nfs4fs

nfs4fs is a Linux fsspec implementation backed by VFSI's vectorized NFSv4 client. Bulk fsspec operations are translated into NFS compounds, reducing round trips for workloads with many small files.

Install

python -m pip install nfs4fs

Published wheels use CPython's stable ABI and support standard CPython 3.9 and newer on Linux x86-64 and AArch64. A separate CPython 3.14 free-threaded wheel is built and tested on x86-64.

A source build additionally requires Rust, Clang, pkg-config, libntirpc 4.3 or newer, Kerberos/GSS development headers, and userspace-RCU development headers. Source and editable builds link the administrator-provided system libntirpc dynamically, so that shared library must be present at runtime.

Release wheels are self-contained. They are built against a checksum-pinned libntirpc source revision, and auditwheel copies libntirpc plus its Kerberos (libgssapi_krb5/libkrb5) and userspace-RCU (liburcu) dependencies into the wheel's .libs/ directory, rewriting RPATH so the package imports without a system libntirpc. CI fails a release whose wheel does not bundle those libraries.

The package installs fsspec.specs entry points for both nfs4 and vfsi, so normal use does not require import nfs4fs before calling fsspec.

NFSv4

Pass the server as an option or as the URL authority:

import fsspec

fs = fsspec.filesystem(
    "nfs4",
    host="nfs.example",
    root="exports/project",
    auth="auth_sys",
    auto_mkdir=False,
)

fs.pipe({"/file-1": b"hello", "/file-2": b"world"})
files = fs.cat(["/file-1", "/file-2"])

with fsspec.open(
    "nfs4://nfs.example/exports/project/file-1", "rb", auth="auth_sys"
) as file:
    print(file.read())

The pipe() call gives both writes to VFSI at once. For these two small files, it encodes both independent write chains in one NFSv4 COMPOUND and completes them in one network round trip. The following cat() also preserves the batch; it currently uses one compound for bounded size discovery and one for the reads.

The client negotiates NFSv4.2 and falls back to NFSv4.1. Set minor_version=1 or minor_version=2 to require a specific version. Select auth="auth_sys" explicitly to use AUTH_SYS. For Kerberos-backed RPCSEC_GSS, obtain a ticket in the process credential cache and select auth="krb5" or auth="krb5i"; the latter integrity-protects RPC payloads. Selecting either mode is explicit and fails closed if Kerberos authentication cannot be established; the client does not downgrade to AUTH_SYS. An optional service_principal overrides the default nfs@<host> principal.

kinit user@EXAMPLE.COM

fs = fsspec.filesystem(
    "nfs4",
    host="nfs.example.com",
    root="exports/project",
    auth="krb5i",
)

Connect from an existing mount

If the directory is already mounted on Linux, infer the connection instead:

fs = fsspec.filesystem("nfs4", mount="/mnt/data/git/some/tree")
fs.pipe({"/file-1": b"hello", "/file-2": b"world"})
files = fs.cat(["/file-1", "/file-2"])

mount must be an absolute path to an existing directory inside the mount. Relative paths are rejected to avoid working-directory-dependent instance caching. If server:/export is mounted at /mnt/data, this example uses /export/git/some/tree as its remote root. Discovery supports NFSv4.1/4.2 TCP sec=sys mounts and infers AUTH_SYS from that explicit mount configuration. Read-only mounts remain read-only through the direct client. Unsupported security modes, bind-root mappings, and directories containing nested mounts fail with a descriptive error. Do not combine mount with host, root, auth, minor_version, or other explicit connection settings; tuning such as timeouts and pool size remains available.

Kerberos mounts are not supported by mount discovery yet. A mount using RPCSEC_GSS (sec=krb5, sec=krb5i, or sec=krb5p) is rejected; mount= never silently falls back to AUTH_SYS. To use Kerberos, configure the server and remote root explicitly with auth="krb5" or auth="krb5i", as shown above. This establishes an independent GSS context using the process's Kerberos credentials, not the kernel mount's context. RPCSEC_GSS privacy (krb5p) is not currently supported by nfs4fs.

