mirror of
https://github.com/lancedb/lancedb.git
synced 2026-07-11 06:50:40 +00:00
c6db80dd0b
# Elastic Streaming Dataloader ## Motivation Training large models on LanceDB tables today requires loading the entire dataset into memory or writing bespoke batching logic. This PR introduces `StreamingDataset`, a PyTorch `IterableDataset` that streams directly from a LanceDB table with two hard guarantees that are difficult to achieve together: **elastic determinism** and **resumability**. ## Goals ### Elastic determinism The dataset partitions the table into a fixed number of *splits* (controlled by `num_splits`, `shuffle_seed`, and `epoch`). Samples are yielded by round-robining over splits one sample per split per cycle. Because the split structure is fixed, the set of samples that makes up each global training step is identical regardless of `world_size` or `num_workers`. You can scale your cluster up or down between runs and the model sees the same data in the same order — no re-sharding, no gradient variance from topology changes. ### Resumability `state_dict()` / `load_state_dict()` capture how many samples each split has consumed. Because all splits are the same size and the round-robin design keeps them in lockstep, the state reduces to a single scalar (`samples_consumed_per_split`) that is topology-independent. A checkpoint saved with 8 GPUs can resume correctly on 4 GPUs or 16 GPUs without any adjustment. ### PyTorch `IterableDataset` / streaming `StreamingDataset` implements the standard PyTorch `IterableDataset` interface, so it drops into any existing `DataLoader` pipeline without modification. Data is fetched lazily from Lance in chunks — only the rows needed for the current batch are ever in memory. Compared to the map dataset this takes more work from pytorch and puts it into the dataset itself (e.g. shuffling, filtering, etc.). We do this because we cannot achieve things like elastic determinism or prefiltering otherwise. ### Multi-worker support DataLoader workers are automatically assigned contiguous sub-blocks of splits (the rank's splits are divided evenly across workers). Each worker is independent: no shared state, no inter-process coordination. The only constraint is that `num_splits` must be divisible by `world_size * num_workers`. That being said, multi-worker is highly discouraged as it relies on multiprocessing which is inefficient. Still, we want to support it. ### Filters as prefilters Filters are applied at *permutation-build time* via `PermutationBuilder.filter()`, not re-evaluated on every fetch. The filtered row IDs are stored in the permutation table so that subsequent reads see only the matching rows. This allows us to avoid loading rows that don't match the filter (which is the default pytorch behavior) ### Prefetching Two parameters control the I/O pipeline: - `read_batch_size` (default 64) — number of rows fetched per `take_offsets` call. Larger values amortise per-request overhead, which is critical on object storage where a single round-trip can cost ~100 ms. - `prefetch_batches` (default 4) — number of batches prefetched in parallel per split via a `ThreadPoolExecutor`. While the model processes the current batch, the next several batches are already in flight, hiding storage latency behind compute. If set correctly then you can get good performance even with num_workers=0 (unless you are bottlenecked on transform). ### Transform parallelism The underlying `Permutation` API supports a `with_transform()` callback for decoding, augmentation, and format conversion. Unfortunately, this is not parallelized. Pytorch typically parallelizes this with num_workers which is multiprocessing which is highly inefficient. For simple transforms we should be able to utilize multithreading and Rust based UDFs. For complex python UDFs we could have a dedicated multiprocessing pipeline for just the transform. Or we could just utilize multithreading. In both cases we would exclude the I/O stage from the multiprocessing because that ends up being very memory hungry and inefficient. --------- Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
LanceDB Documentation
LanceDB docs are available at docs.lancedb.com.
The SDK docs are built and deployed automatically by Github Actions
whenever a commit is pushed to the main branch. So it is possible for the docs to show
unreleased features.
Building the docs
Setup
- Install LanceDB Python. See setup in Python contributing guide.
Run
make developto install the Python package. - Install documentation dependencies. From LanceDB repo root:
pip install -r docs/requirements.txt
Preview the docs
cd docs
mkdocs serve
If you want to just generate the HTML files:
PYTHONPATH=. mkdocs build -f docs/mkdocs.yml
If successful, you should see a docs/site directory that you can verify locally.
Adding examples
To make sure examples are correct, we put examples in test files so they can be run as part of our test suites.
You can see the tests are at:
- Python:
python/python/tests/docs - Typescript:
nodejs/examples/
Checking python examples
cd python
pytest -vv python/tests/docs
Checking typescript examples
The @lancedb/lancedb package must be built before running the tests:
pushd nodejs
npm ci
npm run build
popd
Then you can run the examples by going to the nodejs/examples directory and
running the tests like a normal npm package:
pushd nodejs/examples
npm ci
npm test
popd
API documentation
Python
The Python API documentation is organized based on the file docs/src/python/python.md.
We manually add entries there so we can control the organization of the reference page.
However, this means any new types must be manually added to the file. No additional
steps are needed to generate the API documentation.
Typescript
The typescript API documentation is generated from the typescript source code using typedoc.
When new APIs are added, you must manually re-run the typedoc command to update the API documentation. The new files should be checked into the repository.
pushd nodejs
npm run docs
popd