Skip to content
PhotoroomPublic

About

A natively parallel dataloader for Python, written in Rust. Serving data at GB/s speeds, while covering aspect ratio bucketing, crop and resize for image ML workloads.

Topics

Resources

Stars

128 stars

Watchers

12 watching

Forks

Repository files navigation

datago

Rust Rust-py

A Rust-written data loader which can be used as a python module. Handles several data sources, from local files to webdataset or a VectorDB focused http stack like our open sourced dataroom. Focused on image data at the moment, could also easily be more generic.

Datago handles, outside of the Python GIL

  • per sample IO
  • deserialization (jpg and png decompression)
  • some optional vision processing (aligning different image payloads)
  • optional serialization

Images can be returned as PIL objects or read-only NumPy views over shared Rust storage. File, HTTP, and WebDataset sources support bounded buffering, cancellation, and distributed loading through source-specific rank/world-size settings.

Datago organization

Use it

You can simply install datago with [uv] pip install datago

Use the package from Python

Please note that in all the of the following cases, you can directly get an IterableDataset (torch compatible) with the following code snippet

from dataset import DatagoIterDataset

client_config = {
    "source_type": "file",
    "source_config": {"root_path": "/path/to/images"},
    "limit": 2000,
    "samples_buffer_size": 256,
    "image_format": "numpy",  # Omit for PIL output.
}
datago_dataset = DatagoIterDataset(client_config, return_python_types=True)

With return_python_types=True, samples are dictionaries whose images follow the configured image_format.

Image output

API / setting Result
client.get_sample() Sample and image wrappers with metadata and explicit conversion methods
client.get_sample_auto_convert(), "image_format": "pil" (default) Dictionary containing PIL images
client.get_sample_auto_convert(), "image_format": "numpy" Dictionary containing NumPy arrays for images, masks, and additional images

Set "image_format": "numpy" in the client configuration to avoid materializing PIL images. Raw arrays are read-only views over shared Rust pixels and remain valid after the sample or client is dropped. Use array.copy() for writable storage. For PyTorch, use torch.from_numpy(array.copy()) before any tensor mutation: torch.from_numpy(array) shares the read-only Rust storage, and writing through that tensor is unsupported. torch.tensor(array) also makes an independent copy. Single-channel arrays have shape (H, W); multi-channel arrays use (H, W, C). Their dtype preserves the source: uint8, uint16, or float32.

sample.image.to_numpy_array() and sample.image.to_pil_image() also convert individual images. PIL supports 16-bit grayscale, but wider multi-channel images raise ValueError; use NumPy to preserve their precision, or explicitly set "image_config": {"crop_and_resize": false, "image_to_rgb8": true} to convert to 8-bit RGB. PIL results have independent writable storage. Missing images in metadata-only samples become None in either output format; malformed image payloads and unsupported conversions raise ordinary Python exceptions.

image_format accepts pil and numpy case-insensitively. Invalid configuration errors identify the field; unknown top-level keys remain accepted but are logged as warnings when logging is enabled (initialize_logging("warn")).

Cloning a sample or calling get_payload() shares its immutable pixels. Reading payload.data returns an owned byte copy; assigning to it replaces that wrapper's storage without invalidating existing views. Raw metadata properties (width, height, size, mode, palette) avoid constructing a PIL image. For encoded PNG/JPEG, mode inspects the decoder header without materializing pixels; it raises the same error as PIL conversion if the pixel type cannot be represented in PIL.

Decoded buffers are retained in Arc<Vec<u8>>: sharing ownership moves the existing allocation rather than copying pixels into a new reference-counted slice.

With pre_encode_images=True, the buffer protocol exposes compressed bytes. Use to_numpy_array() or to_pil_image() to decode them. Decoding runs natively outside the GIL and preserves channels and precision; decoded output requires a new allocation rather than a view over the compressed stream. PIL output is an eagerly decoded plain Image, without the former lazy JpegImageFile/PngImageFile type or its .format/.info metadata contract.

