BlockVolume

Package: flyteplugins.union.io

A writable Volume whose content is one ext4 filesystem image.

Use it for single-writer, small-file workloads — build caches (BuildKit), package caches, virtualenvs, node_modules, source trees. mount attaches the image as a loop device and returns the ext4 mount, so file metadata is handled in the kernel instead of a FUSE round trip per operation. Measured on dogfood: BuildKit cold build 67 s vs 125 s on a plain Volume (59 s on local disk); 5,000 small files 0.44 s vs 6.9 s.

Everything else is a Volume: commit snapshots it (after trimming freed blocks), forks share unchanged blocks, and returning it from a task finalizes it. Forks of a block volume are block volumes. A commit is crash-consistent by default – like snapshotting a live disk; ext4 replays its journal when a fork mounts it – and does not stall writers; commit(freeze=True) takes a clean snapshot at the cost of a brief write stall.

Trade-offs versus a plain Volume:

• Trusted workloads only. Attaching makes the node kernel parse the image. Behind the uvol mount broker it is allowed only for pods matching the cluster’s blockAllow rules; a privileged pod without the broker attaches it itself.
• One writer, no read-only mounts, not browsable file by file. Fork it to read it elsewhere.
• A size limit, set here and grown online with grow.
• More object-store traffic on file-heavy work (whole blocks are rewritten).

cache = BlockVolume.new(size="64G")
path = await cache.mount()          # ext4 mount point
...
await cache.grow("128G")             # online
return cache                         # finalized at task return

Parameters

class BlockVolume(
    kind: typing.Literal['flyte.volume/v1'] = 'flyte.volume/v1',
    name: str,
    bucket: str,
    storage: typing.Literal['s3', 'gs', 'wasb'] = 's3',
    region: typing.Optional[str] = None,
    endpoint: typing.Optional[str] = None,
    chunk_prefix: typing.Optional[str] = None,
    index: typing.Optional[flyte.io._file.File] = None,
    metadata_store_type: typing.Optional[str] = None,
    block_size: typing.Optional[str] = None,
    block_journal_blocks: typing.Optional[typing.List[int]] = None,
    metadata_prefix: typing.Optional[str] = None,
    chunk_branch: typing.Optional[str] = None,
    report: typing.Optional[bool] = None,
    used_bytes: typing.Optional[int] = None,
    inode_count: typing.Optional[int] = None,
    index_bytes: typing.Optional[int] = None,
    message: typing.Optional[str] = None,
    produced_by: typing.Optional[flyteplugins.union.io._base_volume.ActionRef] = None,
    parent_produced_by: typing.Optional[flyteplugins.union.io._base_volume.ActionRef] = None,
    artifact: typing.Optional[flyteplugins.union.io._base_volume.VolumeArtifact] = None,
    published_artifact_version: typing.Optional[str] = None,
    last_published_artifact_name: typing.Optional[str] = None,
    last_published_artifact_version: typing.Optional[str] = None,
)

Create a new model by parsing and validating input data from keyword arguments.

Raises ValidationError if the input data cannot be validated to form a valid model.

self is explicitly positional-only to allow self as a field name.

Properties

Methods

abandon()

def abandon(
    timeout: float = 30.0,
)

Let go of a live mount without committing: for a volume whose client died under it, where the sealing teardown would only block.

Best-effort and bounded, every step on its own: detach the block device, stop the client’s process group, retire the broker channel (aborted, never handed out again: the superblock it served is dead), stop the metadata store, close the writer’s slice domain. Nothing is uploaded or published; the last commit stands. Without this a caller that gives up on a dead volume leaked its block attachment and channel in the node’s broker, which refuses new ones past a per-volume cap, so a pod that kept recovering would eventually be unable to remount. Idempotent; a no-op when nothing is mounted.

clone()

def clone(
    destination: str,
    name: Optional[str] = None,
    region: Optional[str] = None,
    metadata_prefix: Optional[str] = None,
    artifact: Any = Ellipsis,
    concurrency: int = 64,
    since: Optional['Volume'] = None,
    publish_artifact: bool = False,
    artifact_version: Optional[str] = None,
) -> 'Volume'

Copy this volume into destination and return the copy.

destination is a bucket URI (s3://other-bucket/volumes) and becomes the clone’s bucket. The published index and the chunk objects it references – nothing else – are copied, and the result is a volume of the same class that reads and writes only destination: the source can be deleted, expired or lose access without affecting it.

