Can't Resolve Path on Network Drive

[RainmakerP]

lance.dataset() fails on Windows when dataset resides on a network share with spaces in the share name

Environment

pylance	4.0.1
lancedb	0.30.2
pyarrow	20.0.0
Python	3.12.3 (CPython, Windows, 64-bit)
OS	Windows 11 (Build 26100)
Storage	SMB network share mounted as a Windows drive letter

---

Description

lance.dataset() and lancedb's open_table() both fail when the Lance dataset resides on a Windows-mapped network drive whose underlying SMB share name contains a space.

The root cause is that Lance's Rust LocalFileSystem resolves any Windows path (drive letter, subst, symlink) down to its underlying UNC path — e.g. \\192.168.x.x\My Share\... — then attempts to construct a file:// URL from it. Because the share name contains a space, the resulting URL (file:///My%20Share/...) cannot be converted back to a valid filesystem path, and Lance errors out.

---

Error

LanceError(IO): Generic LocalFileSystem error: Unable to convert URL
"file:///My%20Share/data/my-dataset.lance/_versions/..."
to filesystem path,
lance-io\src\object_store.rs:716

---

Reproduction

Any Windows machine where:

• A network share is mounted as a drive letter (e.g. Z:)
• The SMB share name contains a space (e.g. \\server\My Share)
• A Lance dataset exists somewhere under that share

import lance
ds = lance.dataset(r"Z:\data\my-dataset.lance")
# raises LanceError(IO): Unable to convert URL "file:///My%20Share/..."

---

What was tried

Every approach below was tested. All fail for the same underlying reason: Windows resolves any path abstraction back to the UNC path before Lance's Rust layer sees it.

1. Drive letter path (baseline)

lance.dataset(r"Z:\data\my-dataset.lance")

Result: Fails. Lance resolves Z: → UNC → URL encoding fails.

---

2. subst virtual drive

subst V: Z:\data

lance.dataset(r"V:\my-dataset.lance")

Result: Fails. subst does not prevent UNC resolution. pathlib.Path(r"V:\...").resolve() returns the full UNC path.

---

3. NTFS symbolic link (mklink /D)

mklink /D C:\LocalLink "Z:\data"

lance.dataset(r"C:\LocalLink\my-dataset.lance")

Result: Fails. NTFS symlinks on Windows are resolved to their UNC target at the kernel level. Lance sees the UNC path.

---

4. RFC-8089 UNC file URI (with host)

lance.dataset("file://192.168.x.x/My%20Share/data/my-dataset.lance")

Result: Fails. Lance strips the host component and produces file:///My%20Share/..., which it cannot convert back to a path.

---

5. Raw UNC path string

lance.dataset(r"\\192.168.x.x\My Share\data\my-dataset.lance")

Result: Fails with same URL conversion error.

---

6. Direct file:// URI with percent-encoded drive path

lance.dataset("file:///Z:/data/my-dataset.lance")
lance.dataset("file:///Z:/data/my%20dataset.lance")

Result: Fails. Lance still resolves the path to UNC before URL construction.

---

7. PyArrow SubTreeFileSystem passed to lance.dataset()

import pyarrow.fs as pafs
fs = pafs.SubTreeFileSystem(r"\\server\My Share\data", pafs.LocalFileSystem())
lance.dataset("my-dataset.lance", filesystem=fs)

Result: lance.dataset() does not accept a filesystem keyword argument in 4.0.1.

---

8. lancedb.connect() + open_table()

import lancedb
db = lancedb.connect(r"C:\LocalLink")       # succeeds — lists tables correctly
db.open_table("my dataset")                 # fails

Result: lancedb.connect() succeeds and list_tables() returns correct table names. However, open_table() panics in Rust with InvalidTableName because the table names contain spaces — spaces are rejected by the name validator even though the underlying directory names contain them. Even when symlinked names without spaces are used, the underlying open_table() call still resolves to UNC and hits the same URL error.

---

9. lancedb.connect() with file:// URI

db = lancedb.connect("file:///C:/LocalLink")
db.open_table("mydataset")

Result: Connect succeeds, open_table() fails with the UNC URL error.

---

10. PyArrow SubTreeFileSystem passed as storage_options to lancedb

db = lancedb.connect(r"\\server\My Share\data", storage_options={"filesystem": fs})

Result: 'SubTreeFileSystem' object cannot be converted to 'PyString' — storage_options does not accept filesystem objects.

---

Workaround

Copying the dataset to a true local NTFS path (no network, no symlink) and opening from there:

import lance, shutil, tempfile, os
SRC  = r"Z:\data\my-dataset.lance"

TEMP = tempfile.mkdtemp(dir=r"C:\Temp")

DST  = os.path.join(TEMP, "dataset.lance")
shutil.copytree(SRC, DST)

ds = lance.dataset(DST)       # works perfectly

print(ds.count_rows())

shutil.rmtree(TEMP)

This confirms Lance reads the format correctly — the failure is exclusively in path/URL resolution for network-backed paths.

---

Expected behaviour

lance.dataset() should be able to open a dataset from any path accessible to the Python process, including Windows-mapped network drives whose UNC share names contain spaces.

At minimum, either:

• The LocalFileSystem URL conversion should correctly handle UNC paths with spaces using the proper RFC-8089 file://host/share/path format, or
• lance.dataset() should accept a pyarrow.fs.FileSystem object so callers can supply their own filesystem abstraction (as pyarrow.dataset.dataset() does), bypassing Lance's internal path resolution entirely.

---

Notes

• The manifests do not contain hardcoded absolute paths — fragment filenames are stored as relative names only. The path issue is entirely in Lance's runtime resolution, not the stored data.
• lancedb.connect() successfully lists tables from the same path, suggesting the directory-listing layer handles UNC paths correctly. Only the dataset-open layer fails.
• The InvalidTableName validator in lancedb rejecting spaces is a separate but related friction point for datasets whose directory names contain spaces.