feat: parallel inserts for local tables (#3062)

When input data is sufficiently large, we automatically split up into
parallel writes using a round-robin exchange operator. We sample the
first batch to determine data width, and target size of 1 million rows
or 2GB, whichever is smaller.
This commit is contained in:
Will Jones
2026-02-24 12:26:51 -08:00
committed by GitHub
parent 0d3fc7860a
commit d5948576b9
5 changed files with 483 additions and 52 deletions

View File

@@ -9,13 +9,6 @@
use std::sync::Arc;
use arrow_array::{ArrayRef, RecordBatch, RecordBatchIterator, RecordBatchReader};
use arrow_schema::{ArrowError, SchemaRef};
use async_trait::async_trait;
use futures::stream::once;
use futures::StreamExt;
use lance_datafusion::utils::StreamingWriteSource;
use crate::arrow::{
SendableRecordBatchStream, SendableRecordBatchStreamExt, SimpleRecordBatchStream,
};
@@ -25,6 +18,12 @@ use crate::embeddings::{
};
use crate::table::{ColumnDefinition, ColumnKind, TableDefinition};
use crate::{Error, Result};
use arrow_array::{ArrayRef, RecordBatch, RecordBatchIterator, RecordBatchReader};
use arrow_schema::{ArrowError, SchemaRef};
use async_trait::async_trait;
use futures::stream::once;
use futures::StreamExt;
use lance_datafusion::utils::StreamingWriteSource;
pub trait Scannable: Send {
/// Returns the schema of the data.
@@ -349,6 +348,133 @@ pub fn scannable_with_embeddings(
Ok(inner)
}
/// A wrapper that buffers the first RecordBatch from a Scannable so we can
/// inspect it (e.g. to estimate data size) without losing it.
pub(crate) struct PeekedScannable {
inner: Box<dyn Scannable>,
peeked: Option<RecordBatch>,
/// The first item from the stream, if it was an error. Stored so we can
/// re-emit it from `scan_as_stream` instead of silently dropping it.
first_error: Option<crate::Error>,
stream: Option<SendableRecordBatchStream>,
}
impl PeekedScannable {
pub fn new(inner: Box<dyn Scannable>) -> Self {
Self {
inner,
peeked: None,
first_error: None,
stream: None,
}
}
/// Reads and buffers the first batch from the inner scannable.
/// Returns a clone of it. Subsequent calls return the same batch.
///
/// Returns `None` if the stream is empty or the first item is an error.
/// Errors are preserved and re-emitted by `scan_as_stream`.
pub async fn peek(&mut self) -> Option<RecordBatch> {
if self.peeked.is_some() {
return self.peeked.clone();
}
// Already peeked and got an error or empty stream.
if self.stream.is_some() || self.first_error.is_some() {
return None;
}
let mut stream = self.inner.scan_as_stream();
match stream.next().await {
Some(Ok(batch)) => {
self.peeked = Some(batch.clone());
self.stream = Some(stream);
Some(batch)
}
Some(Err(e)) => {
self.first_error = Some(e);
self.stream = Some(stream);
None
}
None => {
self.stream = Some(stream);
None
}
}
}
}
impl Scannable for PeekedScannable {
fn schema(&self) -> SchemaRef {
self.inner.schema()
}
fn num_rows(&self) -> Option<usize> {
self.inner.num_rows()
}
fn rescannable(&self) -> bool {
self.inner.rescannable()
}
fn scan_as_stream(&mut self) -> SendableRecordBatchStream {
let schema = self.inner.schema();
// If peek() hit an error, prepend it so downstream sees the error.
let error_item = self.first_error.take().map(Err);
match (self.peeked.take(), self.stream.take()) {
(Some(batch), Some(rest)) => {
let prepend = futures::stream::once(std::future::ready(Ok(batch)));
Box::pin(SimpleRecordBatchStream {
schema,
stream: prepend.chain(rest),
})
}
(Some(batch), None) => Box::pin(SimpleRecordBatchStream {
schema,
stream: futures::stream::once(std::future::ready(Ok(batch))),
}),
(None, Some(rest)) => {
if let Some(err) = error_item {
let stream = futures::stream::once(std::future::ready(err));
Box::pin(SimpleRecordBatchStream { schema, stream })
} else {
