You are viewing a plain text version of this content. The canonical link for it is here.
Posted to commits@arrow.apache.org by yj...@apache.org on 2023/05/01 14:14:12 UTC
[arrow-datafusion] branch main updated: feat: make threshold for using scalar update in aggregate configurable (#6166)
This is an automated email from the ASF dual-hosted git repository.
yjshen pushed a commit to branch main
in repository https://gitbox.apache.org/repos/asf/arrow-datafusion.git
The following commit(s) were added to refs/heads/main by this push:
new 6e8f91b41a feat: make threshold for using scalar update in aggregate configurable (#6166)
6e8f91b41a is described below
commit 6e8f91b41a6ec6a2680357b95b2489d87af33571
Author: Yijie Shen <he...@gmail.com>
AuthorDate: Mon May 1 22:14:07 2023 +0800
feat: make threshold for using scalar update in aggregate configurable (#6166)
* feat: make threshold for using scalar update in aggregate configurable
* fix
* fix test
---
datafusion/common/src/config.rs | 20 ++++++
.../aggregates/bounded_aggregate_stream.rs | 7 +-
.../core/src/physical_plan/aggregates/mod.rs | 3 +
.../core/src/physical_plan/aggregates/row_hash.rs | 7 +-
.../test_files/information_schema.slt | 1 +
datafusion/execution/src/config.rs | 13 ++++
docs/source/user-guide/configs.md | 79 +++++++++++-----------
7 files changed, 89 insertions(+), 41 deletions(-)
diff --git a/datafusion/common/src/config.rs b/datafusion/common/src/config.rs
index 35afabf307..a89bc2f8fd 100644
--- a/datafusion/common/src/config.rs
+++ b/datafusion/common/src/config.rs
@@ -223,6 +223,9 @@ config_namespace! {
/// Parquet options
pub parquet: ParquetOptions, default = Default::default()
+
+ /// Aggregate options
+ pub aggregate: AggregateOptions, default = Default::default()
}
}
@@ -260,6 +263,23 @@ config_namespace! {
}
}
+config_namespace! {
+ /// Options related to aggregate execution
+ pub struct AggregateOptions {
+ /// Specifies the threshold for using `ScalarValue`s to update
+ /// accumulators during high-cardinality aggregations for each input batch.
+ ///
+ /// The aggregation is considered high-cardinality if the number of affected groups
+ /// is greater than or equal to `batch_size / scalar_update_factor`. In such cases,
+ /// `ScalarValue`s are utilized for updating accumulators, rather than the default
+ /// batch-slice approach. This can lead to performance improvements.
+ ///
+ /// By adjusting the `scalar_update_factor`, you can balance the trade-off between
+ /// more efficient accumulator updates and the number of groups affected.
+ pub scalar_update_factor: usize, default = 10
+ }
+}
+
config_namespace! {
/// Options related to query optimization
pub struct OptimizerOptions {
diff --git a/datafusion/core/src/physical_plan/aggregates/bounded_aggregate_stream.rs b/datafusion/core/src/physical_plan/aggregates/bounded_aggregate_stream.rs
index a2ee627eec..41eca116bd 100644
--- a/datafusion/core/src/physical_plan/aggregates/bounded_aggregate_stream.rs
+++ b/datafusion/core/src/physical_plan/aggregates/bounded_aggregate_stream.rs
@@ -103,6 +103,9 @@ pub(crate) struct BoundedAggregateStream {
random_state: RandomState,
/// size to be used for resulting RecordBatches
batch_size: usize,
+ /// threshold for using `ScalarValue`s to update
+ /// accumulators during high-cardinality aggregations for each input batch.
+ scalar_update_factor: usize,
/// if the result is chunked into batches,
/// last offset is preserved for continuation.
