DACS Client Library

This Python library provides interfaces to perform file operations on Azure blobs with caching mechanisms. It includes async operations, a blobfile-compatible synchronous adapter (DACSBlobfileAdapter), and a boostedblob-compatible async adapter (dacs_boostedblob_adapter).

Contents

• Installation
• Authentication
• DACS Cache Client API
  • DACSCacheClient (async)
  • DACSBlobfileAdapter (blobfile-compatible, sync)
  • DACSBoostedBlobAdapter (boostedblob-compatible, async)
  • Blobfile Integration (monkey-patching)
  • Utility API
    • BoundedBufferManager
  • Exception Hierarchy
• Sample Code
• FAQs
• For Dev Operations

---

Installation

Build the DACS Cache Client wheel from this repository or obtain one from your build artifacts. No public package index is currently documented for this component.

The package is built as manylinux wheels for both x86_64 (amd64) and aarch64 (arm64). The wheels are repaired with auditwheel so that all required native shared libraries (e.g. libcurl, libssl, libjsoncpp, libuuid, etc.) are bundled inside the wheel. pip automatically selects the correct wheel for your platform.

Wheel filenames look like:

dacs_client-<version>-py3-none-manylinux_X_Y_x86_64.whl
dacs_client-<version>-py3-none-manylinux_X_Y_aarch64.whl

The exact manylinux_X_Y tag depends on the glibc version of the build image; pip will pick the most compatible wheel.

Install the DACS Cache Client

Install the wheel selected for your platform:

python3 -m pip install /path/to/dacs_client-<version>-*.whl

---

Authentication

The DACS Cache Client supports four authentication modes, resolved automatically at initialization:

Mode	Environment Variable(s)	Description
SAS Token	USE_SAS_TOKEN=true and STORAGE_INTCPT_SAS_TOKEN	For development/testing. Set both env vars.
Access Token	SI_azBlobAccessToken	Provide a pre-fetched Azure access token.
Managed Identity	AZURE_CLIENT_ID, MSI_ENDPOINT and MSI_SECRET	For production. Falls back to IMDS endpoint if MSI vars are not set.
Azure Identity SDK (default)	Fetch token using Defaultazurecredential in the Azure Identity SDK	Default auth mechanism if none of the above env vars are set

Note: No storage account name is needed — dynamic accounts and containers are used automatically.

---

DACS Cache Client API

DACSCacheClient

DACS Cache Client offer both sync and async version of various operations. The sync versions have _sync suffix.

from py_dacs_client.dacs_client import DACSCacheClient

Constructor

DACSCacheClient(executor=None, write_mode=None)

Parameter	Type	Default	Description
executor	concurrent.futures.Executor	None	Thread pool executor. If None, a ThreadPoolExecutor is created with max(4, min(cpu_count * 2, 40)) workers.
write_mode	WriteMode	None	Write mode configuration. If None, writes are disabled. See Write Modes.

Blob Operations

Method	Description
get_file_size(blob_path) -> int	Get the size (in bytes) of a blob.
get_file_stat(blob_path) -> Stat64	Get complete file statistics (size, mtime, atime, ctime, mode, uid, gid, etc.).
get_file_handle(blob_path, flags) -> DACSCacheFileHandle	Get a managed file handle for read/write. Use os.O_RDONLY, os.O_WRONLY | os.O_CREAT | os.O_EXCL, etc.
release_file_handle(file_handle)	Release a managed file handle and close the underlying file.
read_blob(blob_path, buffer, size, offset) -> int	Read data from a blob into a buffer. File open/close is managed automatically. Returns bytes read.
write_blob(blob_path, buffer, size) -> int	Write data from a buffer to a new blob (uses O_CREAT | O_EXCL). File open/close is managed automatically. Returns bytes written.
list_dir(blob_dir_path) -> AsyncIterator[Dict]	Async iterate over directory entries. Each entry is a dict with keys: d_ino, d_off, d_reclen, d_type, d_name, d_size, d_atime, d_mtime.
shutdown()	Close all open file handles and log final metrics (reads, writes, bytes, elapsed time).

