avantgardnerio commented on code in PR #7192: URL: https://github.com/apache/arrow-datafusion/pull/7192#discussion_r1298374221
########## datafusion/core/src/physical_plan/aggregates/priority_map.rs: ########## @@ -0,0 +1,820 @@ +// 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. + +//! A memory-conscious aggregation implementation that limits group buckets to a fixed number + +use crate::physical_plan::aggregates::{ + aggregate_expressions, evaluate_group_by, evaluate_many, AggregateExec, + PhysicalGroupBy, +}; +use crate::physical_plan::{RecordBatchStream, SendableRecordBatchStream}; +use ahash::RandomState; +use arrow::util::pretty::print_batches; +use arrow_array::cast::AsArray; +use arrow_array::downcast_primitive; +use arrow_array::{ + Array, ArrayRef, ArrowNativeTypeOp, ArrowPrimitiveType, PrimitiveArray, RecordBatch, + StringArray, +}; +use arrow_schema::{DataType, SchemaRef, SortOptions}; +use datafusion_common::DataFusionError; +use datafusion_common::Result; +use datafusion_execution::TaskContext; +use datafusion_physical_expr::PhysicalExpr; +use futures::stream::{Stream, StreamExt}; +use hashbrown::raw::RawTable; +use log::{trace, Level}; +use std::cmp::Ordering; +use std::collections::BTreeSet; +use std::hash::{BuildHasher, Hash, Hasher}; +use std::pin::Pin; +use std::sync::Arc; +use std::task::{Context, Poll}; + +pub struct GroupedTopKAggregateStream { + partition: usize, + row_count: usize, + started: bool, + schema: SchemaRef, + input: SendableRecordBatchStream, + aggregate_arguments: Vec<Vec<Arc<dyn PhysicalExpr>>>, + group_by: PhysicalGroupBy, + aggregator: Box<dyn LimitedAggregator>, +} + +impl GroupedTopKAggregateStream { + pub fn new( + agg: &AggregateExec, + context: Arc<TaskContext>, + partition: usize, + limit: usize, + ) -> Result<Self> { + let agg_schema = Arc::clone(&agg.schema); + let group_by = agg.group_by.clone(); + + let input = agg.input.execute(partition, Arc::clone(&context))?; + + let aggregate_arguments = + aggregate_expressions(&agg.aggr_expr, &agg.mode, group_by.expr.len())?; + + let (val_field, _) = agg + .get_minmax_desc() + .ok_or_else(|| DataFusionError::Execution("Min/max required".to_string()))?; + + let vt = val_field.data_type().clone(); + let ag = new_group_values(limit, agg.sort, vt)?; + + Ok(GroupedTopKAggregateStream { + partition, + started: false, + row_count: 0, + schema: agg_schema, + input, + aggregate_arguments, + group_by, + aggregator: ag, + }) + } +} + +pub fn new_group_values( + limit: usize, + sort: SortOptions, + vt: DataType, +) -> Result<Box<dyn LimitedAggregator>> { + macro_rules! downcast_helper { + ($vt:ty, $d:ident) => { + return Ok(Box::new(PrimitiveAggregator::<$vt>::new( + limit, + limit * 10, + sort, + ))) + }; + } + + downcast_primitive! { + vt => (downcast_helper, vt), + _ => {} + } + + Err(DataFusionError::Execution(format!( + "Can't group type: {vt:?}" + ))) +} + +impl RecordBatchStream for GroupedTopKAggregateStream { + fn schema(&self) -> SchemaRef { + self.schema.clone() + } +} + +pub trait LimitedAggregator: Send { + fn intern(&mut self, ids: ArrayRef, vals: ArrayRef) -> Result<()>; + fn emit(&mut self) -> Result<Vec<ArrayRef>>; + fn is_empty(&self) -> bool; +} + +pub trait ValueType: ArrowNativeTypeOp + Clone {} + +impl<T> ValueType for T where T: ArrowNativeTypeOp + Clone {} + +pub trait KeyType: Clone + Eq + Hash {} + +impl<T> KeyType for T where T: Clone + Eq + Hash {} + +struct