Discovery happens during construction. All operations, pooled connections, and reconnects use the same pinned remote configuration. fs.mount exposes the resolved local directory and fs.read_only reports its write restriction. The client uses direct NFS for all operations, with its own caches and open state. It does not share or invalidate the mounted kernel client's caches.

root is a client-side path prefix, not a security boundary. A relative symlink inside it can resolve elsewhere in the same NFS export. Configure a server export rooted at the intended directory when callers must be confined to that directory.

compound_size_limit caps the payload merged into one compound. The default is 1 MiB; lower it when a server has a smaller request limit:

fs = fsspec.filesystem(
    "nfs4", host="nfs.example", auth="auth_sys", compound_size_limit=256 * 1024
)

Connection setup and individual NFS RPCs are independently bounded. Defaults are 10 seconds and 5 seconds, respectively:

fs = fsspec.filesystem(
    "nfs4",
    host="nfs.example",
    auth="auth_sys",
    connect_timeout=5.0,
    request_timeout=15.0,
)

Operations that vectorize

The largest benefit comes from giving fsspec multiple paths at once:

• cat, cat_ranges, and open_files batch reads;
• pipe and put batch writes;
• recursive get, put, cp, rm, and tree walks use vector metadata and mutation operations;
• cp_file/copy retain the low-round-trip vector path for bounded small-file batches. Files beyond the read allocation budget use NFSv4.2 server-side copy when available, in bounded 16 MiB extents, with a streaming client-side fallback.

minor_version(), capabilities(), server_copy_enabled(), compound_stats(), and rpc_stats() are available on the native client for feature inspection and diagnostics.

Small-file benchmark

A local benchmark compared nfs4fs directly with vanilla fsspec LocalFileSystem accessing the same export through a Linux kernel NFS mount. The server was NFS-Ganesha V15.3 over NFSv4.2 on an AArch64 Ubuntu VM, using nfs4fs 0.3.1 and fsspec 2026.7.0. Linux netem added 500 microseconds to each loopback traversal, approximately 1 ms of network round-trip delay; ping averaged 1.43 ms including VM scheduling overhead. Each result is the median of 30 fresh-path trials over 20 files of 4 KiB, with connection setup excluded, fresh paths used for every trial, and client execution order alternated.

The same benchmark was repeated after an untimed warm-up, reusing the exact same files for every measured trial. Ping averaged 1.35 ms during this run:

The kernel's metadata and page caches reduced its repeated-read time, but a normal NFSv4 open and close still involves server state. LocalFileSystem opens the files serially, so those latency-bearing operations remain serial as well. nfs4fs does not retain file data across separate cat() calls; its warm improvement comes from server-side caches while it continues to batch the remote operations. Repeated writes changed little because both clients still have to send the new contents to the server.

The benchmark measures a latency-bound small-file workload, not sequential bandwidth, and results will vary with the server, client, mount options, and compound limits. The exact driver is benchmarks/small_files.py; it verifies returned data and reports VFSI's compound and RPC counters so batching regressions are visible alongside timing changes. Pass --reuse-paths to reproduce the warm-cache variant.

Directory and large-file benchmarks

The benchmark drivers also measure a metadata-heavy directory tree (find and recursive rm) and bounded sequential reading of a single large file. Each compares nfs4fs with fsspec LocalFileSystem on the same kernel NFS mount, checks results, and reports timings alongside nfs4fs RPC counts. The large-file driver explicitly disables per-open read caching so it measures the streaming path.

For stable large files, opt-in read_stream_pipelined() overlaps positional reads across independent native sessions while delivering chunks in order:

fs = fsspec.filesystem(
    "nfs4", host="nfs.example", auth="auth_sys", connection_pool_size=3
)
with open("download.bin", "wb") as output:
    fs.read_stream_pipelined(
        "/large.bin", lambda offset, data: output.write(data), workers=3
    )

Read-ahead is capped at one chunk per worker-owned descriptor, up to max_in_flight chunks and max_buffered_bytes total. With the default three workers and 1 MiB chunks, at most 3 MiB is in flight. This does not change fs.open() and is not a coherent snapshot if another client modifies the file during the read. On the development VM, a 64 MiB warm-file read with approximately 5 ms RTT fell from a 607 ms sequential median to 262 ms with three workers; the kernel client's warm page-cache read was still faster. See the benchmark methodology for the measured workload and limits.