Blob path format: blob_path arguments accept three formats:

• az://account.blob.core.windows.net/container/path (canonical format)
• az://account/container/path (blobfile/boostedblob-compatible short form)
• https://account.blob.core.windows.net/container/path (HTTPS form)

All formats are normalized to the long az:// form internally. Python normalizes HTTPS to az:// and preserves complete hostnames without checking endpoint suffixes. Both native Blob backends enforce one shared Blob-host policy before constructing or reusing a dynamic-account service client: .blob.core.windows.net, .blob.core.usgovcloudapi.net, .blob.core.chinacloudapi.cn, .blob.core.cloudapi.de, .blob.core.microsoft.scloud, and .blob.core.eaglex.ic.gov. AZURE_BLOB_ENDPOINT_SUFFIX adds an operator-approved suffix, with or without its leading dot. Set it before using the client; it is trusted endpoint configuration, not a value to take from an untrusted request. Host matching is case-insensitive, requires a nonempty prefix before the suffix, and rejects empty hostname labels, URL credentials, ports, and suffix lookalikes. Normalization preserves the supplied hostname and path. Short-form az://account/... still defaults to public Azure, even when a custom suffix is configured; use a complete endpoint for sovereign/custom clouds. Unknown or malformed full hostnames are rejected when the native Blob client is reached, not by normalization. Cache hits may not reach that validation. Scheme and basic path-shape checks still prevent an empty or relative-directory authority from escaping the interception mount. Endpoint recognition does not configure sovereign-cloud authentication, imply live qualification of each cloud, or change the independent path parsing/dispatch of the blobfile and boostedblob adapters.

DACSCacheFileHandle

A context manager for managed file handle operations. Each instance owns its own file descriptor, which is opened on enter and closed on exit.

async with DACSCacheFileHandle(open_path, flags, client.executor) as fh:
    bytes_read = await fh.read(buffer, size, offset)
    bytes_written = await fh.write(buffer, size, offset)

Supports both sync (open_sync/close_sync/read_sync/write_sync) and async (context manager with read/write) usage. Async methods delegate to the sync implementation via run_in_executor.

---

Storage-Intercept Configuration

A nonempty STORAGE_INTERCEPT_CONFIG_FILEPATH selects an existing caller-owned regular file. Client initialization preserves that file and its environment setting; the file controls all settings, including write and consistent modes. A missing file or directory path raises FileNotFoundError.

Without a caller-owned configuration, the client writes its per-process storageIntercept.<pid>.config file. Repeated initialization rewrites that generated configuration so the latest client arguments apply.

DACSBlobfileAdapter (blobfile-compatible, sync)

A synchronous, blobfile-compatible adapter that wraps DACSCacheClient. Use this as a drop-in replacement for blobfile in read/write scenarios. Also available as the BlobFile alias.

from py_dacs_client.dacs_blobfile_adapter import DACSBlobfileAdapter
# or use the alias:
from py_dacs_client.dacs_blobfile_adapter import BlobFile

Supported Operations

Function / Method	Description
DACSBlobfileAdapter(path, mode, ...)	Open a file. Modes: "r", "rb", "w", "wb". Append mode is not supported.
.read(size, buffer)	Read data. Returns bytes/str or memoryview if a user buffer is provided.
.write(data, size, buffer)	Write data from str/bytes or a user-provided buffer.
.seek(offset, whence)	Seek to a position. Supported for read-mode remote (az://) paths only. SEEK_SET, SEEK_CUR, and SEEK_END are all supported.
.tell() -> int	Return the current file position. Supported for remote (az://) paths only.
.seekable() -> bool	Returns True for read-mode remote paths, False otherwise.
.close()	Close the file and release internally-allocated write buffers.
copy(src, dst, overwrite, return_md5)	Copy a file between local and remote paths. See Copy Operations.
exists(path) -> bool	Check if a blob exists.
stat(path) -> Stat	Get file stats (size, mtime, ctime, plus extended POSIX fields).
listdir(path) -> Iterator[str]	List filenames in a directory.
scandir(path) -> Iterator[DirEntry]	List directory entries with stat info.
configure(**kwargs)	Configure global settings. Supports log_callback; unsupported options raise OperationNotSupportedError.
shutdown()	Shutdown the client singleton and clean up resources.

Copy Operations

The copy() function supports file copying between local and remote paths:

from py_dacs_client.dacs_blobfile_adapter import copy

copy(src, dst, overwrite=False, return_md5=False)

Scenario	Supported	Notes
Local to Local	Yes	Uses shutil.copyfile. Supports overwrite=True.
Local to Remote (az://)	Yes	Streams entire file through cache layer. overwrite is not supported (always creates new).
Remote (az://) to Local	Yes	Chunked reads (256 MB chunks). Supports overwrite=True.
Remote to Remote	No	Raises OperationNotSupportedError.

