globusfs

An fsspec filesystem for Globus collections.

Status: working. Verified against three live collections — including ALCF Eagle (alcf#dtn_eagle, 747 project directories), where ls, glob, info, open() with mid-file seek, and sparse ranged reads all work against production Lustre. Writes (PUT/DELETE) round-trip on Globus Tutorial Collection 1. Anonymous pyarrow column projection works on a public collection — one column of sixty, a few KB instead of 360 KB — while that collection was intermittently returning backend-fault 404s.

import globusfs

# Browser login once; tokens persist to ~/.globusfs/tokens.json
fs = globusfs.filesystem("<collection-uuid>")

fs.ls("/")
with fs.open("data/file.parquet", "rb") as f:
    ...

Public collections need no credentials:

import fsspec, pyarrow.parquet as pq

fs = fsspec.filesystem(
    "globus",
    collection_id="isaac",
    https_url="https://g-05a4b6.2d513.8443.data.globus.org",
)

with fs.open("isaac/ability/ALL_2007-01.parquet", "rb") as f:
    # Reads only the bytes this column needs, over HTTP range requests.
    table = pq.ParquetFile(f).read(columns=["author"])

Globus is how large scientific datasets actually move between facilities, but there is no fsspec backend for it — so pyarrow, pandas, dask, and grain can't read a Globus collection the way they read s3:// or gs://. This fills that gap with one backend that serves all of them.

Globus Connect Server exposes two services, and this needs both:

The read path subclasses fsspec's HTTPFileSystem, which already speaks exactly the range dialect GCS serves.

Verified against a live collection:

• Backend flakiness surfaces as a 404. GCS load-balances across GridFTP backends; a failing one returns ENDPOINT_ERROR / GCS Manager Internal Error rendered as HTTP 404 — byte-identical in status to a genuinely missing file, and sticky for the life of a connection. Observed failure rates on the public test collection swung from 0/20 to 20/20 within minutes, hitting files, directories, and the collection root alike. Retries need a fresh connection, and the only way to tell a transient error from a real miss is to parse the body. A client that treats 404 as "absent" will report healthy data as missing.
• Suffix ranges (bytes=-8) return 416, which is how parquet readers typically seek to the footer. Because info() knows the true size, readers can use absolute offsets instead.
• HEAD is unusable. A HEAD 404 carries no body — and the body is the only thing distinguishing a backend fault from a real miss. So HEAD results are permanently ambiguous. Size and existence come from the Transfer API, or from a ranged GET (which does return a body and carries the total in Content-Range).
• The HTTPS interface has no directory listings at all — hence the Transfer API for metadata.

Resolved via endpoint_search; ALCF's docs list names, not UUIDs.

Display names work too — globusfs.filesystem("alcf#dtn_eagle") resolves the name to its UUID before building any scopes. (Passing a name to login() directly requires an authenticated client to resolve it; UUIDs never need a lookup.)

Eagle's collection root is already /eagle/projects, so paths are project-relative: fs.ls("/datascience"), not /eagle/projects/datascience.

Expect roughly 2 s per metadata or read operation through the DTN — fine for sparse reads and exploration, not for per-record access in a training loop. See the note on training workloads below.

Token acquisition is pluggable, because it varies more than anything else: a public collection needs nothing, a portal already has a token, an interactive user needs a browser.

Two constraints worth knowing, both from the globus-sdk docs:

• GlobusApp is not thread-safe, but fsspec shares one filesystem across threads. AppCredentials serializes every call through a lock.
• fsspec pickles filesystems to worker processes. A live token in the constructor args would be copied into every worker payload, so StaticToken refuses to pickle; use CallableToken reading from the environment or shared storage so workers re-read rather than receive.

Also set request_refresh_tokens=True — it defaults to False, and without it a long run dies when the access token expires mid-epoch.

A filesystem from globusfs.filesystem() is picklable, so it can be sent to dataloader workers or a dask cluster. Workers reload tokens from disk rather than receiving them, so on a cluster token_path must be reachable from every node:

fs = globusfs.filesystem(COLLECTION, token_path="/eagle/myproject/.globus-tokens.json")

Two consequences worth knowing. A worker that finds no usable tokens fails immediately instead of blocking on a browser prompt nobody can answer — log in once on the head node first. And no bearer token appears in the pickled bytes, so the credential stays where you put it.

