branch already contains all five configurations. Its current commit is ecd2d0de0f5854739ce5c0136f057c99247a6386, generated at 09:56:20 UTC today. Direct anonymous reads of its canonical table succeeded (1,205 rows). Bounded range reads of the paired table verified 1,333 rows, 431 non-null unique work IDs, and the expected 1,205 Standard / 128 Pre-Standard split; image-byte column statistics show 1,205 non-null cells and 128 intentional nulls.

Earlier today at 11:54 UTC, main and companion /first-rows returned 200 at their current revisions. One freshly returned image from each loaded and fully decoded successfully. This verifies those two samples only; it does not establish full image delivery or working main pagination.

Today's ResponseNotReady also affects the video-only control, so the October 18 Xet/Pillow traceback above should not be treated as the established cause of this current failure. The generated-file/public-paging discrepancy is the specific remaining blocker. Related earlier diagnostics: dataset-viewer issue #3476.

Can you identify the blocking dependency and restore paging without requiring changes to the published media or work identities?

Sign up or log in to comment