• Writes are atomic — a temp file is used and moved to the final destination on success.
• If return_md5=True, the MD5 hash of the copied data is returned.
• Raises FileExistsError if destination exists and overwrite=False.
• Raises IsADirectoryError if source or destination is a directory.

Unsupported Operations

The following blobfile operations raise OperationNotSupportedError: glob, scanglob, isdir, makedirs, remove, rmdir, rmtree, walk, basename, dirname, join, get_url, md5, set_mtime, readline, readlines, writelines, flush.

Note: seek() is supported for read-mode remote paths only and raises OperationNotSupportedError for local paths and write-mode files. tell() is supported for all remote paths (both read and write mode) and raises OperationNotSupportedError for local paths only.

Write Modes

WriteMode is shared by DACSCacheClient, DACSBlobfileAdapter, and dacs_boostedblob_adapter. Adapter-specific code can import it from the adapter module; core DACSCacheClient code should import it from py_dacs_client.

from py_dacs_client.dacs_blobfile_adapter import WriteMode

Mode	Value	Description
WRITE_DISABLED	"write_disabled"	Read-only mode. All write operations raise an error.
WRITE_CACHE_ONLY	"write_cache_only"	Write to local cache only, no remote blob writes.
WRITE_BLOB_ONLY	"write_blob_only"	Write directly to remote blob, bypass cache.
WRITE_CACHE_BLOB_ASYNC	"write_cache_blob_async"	Write to cache first, async upload to blob in background.
WRITE_SYNC	"write_sync"	Write to cache and blob synchronously. (Default for DACSBlobfileAdapter)

---

DACSBoostedBlobAdapter (boostedblob-compatible, async)

An async, boostedblob-compatible adapter that wraps DACSCacheClient. Use this as a drop-in replacement for native boostedblob — it provides the same function signatures and submodule structure (read, write, share), but routes all operations through DACSCacheClient.

import py_dacs_client.dacs_boostedblob_adapter as bbb

Supported Operations

Function	Description
await read_single(path, buffer?, size?)	Read entire blob contents. Returns bytes, or int (bytes read) when a user buffer is provided.
await read_byte_range(path, byte_range, buffer?, size?)	Read a byte range from a blob.
read_stream(path, executor?, size?)	Async iterator yielding chunks (streaming read).
read_stream_unordered(path, executor?, size?)	Async iterator yielding chunks in unordered fashion.
read_chunked(path, executor?, size?)	Read using chunked streaming.
await write_single(path, data, overwrite?)	Write data to a new blob.
await exists(path) -> bool	Check if a blob exists.
await stat(path) -> Stat	Get blob metadata (size, mtime, ctime, etc.).
await getsize(path) -> int	Get blob size in bytes.
await isdir(path) -> bool	Check if a path is a directory.
await isfile(path) -> bool	Check if a path is a file.
listdir(path)	Async iterator listing directory contents (yields BasePath).
scandir(path)	Async iterator listing directory entries (yields DirEntry).
listtree(path)	Async iterator listing all files recursively (yields BasePath).
scantree(path)	Async iterator listing all files recursively (yields DirEntry).
await shutdown()	Clean up resources and close all file handles.

Submodule Access (boostedblob API parity)

The adapter exposes synthetic submodules for API parity with native boostedblob:

import py_dacs_client.dacs_boostedblob_adapter as bbb

# Read submodule
data = await bbb.read.read_single("az://account.blob.core.windows.net/container/blob.txt")
data = await bbb.read.read_byte_range("az://account.blob.core.windows.net/container/blob.txt", (0, 1024))

# Write submodule
await bbb.write.write_single("az://account.blob.core.windows.net/container/new_blob.txt", b"data")

Path Classes

Class	Description
AzurePath	Azure blob path — fully supported. Parses az:// and https://*.blob.core.windows.net URLs.
BasePath	Abstract base for path types.
LocalPath, GooglePath	Not supported — DACSCacheClient is Azure-only.