Built for promotion: iterate on a volume in a dev bucket, clone v10 into staging, iterate on, clone v15 into the same destination. Clones of one volume family (the volume, its versions and its forks) into one destination share a filesystem identity there – a JuiceFS name and UUID of their own, recorded under <destination>/_uvol/clones/ – so their chunks are shared too, and each clone copies only the chunks the destination doesn’t already hold. Different families, or different destinations, share nothing.

What the destination holds is judged against a baseline: an earlier clone of the family there, whose index is compared slice by slice with this one’s, so the cost of a clone follows the size of one version’s metadata, not of everything ever cloned into the destination. By default the baseline is the last clone of the family into destination (each clone records itself next to the family’s identity); pass since= – e.g. the latest staging version found through an artifact query – to choose it. Without a usable baseline the destination’s chunks are listed instead. A baseline can only make a clone copy more than it needs, never less than it must.

A clone can be forked and written in the destination like any volume. Its slice allocator is moved past the source family’s, so what is written there can never take the key of a chunk a later clone brings over; and its identity is not the source’s, so not even a node-shared chunk cache mixes the two.

Every chunk is copied server-side (S3 CopyObject, GCS rewrite, Azure Copy Blob): no data passes through this process. Within one bucket that works on every backend; between two S3 buckets it needs boto3. Other pairs raise VolumeCloneUnsupported.

As for every volume, the chunks live at the root of the destination bucket under the volume’s JuiceFS name (<bucket>/<fs-name>/chunks/); the object store client ignores the path part of the bucket URI. The path still scopes what the plugin keeps there itself: the clone identities, the writer-session ledger and, outside a task, the versions.

• name – the clone’s name; defaults to self.name.
• region – the destination bucket’s region (S3); defaults to self.region.
• metadata_prefix – where the clone’s versions are written (see new); not inherited, so by default they follow the clone’s own bucket or task rather than the source’s.
• artifact – the clone’s artifact identity, as for fork.
• concurrency – copy requests in flight.
• since – the baseline: an earlier clone of this volume’s family into destination. A volume that isn’t one is refused.
• publish_artifact / artifact_version – register the clone in the artifact registry now, under its artifact identity, as RWVolume.commit does: the promotion flow looks the latest version up there for the next clone’s since=.

Clones the last published version: on a live mount, writes since the last commit are not included. A failed clone leaves only unreferenced objects behind and can simply be retried; the retry skips what was already copied.

commit()

def commit(
    message: Optional[str] = None,
    mount_path: Optional[str] = None,
    meta_dir: Optional[str] = None,
    publish_artifact: bool = False,
    artifact_version: Optional[str] = None,
    freeze: bool = False,
) -> 'ROBlockVolume'

Snapshot the live image as a new immutable version (see RWVolume.commit).

freeze=False (default): no writer stall; the snapshot is crash-consistent, and ext4 replays its journal when a fork mounts it. freeze=True: freeze the filesystem for the snapshot so it is clean (no replay needed); writers stall until the snapshot is pinned (with a JuiceFS client that announces the pin; older clients hold the freeze through the whole checkpoint, drain included), and the commit fails rather than publish if the freeze had to be released early.

empty()

def empty(
    name: str,
    bucket: Optional[str] = None,
    storage: Optional[StorageBackend] = None,
    region: Optional[str] = None,
    endpoint: Optional[str] = None,
    metadata_store_type: Optional[str] = None,
    metadata_prefix: Optional[str] = None,
) -> 'Volume'

Declare a brand-new volume. The first mount() call will bootstrap the namespace (the underlying client refuses to format over a non-empty bucket prefix).

If bucket is omitted, it is derived from the currently active task context as {raw_data_root}/{project}/{domain}/volumes — following Flyte’s own layout for offloaded data. Must be called from inside a task in that case.

Regardless of store type, mounting the returned Volume needs a FUSE-capable image and pod — the fuse3 apt package and flyte.PodTemplate().allow_fuse() — see mount for the full runtime requirements.