row_group_skip_position: usize,
@@ -126,6 +129,7 @@ impl BoundedAggregateStream {
input: SendableRecordBatchStream,
baseline_metrics: BaselineMetrics,
batch_size: usize,
+ scalar_update_factor: usize,
context: Arc<TaskContext>,
partition: usize,
// Stores algorithm mode and output ordering
@@ -228,6 +232,7 @@ impl BoundedAggregateStream {
baseline_metrics,
random_state: Default::default(),
batch_size,
+ scalar_update_factor,
row_group_skip_position: 0,
indices: [normal_agg_indices, row_agg_indices],
is_end: false,
@@ -747,7 +752,7 @@ impl BoundedAggregateStream {
if matches!(self.mode, AggregateMode::Partial | AggregateMode::Single)
&& normal_aggr_input_values.is_empty()
&& normal_filter_values.is_empty()
- && groups_with_rows.len() >= batch.num_rows() / 10
+ && groups_with_rows.len() >= batch.num_rows() / self.scalar_update_factor
{
self.update_accumulators_using_scalar(
&groups_with_rows,
diff --git a/datafusion/core/src/physical_plan/aggregates/mod.rs b/datafusion/core/src/physical_plan/aggregates/mod.rs
index e1f0bfb752..5c77f3b74a 100644
--- a/datafusion/core/src/physical_plan/aggregates/mod.rs
+++ b/datafusion/core/src/physical_plan/aggregates/mod.rs
@@ -424,6 +424,7 @@ impl AggregateExec {
context: Arc<TaskContext>,
) -> Result<StreamType> {
let batch_size = context.session_config().batch_size();
+ let scalar_update_factor = context.session_config().agg_scalar_update_factor();
let input = self.input.execute(partition, Arc::clone(&context))?;
let baseline_metrics = BaselineMetrics::new(&self.metrics, partition);
@@ -448,6 +449,7 @@ impl AggregateExec {
input,
baseline_metrics,
batch_size,
+ scalar_update_factor,
context,
partition,
aggregation_ordering.clone(),
@@ -463,6 +465,7 @@ impl AggregateExec {
input,
baseline_metrics,
batch_size,
+ scalar_update_factor,
context,
partition,
)?,
diff --git a/datafusion/core/src/physical_plan/aggregates/row_hash.rs b/datafusion/core/src/physical_plan/aggregates/row_hash.rs
index 3f42a0171b..0b98513cce 100644
--- a/datafusion/core/src/physical_plan/aggregates/row_hash.rs
+++ b/datafusion/core/src/physical_plan/aggregates/row_hash.rs
@@ -98,6 +98,9 @@ pub(crate) struct GroupedHashAggregateStream {
random_state: RandomState,
/// size to be used for resulting RecordBatches
batch_size: usize,
+ /// threshold for using `ScalarValue`s to update
+ /// accumulators during high-cardinality aggregations for each input batch.
+ scalar_update_factor: usize,
/// if the result is chunked into batches,
/// last offset is preserved for continuation.
row_group_skip_position: usize,
@@ -119,6 +122,7 @@ impl GroupedHashAggregateStream {
input: SendableRecordBatchStream,
baseline_metrics: BaselineMetrics,
batch_size: usize,
+ scalar_update_factor: usize,
context: Arc<TaskContext>,
partition: usize,
) -> Result<Self> {
@@ -219,6 +223,7 @@ impl GroupedHashAggregateStream {
baseline_metrics,
random_state: Default::default(),
batch_size,
+ scalar_update_factor,
row_group_skip_position: 0,
indices: [normal_agg_indices, row_agg_indices],
})
@@ -555,7 +560,7 @@ impl GroupedHashAggregateStream {
if matches!(self.mode, AggregateMode::Partial | AggregateMode::Single)
&& normal_aggr_input_values.is_empty()
&& normal_filter_values.is_empty()
- && groups_with_rows.len() >= batch.num_rows() / 10
+ && groups_with_rows.len() >= batch.num_rows() / self.scalar_update_factor
{
self.update_accumulators_using_scalar(
&groups_with_rows,
diff --git a/datafusion/core/tests/sqllogictests/test_files/information_schema.slt b/datafusion/core/tests/sqllogictests/test_files/information_schema.slt
index 3ea42583c8..290f7182c8 100644
--- a/datafusion/core/tests/sqllogictests/test_files/information_schema.slt
+++ b/datafusion/core/tests/sqllogictests/test_files/information_schema.slt
@@ -136,6 +136,7 @@ datafusion.catalog.format NULL
datafusion.catalog.has_header false
datafusion.catalog.information_schema true
datafusion.catalog.location NULL
+datafusion.execution.aggregate.scalar_update_factor 10
datafusion.execution.batch_size 8192
datafusion.execution.coalesce_batches true
datafusion.execution.collect_statistics false
diff --git a/datafusion/execution/src/config.rs b/datafusion/execution/src/config.rs
index 730a767826..2867e7bd7c 100644
--- a/datafusion/execution/src/config.rs
+++ b/datafusion/execution/src/config.rs
@@ -237,6 +237,19 @@ impl SessionConfig {
self.options.execution.batch_size
}
+ /// Get the currently configured scalar_update_factor for aggregate
+ pub fn agg_scalar_update_factor(&self) -> usize {
+ self.options.execution.aggregate.scalar_update_factor
+ }
+
+ /// Customize scalar_update_factor for aggregate
+ pub fn with_agg_scalar_update_factor(mut self, n: usize) -> Self {
+ // scalar update factor must be greater than zero
+ assert!(n > 0);
+ self.options.execution.aggregate.scalar_update_factor = n;
+ self
+ }
+
/// Convert configuration options to name-value pairs with values
/// converted to strings.