Unsupported Operations

The following operations raise OperationNotSupportedError: write_stream, write_stream_unordered, write_chunked, remove, rmtree, copyfile, copytree, sync, get_url, list_blobs, glob_scandir.

Write Modes

The boostedblob adapter also uses the shared WriteMode enum:

from py_dacs_client.dacs_boostedblob_adapter import WriteMode

---

Blobfile Integration (monkey-patching)

The dacs_blobfile_patcher module allows you to transparently route all blobfile az:// operations through the DACS Client without changing any existing blobfile code.

from py_dacs_client.dacs_blobfile_patcher import enable_dacs_backend

Enabling the DACS backend

Option A: Explicit initialization

import blobfile as bf
from py_dacs_client.dacs_blobfile_patcher import enable_dacs_backend

enable_dacs_backend()

# All bf.* calls now use DACSBlobfileAdapter for az:// paths
with bf.BlobFile("az://account.blob.core.windows.net/container/file.txt", "rb") as f:
    data = f.read()

Option B: Auto-enable via environment variable

export BLOBFILE_USE_DACS=1

import blobfile as bf
from py_dacs_client.dacs_blobfile_patcher import enable_dacs_backend  # Auto-patches on import

How patching works

• Supported APIs (BlobFile, copy, exists, listdir, scandir, stat) route az:// paths through DACSBlobfileAdapter and non-Azure paths through the original blobfile implementation.
• Unsupported APIs (e.g., glob, remove, walk) raise OperationNotSupportedError when called.
• DACSError exceptions are automatically converted to standard Python exceptions (FileNotFoundError, PermissionError, FileExistsError, TimeoutError, IsADirectoryError, etc.) for blobfile compatibility.

Disabling the patcher

from py_dacs_client.dacs_blobfile_patcher import disable_dacs_backend

disable_dacs_backend()  # Restores original blobfile functions

Checking patch status

from py_dacs_client.dacs_blobfile_patcher import is_patched

if is_patched():
    print("blobfile is using DACS backend")

---

Utility API

from py_dacs_client.dacs_utils import alloc_buffer, release_buffer, BoundedBufferManager, get_memory_limit

Function / Class	Description
alloc_buffer(size) -> ctypes buffer	Allocate a raw C buffer of the specified size (via malloc). Raises MemoryError on failure.
release_buffer(buffer)	Release a previously allocated buffer (via free).
get_memory_limit() -> int	Get the container memory limit in bytes. Uses DACS_MEMORY_MB when set; otherwise derives memory as cpu_core_count * 256 MB (using DACS_CPU_CORES or os.cpu_count()).
BoundedBufferManager(max_memory_limit)	Allocation/free wrapper with synchronous and asynchronous APIs. Memory capping is not currently implemented. See BoundedBufferManager.

BoundedBufferManager

BoundedBufferManager is currently a thin wrapper around alloc_buffer() and release_buffer(). Its synchronous and asynchronous acquire methods allocate immediately, and its release methods free the supplied buffer.

Important: The current implementation does not enforce max_memory_limit, block under memory pressure, prevent out-of-memory conditions, or track allocated bytes. Memory capping is planned for the native C++ layer.

from py_dacs_client.dacs_utils import BoundedBufferManager

Constructor

BoundedBufferManager(max_memory_limit: int = 1024 * 1024 * 1024)

Parameter	Type	Default	Description
max_memory_limit	int	1073741824 (1 GB)	Stored for observability and future native memory-capping support. It does not currently limit allocations.

Methods

Method	Description
acquire_buffer_sync(buf_size) -> ctypes.c_void_p	Allocate a buffer immediately through alloc_buffer().
release_buffer_sync(buffer) -> None	Free a buffer through release_buffer(). Logs a warning and returns when buffer is None.
await acquire_buffer(buf_size) -> ctypes.c_void_p	Async-compatible API that allocates a buffer immediately through alloc_buffer(); it does not wait for memory availability.
await release_buffer(buffer) -> None	Async-compatible API that frees a buffer through release_buffer(). Logs a warning and returns when buffer is None.
current_memory_usage	Always returns 0 because usage accounting is not implemented.

Usage

from py_dacs_client.dacs_utils import BoundedBufferManager

buffer_manager = BoundedBufferManager()

async def process_blob(client, blob_path, blob_size):
    buffer = await buffer_manager.acquire_buffer(blob_size)
    try:
        bytes_read = await client.read_blob(blob_path, buffer, blob_size, 0)
        # Process the data in buffer...
    finally:
        await buffer_manager.release_buffer(buffer)

This example provides paired allocation and release, but callers must control concurrency and memory use themselves.

