# 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>
quick-xml < 0.41.0 has two DoS advisories (quadratic attribute-name
check and unbounded namespace allocation in NsReader). All three
versions in our lockfile (0.26.0, 0.38.4, 0.39.4) are below the patched
threshold.
These are pulled in transitively by inferno (dev-only flame-graph dep),
lance-namespace-impls (git dep from lance), and opendal/reqsign (cloud
storage XML parsing). None of these paths expose attacker- controlled
XML; clearing them requires upstream to upgrade to quick-xml >= 0.41.0.
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
This PR is part cleanup, part feature, part example.
It removes `IntoArrow` and `IntoArrowStream`. There was only one
redundant call site between the two. Once we moved everything to
`Scannable` these traits no longer serve any purpose.
It adds a `Scannable` impl for a polars DataFrame. We used to have this
at one point for `IntoArrow` so this is more like a regression fix than
anything.
It adds an example (and unit test) which ensures we can ingest from a
Polars DataFrame and export to one. LazyFrame support would be a
follow-up (though a pretty straightforward one) but we've never had
proper LazyFrame support before.
Fixes#3136.
## Summary
- `WithEmbeddingsScannable::scan_as_stream` matched columns positionally
against the table schema, so a `CastError` was raised whenever the
computed batch order differed from the table schema order.
- The mismatch surfaced when `add_columns` added a new physical column
**after** an embedding column: the table schema became
`[..., embedding, extra]`, but `compute_embeddings_for_batch` always
appends embeddings at the end, producing `[..., extra, embedding]`.
Position 2 then tried to cast e.g. `score: Float64` →
`embedding: FixedSizeList` and failed.
- Now we look each output column up by name in the result batch, which
is order-independent. If a non-embedding column required by the table
schema is missing from the input, we return a clear `InvalidInput`
error instead of a confusing cast error.
## Reproduction (from the issue)
```text
Table created with: [id, text, text_vec(embedding)]
add_columns("score") → schema: [id, text, text_vec, score]
table.add([id, text, score]) → BEFORE: CastError on position 2
AFTER: succeeds, embedding is computed
```
## Tests
-
`data::scannable.rs::test_with_embeddings_scannable_column_added_after_embedding`
— unit test exercising the exact column-order mismatch via
`WithEmbeddingsScannable::with_schema`.
-
`data::scannable.rs::test_with_embeddings_scannable_missing_required_column`
— covers the new "missing column" error path.
- `table::add_data.rs::test_add_with_embeddings_after_add_columns`
— end-to-end regression test mirroring the reproduction in the issue
(create table with embedding → `add_columns` → `table.add`).
## Test plan
- [x] `cargo check --quiet --features remote --tests --examples`
- [x] `cargo clippy --quiet --features remote --tests --examples`
- [x] `cargo fmt --all`
- [x] `cargo test --quiet --features remote -p lancedb embedding` — 18
tests pass
🤖 Generated with [Claude Code](https://claude.com/claude-code)
---------
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The 0.29 release happened on a branch because the main line had already
moved past the 6.0.0 stable lance release. As a result the version bump
commits ended up on the branch. This merges those commits back into
main.
---------
Co-authored-by: Lance Release <lance-dev@lancedb.com>
## Summary
PyTorch's `DataLoader` uses fork-based multiprocessing by default on
Linux, but threads do not survive `fork()`. LanceDB's Python bindings
drive async work through two threaded layers, both of which become inert
in a forked child:
- `BackgroundEventLoop` runs an asyncio loop on a Python
`threading.Thread`.
- `pyo3-async-runtimes::tokio` holds a global multi-threaded tokio
runtime whose worker threads also die on fork — and its runtime lives in
a `OnceLock` that cannot be replaced after first use.
As a result, any `Permutation` (or other async API) used inside a
fork-based `DataLoader` worker hangs indefinitely. This PR makes both
layers fork-safe so `Permutation` works as a `torch.utils.data.Dataset`
with `num_workers > 0`.
## Approach
### Rust — new `python/src/runtime.rs`
Mirrors the pattern used in [Lance's Python
bindings](https://github.com/lance-format/lance/blob/456198cd6f42be07f99617a6d7e39d6209cdf3cc/python/src/lib.rs#L139),
adapted for the async-bridge use case.