The namespace lives in an embedded BadgerDB directory, published as a store archive. It needs the forked client binary (checkpoint verbs and --slice-domain), which flyteplugins-union bundles, and a writable mount needs a claimed writer-session domain (the default domain-claim path provides one). metadata_store_type is accepted for compatibility and must be one of METADATA_STORE_TYPES.

storage is the JuiceFS object-store backend. When omitted it is inferred from the bucket URI scheme (s3:// → s3, gs:// → gs, abfs(s):// → wasb), so a GCS/Azure bucket — including the default one derived from raw_data_path — gets the right backend without the caller spelling it out.

metadata_prefix is where this volume’s new versions (their index and metadata objects) are written. Leave it unset inside a task (the action’s output path is used); outside one – an app, a sidecar, a script – it defaults to $UNION_VOLUME_METADATA_PREFIX and then to <bucket>/versions, so a Volume works anywhere with no task context.

region pins the object-store region onto the Volume (S3 only — it forms the endpoint host). When omitted it’s derived from the ambient AWS_REGION / AWS_DEFAULT_REGION at mount time; pass it to make the Volume self-describing so a cross-region remount doesn’t depend on the consumer’s env.

finalize()

def finalize(
    message: Optional[str] = None,
    timeout: float = 60.0,
    publish_artifact: bool = False,
    artifact_version: Optional[str] = None,
) -> 'ROBlockVolume'

Detach, unmount and publish as an ROBlockVolume (see RWVolume.finalize).

fork()

def fork(
    name: str,
    as_: Literal['rw', 'ro'] = 'rw',
    mount_path: Optional[str] = None,
    meta_dir: Optional[str] = None,
    timeout: float = 60.0,
    artifact: object = Ellipsis,
) -> Union['BlockVolume', 'ROBlockVolume']

Branch this block volume (see RWVolume.fork): a BlockVolume, or an ROBlockVolume snapshot with as_="ro".

from_block()

def from_block(
    source: Volume,
    name: Optional[str] = None,
    workers: int = 8,
    mount_path: Optional[str] = None,
) -> 'RWVolume'

A new regular volume holding a copy of a block volume’s files – the reverse of BlockVolume.from_volume, with the same copy semantics. ext4’s lost+found is not copied.

A block source that is mounted by this task is read from its live ext4 mount (do not write to it meanwhile); otherwise a temporary fork of it is mounted for the copy and torn down after (a block image mounts read-write only). Returns the new volume mounted.

from_locator()

def from_locator(
    locator: str,
) -> 'ROVolume'

Load a previously published volume version by its locator.

The inverse of locator: download the metadata object at locator (the full serialized Volume value) and reconstruct it as an immutable ROVolume — index, bucket, metadata_store_type, stats and lineage all recovered, so the result is mountable read-only (or fork-able to branch + write) without any other arguments. This is how you reference a specific volume version across runs: stash vol.locator somewhere, then Volume.from_locator it back later.

Reads only the metadata object; the index and chunks are fetched lazily by mount. Needs object-store credentials for locator but no active task context. Raises VolumeError if locator is empty.

from_volume()

def from_volume(
    source: Volume,
    name: Optional[str] = None,
    size: Optional[str] = None,
    workers: int = 8,
    mount_path: Optional[str] = None,
) -> 'BlockVolume'

A new block volume holding a copy of a regular volume’s files.

There is no in-place conversion: a regular volume stores each file as its own JuiceFS file and a block volume stores one ext4 image, and neither layout’s data can be reused by the other. This mounts both and copies the tree (modes, timestamps, symlinks and hard links, as cp -a; see workers for hard links across top-level entries).

size defaults to the source’s recorded usage with headroom – by bytes and by file count, since ext4 has one inode per 16 KiB of image (block_size_for). A source that is mounted by this task is read from its live mount (do not write to it meanwhile); otherwise it is mounted read-only for the copy and unmounted after.

Returns the new volume mounted: keep working in it, commit it, or return it from the task (which publishes it).

get_artifact_metadata()

def get_artifact_metadata()

Artifact declaration for the declarative (returned-as-output) publish path — flyte-sdk’s output conversion calls this on every top-level task output exposing it and, when it returns metadata, emits a ProducedArtifact on the Outputs envelope; the backend registers the artifact atomically with the action record (no RPC from the task).

Returns None — no declaration — when this volume carries no artifact intent, or when this exact seal is already in the registry (it was published imperatively via publish_artifact=True; re-declaring it would duplicate the version and emit a self-parent).