///
diff --git a/docs/source/user-guide/configs.md b/docs/source/user-guide/configs.md
index 4b754bcd8a..decab6a719 100644
--- a/docs/source/user-guide/configs.md
+++ b/docs/source/user-guide/configs.md
@@ -35,42 +35,43 @@ Values are parsed according to the [same rules used in casts from Utf8](https://
If the value in the environment variable cannot be cast to the type of the configuration option, the default value will be used instead and a warning emitted.
Environment variables are read during `SessionConfig` initialisation so they must be set beforehand and will not affect running sessions.
-| key | default | description [...]
-| ---------------------------------------------------------- | ---------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- [...]
-| datafusion.catalog.create_default_catalog_and_schema | true | Whether the default catalog and schema should be created automatically. [...]
-| datafusion.catalog.default_catalog | datafusion | The default catalog name - this impacts what SQL queries use if not specified [...]
-| datafusion.catalog.default_schema | public | The default schema name - this impacts what SQL queries use if not specified [...]
-| datafusion.catalog.information_schema | false | Should DataFusion provide access to `information_schema` virtual tables for displaying schema information [...]
-| datafusion.catalog.location | NULL | Location scanned to load tables for `default` schema [...]
-| datafusion.catalog.format | NULL | Type of `TableProvider` to use when loading `default` schema [...]
-| datafusion.catalog.has_header | false | If the file has a header [...]
-| datafusion.execution.batch_size | 8192 | Default batch size while creating new batches, it's especially useful for buffer-in-memory batches since creating tiny batches would result in too much metadata memory consumption [...]
-| datafusion.execution.coalesce_batches | true | When set to true, record batches will be examined between each operator and small batches will be coalesced into larger batches. This is helpful when there are highly selective filters or joins that could produce tiny output batches. The target batch size is determined by the configuration setting [...]
-| datafusion.execution.collect_statistics | false | Should DataFusion collect statistics after listing files [...]
-| datafusion.execution.target_partitions | 0 | Number of partitions for query execution. Increasing partitions can increase concurrency. Defaults to the number of CPU cores on the system [...]
-| datafusion.execution.time_zone | +00:00 | The default time zone Some functions, e.g. `EXTRACT(HOUR from SOME_TIME)`, shift the underlying datetime according to this time zone, and then extract the hour [...]
-| datafusion.execution.parquet.enable_page_index | false | If true, uses parquet data page level metadata (Page Index) statistics to reduce the number of rows decoded. [...]
-| datafusion.execution.parquet.pruning | true | If true, the parquet reader attempts to skip entire row groups based on the predicate in the query and the metadata (min/max values) stored in the parquet file [...]
-| datafusion.execution.parquet.skip_metadata | true | If true, the parquet reader skip the optional embedded metadata that may be in the file Schema. This setting can help avoid schema conflicts when querying multiple parquet files with schemas containing compatible types but different metadata [...]
-| datafusion.execution.parquet.metadata_size_hint | NULL | If specified, the parquet reader will try and fetch the last `size_hint` bytes of the parquet file optimistically. If not specified, two reads are required: One read to fetch the 8-byte parquet footer and another to fetch the metadata length encoded in the footer [...]
-| datafusion.execution.parquet.pushdown_filters | false | If true, filter expressions are be applied during the parquet decoding operation to reduce the number of rows decoded [...]
-| datafusion.execution.parquet.reorder_filters | false | If true, filter expressions evaluated during the parquet decoding operation will be reordered heuristically to minimize the cost of evaluation. If false, the filters are applied in the same order as written in the query [...]
-| datafusion.optimizer.enable_round_robin_repartition | true | When set to true, the physical plan optimizer will try to add round robin repartitioning to increase parallelism to leverage more CPU cores [...]
-| datafusion.optimizer.filter_null_join_keys | false | When set to true, the optimizer will insert filters before a join between a nullable and non-nullable column to filter out nulls on the nullable side. This filter can add additional overhead when the file format does not fully support predicate push down. [...]