PrimitiveAggregator<VAL: ArrowPrimitiveType> +where + <VAL as ArrowPrimitiveType>::Native: Clone, +{ + priority_map: PriorityMap<Option<String>, VAL::Native>, +} + +impl<VAL: ArrowPrimitiveType> PrimitiveAggregator<VAL> +where + <VAL as ArrowPrimitiveType>::Native: Clone, +{ + pub fn new(limit: usize, capacity: usize, sort: SortOptions) -> Self { + Self { + priority_map: PriorityMap::new(limit, capacity, sort), + } + } +} + +unsafe impl<VAL: ArrowPrimitiveType> Send for PrimitiveAggregator<VAL> where + <VAL as ArrowPrimitiveType>::Native: Clone +{ +} + +impl<VAL: ArrowPrimitiveType> LimitedAggregator for PrimitiveAggregator<VAL> +where + <VAL as ArrowPrimitiveType>::Native: Clone, +{ + fn intern(&mut self, ids: ArrayRef, vals: ArrayRef) -> Result<()> { + let ids = ids.as_any().downcast_ref::<StringArray>().ok_or_else(|| { + DataFusionError::Execution("Expected StringArray".to_string()) + })?; + let sort = self.priority_map.sort; + let vals = vals.as_primitive::<VAL>(); + for row_idx in 0..ids.len() { + let val = if vals.is_null(row_idx) { + None + } else { + Some(vals.value(row_idx)) + }; + let id = if ids.is_null(row_idx) { + None + } else { + // Check goes here, because it is generalizable between str/String and Row/OwnedRow + let id = ids.value(row_idx); + if let Some(worst) = self.priority_map.worst_val() { + if cmp(&sort, &val, worst) == Ordering::Greater { + continue; + } + } + Some(id.to_string()) + }; + + self.priority_map.insert(id, val)?; + } + Ok(()) + } + + fn emit(&mut self) -> Result<Vec<ArrayRef>> { + let (keys, vals): (Vec<_>, Vec<_>) = + self.priority_map.drain().into_iter().unzip(); + let keys = Arc::new(StringArray::from(keys)); + let vals = Arc::new(PrimitiveArray::<VAL>::from_iter(vals)); + Ok(vec![keys, vals]) + } + + fn is_empty(&self) -> bool { + self.priority_map.is_empty() + } +} + +impl Stream for GroupedTopKAggregateStream { + type Item = Result<RecordBatch>; + + fn poll_next( + mut self: Pin<&mut Self>, + cx: &mut Context<'_>, + ) -> Poll<Option<Self::Item>> { + while let Poll::Ready(res) = self.input.poll_next_unpin(cx) { + match res { + // got a batch, convert to rows and append to our TreeMap + Some(Ok(batch)) => { + self.started = true; + trace!( + "partition {} has {} rows and got batch with {} rows", + self.partition, + self.row_count, + batch.num_rows() + ); + if log::log_enabled!(Level::Trace) && batch.num_rows() < 20 { + print_batches(&[batch.clone()])?; + } + self.row_count += batch.num_rows(); + let batches = &[batch]; + let group_by_values = + evaluate_group_by(&self.group_by, batches.first().unwrap())?; + let group_by_values = + group_by_values.into_iter().last().expect("values"); + let group_by_values = + group_by_values.into_iter().last().expect("values"); + let input_values = evaluate_many( + &self.aggregate_arguments, + batches.first().unwrap(), + )?; + let input_values = match input_values.as_slice() { + [] => { + Err(DataFusionError::Execution("vals required".to_string()))? + } + [vals] => vals, + _ => { + Err(DataFusionError::Execution("1 val required".to_string()))? + } + }; + let input_values = match input_values.as_slice() { + [] => { + Err(DataFusionError::Execution("vals required".to_string()))? + } + [vals] => vals, + _ => { + Err(DataFusionError::Execution("1 val required".to_string()))? + } + } + .clone(); + + // iterate over each column of group_by values + (*self.aggregator).intern(group_by_values, input_values)?; + } + // inner is done, emit all rows and switch to producing output + None => { + if self.aggregator.is_empty() { + trace!