All data-transfer entry points accept fsspec's callback= argument. Bulk operations (cat, cat_ranges, pipe, get, put, and copy) report item progress on the parent callback and byte progress on a branched callback for each file. Single-file operations report bytes directly, and oversized streaming transfers update after every chunk.

File buffering

Read handles use fsspec's standard per-open caches by default. The default 1 MiB block size matches the VFSI range-I/O path and can be changed per filesystem or per direct fs.open() call:

fs = fsspec.filesystem(
    "nfs4",
    host="nfs.example",
    auth="auth_sys",
    block_size=256 * 1024,
    cache_type="readahead",
)

with fs.open("/large.parquet", "rb", cache_type="blockcache") as file:
    header = file.read(4096)
    file.seek(-8192, 2)
    footer = file.read()

All cache types registered by fsspec are accepted. Use cache_type="none" for unbuffered reads, especially when another client can modify an already open file. Buffered data belongs to one open handle and is discarded on close. Blocks already fetched by that handle remain cached, but fsspec does not revalidate the source while the handle is open, so uncached blocks may reflect a concurrent writer. This is a per-open cache, not a persistent client cache.

fsspec also provides a persistent sparse-file wrapper under the same blockcache name:

cached = fsspec.filesystem(
    "blockcache",
    target_protocol="nfs4",
    target_options={"host": "nfs.example", "auth": "auth_sys"},
    cache_storage="/var/tmp/nfs4fs-cache",
    check_files=True,
)

With check_files=True, each open compares the saved source identity with nfs4fs's ukey() and starts a fresh local cache generation when it changes. For NFS this identity uses FATTR4_CHANGE and the file ID; the dummy backend uses file ID, nanosecond timestamps, and size. Validation happens at open, not on every read. With fsspec's default check_files=False, persistent entries are reused without source validation until expiry_time, so that mode is appropriate only when external clients use immutable or versioned paths. Mutations made through the same nfs4fs target invalidate registered persistent block-cache generations automatically, including renamed or recursively deleted subtrees. Already-open handles retain their per-handle cached blocks; new persistent generations use separate backing files so refreshes cannot truncate an active mmap. Retired files are unlinked, with their storage released when existing handles close. cache_check controls how often local cache metadata is reloaded and does not validate the server. invalidate_cache() controls directory listings; use pop_from_cache() to evict persistent file data explicitly.

Files opened together with fsspec.open_files() share a coordinator. A cache miss reads the same aligned range from a bounded group of sibling descriptors through one VFSI vector operation. A requested whole-file read keeps its fast path for the default read-ahead cache; cache_type="blockcache" fills its bounded LRU block-by-block. Speculative siblings are capped to one block (which may contain a whole small file). Retained group prefetch is also bounded by the smaller of max_batch_bytes and read_all_max_total_bytes; older speculative entries are evicted when necessary. Set vectorized_buffering=False to disable this adaptive fan-out. The persistent blockcache wrapper participates in the same coordinator: it validates source identities in a metadata batch, serves complete hits locally, and fills sparse misses with vector reads.

Scalar fs.open() pins a native descriptor immediately so its handle remains bound to the opened file after a rename or pathname replacement. Path-based cat() and grouped open_files() retain their batched read paths.

Writes remain write-through by default. Set write_buffering=True to delay small writes until flush() or close(); an open_files() write group drains one block from each file per vector operation. Explicit flush() always sends that file's staged data to the server. Update modes containing + retain the unbuffered seekable implementation, and fsspec transactions retain their existing disk-spooled commit behavior.

A failed buffered flush makes the writer unusable, including semantic errors such as a full server. Close it to release its descriptor; close() reports the failed writer without replaying any bytes that might already have been written.

Directory listing cache

Directory listing caching is disabled by default so that NFS namespaces modified by other clients remain immediately visible. Enable a bounded-staleness cache with the standard fsspec options:

fs = fsspec.filesystem(
    "nfs4",
    host="nfs.example",
    root="exports/project",
    auth="auth_sys",
    use_listings_cache=True,
    listings_expiry_time=1.0,
    max_paths=1024,
)

listings_expiry_time is the TTL in seconds and max_paths bounds the number of directory paths tracked by fsspec's DirCache; it does not bound the number of entries inside one directory. ls(..., refresh=True) bypasses the cached listing. Native vectorized walks populate the cache for the complete subtree, so repeated walk, find, glob, and du operations can reuse it without turning a miss into one request per directory.

nfs4fs invalidates affected listings after its own writes and namespace mutations, including uncertain failures. Changes made by another client become visible when the TTL expires or when the caller requests a refresh. Use use_listings_cache=False when any period of staleness is unacceptable.