- `LanceRuntime` implements `pyo3_async_runtimes::generic::Runtime +
ContextExt`, backed by an `AtomicPtr<tokio::runtime::Runtime>` we own
(sidestepping `pyo3-async-runtimes`'s frozen `OnceLock` global).
- A `pthread_atfork(after_in_child)` handler nulls the pointer; the next
`spawn` rebuilds the runtime in the child. The previous runtime is
intentionally **leaked** — calling `Drop` would try to join now-dead
worker threads and hang.
- `runtime::future_into_py` is a drop-in for
`pyo3_async_runtimes::tokio::future_into_py`. All ~80 call sites in
`arrow.rs` / `connection.rs` / `permutation.rs` / `query.rs` /
`table.rs` are updated to route through it.
- `python/Cargo.toml` adds `libc = "0.2"` and the tokio
`rt-multi-thread` feature.
### Python — `lancedb/background_loop.py`
- Refactors `BackgroundEventLoop.__init__` to a reusable `_start()`
method.
- An `os.register_at_fork(after_in_child=…)` hook calls `LOOP._start()`
to give the singleton a fresh asyncio loop and thread **in place**. This
matters because the rest of the codebase imports `LOOP` via `from
.background_loop import LOOP` — rebinding the module attribute would
leave those references holding the dead loop.
### Python — `lancedb/__init__.py`
Removes the `__warn_on_fork` pre-fork warning (and the now-unused
`import warnings`). Fork is supported.
## Test plan
- [x] New `test_permutation_dataloader_fork_workers` in
`python/tests/test_torch.py`: runs a `Permutation` through
`torch.utils.data.DataLoader(num_workers=2,
multiprocessing_context="fork")` inside a spawn-isolated child with a
30s hang detector. **Pre-fix**: timed out at 36s. **Post-fix**: passes
in ~3.6s.
- [x] New `test_remote_connection_after_fork` in
`python/tests/test_remote_db.py`: forks a child that creates a fresh
`lancedb.connect(...)` against a mock HTTP server and calls
`table_names()`; passes in <1s, validates the runtime reset is
sufficient for fresh remote clients.
- [x] All 62 tests in `test_torch.py` + `test_permutation.py` pass.
- [x] All 35 tests in `test_remote_db.py` pass.
- [x] `test_table.py` (87) + `test_db.py` + `test_query.py` (157, minus
one unrelated `sentence_transformers` import skip) — 244 passing.
- [x] `cargo clippy -p lancedb-python --tests` clean.
- [x] `cargo fmt`, `ruff check`, `ruff format` all clean.
## Known limitation (follow-up)
This PR makes a **freshly-built** `lancedb.connect(...)` work in a
forked child. An **inherited** `Connection` from the parent still
carries an inherited `reqwest::Client` whose hyper connection pool
references socket FDs and TCP/TLS state shared with the parent — using
it from the child after fork is unsafe (especially with HTTP/1.1
keep-alive). The recommended pattern for fork-based `DataLoader` workers
that hit a remote DB is to construct a new connection inside the worker.
Auto-clearing inherited HTTP client pools on fork would require tracking
live `Connection` instances in `lancedb` core and is left for a
follow-up PR.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
---------
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
## Summary
When pytorch is used with multiprocessing and the mp mode is spawn then
the Permutation needs to be pickled. It could not be pickled because
`Table` and `Connection` are not serializable. This PR adds pickle
support to Permutation without adding general pickle support to `Table`
or `Connection`. To add general support we probably need to start by
adding serialization in the namespace client.
In the meantime this PR enable pickling by adding special cases for:
* In-memory tables (just serialize as Arrow IPC)
* Native tables (serialize the URI)
If a user is not using one of the above cases (e.g. using a remote
connection) then they will need to provide a connection factory that can
be pickled.
## Breaking change
`PermutationBuilder.persist(...)` is removed from the Python bindings;
the permutation table is now always in-memory. The underlying Rust
`PermutationBuilder::persist` API is untouched and can be re-exposed
later if needed. It probably won't make sense to do that until we have a
way to serialize `Table` and `Connection`.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
---------
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
## Summary
- Update `rustls-webpki` 0.103.10 → 0.103.13 to fix RUSTSEC-2026-0104
(reachable panic in CRL parsing)
- Add advisory ignore for the legacy `rustls-webpki` 0.101.7 copy pinned
to the aws-smithy/rustls 0.21 chain (same chain already exempted for
RUSTSEC-2026-0098/0099)
Fixes the `deny` CI job failure seen in #3325.