rest
}
}
(None, None) => {
// peek() was never called — just delegate
self.inner.scan_as_stream()
}
}
}
}
/// Compute the number of write partitions based on data size estimates.
///
/// `sample_bytes` and `sample_rows` come from a representative batch and are
/// used to estimate per-row size. `total_rows_hint` is the total row count
/// when known; otherwise `sample_rows` row count is used as a lower bound
/// estimate.
///
/// Targets roughly 1 million rows or 2 GB per partition, capped at
/// `max_partitions` (typically the number of available CPU cores).
pub(crate) fn estimate_write_partitions(
sample_bytes: usize,
sample_rows: usize,
total_rows_hint: Option<usize>,
max_partitions: usize,
) -> usize {
if sample_rows == 0 {
return 1;
}
let bytes_per_row = sample_bytes / sample_rows;
let total_rows = total_rows_hint.unwrap_or(sample_rows);
let total_bytes = total_rows * bytes_per_row;
let by_rows = total_rows.div_ceil(1_000_000);
let by_bytes = total_bytes.div_ceil(2 * 1024 * 1024 * 1024);
by_rows.max(by_bytes).max(1).min(max_partitions)
}
#[cfg(test)]
mod tests {
use super::*;
@@ -445,6 +571,231 @@ mod tests {
assert!(result2.unwrap().is_err());
}
mod peeked_scannable_tests {
use crate::test_utils::TestCustomError;
use super::*;
#[tokio::test]
async fn test_peek_returns_first_batch() {
let batch = record_batch!(("id", Int64, [1, 2, 3])).unwrap();
let mut peeked = PeekedScannable::new(Box::new(batch.clone()));
let first = peeked.peek().await.unwrap();
assert_eq!(first, batch);
}
#[tokio::test]
async fn test_peek_is_idempotent() {
let batch = record_batch!(("id", Int64, [1, 2, 3])).unwrap();
let mut peeked = PeekedScannable::new(Box::new(batch.clone()));
let first = peeked.peek().await.unwrap();
let second = peeked.peek().await.unwrap();
assert_eq!(first, second);
}
#[tokio::test]
async fn test_scan_after_peek_returns_all_data() {
let batches = vec![
record_batch!(("id", Int64, [1, 2])).unwrap(),
record_batch!(("id", Int64, [3, 4, 5])).unwrap(),
];
let mut peeked = PeekedScannable::new(Box::new(batches.clone()));
let first = peeked.peek().await.unwrap();
assert_eq!(first, batches[0]);
let result: Vec<RecordBatch> = peeked.scan_as_stream().try_collect().await.unwrap();
assert_eq!(result.len(), 2);
assert_eq!(result[0], batches[0]);
assert_eq!(result[1], batches[1]);
}
#[tokio::test]
async fn test_scan_without_peek_passes_through() {
let batch = record_batch!(("id", Int64, [1, 2, 3])).unwrap();
let mut peeked = PeekedScannable::new(Box::new(batch.clone()));
let result: Vec<RecordBatch> = peeked.scan_as_stream().try_collect().await.unwrap();
assert_eq!(result.len(), 1);
assert_eq!(result[0], batch);
}
#[tokio::test]
async fn test_delegates_num_rows() {
let batches = vec![
record_batch!(("id", Int64, [1, 2])).unwrap(),
record_batch!(("id", Int64, [3])).unwrap(),
];
let peeked = PeekedScannable::new(Box::new(batches));
assert_eq!(peeked.num_rows(), Some(3));
}
#[tokio::test]
async fn test_non_rescannable_stream_data_preserved() {
let batches = vec![
record_batch!(("id", Int64, [1, 2])).unwrap(),
record_batch!(("id", Int64, [3])).unwrap(),
];
let schema = batches[0].schema();
let inner = futures::stream::iter(batches.clone().into_iter().map(Ok));
let stream: SendableRecordBatchStream = Box::pin(SimpleRecordBatchStream {
schema,
stream: inner,
});
let mut peeked = PeekedScannable::new(Box::new(stream));
assert!(!peeked.rescannable());
assert_eq!(peeked.num_rows(), None);
let first = peeked.peek().await.unwrap();
assert_eq!(first, batches[0]);
// All data is still available via scan_as_stream
let result: Vec<RecordBatch> = peeked.scan_as_stream().try_collect().await.unwrap();
assert_eq!(result.len(), 2);
assert_eq!(result[0], batches[0]);
assert_eq!(result[1], batches[1]);
}
#[tokio::test]