Lifecycle and cancellation

Python sample reads try the ready queue without releasing the GIL, then release it while waiting and poll Python signals every 100 ms. Python delivers SIGINT handlers on the main thread only. If a reader runs in a background thread, handle Ctrl-C on the main thread by calling client.stop(); that cancellation is safe while the reader is waiting. stop() closes both pipeline queues, releases the GIL while joining, and returns after a short grace period if native work is still in flight. Remaining thread handles are retained by a cleanup reaper rather than detached. Prefer try/finally with client.stop() for explicit cleanup. Cancellation remains active after metadata ends and while the output queue is full. Async downloads and queued work are cancelled; already-running blocking IO or computation may continue in the background until the native operation returns. The native Rust client retains its blocking API. Reads do not silently restart: once the stream is exhausted (or stop() has been called) get_sample() returns None, and you must call start() again to begin a new pass.

Waiting on an empty queue also lets other Python threads run. For very cheap samples, GIL contention can reduce throughput compared with a reader that keeps the GIL; queued samples still use the fast path without releasing it. live_cleanup_threads() counts cleanup reapers, not all native threads or blocking-pool tasks that outlive runtime shutdown.

Workers count successfully published samples: limit is an upper bound on delivery, including under concurrency. Unreadable items do not consume that budget, although a finite source may end before reaching the limit. File and DB feeders keep supplying candidates until that limit is satisfied or their source is exhausted; there is no fixed allowance for bad images. In-flight tasks are also capped by the remaining delivery budget, so a small limit does not start a full worker pool. WebDataset samples require a readable reference image (reference_image_type); missing or corrupt references are skipped. Additional images use the reference image's target aspect ratio, regardless of their order in the archive. WebDataset keys retain directory prefixes (for example, train/0000.seg.png belongs to train/0000). reference_image_type selects the full field name, case-insensitively: png selects 0000.png, while right.png selects 0000.right.png. Other images, including those with the same final extension, are returned in additional_images under their archive filenames. Text attribute keys retain the full field name in lowercase, such as caption.json and json.

Sequential single-rank filesystem reads stream paths in sorted directory order. Global shuffling and distributed filesystem reads enumerate all candidate paths before delivering samples. Distributed ranks take balanced, disjoint contiguous slices of this deterministic ordering, then shuffle within their slice if requested. Enumeration remains cancellable, but these modes need memory proportional to the number of candidate paths and have a longer startup on large trees.

Read results and errors

Outcome Rust (try_get_sample) Python (get_sample / get_sample_auto_convert)
Sample available Ok(Some(sample)) Return the sample
Normal exhaustion or explicit stop Ok(None) Return None
Fatal source or worker failure Err(DatagoError::Source(_)) / Err(DatagoError::Worker(_)) Raise RuntimeError
Read timeout Err(DatagoError::Timeout) Raise TimeoutError
Invalid configuration or timeout argument Err(DatagoError::Configuration(_)) Raise ValueError

Unreadable individual images (including decoder panics) and unreadable nested directories are logged and skipped. An unreadable filesystem root, failed metadata request/parse, or worker failure ends the pass with an error. WDS listing errors and shard download/unpacking errors after retries are fatal: Datago does not silently skip a missing or corrupt shard and drop its samples. A fatal failure cancels outstanding work and discards queued samples. Concurrent stop and failure use the first recorded terminal outcome. After a fatal source/worker error, subsequent reads return None; call start() explicitly to begin a new pass. Errors belong to their engine generation and cannot stop a restarted pass.