## Test plan
- [x] `cargo deny check advisories` passes locally
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
When Lance 3.0.0 released the check_lance_release.py script did not make
a PR for it because it was a pre-release. This change may not be perfect
but it always ranks stable releases above non-stable releases.
## Summary
- Update all 14 lance crates from `3.0.0-rc.3` (git source) to `3.0.0`
(crates.io release)
- Remove git/tag source references since 3.0.0 is published on crates.io
## Test plan
- [x] `cargo check --features remote --tests --examples` passes
- [x] `cargo clippy --features remote --tests --examples` passes
- [ ] CI passes
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
## Summary
- Upgrade LocalStack from 3.3 to 4.0 in `docker-compose.yml` to fix S3
integration test failures in CI
- Version 3.3 has compatibility issues with newer Python 3.13 and
updated boto3 dependencies
- Matches the LocalStack version used successfully in the lance
repository
## Test plan
- [ ] Verify `docker compose up --detach --wait` completes successfully
in CI
- [ ] All tests in `test_s3.py` pass (5 tests)
- [ ] All `@pytest.mark.s3_test` tests in
`test_namespace_integration.py` pass (7 tests)
- [ ] No regressions in non-integration test jobs (Mac, Windows)
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
## Summary
- Removes the "Experimental API" section from `optimize` method
documentation across Rust, Python, and TypeScript
- Adds a warning to `delete_unverified` documentation in all bindings:
this should only be set to true if you can guarantee no other process is
working on the dataset, otherwise it could be corrupted
- Fixes a typo ("shoudl" → "should")
Closes#3125🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
When we create tables without using Arrow by parsing JS records we
always infer to float64. Many times embeddings are not float64 and it
would be nice to be able to use the native type without requiring users
to pull in Arrow. We can utilize JS's builtin Float32Array to do this.
This PR also adds support for UInt8/16/32 and Int8/16/32 arrays as well.
Closes#3115
There were two issues:
1. The python code needs to get access to the underlying rust table to
setup the permutation reader and the attributes involved in this differ
between the python local table and remote table objects.
~~2. The remote table was sending projection dictionaries as arrays of
tuples and (on LanceDB cloud at least) it does not appear this is how
rest servers are setup to receive them.~~ (this is now fixed as #3023)
~~Leaving as draft as this is built on
https://github.com/lancedb/lancedb/pull/3016~~
This changes around the output format of `Permutation` in some breaking
ways but I think the API is still new enough to be considered
experimental.
1. In order to align with both huggingface's dataset and torch's
expectations the default output format is now a list of dicts
(row-major) instead of a dict of lists (column-major). I've added a
python_col option which will return the dict of lists.
2. In order to align with pytorch's expectation the `torch` format is
now a list of tensors (row-major) instead of a 2D tensor (column-major).
I've added a torch_col option which will return the 2D tensor instead.
Added tests for torch integration with Permutation
~~Leaving draft until https://github.com/lancedb/lancedb/pull/3013
merges as this is built on top of that~~
The permutation table was always intended to be a small table of row id
pointers (and split id). However, it was accidentally doing a full
materialization of the base table 🤦
This PR changes the permutation builder to only store row id and split
id.
I'm working on a lancedb version of pytorch data loading (and hopefully
addressing https://github.com/lancedb/lance/issues/3727).
However, rather than rely on pytorch for everything I'm moving some of
the things that pytorch does into rust. This gives us more control over
data loading (e.g. using shards or a hash-based split) and it allows
permutations to be persistent. In particular I hope to be able to:
* Create a persistent permutation
* This permutation can handle splits, filtering, shuffling, and sharding
* Create a rust data loader that can read a permutation (one or more
splits), or a subset of a permutation (for DDP)
* Create a python data loader that delegates to the rust data loader
Eventually create integrations for other data loading libraries,
including rust & node
We had previously prototyped a `Catalog` trait anticipating a
three-tiered Catalog-Database-Table structure. Now that we have
namespaces in the `Database` we can support any tiering scheme and the
`Catalog` trait is no longer needed.
These operations have existed in lance for a long while and many users
need to drop down to lance for this capability. This PR adds the API and
implements it using filters (e.g. `_rowid IN (...)`) so that in doesn't
currently add any load to `BaseTable`. I'm not sure that is sustainable
as base table implementations may want to specialize how they handle
this method. However, I figure it is a good starting point.