async fn test_error_in_first_batch_propagates() {
let schema = Arc::new(arrow_schema::Schema::new(vec![arrow_schema::Field::new(
"id",
arrow_schema::DataType::Int64,
false,
)]));
let inner = futures::stream::iter(vec![Err(Error::External {
source: Box::new(TestCustomError),
})]);
let stream: SendableRecordBatchStream = Box::pin(SimpleRecordBatchStream {
schema,
stream: inner,
});
let mut peeked = PeekedScannable::new(Box::new(stream));
// peek returns None for errors
assert!(peeked.peek().await.is_none());
// But the error should come through when scanning
let mut stream = peeked.scan_as_stream();
let first = stream.next().await.unwrap();
assert!(first.is_err());
let err = first.unwrap_err();
assert!(
matches!(&err, Error::External { source } if source.downcast_ref::<TestCustomError>().is_some()),
"Expected TestCustomError to be preserved, got: {err}"
);
}
#[tokio::test]
async fn test_error_in_later_batch_propagates() {
let good_batch = record_batch!(("id", Int64, [1, 2])).unwrap();
let schema = good_batch.schema();
let inner = futures::stream::iter(vec![
Ok(good_batch.clone()),
Err(Error::External {
source: Box::new(TestCustomError),
}),
]);
let stream: SendableRecordBatchStream = Box::pin(SimpleRecordBatchStream {
schema,
stream: inner,
});
let mut peeked = PeekedScannable::new(Box::new(stream));
// peek succeeds with the first batch
let first = peeked.peek().await.unwrap();
assert_eq!(first, good_batch);
// scan_as_stream should yield the first batch, then the error
let mut stream = peeked.scan_as_stream();
let batch1 = stream.next().await.unwrap().unwrap();
assert_eq!(batch1, good_batch);
let batch2 = stream.next().await.unwrap();
assert!(batch2.is_err());
let err = batch2.unwrap_err();
assert!(
matches!(&err, Error::External { source } if source.downcast_ref::<TestCustomError>().is_some()),
"Expected TestCustomError to be preserved, got: {err}"
);
}
#[tokio::test]
async fn test_empty_stream_returns_none() {
let schema = Arc::new(arrow_schema::Schema::new(vec![arrow_schema::Field::new(
"id",
arrow_schema::DataType::Int64,
false,
)]));
let inner = futures::stream::empty();
let stream: SendableRecordBatchStream = Box::pin(SimpleRecordBatchStream {
schema,
stream: inner,
});
let mut peeked = PeekedScannable::new(Box::new(stream));
assert!(peeked.peek().await.is_none());
// Scanning an empty (post-peek) stream should yield nothing
let result: Vec<RecordBatch> = peeked.scan_as_stream().try_collect().await.unwrap();
assert!(result.is_empty());
}
}
mod estimate_write_partitions_tests {
use super::*;
#[test]
fn test_small_data_single_partition() {
// 100 rows * 24 bytes/row = 2400 bytes — well under both thresholds
assert_eq!(estimate_write_partitions(2400, 100, Some(100), 8), 1);
}
#[test]
fn test_scales_by_row_count() {
// 2.5M rows at 24 bytes/row — row threshold dominates
// ceil(2_500_000 / 1_000_000) = 3
assert_eq!(estimate_write_partitions(72, 3, Some(2_500_000), 8), 3);
}
#[test]
fn test_scales_by_byte_size() {
// 100k rows at 40KB/row = ~4GB total → ceil(4GB / 2GB) = 2
let sample_bytes = 40_000 * 10;
assert_eq!(
estimate_write_partitions(sample_bytes, 10, Some(100_000), 8),
2
);
}
#[test]
fn test_capped_at_max_partitions() {
// 10M rows would want 10 partitions, but capped at 4
assert_eq!(estimate_write_partitions(72, 3, Some(10_000_000), 4), 4);
}
#[test]
fn test_zero_sample_rows_returns_one() {
assert_eq!(estimate_write_partitions(0, 0, Some(1_000_000), 8), 1);
}
#[test]
fn test_no_row_hint_uses_sample_size() {
// Without a hint, uses sample_rows (3), which is small
assert_eq!(estimate_write_partitions(72, 3, None, 8), 1);
}
#[test]
fn test_always_at_least_one() {
assert_eq!(estimate_write_partitions(24, 1, Some(1), 8), 1);
}
}
mod embedding_tests {
use super::*;
use crate::embeddings::MemoryRegistry;