The version is left to the literal’s content hash (version_from_content — the same identity hash an imperative publish defaults to), because for a still-mounted RWVolume the final seal doesn’t exist until the transformer auto-finalizes during conversion. For that same reason stats attrs reflect this handle’s last seal (or are absent on a fresh handle) rather than the final one; the registered value itself always carries the authoritative numbers.

grow()

def grow(
    size: str,
)

Grow the image and its filesystem online, while mounted. Never shrinks.

The new size is recorded on this handle, so later commits and forks carry it.

migrate_metadata_store_type()

def migrate_metadata_store_type(
    new_metadata_store_type: str,
    meta_dir: Optional[str] = None,
    new_meta_dir: Optional[str] = None,
) -> 'Volume'

Re-host this Volume’s metadata on new_metadata_store_type without copying data chunks.

new_metadata_store_type must differ from the current store type. The full namespace is exported and re-imported into a fresh meta store pointing at the same bucket. Chunks are addressed by stable IDs that are preserved across the migration, so no chunk traffic is required.

The returned Volume has a fresh index (snapshot of the loaded meta store), parent_produced_by linking to the pre-migration version, the same bucket / storage / name, and the new metadata_store_type.

Intent is migration, not fork: the old store is meant to be retired. As a defense against accidental concurrent use, the loaded store’s chunk-slice / inode / session counters are advanced by a random offset (same mechanism as fork) so that even if the old store is still mounted somewhere, its writes can’t collide with the migrated store’s writes in shared object-store keys.

Does not require a FUSE mount on either side. Safe to call from any task pod running the Volume runtime (Redis tooling is only needed if one of the stores is "redis").

model_post_init()

def model_post_init(
    context: Any,
)

This function is meant to behave like a BaseModel method to initialize private attributes.

It takes context as an argument since that’s what pydantic-core passes when calling it.

mount()

def mount(
    mount_path: Optional[str] = None,
    meta_dir: Optional[str] = None,
    cache_dir: Optional[str] = None,
    timeout: float = 120.0,
    writeback: bool = True,
    upload_delay: Optional[str] = None,
    max_uploads: int = 50,
    attr_cache: float = 60.0,
    entry_cache: float = 60.0,
    dir_entry_cache: float = 60.0,
    read_only: bool = False,
    shared_node_cache: Optional[bool] = None,
    cache_size_mb: Optional[int] = None,
    passthrough: Optional[bool] = None,
    enable_xattr: bool = False,
    transport: str = 'auto',
    credentials: str = 'auto',
    log_path: Optional[str] = None,
    on_existing_index: str = 'replace',
) -> Path

Format (if fresh) and mount the volume at mount_path in this process, and return the resolved mount point as a Path (also available afterwards via the mount_path property).

Call once near the top of a task body before reading or writing under mount_path.

shared_node_cache selects the chunk cache shared by every pod on the node, which the pod template exposes as $UNION_VOLUME_SHARED_NODE_CACHE_DIR (see allow_volumes). Every pod on the node mounting this volume family then fetches each chunk from object storage once per node rather than once per pod. The default None uses it automatically whenever it is present and the mount is read-only – a pod that never heard of the flag still benefits, and still contributes chunks. True requires it (an error if the pod has none, or if the mount is writable); False never uses it. Writable mounts always get a per-pod cache: the writeback staging queue lives inside the cache directory and is per-writer. Clients sharing a directory each run their own eviction against their own budget, so they can evict one another; the hit rate depends on what else is mounted on that node.

cache_size_mb caps the on-disk chunk cache for this mount. When omitted, a mount whose cache lives on the volume that allow_volumes(cache_size=...) attached takes 90% of that volume’s size – per mount: two volumes mounted in one task each get the full budget, so a task that mounts several must split it here, or size the volume for the sum, or kubelet evicts the pod for exceeding the emptyDir. The default cache anywhere else is on the container’s writable layer, which counts against the pod’s ephemeral-storage limit while the client’s free-space guard reads the node’s disk; there the mount leases a budget from the limit allow_volumes publishes: half of the ephemeral share not already leased by other mounts (the share is half the limit), at least 1 GiB, returned at unmount. So a 30 GiB pod’s first mount gets 7.5 GiB and the next 3.75 GiB. An explicit cache_dir and the node cache are left on the client’s own default and its free-space guard.