This is built for remote and sparse reads: column projection, exploration, data not yet staged. For distributed training, staging with Globus Transfer to node-local scratch beats per-record HTTPS on every axis — no per-record latency, no token expiry mid-epoch, and it works with formats like ArrayRecord whose readers do their own seeking.

Grain wants a RandomAccessDataSource — __len__ plus __getitem__, thread-safe, picklable, and __repr__ for checkpointing. globusfs supplies the bytes; a small source supplies the indexing. This one indexes parquet row groups:

import fsspec
import pyarrow.parquet as pq
import grain


class ParquetSource:
    def __init__(self, path, **so):
        self._path, self._so = path, so
        self._fs = None
        with self._open() as f:
            self._n = pq.ParquetFile(f).metadata.num_row_groups

    def _open(self):
        # Built lazily so the source stays picklable for grain workers.
        if self._fs is None:
            self._fs = fsspec.filesystem("globus", **self._so)
        return self._fs.open(self._path, "rb")

    def __len__(self):
        return self._n

    def __getitem__(self, i):
        with self._open() as f:
            return pq.ParquetFile(f).read_row_group(int(i), columns=["author"])

    def __getstate__(self):
        return {**self.__dict__, "_fs": None}  # workers rebuild it

    def __repr__(self):
        return f"ParquetSource({self._path!r}, n={self._n})"


src = ParquetSource(
    "isaac/ability/ALL_2007-01.parquet",
    collection_id="isaac",
    https_url="https://g-05a4b6.2d513.8443.data.globus.org",
)
ds = grain.MapDataset.source(src).shuffle(seed=10)

Index row groups rather than individual records. Each __getitem__ is at least one HTTP round-trip, so per-record indexing over a wide-area link means one round-trip per record — see the note on training workloads.

Note also that ArrayRecordDataSource will not work over globus://: its reader is C++ and does its own file seeks, so it cannot route through a Python filesystem. Formats with Python-level readers (parquet, or your own offset index) are the ones that work.

Globus Streaming solves a different problem, and the two compose rather than compete.

Streaming intercepts only connection establishment — bind(), connect(), getaddrinfo — so unmodified binaries tunnel through firewalls with no per-byte overhead, but there is no file to open and no offset to seek to. It cannot back an fsspec filesystem. Conversely globusfs pays an HTTP round-trip per read and cannot see a live feed.

The natural split is streaming writes it, globusfs reads it back: an instrument ships frames to a facility during an experiment, and analysis addresses the resulting files afterward.

During the experiment — on the instrument host, pipe the detector into a socket that Globus tunnels to the facility:

# One-time: create a tunnel at https://app.globus.org/streams
globus-streams environment initialize --globus-contact dtn.alcf.anl.gov:8888 "$TUNNEL_ID"

# Frames leave the instrument through an ordinary TCP socket.
globus-streams-launch.sh "$TUNNEL_ID" ./detector-stream --host dtn.alcf.anl.gov --port 8888

On the facility side a listener accepts the stream and writes it to the collection's filesystem — plain files on Lustre, nothing Globus-specific:

globus-streams environment initialize --listener-contact-string 10.0.2.164:8888 "$TUNNEL_ID"
globus-streams-launch.sh "$TUNNEL_ID" ./ingest --out /eagle/myproject/run042/

Afterwards — the frames are now stored files, so globusfs addresses them from anywhere, no tunnel and no login on the instrument:

import globusfs

fs = globusfs.filesystem("05d2c76a-e867-4f67-aa57-76edeb0beda0")  # alcf#dtn_eagle

frames = fs.glob("/myproject/run042/*.h5")
print(f"{len(frames)} frames landed")

# Read only what you need: headers first, before pulling any bulk data.
for path in frames[:10]:
    with fs.open(path, "rb") as f:
        magic = f.read(8)
        f.seek(fs.info(path)["size"] - 4)
        footer = f.read(4)
    print(path, magic, footer)

Why the handoff is clean: streaming's job ends once bytes are on disk, and that is exactly where globusfs starts. Neither knows about the other.

Two caveats. Streaming authenticates every leg of the route but leaves end-to-end encryption to your application. And for the analysis pass, if you are reading every byte of every frame repeatedly — training rather than inspection — stage with Globus Transfer instead, per the note above.

Characterization tests hit a real public collection and are marked network:

pytest                    # everything
pytest -m "not network"   # offline only

MIT