Python reads accept a keyword-only timeout in seconds (default: 300). Rust callers can use try_get_sample_timeout(Duration) for a per-read timeout. A timeout only ends that read, including at the default 300-second deadline: retry to keep reading the same pass, or call stop() to cancel it. Zero returns a ready sample or raises TimeoutError immediately if the queue is empty and still open; an exhausted or stopped stream returns None. Cleanup gets a short grace period; stalled native operations may finish under a background reaper. Python KeyboardInterrupt remains an interrupt rather than a pipeline error.

try:
    while (sample := client.get_sample(timeout=30.0)) is not None:
        process(sample)
finally:
    client.stop()

The existing Rust get_sample() remains a compatibility wrapper: it logs failures and returns None. Use try_get_sample() when failure must be distinguished from normal exhaustion. Python propagates failures through its existing read methods.

Release notes: 2026.11.1

  • Breaking change: encoded image conversion now produces an eagerly decoded plain PIL Image, without the previous lazy file subclasses or their .format/.info metadata contract. Native decoding preserves pixel precision.
  • Breaking change: PIL conversion rejects multichannel uint16/float32 images, including pre-encoded RGB16 PNGs, instead of silently narrowing them. Use image_format="numpy" to retain precision, or explicitly enable image_config.image_to_rgb8 for 8-bit RGB output.
  • Breaking change: single-channel NumPy arrays use (H, W) rather than (H, W, 1). Use array[..., None] to restore the singleton channel axis.
  • Raw NumPy arrays and memoryviews share immutable Rust pixels and retain their allocation after sample destruction or payload reassignment. Copy before writing, including through torch.from_numpy(array.copy()). PIL images have independent writable storage, including 16-bit grayscale.
  • File and DB candidate enumeration no longer stops at a fixed multiple of limit. Failed items are replaced until the successful-delivery limit or true source exhaustion. Filesystem rank partitions now cover the complete sorted candidate list; shuffle/distributed modes enumerate that list before delivery.
  • Metadata-only samples convert missing images to None. Invalid image payloads still raise conversion errors. mode agrees with PIL representability, and encoded PNG/JPEG mode inspection does not decode pixel data.
  • image_format is case-insensitive; configuration failures retain field details. Unknown top-level keys are logged and ignored for compatibility. Obsolete prefetch_buffer_size examples and the temporary stage benchmark were removed.

Release notes: 2026.10.2

  • Breaking change: Python get_sample() and get_sample_auto_convert() now raise RuntimeError for fatal source/worker failures instead of returning None. This includes DB outages and WDS shard failures after retries.
  • Python start() and reads raise ValueError for invalid client/source configuration; invalid read timeouts also raise ValueError.
  • Reads raise TimeoutError on their deadline (300 seconds by default). The pass stays active: retry the read to continue, or call stop() to cancel it.
  • Rust callers can use try_get_sample() / try_get_sample_timeout() for typed errors. The legacy Rust get_sample() still logs errors and returns None.
  • Low-level Rust source orchestrate() functions now return Result<DatagoEngine, DatagoError>. Use engine.close() or drop the engine to cancel it; its metadata queue and cancellation token are now internal.

Release notes: 2026.10.1

  • Breaking change: reads after end-of-stream or stop() return None rather than silently starting another epoch. Call start() explicitly for each new pass.
  • DB processing now defaults to min(CPU count, connection budget, sample limit) in-flight tasks, rather than accidentally serializing when DATAGO_MAX_TASKS is unset. WDS processing defaults to one task per CPU. DATAGO_MAX_TASKS=0 retains serial processing; explicit overrides retain their existing source-specific caps.
  • Python waits release the GIL and check signals; cancellation also covers DB page requests and WDS shard listing/downloads. Already-running blocking filesystem calls may still outlive shutdown.
  • Source builds require maturin>=1.9.4,<2. Release builds no longer use panic = "abort", allowing worker panics to be reported instead of aborting the process.
Dataroom
from datago import DatagoClient, initialize_logging
import os
import json

# Respects RUST_LOG=INFO env var for setting log level
# If omitted the logger will be initialized when the client starts.
initialize_logging()

config = {
    "source_config": {
        "sources": os.environ.get("DATAROOM_TEST_SOURCE", ""),
        "page_size": 500,
        "rank": 0,
        "world_size": 1,
    },
    "limit": 200,
    "samples_buffer_size": 32,
}

client = DatagoClient(json.dumps(config))

for _ in range(10):
    sample = client.get_sample()

get_sample() returns wrappers with explicit image conversion methods; see Image output.

