mbutrovich commented on code in PR #3845: URL: https://github.com/apache/datafusion-comet/pull/3845#discussion_r3065910873
########## native/shuffle/src/partitioners/immediate_mode.rs: ########## @@ -0,0 +1,1079 @@ +// Licensed to the Apache Software Foundation (ASF) under one +// or more contributor license agreements. See the NOTICE file +// distributed with this work for additional information +// regarding copyright ownership. The ASF licenses this file +// to you under the Apache License, Version 2.0 (the +// "License"); you may not use this file except in compliance +// with the License. You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, +// software distributed under the License is distributed on an +// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +// KIND, either express or implied. See the License for the +// specific language governing permissions and limitations +// under the License. + +use crate::metrics::ShufflePartitionerMetrics; +use crate::partitioners::partition_id::{assign_hash_partition_ids, assign_range_partition_ids}; +use crate::partitioners::ShufflePartitioner; +use crate::{CometPartitioning, CompressionCodec}; +use arrow::array::builder::{ + make_builder, ArrayBuilder, BinaryBuilder, BinaryViewBuilder, BooleanBuilder, + LargeBinaryBuilder, LargeStringBuilder, NullBuilder, PrimitiveBuilder, StringBuilder, + StringViewBuilder, +}; +use arrow::array::{ + Array, ArrayRef, AsArray, BinaryViewArray, RecordBatch, StringViewArray, UInt32Array, +}; +use arrow::compute::take; +use arrow::datatypes::{ + DataType, Date32Type, Date64Type, Decimal128Type, Decimal256Type, Float32Type, Float64Type, + Int16Type, Int32Type, Int64Type, Int8Type, SchemaRef, TimeUnit, TimestampMicrosecondType, + TimestampMillisecondType, TimestampNanosecondType, TimestampSecondType, UInt16Type, UInt32Type, + UInt64Type, UInt8Type, +}; +use arrow::ipc::writer::StreamWriter; +use datafusion::common::{DataFusionError, Result}; +use datafusion::execution::memory_pool::{MemoryConsumer, MemoryLimit, MemoryReservation}; +use datafusion::execution::runtime_env::RuntimeEnv; +use datafusion_comet_spark_expr::murmur3::create_murmur3_hashes; +use std::fs::{File, OpenOptions}; +use std::io::{BufWriter, Seek, Write}; +use std::sync::Arc; +use tokio::time::Instant; + +macro_rules! scatter_byte_array { + ($builder:expr, $source:expr, $indices:expr, $offset_type:ty, $builder_type:ty, $cast:ident) => {{ + let src = $source.$cast::<$offset_type>(); + let dst = $builder + .as_any_mut() + .downcast_mut::<$builder_type>() + .expect("builder type mismatch"); + if src.null_count() == 0 { + for &idx in $indices { + dst.append_value(src.value(idx)); + } + } else { + for &idx in $indices { + dst.append_option(src.is_valid(idx).then(|| src.value(idx))); + } + } + }}; +} + +macro_rules! scatter_byte_view { + ($builder:expr, $source:expr, $indices:expr, $array_type:ty, $builder_type:ty) => {{ + let src = $source + .as_any() + .downcast_ref::<$array_type>() + .expect("array type mismatch"); + let dst = $builder + .as_any_mut() + .downcast_mut::<$builder_type>() + .expect("builder type mismatch"); + if src.null_count() == 0 { + for &idx in $indices { + dst.append_value(src.value(idx)); + } + } else { + for &idx in $indices { + dst.append_option(src.is_valid(idx).then(|| src.value(idx))); + } + } + }}; +} + +macro_rules! scatter_primitive { + ($builder:expr, $source:expr, $indices:expr, $arrow_type:ty) => {{ + let src = $source.as_primitive::<$arrow_type>(); + let dst = $builder + .as_any_mut() + .downcast_mut::<PrimitiveBuilder<$arrow_type>>() + .expect("builder type mismatch"); + if src.null_count() == 0 { + for &idx in $indices { + dst.append_value(src.value(idx)); + } + } else { + for &idx in $indices { + dst.append_option(src.is_valid(idx).then(|| src.value(idx))); + } + } + }}; +} + +/// Scatter-append selected rows from `source` into `builder`. +fn scatter_append( + builder: &mut dyn ArrayBuilder, + source: &dyn Array, + indices: &[usize], +) -> Result<()> { + use DataType::*; + match source.data_type() { + Boolean => { + let src = source.as_boolean(); + let dst = builder + .as_any_mut() + .downcast_mut::<BooleanBuilder>() + .unwrap(); + if src.null_count() == 0 { + for &idx in indices { + dst.append_value(src.value(idx)); + } + } else { + for &idx in indices { + dst.append_option(src.is_valid(idx).then(|| src.value(idx))); + } + } + } + Int8 => scatter_primitive!