("partition {} emit None", self.partition); + return Poll::Ready(None); + } + let cols = self.aggregator.emit()?; + let batch = RecordBatch::try_new(self.schema.clone(), cols)?; + trace!( + "partition {} emit batch with {} rows", + self.partition, + batch.num_rows() + ); + if log::log_enabled!(Level::Trace) { + print_batches(&[batch.clone()])?; + } + return Poll::Ready(Some(Ok(batch))); + } + // inner had error, return to caller + Some(Err(e)) => { + return Poll::Ready(Some(Err(e))); + } + } + } + Poll::Pending + } +} + +pub struct PriorityMap<ID: KeyType, VAL: ValueType> { + limit: usize, + sort: SortOptions, + rnd: RandomState, + id_to_val: RawTable<MapItem<ID, VAL>>, + val_to_idx: BTreeSet<HeapItem<VAL>>, +} + +pub struct MapItem<ID: KeyType, VAL: ValueType> { + hash: u64, + pub id: ID, + pub val: Option<VAL>, +} + +impl<ID: KeyType, VAL: ValueType> MapItem<ID, VAL> { + pub fn new(hash: u64, id: ID, val: Option<VAL>) -> Self { + Self { hash, id, val } + } +} + +struct HeapItem<VAL: ValueType> { + val: Option<VAL>, + buk_idx: usize, + sort: SortOptions, +} + +impl<VAL: ValueType> HeapItem<VAL> { + pub fn new(val: Option<VAL>, buk_idx: usize, sort: SortOptions) -> Self { + Self { val, buk_idx, sort } + } +} + +impl<VAL: ValueType> Eq for HeapItem<VAL> {} + +impl<VAL: ValueType> PartialEq<Self> for HeapItem<VAL> { + fn eq(&self, other: &Self) -> bool { + self.cmp(other) == Ordering::Equal + } +} + +impl<VAL: ValueType> PartialOrd<Self> for HeapItem<VAL> { + fn partial_cmp(&self, other: &Self) -> Option<Ordering> { + Some(self.cmp(other)) + } +} + +impl<VAL: ValueType> Ord for HeapItem<VAL> { + fn cmp(&self, other: &Self) -> Ordering { + let res = cmp(&self.sort, &self.val, &other.val); + if res != Ordering::Equal { + return res; + } + self.buk_idx.cmp(&other.buk_idx) + } +} + +impl<ID: KeyType, VAL: ValueType> PriorityMap<ID, VAL> +where + VAL: PartialEq<VAL>, +{ + pub fn new(limit: usize, capacity: usize, sort: SortOptions) -> Self { + Self { + limit, + sort, + rnd: Default::default(), + id_to_val: RawTable::with_capacity(capacity), + val_to_idx: Default::default(), + } + } + + pub fn insert(&mut self, new_id: ID, new_val: Option<VAL>) -> Result<()> { + let is_full = self.is_full(); + assert!(self.len() <= self.limit, "Overflow"); + let (id_to_val, val_to_idx) = (&mut self.id_to_val, &mut self.val_to_idx); + + // if we're full, and the new val is worse than all our values, just bail + if is_full { + let worst_entry = val_to_idx.last().expect("Missing value!"); + if cmp(&self.sort, &new_val, &worst_entry.val) == Ordering::Greater { + return Ok(()); + } + } + + // handle new groups we haven't seen yet + let mut hasher = self.rnd.build_hasher(); + new_id.hash(&mut hasher); + let new_hash = hasher.finish(); + let old_bucket = match id_to_val.find(new_hash, |mi| new_id == mi.id) { + None => { + // we're full and this is a better value, so remove the worst + if is_full { + let worst_hi = val_to_idx.pop_last().expect("Missing value!"); + unsafe { id_to_val.erase(id_to_val.bucket(worst_hi.buk_idx)) }; Review Comment: 💯 % agree. I was struggling with where and how to best document this though, as I have 8 unsafe blocks in this file. The reason is that this was a significant performance boost (the ~15%) that got things on par with `main` and should allow the optimizer to be simple and just