With DACSBlobfileAdapter

DACSBlobfileAdapter creates a shared BoundedBufferManager and uses it for buffer allocation and release. The shared manager does not currently bound aggregate memory use:

• Reads without a user buffer: The adapter allocates a buffer internally, copies data to Python bytes, and releases the buffer immediately.
• Reads with a user buffer: The user allocates via BoundedBufferManager.acquire_buffer() and is responsible for calling release_buffer() after use.
• Writes without a user buffer: The adapter allocates internal buffers that are automatically released on close().
• Writes with a user buffer: The user must keep the buffer alive until close() and release it via BoundedBufferManager.release_buffer() afterward.

---

Exception Hierarchy

All DACS-specific exceptions are importable from py_dacs_client:

DACSError (base)
├── DACSStorageError          - Storage operation failures
│   ├── DACSAuthenticationError  - Auth failures (EPERM)
│   ├── DACSPermissionError      - Access denied (EACCES)
│   ├── DACSNotFoundError        - Blob/container not found (ENOENT)
│   ├── DACSExistsError          - Resource already exists (EEXIST)
│   ├── DACSBusyError            - Resource busy/locked (EBUSY)
│   ├── DACSTimeoutError         - Operation timed out (ETIMEDOUT)
│   ├── DACSNetworkError         - Network issues (ECONNREFUSED, etc.)
│   ├── DACSQuotaError           - Quota exceeded (EDQUOT, EFBIG)
│   ├── DACSInvalidInputError    - Invalid parameters (EINVAL)
│   └── DACSServiceUnavailableError - Temporarily unavailable (EAGAIN)
└── OperationNotSupportedError   - Unsupported operations (ENOTSUP)

Usage:

from py_dacs_client.dacs_exceptions import DACSNotFoundError, handle_native_error

try:
    size = await client.get_file_size(blob_path)
except DACSNotFoundError:
    print("Blob does not exist")

---

Sample Code

Using DACSCacheClient (async)

Import dependencies

from py_dacs_client.dacs_client import DACSCacheClient
from py_dacs_client.dacs_utils import alloc_buffer, release_buffer

Initialize the Cache Client

Create an instance of DACSCacheClient. No storage account name is needed — authentication is resolved from environment variables (see Authentication).

client = DACSCacheClient()

Or with write support and a custom write mode:

from py_dacs_client import WriteMode

client = DACSCacheClient(write_mode=WriteMode.WRITE_SYNC)

Get size of a blob

filesize = await client.get_file_size("az://account.blob.core.windows.net/container/path/to/blob")

Get file statistics

stat = await client.get_file_stat("az://account.blob.core.windows.net/container/path/to/blob")
print(f"Size: {stat.st_size}, Modified: {stat.st_mtime}")

List entries of a remote storage path

async for entry in client.list_dir("az://account.blob.core.windows.net/container/dir/"):
    blob_name = entry.get("d_name")
    blob_size = entry.get("d_size")
    blob_mtime = entry.get("d_mtime")
    blob_type = entry.get("d_type")

Read blob contents

Using the async-compatible BoundedBufferManager allocation/free wrapper:

from py_dacs_client.dacs_utils import BoundedBufferManager

buffer_manager = BoundedBufferManager()

read_size = 33554432  # 32 MB
buffer = await buffer_manager.acquire_buffer(read_size)

try:
    bytes_read = await client.read_blob(
        "az://account.blob.core.windows.net/container/path/to/blob",
        buffer, read_size, chunk_offset
    )
    # Process the data in buffer...
finally:
    await buffer_manager.release_buffer(buffer)

Using raw alloc_buffer / release_buffer (simple single-buffer use):

read_size = 33554432  # 32 MB
buffer = alloc_buffer(read_size)

if not buffer:
    raise MemoryError("Failed to allocate buffer for reading blob")

try:
    bytes_read = await client.read_blob(
        "az://account.blob.core.windows.net/container/path/to/blob",
        buffer, read_size, chunk_offset
    )
    # Process the data in buffer...
finally:
    release_buffer(buffer)