View File

@@ -97,10 +97,7 @@ pub type Result<T> = std::result::Result<T, Error>;
impl From<ArrowError> for Error {
fn from(source: ArrowError) -> Self {
match source {
ArrowError::ExternalError(source) => match source.downcast::<Self>() {
Ok(e) => *e,
Err(source) => Self::External { source },
},
ArrowError::ExternalError(source) => Self::from_box_error(source),
_ => Self::Arrow { source },
}
}
@@ -110,15 +107,7 @@ impl From<DataFusionError> for Error {
fn from(source: DataFusionError) -> Self {
match source {
DataFusionError::ArrowError(source, _) => (*source).into(),
DataFusionError::External(source) => match source.downcast::<Self>() {
Ok(e) => *e,
Err(source) => match source.downcast::<ArrowError>() {
Ok(arrow_error) => Self::Arrow {
source: *arrow_error,
},
Err(source) => Self::External { source },
},
},
DataFusionError::External(source) => Self::from_box_error(source),
other => Self::External {
source: Box::new(other),
},
@@ -130,15 +119,52 @@ impl From<lance::Error> for Error {
fn from(source: lance::Error) -> Self {
// Try to unwrap external errors that were wrapped by lance
match source {
lance::Error::Wrapped { error, .. } => match error.downcast::<Self>() {
Ok(e) => *e,
Err(source) => Self::External { source },
},
lance::Error::Wrapped { error, .. } => Self::from_box_error(error),
lance::Error::External { source } => Self::from_box_error(source),
_ => Self::Lance { source },
}
}
}
impl Error {
fn from_box_error(mut source: Box<dyn std::error::Error + Send + Sync>) -> Self {
source = match source.downcast::<Self>() {
Ok(e) => match *e {
Self::External { source } => return Self::from_box_error(source),
other => return other,
},
Err(source) => source,
};
source = match source.downcast::<lance::Error>() {
Ok(e) => match *e {
lance::Error::Wrapped { error, .. } => return Self::from_box_error(error),
other => return other.into(),
},
Err(source) => source,
};
source = match source.downcast::<ArrowError>() {
Ok(e) => match *e {
ArrowError::ExternalError(source) => return Self::from_box_error(source),
other => return other.into(),
},
Err(source) => source,
};
source = match source.downcast::<DataFusionError>() {
Ok(e) => match *e {
DataFusionError::ArrowError(source, _) => return (*source).into(),
DataFusionError::External(source) => return Self::from_box_error(source),
other => return other.into(),
},
Err(source) => source,
};
Self::External { source }
}
}
impl From<object_store::Error> for Error {
fn from(source: object_store::Error) -> Self {
Self::ObjectStore { source }

View File

@@ -6,11 +6,12 @@
use arrow_array::{RecordBatch, RecordBatchReader};
use arrow_schema::{DataType, Field, Schema, SchemaRef};
use async_trait::async_trait;
use datafusion_execution::TaskContext;
use datafusion_expr::Expr;
use datafusion_physical_plan::display::DisplayableExecutionPlan;
use datafusion_physical_plan::ExecutionPlan;
use futures::stream::FuturesUnordered;
use futures::StreamExt;
use futures::TryStreamExt;
use lance::dataset::builder::DatasetBuilder;
pub use lance::dataset::ColumnAlteration;
pub use lance::dataset::NewColumnTransform;
@@ -21,7 +22,6 @@ use lance::dataset::{InsertBuilder, WriteParams};
use lance::index::vector::utils::infer_vector_dim;
use lance::index::vector::VectorIndexParams;
use lance::io::{ObjectStoreParams, WrappingObjectStore};
use lance_datafusion::exec::execute_plan;
use lance_datafusion::utils::StreamingWriteSource;
use lance_index::scalar::{BuiltinIndexType, ScalarIndexParams};
use lance_index::vector::bq::RQBuildParams;
@@ -43,7 +43,7 @@ use std::format;
use std::path::Path;
use std::sync::Arc;
use crate::data::scannable::Scannable;
use crate::data::scannable::{estimate_write_partitions, PeekedScannable, Scannable};
use crate::database::Database;
use crate::embeddings::{EmbeddingDefinition, EmbeddingRegistry, MemoryRegistry};
use crate::error::{Error, Result};
@@ -2113,7 +2113,7 @@ impl BaseTable for NativeTable {
}
}
async fn add(&self, add: AddDataBuilder) -> Result<AddResult> {
async fn add(&self, mut add: AddDataBuilder) -> Result<AddResult> {
let table_def = self.table_definition().await?;
self.dataset.ensure_mutable()?;
@@ -2122,6 +2122,22 @@ impl BaseTable for NativeTable {
let table_schema = Schema::from(&ds.schema().clone());
// Peek at the first batch to estimate a good partition count for
// write parallelism.
let mut peeked = PeekedScannable::new(add.data);
let num_partitions = if let Some(first_batch) = peeked.peek().await {
let max_partitions = lance_core::utils::tokio::get_num_compute_intensive_cpus();
estimate_write_partitions(
first_batch.get_array_memory_size(),
first_batch.num_rows(),
peeked.num_rows(),
max_partitions,
)
} else {
1
};
add.data = Box::new(peeked);
let output = add.into_plan(&table_schema, &table_def)?;
let lance_params = output
@@ -2135,18 +2151,41 @@ impl BaseTable for NativeTable {
..Default::default()
});
let plan = Arc::new(InsertExec::new(
ds_wrapper.clone(),
ds,
output.plan,
lance_params,
));
// Repartition for write parallelism if beneficial.
let plan = if num_partitions > 1 {
Arc::new(
datafusion_physical_plan::repartition::RepartitionExec::try_new(
output.plan,
datafusion_physical_plan::Partitioning::RoundRobinBatch(num_partitions),
)?,
) as Arc<dyn ExecutionPlan>
} else {
output.plan
};
let stream = execute_plan(plan, Default::default())?;
stream
.try_collect::<Vec<_>>()
.await
.map_err(crate::Error::from)?;
let insert_exec = Arc::new(InsertExec::new(ds_wrapper.clone(), ds, plan, lance_params));
// Execute all partitions in parallel.
let task_ctx = Arc::new(TaskContext::default());
let handles = FuturesUnordered::new();
for partition in 0..num_partitions {
let exec = insert_exec.clone();
let ctx = task_ctx.clone();
handles.push(tokio::spawn(async move {
let mut stream = exec
.execute(partition, ctx)
.map_err(|e| -> Error { e.into() })?;
while let Some(batch) = stream.next().await {
batch.map_err(|e| -> Error { e.into() })?;
}
Ok::<_, Error>(())
}));
}
for handle in handles {
handle.await.map_err(|e| Error::Runtime {
message: format!("Insert task panicked: {}", e),
})??;
}
let version = ds_wrapper.get().await?.manifest().version;
Ok(AddResult { version })