Local files

To test datago while serving local files (jpg, png, ..), code would look like the following. Note that datago serving files with a lot of concurrent threads means that, even if random_sampling is not set, there will be some randomness in the sample ordering.

from datago import DatagoClient, initialize_logging
import os
import json

# Can also set the log level directly instead of using RUST_LOG env var
initialize_logging(log_level="warn")

config = {
    "source_type": "file",
    "source_config": {
        "root_path": "myPath",
        "random_sampling": False,  # True if used directly for training
        "rank": 0,  # Optional, distributed workloads are possible
        "world_size": 1,
    },
    "limit": 200,
    "samples_buffer_size": 32,
}

client = DatagoClient(json.dumps(config))

for _ in range(10):
    sample = client.get_sample()
[experimental] Webdataset

Please note that this implementation is very new, and probably has significant limitations still. It has not yet been tested at scale. Please also note that you can find a better example in /python/benchmark_webdataset.py, which will show how to convert everything to more pythonic types (PIL images).

from datago import DatagoClient, initialize_logging
import os
import json

# Can also set the log level directly instead of using RUST_LOG env var
initialize_logging(log_level="warn")

# URL of the test bucket
bucket = "https://storage.googleapis.com/webdataset/fake-imagenet"
dataset = "/imagenet-train-{000000..001281}.tar"
url = bucket + dataset

client_config = {
    "source_type": "webdataset",
    "source_config": {
        "url": url,
        "random_sampling": False,
        "concurrent_downloads": 8,  # The number of TarballSamples which should be handled concurrently
        "rank": 0,
        "world_size": 1,
    },
    "samples_buffer_size": 64,
    "limit": 1_000_000,  # Dummy example, max number of samples you would like to serve
}

client = DatagoClient(json.dumps(client_config))

for _ in range(10):
    sample = client.get_sample()

Process images on the fly

Datago can also process images on the fly, for instance to align different image payloads. This is done by adding an image_config to the configuration. The following example shows how to align different image payloads.

Processing can be very CPU heavy, but it will be distributed over all CPU cores wihout requiring multiple python processes. I.e., you can keep a single python process using get_sample() on the client and still saturate all CPU cores.

There are three main processing topics that you can choose from:

  • crop the images to within an aspect ratio bucket (which is very handy for all Transformer / patch based architectures)
  • resize the images (setting here will be related to the square aspect ratio bucket, other buckets will differ of course)
  • pre-encode the images to a specific format (jpg, png, ...)
config = {
    "source_type": "file",
    "source_config": {
        "root_path": "myPath",
        "random_sampling": False,  # True if used directly for training
    },
    # Optional pre-processing of the images, placing them in an aspect ratio bucket to preserve as much as possible of the original content
    "image_config": {
        "crop_and_resize": True,  # False to turn it off, or just omit this part of the config
        "default_image_size": 1024,
        "downsampling_ratio": 32,
        "min_aspect_ratio": 0.5,
        "max_aspect_ratio": 2.0,
        "pre_encode_images": False,
    },
    "limit": 200,
    "samples_buffer_size": 32,
}

Match the raw exported buffers with typical python types

Use sample.image.to_numpy_array() or sample.image.to_pil_image() for a single image, or client.get_sample_auto_convert() for a complete dictionary. See Image output for storage, dtype, and lifetime guarantees.

Logging

We are using the log crate with env_logger. You can set the log level using the RUST_LOG environment variable. E.g. RUST_LOG=INFO.