Write blob contents

Using the async-compatible BoundedBufferManager allocation/free wrapper:

import ctypes
from py_dacs_client.dacs_utils import BoundedBufferManager

buffer_manager = BoundedBufferManager()

write_size = len(data_bytes)
buffer = await buffer_manager.acquire_buffer(write_size)
ctypes.memmove(buffer, data_bytes, write_size)

try:
    bytes_written = await client.write_blob(
        "az://account.blob.core.windows.net/container/path/to/new_blob",
        buffer, write_size
    )
finally:
    await buffer_manager.release_buffer(buffer)

Using raw alloc_buffer / release_buffer (simple single-buffer use):

write_size = len(data_bytes)
buffer = alloc_buffer(write_size)
import ctypes
ctypes.memmove(buffer, data_bytes, write_size)

try:
    bytes_written = await client.write_blob(
        "az://account.blob.core.windows.net/container/path/to/new_blob",
        buffer, write_size
    )
finally:
    release_buffer(buffer)

Managed file handle operations

import os
from py_dacs_client.dacs_client import DACSCacheFileHandle

# Using context manager for automatic handle management
open_path = "/mnt/dacs_cache/account.blob.core.windows.net/container/blob"
async with DACSCacheFileHandle(open_path, os.O_RDONLY, client.executor) as fh:
    bytes_read = await fh.read(buffer, size, offset)

Shutdown

await client.shutdown()

---

Using DACSBlobfileAdapter (blobfile-compatible, sync)

Read a remote blob

from py_dacs_client.dacs_blobfile_adapter import DACSBlobfileAdapter

with DACSBlobfileAdapter("az://account.blob.core.windows.net/container/blob", "rb") as f:
    data = f.read()

Read with a user-provided buffer (zero-copy)

Using the async-compatible BoundedBufferManager allocation/free wrapper:

import asyncio
from py_dacs_client.dacs_blobfile_adapter import DACSBlobfileAdapter
from py_dacs_client.dacs_utils import BoundedBufferManager

buffer_manager = BoundedBufferManager()

buffer = asyncio.run(buffer_manager.acquire_buffer(1024))
try:
    with DACSBlobfileAdapter("az://account.blob.core.windows.net/container/blob", "rb") as f:
        mv = f.read(size=1024, buffer=buffer)
        # Use memoryview — buffer must stay alive
        data = bytes(mv)  # Copy if needed after release
finally:
    asyncio.run(buffer_manager.release_buffer(buffer))  # User responsible for releasing

Using raw alloc_buffer / release_buffer:

from py_dacs_client.dacs_blobfile_adapter import DACSBlobfileAdapter
from py_dacs_client.dacs_utils import alloc_buffer, release_buffer

buffer = alloc_buffer(1024)
try:
    with DACSBlobfileAdapter("az://account.blob.core.windows.net/container/blob", "rb") as f:
        mv = f.read(size=1024, buffer=buffer)
        # Use memoryview — buffer must stay alive
        data = bytes(mv)  # Copy if needed after release
finally:
    release_buffer(buffer)  # User responsible for freeing

Write a new remote blob

from py_dacs_client.dacs_blobfile_adapter import DACSBlobfileAdapter, WriteMode

with DACSBlobfileAdapter("az://account.blob.core.windows.net/container/new_blob", "wb", write_mode=WriteMode.WRITE_SYNC) as f:
    f.write(data=b"Hello, World!")

Check if a blob exists

from py_dacs_client.dacs_blobfile_adapter import exists

if exists("az://account.blob.core.windows.net/container/blob"):
    print("Blob exists!")

List directory contents

from py_dacs_client.dacs_blobfile_adapter import listdir

for name in listdir("az://account.blob.core.windows.net/container/dir/"):
    print(name)

Get file statistics

from py_dacs_client.dacs_blobfile_adapter import stat

s = stat("az://account.blob.core.windows.net/container/blob")
print(f"Size: {s.size}, Modified: {s.mtime}")

Copy files

from py_dacs_client.dacs_blobfile_adapter import copy

# Remote to local
copy("az://account.blob.core.windows.net/container/blob", "/local/path/file.bin")

# Local to remote
copy("/local/path/file.bin", "az://account.blob.core.windows.net/container/new_blob")

# With MD5 hash
md5_hash = copy("az://account/container/blob", "/local/file", return_md5=True)
print(f"MD5: {md5_hash}")

Using blobfile with DACS backend

import blobfile as bf
from py_dacs_client.dacs_blobfile_patcher import enable_dacs_backend

enable_dacs_backend()