always use this rule when possible, vs having to do heuristics of some sort. This optimization takes advantage of the fact that Swiss tables (the implementation in `hashbrown`) don't re-index entries unless the `RawTable` is resized - this is a specialized case that suits us nicely, because we are always accumulating a fixed size (`limit`). Because the nature of this PR is implementing bi-directionally linked `Map` and `Heap`, this allows us to go from `Heap` -> `Map` with an index directly into the `RawTable`, rather than having to look it up by hash code, then index (which the profiler showed was how that ~15% was being spent). So in regards to safety, the reason they mark `bucket()` and `bucket_index()` unsafe is because it goes from `index` to memory pointer, and if that `index` is invalid because the `RawTable` resized, then memory corruption could occur. We catch this however by using `try_insert_no_grow()` and catching the error in case the `RawTable` decides to grow, and in that event we rebuild the heap (BTreeSet) so that we always have valid pointers. This operation is relatively expensive, but we allocate a higher fill factor: ``` return Ok(Box::new(PrimitiveAggregator::<$vt>::new( limit, limit * 10, sort, ))) ``` So that this is statistically unlikely. Maybe it will happen on a bad day with some keys that tend to collide, but an exponential resize or two will be amortized away, and even with larger pre-allocated tables (i.e. limit = 1000, 10x) we still use way less memory than when we were storing accumulators for every distinct group prior to this PR. ---- So what do I do with the description above? Just throw it at the top of the file in a doc section? No additional comments needed at the unsafe blocks themselves? ########## datafusion/core/src/physical_plan/aggregates/priority_map.rs: ########## @@ -0,0 +1,820 @@ +// 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. + +//! A memory-conscious aggregation implementation that limits group buckets to a fixed number + +use crate::physical_plan::aggregates::{ + aggregate_expressions, evaluate_group_by, evaluate_many, AggregateExec, + PhysicalGroupBy, +}; +use crate::physical_plan::{RecordBatchStream, SendableRecordBatchStream}; +use ahash::RandomState; +use arrow::util::pretty::print_batches; +use arrow_array::cast::AsArray; +use arrow_array::downcast_primitive; +use arrow_array::{ + Array, ArrayRef, ArrowNativeTypeOp, ArrowPrimitiveType, PrimitiveArray, RecordBatch, + StringArray, +}; +use arrow_schema::{DataType, SchemaRef, SortOptions}; +use datafusion_common::DataFusionError; +use datafusion_common::Result; +use datafusion_execution::TaskContext; +use datafusion_physical_expr::PhysicalExpr; +use futures::stream::{Stream, StreamExt}; +use hashbrown::raw::RawTable; +use log::{trace, Level}; +use std::cmp::Ordering; +use std::collections::BTreeSet; +use std::hash::{BuildHasher, Hash, Hasher}; +use std::pin::Pin; +use std::sync::Arc; +use std::task::{Context, Poll}; + +pub struct GroupedTopKAggregateStream { + partition: usize, + row_count: usize, + started: bool, + schema: SchemaRef, + input: SendableRecordBatchStream, + aggregate_arguments: Vec<Vec<Arc<dyn PhysicalExpr>>>, + group_by: PhysicalGroupBy, + aggregator: Box<dyn LimitedAggregator>, +} + +impl GroupedTopKAggregateStream { + pub fn new( + agg: &AggregateExec, + context: Arc<TaskContext>, + partition: usize, + limit: usize, + ) -> Result<Self> { + let agg_schema = Arc::clone(&agg.schema); + let group_by = agg.group_by.clone(); + + let input = agg.input.execute(partition, Arc::clone(&context))?; + + let aggregate_arguments = + aggregate_expressions(&agg.aggr_expr, &agg.mode, group_by.expr.len())?; + + let (val_field, _) = agg + .get_minmax_desc() + .ok_or_else(|| DataFusionError::Execution("Min/max required".to_string()))?; + + let vt = val_field.data_type().clone(); + let ag = new_group_values(limit, agg.sort, vt)?; + + Ok(GroupedTopKAggregateStream { + partition, + started: false, + row_count: 0, + schema: agg_schema, + input, + aggregate_arguments, + group_by, + aggregator: ag, + }) + } +} + +pub fn new_group_values( + limit: usize, + sort: SortOptions, + vt: DataType, +) -> Result<Box<dyn LimitedAggregator>> { + macro_rules! downcast_helper { + ($vt:ty, $d:ident) => { + return Ok(Box::new(PrimitiveAggregator::<$vt>::new( + limit, + limit * 10, + sort, + ))) + }; + } + + downcast_primitive! { + vt => (downcast_helper, vt), + _ => {} + } + + Err(DataFusionError::Execution(format!( + "Can't group type: {vt:?}" + ))) +} + +impl RecordBatchStream for GroupedTopKAggregateStream { + fn schema(&self) -> SchemaRef { + self.schema.clone() + } +} + +pub trait LimitedAggregator: Send { + fn intern(&mut self, ids: ArrayRef, vals: ArrayRef) -> Result<()>; + fn emit(&mut self) -> Result<Vec<ArrayRef>>; + fn is_empty(&self) -> bool; +} + +pub trait ValueType: ArrowNativeTypeOp + Clone {} + +impl<T> ValueType for T where T: ArrowNativeTypeOp + Clone {} + +pub trait KeyType: Clone + Eq + Hash {} + +impl<T> KeyType for T where T: Clone + Eq + Hash {} + +struct PrimitiveAggregator<VAL: ArrowPrimitiveType> +where + <VAL as ArrowPrimitiveType>::Native: Clone, +{ + priority_map: PriorityMap<Option<String>, VAL::Native>, +} + +impl<VAL: ArrowPrimitiveType> PrimitiveAggregator<VAL> +where + <VAL as ArrowPrimitiveType>::Native: Clone, +{ + pub fn new(limit: usize, capacity: usize, sort: SortOptions) -> Self { + Self { + priority_map: PriorityMap::new(limit, capacity, sort), + } + } +} + +unsafe impl<VAL: ArrowPrimitiveType> Send for PrimitiveAggregator<VAL> where + <VAL as ArrowPrimitiveType>::Native: Clone +{ +} + +impl<VAL: ArrowPrimitiveType> LimitedAggregator for PrimitiveAggregator<VAL> +where + <VAL as ArrowPrimitiveType>::Native: Clone, +{ + fn intern(&mut self, ids: ArrayRef, vals: ArrayRef) -> Result<()> { + let ids = ids.as_any().downcast_ref::<StringArray>().ok_or_else(|| { + DataFusionError::Execution("Expected StringArray".to_string()) + })?; + let sort = self.priority_map.sort; + let vals = vals.as_primitive::<VAL>(); + for row_idx in 0..ids.len() { + let val = if vals.is_null(row_idx) { + None + } else { + Some(vals.value(row_idx)) + }; + let id = if ids.is_null(row_idx) { + None + } else { + // Check goes here, because it is generalizable between str/String and Row/OwnedRow + let id = ids.value(row_idx); + if let Some(worst) = self.priority_map.worst_val() { + if cmp(&sort, &val, worst) == Ordering::Greater { + continue; + } + } + Some(id.to_string()) + }; + + self.priority_map.insert(id, val)?; + } + Ok(()) + } + + fn emit(&mut self) -> Result<Vec<ArrayRef>> { + let (keys, vals): (Vec<_>, Vec<_>) = + self.priority_map.drain().into_iter().unzip(); + let keys = Arc::new(StringArray::from(keys)); + let vals = Arc::new(PrimitiveArray::<VAL>::from_iter(vals)); + Ok(vec![keys, vals]) + } + + fn is_empty(&self) -> bool { + self.priority_map.is_empty() + } +} + +impl Stream for GroupedTopKAggregateStream { + type Item = Result<RecordBatch>; + + fn poll_next( + mut self: Pin<&mut Self>, + cx: &mut Context<'_>, + ) -> Poll<Option<Self::Item>> { + while let Poll::Ready(res) = self.input.poll_next_unpin(cx) { + match res { + // got a batch, convert to rows and append to our TreeMap + Some(Ok(batch)) => { + self.started = true; + trace!