When using the library from Python, env_logger will be initialized automatically when creating a DatagoClient. There is also a initialize_logging function in the datago module, which if called before using a client, allows to customize the log level. This only works if RUST_LOG is not set.

Env variables

There are a couple of env variables which will change the behavior of the library, for settings which felt too low level to be exposed in the config.

  • DATAGO_MAX_TASKS: maximum in-flight processing tasks, not native thread count. Defaults to CPU count for files/WDS and the smaller of CPU count and connection budget for DB sources, with source-specific caps. 0 means serial processing.
  • RUST_LOG: see above, will change the level of logging for the whole library, could be useful for debugging or to report an issue here.
  • DATAGO_MAX_RETRIES: number of retries for a failed sample load, defaults to 3.
Build it

Preamble

Just install the rust toolchain via rustup

[Apple Silicon MacOS only]

If you are using an Apple Silicon Mac OS machine, create a .cargo/config.toml file and paste the following:

[target.x86_64-apple-darwin]
rustflags = [
  "-C", "link-arg=-undefined",
  "-C", "link-arg=dynamic_lookup",
]

[target.aarch64-apple-darwin]
rustflags = [
  "-C", "link-arg=-undefined",
  "-C", "link-arg=dynamic_lookup",
]

Code organization

  • client.rs owns the engine lifecycle, Python signal polling, and cleanup.
  • generator_*.rs enumerate source metadata; worker_*.rs prepare individual samples.
  • worker_utils.rs coordinates bounded concurrency, publication, sample limits, and cancellation.
  • image_payload.rs owns shared pixels and validates their layout for image access.
  • python_buffer.rs contains the buffer protocol boundary and its ownership invariants.
  • python_conversion.rs converts samples into Python dictionaries and propagates errors.
  • config.rs and http_client.rs contain configuration and connection policy.
  • structs.rs holds pipeline data types and re-exports the original public type paths.

Tasks return prepared samples to the coordinator. Only the coordinator writes to the output queue, so counting, cancellation, and end-of-stream behavior are shared by all sources. Encoded bytes and raw pixel layouts are interpreted separately; public payload fields remain compatible, with validation before pixel conversion.

DB benchmark CLI

Run cargo run --release -- -h for the existing DB benchmark options.

Native filesystem benchmark

Run the filesystem pipeline directly from Rust, including image decoding and optional preprocessing, without Python conversion:

cargo run --release --example benchmark_filesystem -- /path/to/images --limit 10000 --num-workers 16

Add --crop-and-resize for a 1024 × 1024 center crop. --num-workers controls in-flight tasks and defaults to CPU count; 0 means serial processing. The benchmark uses a 256-sample buffer and requires the requested number of samples. Image payloads are borrowed when collecting metrics, avoiding a pixel-buffer copy.

The JSON report includes FPS, output MP/s and MB/s, time to the first sample, and FPS after the first sample (steady_fps, or null for one sample). Elapsed time includes construction and startup and excludes cleanup. Compare repeated release runs with identical options and comparable filesystem cache state.

Run the rust test suite

From the datago folder

cargo test

Generate the python package binaries manually

Build a wheel useable locally

maturin build -i python3.11 --release --target "x86_64-unknown-linux-gnu"

Build a wheel which can be uploaded to pypi or related

  • either use a manylinux docker image

  • or cross compile using zip

maturin build -i python3.11 --release --target "x86_64-unknown-linux-gnu" --manylinux 2014 --zig

then you can pip install from target/wheels

Update the pypi release (maintainers)

Create a new tag and a new release in this repo, a new package will be pushed automatically.

Benchmarks As usual, benchmarks are a tricky game, and you shouldn't read too much into the following plots but do your own tests. Some python benchmark examples are provided in the [python](./python/) folder.

In general, Datago will be impactful if you want to load a lot of images very fast, but if you consume them as you go at a more leisury pace then it's not really needed. The more CPU work there is with the images and the higher quality they are, the more Datago will shine.