In addition, unlike Lance, this API does not currently guarantee
anything about the order of the take results. This is necessary for the
fallback filter approach to work (SQL filters cannot guarantee result
order)
The `MetadataEraserExec` is super lightweight and doesn't really justify
partitioning. I had a plan recently that was partitioning just for this
node and that seems wasteful.
I can't find any reason for pinning this dependency and the fact that it
is pinned can be kind of annoying to use downstream (e.g. datafusion
currently requires >= 2.6).
<!-- This is an auto-generated comment: release notes by coderabbit.ai
-->
## Summary by CodeRabbit
* **Chores**
* Updated dependencies to newer versions for improved compatibility and
stability.
* **Refactor**
* Improved internal handling of data ranges and stream lifetimes for
enhanced performance and reliability.
* Simplified code style for Python query object conversions without
affecting functionality.
<!-- end of auto-generated comment: release notes by coderabbit.ai -->
This exposes the maximum_nprobes and minimum_nprobes feature that was
added in https://github.com/lancedb/lance/pull/3903
<!-- This is an auto-generated comment: release notes by coderabbit.ai
-->
## Summary by CodeRabbit
- **New Features**
- Added support for specifying minimum and maximum probe counts in
vector search queries, allowing finer control over search behavior.
- Users can now independently set minimum and maximum probes for vector
and hybrid queries via new methods and parameters in Python, Node.js,
and Rust APIs.
- **Bug Fixes**
- Improved parameter validation to ensure correct usage of minimum and
maximum probe values.
- **Tests**
- Expanded test coverage to validate correct handling, serialization,
and error cases for the new probe parameters.
<!-- end of auto-generated comment: release notes by coderabbit.ai -->
<!-- This is an auto-generated comment: release notes by coderabbit.ai
-->
## Summary by CodeRabbit
- **New Features**
- Added the ability to prewarm (load into memory) table indexes via new
methods in Python, Node.js, and Rust APIs, potentially reducing
cold-start query latency.
- **Bug Fixes**
- Ensured prewarming an index does not interfere with subsequent search
operations.
- **Tests**
- Introduced new test cases to verify full-text search index creation,
prewarming, and search functionalities in both Python and Node.js.
- **Chores**
- Updated dependencies for improved compatibility and performance.
<!-- end of auto-generated comment: release notes by coderabbit.ai -->
---------
Co-authored-by: Lu Qiu <luqiujob@gmail.com>
<!-- This is an auto-generated comment: release notes by coderabbit.ai
-->
## Summary by CodeRabbit
- **Chores**
- Updated core dependency versions to v0.25.3-beta.2.
- Enabled additional functionality with a new "dynamodb" feature.
<!-- end of auto-generated comment: release notes by coderabbit.ai -->
We recently allowed this for lance but there was a check in lancedb as
well that was preventing it
<!-- This is an auto-generated comment: release notes by coderabbit.ai
-->
## Summary by CodeRabbit
- **New Features**
- Added support for indexing fixed-size binary data using B-tree
structures for efficient data storage and retrieval.
- **Tests**
- Implemented automated tests to ensure the new binary indexing works
correctly and meets the expected configuration.
<!-- end of auto-generated comment: release notes by coderabbit.ai -->
<!-- This is an auto-generated comment: release notes by coderabbit.ai
-->
## Summary by CodeRabbit
- **Chores**
- Updated dependency versions for improved performance and
compatibility.
- **New Features**
- Added support for structured full-text search with expanded query
types (e.g., match, phrase, boost, multi-match) and flexible input
formats.
- Introduced a new method to check server support for structural
full-text search features.
- Enhanced the query system with new classes and interfaces for handling
various full-text queries.
- Expanded the functionality of existing methods to accept more complex
query structures, including updates to method signatures.
- **Bug Fixes**
- Improved error handling and reporting for full-text search queries.
- **Refactor**
- Enhanced query processing with streamlined input handling and improved
error reporting, ensuring more robust and consistent search results
across platforms.
<!-- end of auto-generated comment: release notes by coderabbit.ai -->
---------
Signed-off-by: BubbleCal <bubble-cal@outlook.com>
Co-authored-by: BubbleCal <bubble-cal@outlook.com>
This PR adds a `to_query_object` method to the various query builders
(except not hybrid queries yet). This makes it possible to inspect the
query that is built.