( + "partition {} has {} rows and got batch with {} rows", + self.partition, + self.row_count, + batch.num_rows() + ); + if log::log_enabled!(Level::Trace) && batch.num_rows() < 20 { + print_batches(&[batch.clone()])?; + } + self.row_count += batch.num_rows(); + let batches = &[batch]; + let group_by_values = + evaluate_group_by(&self.group_by, batches.first().unwrap())?; + let group_by_values = + group_by_values.into_iter().last().expect("values"); + let group_by_values = + group_by_values.into_iter().last().expect("values"); + let input_values = evaluate_many( + &self.aggregate_arguments, + batches.first().unwrap(), + )?; + let input_values = match input_values.as_slice() { + [] => { + Err(DataFusionError::Execution("vals required".to_string()))? + } + [vals] => vals, + _ => { + Err(DataFusionError::Execution("1 val required".to_string()))? + } + }; + let input_values = match input_values.as_slice() { + [] => { + Err(DataFusionError::Execution("vals required".to_string()))? + } + [vals] => vals, + _ => { + Err(DataFusionError::Execution("1 val required".to_string()))? + } + } + .clone(); + + // iterate over each column of group_by values + (*self.aggregator).intern(group_by_values, input_values)?; + } + // inner is done, emit all rows and switch to producing output + None => { + if self.aggregator.is_empty() { + trace!("partition {} emit None", self.partition); + return Poll::Ready(None); + } + let cols = self.aggregator.emit()?; + let batch = RecordBatch::try_new(self.schema.clone(), cols)?; + trace!( + "partition {} emit batch with {} rows", + self.partition, + batch.num_rows() + ); + if log::log_enabled!(Level::Trace) { + print_batches(&[batch.clone()])?; + } + return Poll::Ready(Some(Ok(batch))); + } + // inner had error, return to caller + Some(Err(e)) => { + return Poll::Ready(Some(Err(e))); + } + } + } + Poll::Pending + } +} + +pub struct PriorityMap<ID: KeyType, VAL: ValueType> { + limit: usize, + sort: SortOptions, + rnd: RandomState, + id_to_val: RawTable<MapItem<ID, VAL>>, + val_to_idx: BTreeSet<HeapItem<VAL>>, +} + +pub struct MapItem<ID: KeyType, VAL: ValueType> { + hash: u64, + pub id: ID, + pub val: Option<VAL>, +} + +impl<ID: KeyType, VAL: ValueType> MapItem<ID, VAL> { + pub fn new(hash: u64, id: ID, val: Option<VAL>) -> Self { + Self { hash, id, val } + } +} + +struct HeapItem<VAL: ValueType> { + val: Option<VAL>, + buk_idx: usize, + sort: SortOptions, +} + +impl<VAL: ValueType> HeapItem<VAL> { + pub fn new(val: Option<VAL>, buk_idx: usize, sort: SortOptions) -> Self { + Self { val, buk_idx, sort } + } +} + +impl<VAL: ValueType> Eq for HeapItem<VAL> {} + +impl<VAL: ValueType> PartialEq<Self> for HeapItem<VAL> { + fn eq(&self, other: &Self) -> bool { + self.cmp(other) == Ordering::Equal + } +} + +impl<VAL: ValueType> PartialOrd<Self> for HeapItem<VAL> { + fn partial_cmp(&self, other: &Self) -> Option<Ordering> { + Some(self.cmp(other)) + } +} + +impl<VAL: ValueType> Ord for HeapItem<VAL> { + fn cmp(&self, other: &Self) -> Ordering { + let res = cmp(&self.sort, &self.val, &other.val); + if