View File

@@ -219,6 +219,7 @@ mod tests {
use crate::table::add_data::NaNVectorBehavior;
use crate::table::{ColumnDefinition, ColumnKind, Table, TableDefinition, WriteOptions};
use crate::test_utils::embeddings::MockEmbed;
use crate::test_utils::TestCustomError;
use crate::Error;
use super::AddDataMode;
@@ -283,17 +284,20 @@ mod tests {
test_add_with_data(stream).await;
}
#[derive(Debug)]
struct MyError;
impl std::fmt::Display for MyError {
fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
write!(f, "MyError occurred")
}
fn assert_preserves_external_error(err: &Error) {
assert!(
matches!(err, Error::External { source } if source.downcast_ref::<TestCustomError>().is_some()),
"Expected Error::External, got: {err:?}"
);
// The original TestCustomError message should be preserved through the
// error chain, even if the error gets wrapped multiple times by
// lance's insert pipeline.
assert!(
err.to_string().contains("TestCustomError occurred"),
"Expected original error message to be preserved, got: {err}"
);
}
impl std::error::Error for MyError {}
#[tokio::test]
async fn test_add_preserves_reader_error() {
let table = create_test_table().await;
@@ -301,7 +305,7 @@ mod tests {
let schema = first_batch.schema();
let iterator = vec![
Ok(first_batch),
Err(ArrowError::ExternalError(Box::new(MyError))),
Err(ArrowError::ExternalError(Box::new(TestCustomError))),
];
let reader: Box<dyn arrow_array::RecordBatchReader + Send> = Box::new(
RecordBatchIterator::new(iterator.into_iter(), schema.clone()),
@@ -309,7 +313,7 @@ mod tests {
let result = table.add(reader).execute().await;
assert!(result.is_err());
assert_preserves_external_error(&result.unwrap_err());
}
#[tokio::test]
@@ -320,7 +324,7 @@ mod tests {
let iterator = vec![
Ok(first_batch),
Err(Error::External {
source: Box::new(MyError),
source: Box::new(TestCustomError),
}),
];
let stream = futures::stream::iter(iterator);
@@ -331,7 +335,7 @@ mod tests {
let result = table.add(stream).execute().await;
assert!(result.is_err());
assert_preserves_external_error(&result.unwrap_err());
}
#[tokio::test]

View File

@@ -4,3 +4,14 @@
pub mod connection;
pub mod datagen;
pub mod embeddings;
#[derive(Debug)]
pub struct TestCustomError;
impl std::fmt::Display for TestCustomError {
fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
write!(f, "TestCustomError occurred")
}
}
impl std::error::Error for TestCustomError {}