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.
Use it
You can simply install datago with [uv] pip install datago
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.
| 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.
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.
| 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.
- Breaking change: encoded image conversion now produces an eagerly decoded
plain PIL
Image, without the previous lazy file subclasses or their.format/.infometadata contract. Native decoding preserves pixel precision. - Breaking change: PIL conversion rejects multichannel
uint16/float32images, including pre-encoded RGB16 PNGs, instead of silently narrowing them. Useimage_format="numpy"to retain precision, or explicitly enableimage_config.image_to_rgb8for 8-bit RGB output. - Breaking change: single-channel NumPy arrays use
(H, W)rather than(H, W, 1). Usearray[..., 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.modeagrees with PIL representability, and encoded PNG/JPEG mode inspection does not decode pixel data. image_formatis case-insensitive; configuration failures retain field details. Unknown top-level keys are logged and ignored for compatibility. Obsoleteprefetch_buffer_sizeexamples and the temporary stage benchmark were removed.
- Breaking change: Python
get_sample()andget_sample_auto_convert()now raiseRuntimeErrorfor fatal source/worker failures instead of returningNone. This includes DB outages and WDS shard failures after retries. - Python
start()and reads raiseValueErrorfor invalid client/source configuration; invalid read timeouts also raiseValueError. - Reads raise
TimeoutErroron their deadline (300 seconds by default). The pass stays active: retry the read to continue, or callstop()to cancel it. - Rust callers can use
try_get_sample()/try_get_sample_timeout()for typed errors. The legacy Rustget_sample()still logs errors and returnsNone. - Low-level Rust source
orchestrate()functions now returnResult<DatagoEngine, DatagoError>. Useengine.close()or drop the engine to cancel it; its metadata queue and cancellation token are now internal.
- Breaking change: reads after end-of-stream or
stop()returnNonerather than silently starting another epoch. Callstart()explicitly for each new pass. - DB processing now defaults to
min(CPU count, connection budget, sample limit)in-flight tasks, rather than accidentally serializing whenDATAGO_MAX_TASKSis unset. WDS processing defaults to one task per CPU.DATAGO_MAX_TASKS=0retains 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 usepanic = "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()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,
}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.
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.
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.0means 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
Just install the rust toolchain via rustup
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",
]client.rsowns the engine lifecycle, Python signal polling, and cleanup.generator_*.rsenumerate source metadata;worker_*.rsprepare individual samples.worker_utils.rscoordinates bounded concurrency, publication, sample limits, and cancellation.image_payload.rsowns shared pixels and validates their layout for image access.python_buffer.rscontains the buffer protocol boundary and its ownership invariants.python_conversion.rsconverts samples into Python dictionaries and propagates errors.config.rsandhttp_client.rscontain configuration and connection policy.structs.rsholds 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.
Run cargo run --release -- -h for the existing DB benchmark options.
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 16Add --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.
From the datago folder
cargo testBuild 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 --zigthen you can pip install from target/wheels
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.
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-torchThe 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.
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.
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 500This 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.
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.
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.
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.