(builder, source, indices, Int8Type), + Int16 => scatter_primitive!(builder, source, indices, Int16Type), + Int32 => scatter_primitive!(builder, source, indices, Int32Type), + Int64 => scatter_primitive!(builder, source, indices, Int64Type), + UInt8 => scatter_primitive!(builder, source, indices, UInt8Type), + UInt16 => scatter_primitive!(builder, source, indices, UInt16Type), + UInt32 => scatter_primitive!(builder, source, indices, UInt32Type), + UInt64 => scatter_primitive!(builder, source, indices, UInt64Type), + Float32 => scatter_primitive!(builder, source, indices, Float32Type), + Float64 => scatter_primitive!(builder, source, indices, Float64Type), + Date32 => scatter_primitive!(builder, source, indices, Date32Type), + Date64 => scatter_primitive!(builder, source, indices, Date64Type), + Timestamp(TimeUnit::Second, _) => { + scatter_primitive!(builder, source, indices, TimestampSecondType) + } + Timestamp(TimeUnit::Millisecond, _) => { + scatter_primitive!(builder, source, indices, TimestampMillisecondType) + } + Timestamp(TimeUnit::Microsecond, _) => { + scatter_primitive!(builder, source, indices, TimestampMicrosecondType) + } + Timestamp(TimeUnit::Nanosecond, _) => { + scatter_primitive!(builder, source, indices, TimestampNanosecondType) + } + Decimal128(_, _) => scatter_primitive!(builder, source, indices, Decimal128Type), + Decimal256(_, _) => scatter_primitive!(builder, source, indices, Decimal256Type), + Utf8 => scatter_byte_array!(builder, source, indices, i32, StringBuilder, as_string), + LargeUtf8 => { + scatter_byte_array!(builder, source, indices, i64, LargeStringBuilder, as_string) + } + Binary => scatter_byte_array!(builder, source, indices, i32, BinaryBuilder, as_binary), + LargeBinary => { + scatter_byte_array!(builder, source, indices, i64, LargeBinaryBuilder, as_binary) + } + Utf8View => { + scatter_byte_view!(builder, source, indices, StringViewArray, StringViewBuilder) + } + BinaryView => { + scatter_byte_view!(builder, source, indices, BinaryViewArray, BinaryViewBuilder) + } + Null => { + let dst = builder.as_any_mut().downcast_mut::<NullBuilder>().unwrap(); + dst.append_nulls(indices.len()); + } + dt => { + return Err(DataFusionError::NotImplemented(format!( + "Scatter append not implemented for {dt}" + ))); + } + } + Ok(()) +} + +/// Per-column strategy: scatter-write via builder for primitive/string types, +/// or accumulate taken sub-arrays for complex types (List, Map, Struct, etc.). +enum ColumnBuffer { + /// Fast path: direct scatter into a pre-allocated builder. Review Comment: This is something we can experiment with the shuffle benchmark or other workloads in the future, just documenting so we can maybe make an issue to track it. Just on my brain after reviewing DataFusion PRs today that optimized NULL buffer building. The `Builder` path is a hand-rolled `take` that writes into pre-allocated builders, requiring ~150 lines of macros and type dispatch (`scatter_primitive`, `scatter_byte_array`, `scatter_byte_view`, `scatter_append`, `has_scatter_support`, `ColumnBuffer` enum). Arrow's `take` kernel handles all types automatically, uses SIMD for primitives, and does null propagation via bulk bitmap operations rather than per-row `is_valid` checks (see [datafusion#21538](https://github.com/apache/datafusion/pull/21538), [datafusion#21468](https://github.com/apache/datafusion/pull/21468)). Using `take` for all types and accumulating `RecordBatch`es per partition (flushed via `concat_batches` at target size) would eliminate the scatter macros and `ColumnBuffer` enum. The trade-off depends on partition count: with few partitions, `take`'s SIMD and bulk null handling likely win. With many partitions, each gets only a handful of rows per batch, so you'd accumulate thousands of tiny batches and `concat_batches` over many small inputs has real overhead. In that regime the builder approach avoids intermediate allocations and is probably better. Not a blocker for this PR, but as a followup it might be worth benchmarking `take` + `concat_batches` against the builder scatter for the partition counts where immediate mode is expected to be used. If `take` wins there, it would simplify the code considerably. -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