# Existing blobfile code works transparently with DACS
with bf.BlobFile("az://account.blob.core.windows.net/container/blob", "rb") as f:
    data = f.read()

if bf.exists("az://account.blob.core.windows.net/container/blob"):
    print("Blob exists!")

for name in bf.listdir("az://account.blob.core.windows.net/container/dir/"):
    print(name)

---

Using DACSBoostedBlobAdapter (boostedblob-compatible, async)

Read a blob

import asyncio
import py_dacs_client.dacs_boostedblob_adapter as bbb

async def main():
    data = await bbb.read.read_single("az://account.blob.core.windows.net/container/blob")
    print(f"Read {len(data)} bytes")

    await bbb.shutdown()

asyncio.run(main())

Read a byte range

import asyncio
import py_dacs_client.dacs_boostedblob_adapter as bbb

async def main():
    data = await bbb.read.read_byte_range(
        "az://account.blob.core.windows.net/container/blob",
        byte_range=(0, 1024)
    )
    print(f"Read {len(data)} bytes")

    await bbb.shutdown()

asyncio.run(main())

Write a blob

import asyncio
import py_dacs_client.dacs_boostedblob_adapter as bbb

async def main():
    await bbb.write.write_single(
        "az://account.blob.core.windows.net/container/new_blob",
        b"Hello, World!"
    )

    await bbb.shutdown()

asyncio.run(main())

List and scan directories

import asyncio
import py_dacs_client.dacs_boostedblob_adapter as bbb

async def main():
    # List filenames
    async for path in bbb.listdir("az://account.blob.core.windows.net/container/dir/"):
        print(path)

    # Scan with stat info
    async for entry in bbb.scandir("az://account.blob.core.windows.net/container/dir/"):
        print(f"{entry.path}: size={entry.stat.size}")

    await bbb.shutdown()

asyncio.run(main())

Streaming read

import asyncio
import py_dacs_client.dacs_boostedblob_adapter as bbb

async def main():
    async for chunk in bbb.read.read_stream("az://account.blob.core.windows.net/container/large_blob"):
        process(chunk)

    await bbb.shutdown()

asyncio.run(main())

Using AzurePath

import asyncio
from py_dacs_client.dacs_boostedblob_adapter import AzurePath, stat
import py_dacs_client.dacs_boostedblob_adapter as bbb

async def main():
    path = AzurePath.from_str("az://account.blob.core.windows.net/container/blob")
    s = await stat(path)
    print(f"Size: {s.size}, Modified: {s.mtime}")

    await bbb.shutdown()

asyncio.run(main())

---

Environment Variable Configuration

Variable	Required	Description
USE_SAS_TOKEN	No	Set to "true" to enable SAS token auth.
STORAGE_INTCPT_SAS_TOKEN	When USE_SAS_TOKEN=true	The SAS token for blob access.
SI_azBlobAccessToken	No	Pre-fetched Azure access token.
AZURE_CLIENT_ID	For managed identity	Azure managed identity client ID.
MSI_ENDPOINT	No	MSI endpoint URL. Defaults to IMDS.
MSI_SECRET	No	MSI secret. Defaults to "dummy".
CACHE_SERVER_LIST	No	Comma-separated list of cache servers. Enables caching when set with CACHE_SERVER_PORT.
CACHE_DISCOVERY_URL	No	Cache server discovery endpoint URL. Enables dynamic cache server discovery when set with CACHE_SERVER_PORT. Takes precedence over CACHE_SERVER_LIST.
CACHE_SERVER_PORT	No	Cache server port. Required for both static and discovery-based cache configurations.
BLOBFILE_USE_DACS	No	Set to "1" to auto-enable blobfile monkey-patching on import. See Blobfile Integration.
DACS_MEMORY_MB	No	Container memory limit in MB returned by get_memory_limit(). If unset/invalid, the value defaults to cpu_core_count * 256 MB (e.g., 8 cores -> 2 GB, 128 cores -> 32 GB). BoundedBufferManager does not currently enforce this value.
DACS_PY_CACHE_CLIENT_THREADPOOL_SIZE	No	Thread pool size for parallel downloads. Defaults to min(cpu_count * 4, 48).
DACS_PY_CACHE_CLIENT_CHUNK_SIZE	No	Chunk size in bytes for chunked blob downloads. Defaults to 268435456 (256 MB).
DACS_DATA_LOG_LEVEL	No	Logging level (DEBUG, INFO, WARNING, etc.). Defaults to INFO.
DACS_ENABLE_IB_ZERO_COPY	No	Set to true to enable the zero-copy InfiniBand upload path (registers the caller's buffer directly instead of copying it into the send buffer). Disabled by default; falls back to the copy-based path when unset or not true.