Filesystem throughput: Python

The Python filesystem benchmark reports FPS, MP/s, MB/s, time to first sample, and post-start FPS for Datago and optionally Torch. Its returned results and sweep JSON include pixels and megapixels_per_second. Pixels count output spatial positions (width * height), independent of channel count or bit depth.

Compare NumPy and PIL output with the native Rust filesystem benchmark (examples/benchmark_filesystem.rs) to assess Python conversion overhead:

# Full Python NumPy output; remove --no-compare-torch to include Torch.
uv run --group dev python/benchmark_filesystem.py --root-path /path/to/images --limit 2000 --num-workers 8 --image-format numpy --no-compare-torch
# PIL output with the same pipeline settings.
uv run --group dev python/benchmark_filesystem.py --root-path /path/to/images --limit 2000 --num-workers 8 --image-format pil --no-compare-torch

The Python benchmark accepts --crop-and-resize for 1024 × 1024 output. Resized runs count output pixels. It uses a 256-sample buffer, includes construction/startup in elapsed time, excludes cleanup, and reports an error if the source cannot supply the limit. The full Python report uses startup_seconds for time to the first sample and steady_fps for the rate after it (null for a single-sample run).

Use release builds for the extension, identical options and an RGB8 corpus, and repeated runs with comparable filesystem cache state. Raw versus converted Python output includes conversion and object creation. Python progress reporting also adds work. num_workers controls Datago tasks and Torch processes; native thread counts can differ.

From disk: ImageNet

The following benchmarks are using ImageNet 1k, which is very low resolution and thus kind of a worst case scenario. Data is served from cache (i.e. the OS cache) and the images are not pre-processed. In this case the receiving python process is typically the bottleneck, and caps at around 3000 images per second.

AMD Zen3 laptop - IN1k - disk - no processing

AMD Zen3 laptop & M2 SSD

AMD EPYC 9454 - IN1k - disk - no processing

AMD EPYC 9454

One liner to repro locally (rebuilding the package):

DATAGO_TEST_FILESYSTEM=$PATH_TO_YOUR_TEST_DATA uv run --reinstall-package datago --python 3.14 --group dev python/benchmark_filesystem.py --sweep --limit 500

Webdataset: FakeIN

This benchmark is using low resolution images. It's accessed through the webdataset front end, datago is compared with the popular python webdataset library. Note that datago will start streaming the images faster here (almost instantly !), which emphasizes throughput differences depending on how long you test it for.

Of note is also that this can be bottlenecked by your external bandwidth to the remote storage where WDS is hosted, in which case both solution would yield comparable numbers.

AMD Zen3 laptop - webdataset - no processing

AMD EPYC 9454

Webdataset: PD12M

This benchmark is using high resolution images. It's accessed through the webdataset front end, datago is compared with the popular python webdataset library. Note that datago will start streaming the images faster here (almost instantly !), which emphasizes throughput differences depending on how long you test it for.

Of note is also that this can be bottlenecked by your external bandwidth to the remote storage where WDS is hosted, in which case both solution would yield comparable numbers.

AMD Zen3 laptop - webdataset - no processing

AMD Zen3 laptop

AMD EPYC 9454 - pd12m - webdataset - no processing

AMD EPYC 9454

AMD Zen3 laptop - webdataset - processing

Adding image processing (crop and resize to Transformer compatible size buckets) to the equation changes the picture, as the work spread becomes more important. If you're training a diffusion model or an image encoder from a diverse set of images, this is likely to be the most realistic micro-benchmark.

AMD Zen3 laptop

License

MIT License

Copyright (c) 2025 Photoroom

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.

About

A natively parallel dataloader for Python, written in Rust. Serving data at GB/s speeds, while covering aspect ratio bucketing, crop and resize for image ML workloads.

Topics

Resources

Stars

128 stars

Watchers

12 watching

Forks

Releases

Packages

Used by

Contributors

Languages