res != Ordering::Equal { + return res; + } + self.buk_idx.cmp(&other.buk_idx) + } +} + +impl<ID: KeyType, VAL: ValueType> PriorityMap<ID, VAL> +where + VAL: PartialEq<VAL>, +{ + pub fn new(limit: usize, capacity: usize, sort: SortOptions) -> Self { + Self { + limit, + sort, + rnd: Default::default(), + id_to_val: RawTable::with_capacity(capacity), + val_to_idx: Default::default(), + } + } + + pub fn insert(&mut self, new_id: ID, new_val: Option<VAL>) -> Result<()> { + let is_full = self.is_full(); + assert!(self.len() <= self.limit, "Overflow"); + let (id_to_val, val_to_idx) = (&mut self.id_to_val, &mut self.val_to_idx); + + // if we're full, and the new val is worse than all our values, just bail + if is_full { + let worst_entry = val_to_idx.last().expect("Missing value!"); + if cmp(&self.sort, &new_val, &worst_entry.val) == Ordering::Greater { + return Ok(()); + } + } + + // handle new groups we haven't seen yet + let mut hasher = self.rnd.build_hasher(); + new_id.hash(&mut hasher); + let new_hash = hasher.finish(); + let old_bucket = match id_to_val.find(new_hash, |mi| new_id == mi.id) { + None => { + // we're full and this is a better value, so remove the worst + if is_full { + let worst_hi = val_to_idx.pop_last().expect("Missing value!"); + unsafe { id_to_val.erase(id_to_val.bucket(worst_hi.buk_idx)) }; Review Comment: 💯 % agree. I was struggling with where and how to best document this though, as I have 8 unsafe blocks in this file. The reason is that this was a significant performance boost (the ~15%) that got things on par with `main` and should allow the optimizer to be simple and just always use this rule when possible, vs having to do heuristics of some sort. This optimization takes advantage of the fact that Swiss tables (the implementation in `hashbrown`) don't re-index entries unless the `RawTable` is resized - this is a specialized case that suits us nicely, because we are always accumulating a fixed size (`limit`). Because the nature of this PR is implementing bi-directionally linked `Map` and `Heap`, this allows us to go from `Heap` -> `Map` with an index directly into the `RawTable`, rather than having to look it up by hash code, then index (which the profiler showed was how that ~15% was being spent). So in regards to safety, the reason they mark `bucket()` and `bucket_index()` unsafe is because it goes from `index` to memory pointer, and if that `index` is invalid because the `RawTable` resized, then memory corruption could occur. We catch this however by using `try_insert_no_grow()` and catching the error in case the `RawTable` decides to grow, and in that event we rebuild the heap (BTreeSet) so that we always have valid pointers. This operation is relatively expensive, but we allocate a higher fill factor: ``` return Ok(Box::new(PrimitiveAggregator::<$vt>::new( limit, limit * 10, sort, ))) ``` So that this is statistically unlikely. Maybe it will happen on a bad day with some keys that tend to collide, but an exponential resize or two will be amortized away, and even with larger pre-allocated tables (i.e. limit = 1000, 10x) we still use way less memory than when we were storing accumulators for every distinct group prior to this PR. ---- So what do I do with the description above? Just throw it at the top of the file in a doc section? No additional comments needed at the unsafe blocks themselves? -- 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]