---

FAQs

Missing native library at runtime

If you see:

FATAL: Required library 'libStorageDirect.so' not found!

Ensure libStorageDirect.so is present in the same directory as the py_dacs_client package. This shared library provides the native storage intercept layer.

---

For Dev Operations

Throughput Benchmarking

For Kubernetes throughput benchmarking, use the benchmark scripts in aks_stress_test/:

cd aks_stress_test
./run_benchmark.sh

The benchmark deploys DACS clients as a DaemonSet and measures read/write throughput. Results include per-chunk latency percentiles and overall throughput metrics.

To Build

The wheel is built by the helper script .pipelines/scripts/build/build-python-package.sh. It can also be run locally.

Prerequisites

1. Build the native components first so that libStorageDirect.so exists at one of the paths setup.py looks for:

   • bin/Debug/<arch>/native-storage/storage/libStorageDirect.so
   • bin/Release/<arch>/native-storage/storage/libStorageDirect.so

   make build BUILD_TYPE=Release
   

2. Install the Python build dependencies. On modern distros (Ubuntu 23.04+, Debian 12+) Python is "externally managed" per PEP 668, so do this inside a virtualenv:

   python3 -m venv .venv
   source .venv/bin/activate
   python3 -m pip install --require-hashes \
     -r .pipelines/requirements/python-build.txt
   

   auditwheel is pinned to 6.7.0 because 6.8.0 can produce invalid DT_NEEDED and RPATH entries, causing repaired wheels to fail when imported.

Regenerate all Python lock files with Python 3.10 and the locked compiler environment:

python3.10 -m venv .venv-requirements
source .venv-requirements/bin/activate
python -m pip install --require-hashes \
  -r .pipelines/requirements/compile.txt
./.pipelines/scripts/update-python-requirements.sh

Build via the helper script (recommended)

export BUILD_ARCHITECTURE=amd64 # 'amd64' or 'arm64' (auto-detected if unset)

./.pipelines/scripts/build/build-python-package.sh
ls -al src/subsystems/storage/dacs_client/dist/

The package version is derived from Git according to VERSIONING.md.

The script:

1. Substitutes the derived version in py_dacs_client/__init__.py; pyproject.toml keeps the fixed dacs-client distribution name.
2. Runs python -m build, producing a wheel named dacs_client-<version>-py3-none-linux_{x86_64,aarch64}.whl.
3. Restores py_dacs_client/__init__.py.
4. Runs auditwheel repair on the produced wheel. This bundles all required native shared libraries into the wheel and re-tags it as manylinux_X_Y_{x86_64,aarch64}.
5. Deletes the unrepaired linux_* wheel so only the manylinux wheel remains.

The final artifact in dist/ looks like (the manylinux_X_Y tag is set by auditwheel based on the build image's glibc, e.g. manylinux_2_31 on Ubuntu 20.04):

dacs_client-0.1.0-py3-none-manylinux_X_Y_x86_64.whl

Build manually

Use the helper script above so the wheel receives the Git-derived version and the staged package metadata is restored even when the build fails.

To Publish

The release pipeline produces repaired manylinux wheels for both amd64 and arm64. Public package publication is configured separately from the source tree; do not commit repository credentials or feed URLs.

To Test

Set authentication (choose one method):

For SAS Token (Development):
export STORAGE_INTCPT_SAS_TOKEN="your-sas-token"

For Managed Identity (Production):
export AZURE_CLIENT_ID="your-client-id"
# Optional: MSI_ENDPOINT and MSI_SECRET default to IMDS values
export MSI_ENDPOINT="http://169.254.169.254/metadata/identity/oauth2/token"
export MSI_SECRET="dummy"

Validate installation:
python3 -c "from py_dacs_client import dacs_client; print(dacs_client.help())"