Runtime requirements (both needed, or the mount fails):

• Image — the fuse3 apt package. JuiceFS execs the fusermount3 userspace helper to mount; the default minimal images don’t ship it, so the mount client exits immediately with fuse: fuse is not installed. Add it with flyte.Image...with_apt_packages("fuse3").
• Pod — pod_template=flyte.PodTemplate().allow_fuse() on the flyte.TaskEnvironment, which grants the /dev/fuse device and the capability an unprivileged mount needs (the cluster must run a FUSE device plugin; the Union data plane ships one).

The mount point, meta_dir and cache_dir must also be writable by the task user; the name-keyed defaults live under $HOME, which the default image owns.

passthrough selects the FUSE passthrough fast-write path: None (default) means on for writable mounts unless UNION_JUICEFS_PASSTHROUGH=0 is set; True/False force it. Read-only mounts never use it. enable_xattr turns on extended attributes for the mount (off by JuiceFS default). Both matter when the volume is used as an overlayfs layer (for example as a container snapshotter’s root): overlayfs needs trusted.overlay.* xattrs, and the kernel refuses a passthrough FUSE superblock as a layer (“maximum fs stacking depth exceeded”), so such a mount needs enable_xattr=True, passthrough=False.

When writeback=True (default), writes land in the local cache directory first and are uploaded asynchronously in the background. This decouples write latency from object-store round-trips. The pending upload queue is drained on commit(); if the pod dies before commit, in-flight chunks are lost — but that’s fine because the Volume itself is never published in that case.

upload_delay (e.g. "1h", "30m", "5s") defers uploads by the given duration. Useful for write-scratchy workloads — files that are written and then overwritten / deleted within the delay window are never uploaded at all. Default (None) is no extra delay; the background uploader starts as soon as a chunk is written. Has no effect without writeback=True.

max_uploads caps concurrent S3 PUTs (default 50; underlying client default is 20). Bumping helps write-burst phases when the chunks are small enough that 20 streams can’t saturate the link.

attr_cache / entry_cache / dir_entry_cache are kernel-side TTLs in seconds for file attributes, name-to-inode lookups, and directory listings respectively. Defaults are 60.0 for all three, which collapses stat / getattr / lookup storms by an order of magnitude on directory-heavy workloads (Go toolchain, package managers, codegen). This is safe because a Volume is single-writer for the duration of its mount — no external mutator is supported by this mechanism. Concurrent-writer scenarios will be opt-in via a separate API when added.

Periodic checkpointing and crash recovery are intentionally not part of this method. Callers that want a keep-alive loop drive it themselves with RWVolume.commit on whatever cadence they choose, and own any resume-from-checkpoint policy at the usage layer.

new()

def new(
    name: Optional[str] = None,
    bucket: Optional[str] = None,
    size: str,
    storage: Optional[StorageBackend] = None,
    region: Optional[str] = None,
    endpoint: Optional[str] = None,
    metadata_store_type: Optional[str] = None,
    artifact: Union[bool, str, 'ArtifactMetadata', 'VolumeArtifact', None] = None,
    metadata_prefix: Optional[str] = None,
) -> 'BlockVolume'

Declare a new, empty block volume of size (e.g. "64G", binary units).

The image is created sparse on the first mount — only what ext4 writes is ever stored or uploaded.

recover_mount()

def recover_mount(
    timeout: float = 120.0,
) -> Path

Remount this volume after its mount daemon died under it.

The case: the client process is gone (crashed, or torn down without a successor) and its FUSE connection was aborted — by the broker’s auto-abort or by _broker.retire_channel — so every access to the old mount fails with ENOTCONN. The metadata store and cache directories are local to this pod and intact, so the same volume can be served again from a fresh channel with nothing lost that had reached them: writeback resumes uploading what it had staged.

Steps: stop whatever is left of the old adopter and drop its bookkeeping; retire its channel lease (the main premount stays dead for the pod’s lifetime, so a later mount leases a runtime channel); lease a fresh channel; start the client against the SAME meta and cache dirs; repoint the requested-path symlink; restart the report. Returns the new mount path. Only the broker path is supported — a privileged in-pod mount would need fusermount on a dead superblock.

Raises VolumeMountError if this instance has no live mount to recover, or the metadata dir is gone.