Production notes

• One filesystem instance owns connection_pool_size native sessions (one by default), each protected by its own mutex. Operations on one session are serialized, while a larger pool permits concurrent independent requests. Use separate instances with skip_instance_cache=True when independent filesystem lifetimes or configuration are required.
• Native calls release the CPython interpreter lock while waiting for storage. The extension also declares free-threaded CPython support; each filesystem's native-session mutex still serializes operations assigned to that session.
• On a transport failure, safe idempotent path reads and positional reads in read_stream_pipelined() reconnect their affected session and retry once by default. Mutations are never replayed automatically because the server may already have completed an ambiguously failed request. Set auto_reconnect=False to disable automatic read recovery.
• A process fork is detected before the next operation and creates a fresh native session in the child without sending protocol teardown over the parent's inherited connection. Do not fork while an operation is in flight or share an already-open Nfs4File across a fork. Once application or RPC helper threads exist, use multiprocessing's spawn method. Prefer creating filesystem instances after worker processes start.
• Call close() or use with fsspec.filesystem(...) as fs: to release the native session deterministically. Closing also removes the instance from fsspec's instance cache.
• exists, isfile, and isdir return False for missing paths but propagate authentication and connectivity failures. This prevents outages from being mistaken for absent data.
• auto_mkdir is False by default. Enable it only when write operations are allowed to create missing parents.
• root rejects . and .. components. Use it to keep all paths under an export-relative prefix; it is not a substitute for server-side authorization.
• Vector batches are bounded by both batch_size (128 items) and max_batch_bytes (64 MiB). Files larger than that byte threshold stream in transfer_chunk_size chunks (8 MiB). Tune these per server and workload; every value must be positive.
• One native session preserves maximum batching by default. For thread-pool workloads, set connection_pool_size above one; independent path operations are distributed across sessions while every open descriptor remains pinned to its owning session. Each additional session consumes one server client identity and connection, so size the pool deliberately.
• APIs that materialize complete remote results have independent safety limits. read_all_max_total_bytes bounds every bytes-returning read, including whole-file reads, buffered read(-1), cat, and cat_ranges (16 MiB); directory_max_entries (100,000), directory_max_path_bytes (16 MiB), and walk_max_depth (128) bound listings and recursive walks. These limits are configurable per filesystem instance and survive reconnects.
• Transaction writes use a disk-backed spool after transaction_spool_threshold (8 MiB), upload to unpredictable temporary names in each destination directory, then expose the batch with a vectorized rename. This prevents partial file contents from becoming visible. Like fsspec transactions generally, a multi-file commit is staged rather than a server-wide atomic transaction: a server failure during the final rename batch can expose a prefix of the batch.
• pipe_file(..., mode="create") and create-mode put use an exclusive server create, so concurrent creators cannot silently overwrite one another.

Local development

On Ubuntu/Debian:

sudo apt-get install clang libclang-dev pkg-config libntirpc-dev \
  libkrb5-dev libgssglue-dev
python -m venv .venv
.venv/bin/pip install hypothesis maturin fsspec pytest
cd adapters/nfs4fs
../../.venv/bin/maturin develop
../../.venv/bin/python -m pytest tests

Tests use the local dummy backend unless NFS integration variables are provided. CI runs the same suite against NFSv4.1, NFSv4.2, and patched server-side COPY. Samba coverage lives with the vsmbfs adapter.

The default suite includes a deterministic Hypothesis state machine that runs randomized filesystem histories against both nfs4fs's dummy backend and fsspec's LocalFileSystem oracle. It compares public results, exception classes, directory trees, and file contents after every step. Increase its budget for a longer local fuzz run:

NFS4FS_FUZZ_EXAMPLES=1000 NFS4FS_FUZZ_STEPS=100 \
  ../../.venv/bin/python -m pytest tests/test_fuzz_parity.py