In addition this PR does some normalization between the sync and async
query paths. A few custom defaults were removed in favor of None (with
the default getting set once, in rust).
Also, the synchronous to_batches method will now actually stream results
Also, the remote API now defaults to prefiltering
The `ConnectRequest` has a set of properties that only make sense for
listing databases / catalogs and a set of properties that only make
sense for remote databases.
This PR reduces all options to a single `HashMap<String, String>`. This
makes it easier to add new database / catalog implementations and makes
it clearer to users which options are applicable in which situations.
I don't believe there are any breaking changes here. The closest thing
is that I placed the `ConnectBuilder` methods `api_key`, `region`, and
`host_override` behind a `remote` feature gate. This is not strictly
needed and I could remove the feature gate but it seemed appropriate.
Since using these methods without the remote feature would have been
meaningless I don't feel this counts as a breaking change.
We could look at removing these methods entirely from the
`ConnectBuilder` (and encouraging users to use `RemoteDatabaseOptions`
instead) but I'm not sure how I feel about that.
Another approach we could take is to move these methods into a
`RemoteConnectBuilderExt` trait (and there could be a similar
`ListingConnectBuilderExt` trait to add methods for the listing database
/ catalog).
For now though my main goal is to simplify `ConnectRequest` as much as
possible (I see this being part of the key public API for database /
catalog integrations, similar to the `BaseTable`, `Catalog`, and
`Database` traits and I'd like it to be simple).
We attempted to make pylance optional in
https://github.com/lancedb/lancedb/pull/2156 but it appears this did not
quite work. Users are unable to use lancedb from a fresh install. This
reverts the optional-ness so we can get back in a working state while we
fix the issue.
Datafusion makes the batch size available as part of the `SessionState`.
We should use that to set the `max_batch_length` property in the
`QueryExecutionOptions`.
This PR makes it possible to create a table using an asynchronous stream
of input data. Currently only a synchronous iterator is supported. There
are a number of follow-ups not yet tackled:
* Support for embedding functions (the embedding functions wrapper needs
to be re-written to be async, should be an easy lift)
* Support for async input into the remote table (the make_ipc_batch
needs to change to accept async input, leaving undone for now because I
think we want to support actual streaming uploads into the remote table
soon)
* Support for async input into the add function (pretty essential, but
it is a fairly distinct code path, so saving for a different PR)
It seems that `RecordBatch::with_schema` is unable to remove schema
metadata from a batch. It fails with the error `target schema is not
superset of current schema`.
I'm not sure how the `test_metadata_erased` test is passing. Strangely,
the metadata was not present by the time the batch arrived at the
metadata eraser. I think maybe the schema metadata is only present in
the batch if there is a filter.
I've created a new unit test that makes sure the metadata is erased if
we have a filter also
This also changes the pylance pin from `==0.23.2` to `~=0.23.2` which
should allow the pylance dependency to float a little. The pylance
dependency is actually not used for much anymore and so it should be
tolerant of patch changes.
If we start supporting external catalogs then "drop database" may be
misleading (and not possible). We should be more clear that this is a
utility method to drop all tables. This is also a nice chance for some
consistency cleanup as it was `drop_db` in rust, `drop_database` in
python, and non-existent in typescript.
This PR also adds a public accessor to get the database trait from a
connection.
BREAKING CHANGE: the `drop_database` / `drop_db` methods are now
deprecated.
Similar to
https://github.com/lancedb/lancedb/commit/c269524b2f394a6c8508436840bce17d884a42c0
this PR reworks and exposes an internal trait (this time
`TableInternal`) to be a public trait. These two PRs together should
make it possible for others to integrate LanceDB on top of other
catalogs.
This PR also adds a basic `TableProvider` implementation for tables,
although some work still needs to be done here (pushdown not yet
enabled).
This opens up the door for more custom database implementations than the
two we have today. The biggest change should be inivisble:
`ConnectionInternal` has been renamed to `Database`, made public, and
refactored
However, there are a few breaking changes. `data_storage_version` and
`enable_v2_manifest_paths` have been moved from options on
`create_table` to options for the database which are now set via
`storage_options`.
