I fixed the tests but i'm not happy with the current implementation.

The small caveat we have now that you want to know about: when the input is hdfs:///path, the asset is now stored as hdfs:/path. This is a consequence of how urllib.parse represents and serializes URIs, not a deliberate design choice on my part:

>>> from urllib.parse import urlsplit, urlunsplit
>>> urlsplit("hdfs:///apps/x") == urlsplit("hdfs:/apps/x")
True
>>> urlunsplit(urlsplit("hdfs:///apps/x"))
'hdfs:/apps/x'

urlsplit cannot distinguish the two forms: they parse to identical SplitResult instances with an empty netloc and urlunsplit cannot emit // for an empty authority; it omits the // entirely. Since the provider hook only sees and returns a SplitResult, and the surrounding _sanitize_uri in task-sdk calls urlunsplit on the result, there is no way for the provider on its own to preserve the /// form in the stored URI.

Round-trip identity is still stable: both hdfs:///apps/x and hdfs:/apps/x normalize to hdfs:/apps/x, so asset matching remains consistent across re-parses. For the OpenLineage conversion, an empty netloc produces a hdfs:// namespace, which is what consumers expect for fs.defaultFS-resolved paths.

If preserving the literal hdfs:///path form in storage is considered worth the additional surface area, it could be done with a small opt-in in _sanitize_uri (for example, a normalizer.preserve_empty_authority = True attribute set when we want the empty-authority form retained, and a corresponding string-level fix-up after urlunsplit). That keeps the behavior change scoped to opt-in normalizers (only HDFS would set it today), and leaves others untouched.

I deliberately left this out of the current PR to keep the change minimal and provider-local. Happy to follow up with that task-sdk change if someone thinks it is worth.

Update after digging into normalization.

Dropping the host check alone is not enough: the SDK normalizes asset URIs with urllib.parse.urlunsplit, which only keeps the // authority for schemes in uses_netloc. hdfs is not in that list, so hdfs:///apps/x was being silently rewritten to hdfs:/apps/x. Since asset URIs are primary keys and feed OpenLineage that silent rewrite is not acceptable.

Current branch:

Register hdfs in uses_netloc so hdfs:///path round-trips intact, like file already does.
hdfs://namenode:8020/path is unaffected;

In current behaviour hdfs:/path canonicalizes to hdfs:///path. Is this solution acceptable knowing that Hadoop treats them differently at the URI level (hdfs:/path and hdfs:///path could be distinct assets).