Before:
```
db = connect(uri)
tbl = db.create_table("my_table", data, data_storage_version="legacy", enable_v2_manifest_paths=True)
```
After:
```
db = connect(uri, storage_options={
"new_table_enable_v2_manifest_paths": "true",
"new_table_data_storage_version": "legacy"
})
tbl = db.create_table("my_table", data)
```
BREAKING CHANGE: the data_storage_version, enable_v2_manifest_paths
options have moved from options to create_table to storage_options.
BREAKING CHANGE: the use_legacy_format option has been removed,
data_storage_version has replaced it for some time now
This currently only works for local tables (remote tables cannot be
queried)
This is also exclusive to the sync interface. However, since the pyarrow
dataset interface is synchronous I am not sure if there is much value in
making an async-wrapping variant.
In addition, I added a `to_batches` method to the base query in the sync
API. This already exists in the async API. In the sync API this PR only
adds support for vector queries and scalar queries and not for hybrid or
FTS queries.
This doesn't actually block a python-only release since this step runs
after the version bump has been pushed but it still would be nice for
the git job to finish successfully.
Attempting to create a new minor version failed with:
```
Specified version (0.4.21-beta.0) does not match last tagged version (0.4.20)
```
It seems the last release commit for rust/node was made without the new
process and did not adjust bumpversion.toml correctly (or maybe
bumpversion.toml did not exist at that time)
I tried to do a stable release and it failed with:
```
Traceback (most recent call last):
File "/home/runner/work/lancedb/lancedb/ci/check_breaking_changes.py", line 20, in <module>
commits = repo.compare(args.base, args.head).commits
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/opt/hostedtoolcache/Python/3.11.9/x64/lib/python3.11/site-packages/github/Repository.py", line 1133, in compare
headers, data = self._requester.requestJsonAndCheck("GET", f"{self.url}/compare/{base}...{head}", params)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/opt/hostedtoolcache/Python/3.11.9/x64/lib/python3.11/site-packages/github/Requester.py", line 548, in requestJsonAndCheck
return self.__check(*self.requestJson(verb, url, parameters, headers, input, self.__customConnection(url)))
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/opt/hostedtoolcache/Python/3.11.9/x64/lib/python3.11/site-packages/github/Requester.py", line 609, in __check
raise self.createException(status, responseHeaders, data)
github.GithubException.UnknownObjectException: 404 {"message": "Not Found", "documentation_url": "https://docs.github.com/rest/commits/commits#compare-two-commits"}
```
I believe the problem is that we are calculating the
`LAST_STABLE_RELEASE` after we have run bump version and so the newly
created tag is in the list of tags we search and it is the most recent
one and so it gets included as `LAST_STABLE_RELEASE`. Then, the call to
github fails because we haven't pushed the tag yet. This changes the
logic to grab `LAST_STABLE_RELEASE` before we create any new tags.
The optimize function is pretty crucial for getting good performance
when building a large scale dataset but it was only exposed in rust
(many sync python users are probably doing this via to_lance today)
This PR adds the optimize function to nodejs and to python.
I left the function marked experimental because I think there will
likely be changes to optimization (e.g. if we add features like
"optimize on write"). I also only exposed the `cleanup_older_than`
configuration parameter since this one is very commonly used and the
rest have sensible defaults and we don't really know why we would
recommend different values for these defaults anyways.
This was already configurable in the rust API but it wasn't actually
being passed down to the underlying dataset. I added this option to both
the async python API and the new nodejs API.
I also added this option to the synchronous python API.
I did not add the option to vectordb.
The rust implementation of the remote client is not yet ready. This is
understandably confusing for users since it is enabled by default. This
PR disables it by default. We can re-enable it when we are ready (even
then it is not clear this is something that should be a default
feature).
---------
Co-authored-by: Will Jones <willjones127@gmail.com>
We aren't yet ready to switch over the examples since almost all JS
examples rely on embeddings and we haven't yet ported those over.
However, this makes it possible for those that are interested to start
using `@lancedb/lancedb`
In
https://github.com/lancedb/lancedb/commit/2de226220b78c51b22319fb4fbc4678f4d269d65
I added a new `IntoArrow` trait for adding data into a table.
Unfortunately, it seems my approach for implementing the trait for
"things that are already record batch readers" was flawed. This PR
corrects that flaw and, conveniently, removes the need to box readers at
all (though it is ok if you do).
Weston Pace