-| datafusion.optimizer.repartition_aggregations | true | Should DataFusion repartition data using the aggregate keys to execute aggregates in parallel using the provided `target_partitions` level [...]
-| datafusion.optimizer.repartition_file_min_size | 10485760 | Minimum total files size in bytes to perform file scan repartitioning. [...]
-| datafusion.optimizer.repartition_joins | true | Should DataFusion repartition data using the join keys to execute joins in parallel using the provided `target_partitions` level [...]
-| datafusion.optimizer.allow_symmetric_joins_without_pruning | true | Should DataFusion allow symmetric hash joins for unbounded data sources even when its inputs do not have any ordering or filtering If the flag is not enabled, the SymmetricHashJoin operator will be unable to prune its internal buffers, resulting in certain join types - such as Full, Left, LeftAnti, LeftSemi, Right, RightAnti, and RightSemi - being produced only at the end of the execution. This is not typical in [...]
-| datafusion.optimizer.repartition_file_scans | true | When set to true, file groups will be repartitioned to achieve maximum parallelism. Currently supported only for Parquet format in which case multiple row groups from the same file may be read concurrently. If false then each row group is read serially, though different files may be read in parallel. [...]
-| datafusion.optimizer.repartition_windows | true | Should DataFusion repartition data using the partitions keys to execute window functions in parallel using the provided `target_partitions` level [...]
-| datafusion.optimizer.repartition_sorts | true | Should DataFusion execute sorts in a per-partition fashion and merge afterwards instead of coalescing first and sorting globally. With this flag is enabled, plans in the form below `text "SortExec: [a@0 ASC]", " CoalescePartitionsExec", " RepartitionExec: partitioning=RoundRobinBatch(8), input_partitions=1", ` would turn into the plan below which performs better in multithreaded environments `text "SortPreserving [...]
-| datafusion.optimizer.skip_failed_rules | true | When set to true, the logical plan optimizer will produce warning messages if any optimization rules produce errors and then proceed to the next rule. When set to false, any rules that produce errors will cause the query to fail [...]
-| datafusion.optimizer.max_passes | 3 | Number of times that the optimizer will attempt to optimize the plan [...]
-| datafusion.optimizer.top_down_join_key_reordering | true | When set to true, the physical plan optimizer will run a top down process to reorder the join keys [...]
-| datafusion.optimizer.prefer_hash_join | true | When set to true, the physical plan optimizer will prefer HashJoin over SortMergeJoin. HashJoin can work more efficiently than SortMergeJoin but consumes more memory [...]
-| datafusion.optimizer.hash_join_single_partition_threshold | 1048576 | The maximum estimated size in bytes for one input side of a HashJoin will be collected into a single partition [...]
-| datafusion.explain.logical_plan_only | false | When set to true, the explain statement will only print logical plans [...]
-| datafusion.explain.physical_plan_only | false | When set to true, the explain statement will only print physical plans [...]
-| datafusion.sql_parser.parse_float_as_decimal | false | When set to true, SQL parser will parse float as decimal type [...]
-| datafusion.sql_parser.enable_ident_normalization | true | When set to true, SQL parser will normalize ident (convert ident to lowercase when not quoted) [...]
-| datafusion.sql_parser.dialect | generic | Configure the SQL dialect used by DataFusion's parser; supported values include: Generic, MySQL, PostgreSQL, Hive, SQLite, Snowflake, Redshift, MsSQL, ClickHouse, BigQuery, and Ansi. [...]
+| key | default | description [...]
+| ---------------------------------------------------------- | ---------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- [...]
+| datafusion.catalog.create_default_catalog_and_schema | true | Whether the default catalog and schema should be created automatically. [...]
+| datafusion.catalog.default_catalog | datafusion | The default catalog name - this impacts what SQL queries use if not specified [...]
+| datafusion.catalog.default_schema | public | The default schema name - this impacts what SQL queries use if not specified [...]
+| datafusion.catalog.information_schema | false | Should DataFusion provide access to `information_schema` virtual tables for displaying schema information [...]
+| datafusion.catalog.location | NULL | Location scanned to load tables for `default` schema [...]
+| datafusion.catalog.format | NULL | Type of `TableProvider` to use when loading `default` schema [...]
+| datafusion.catalog.has_header | false | If the file has a header [...]
+| datafusion.execution.batch_size | 8192 | Default batch size while creating new batches, it's especially useful for buffer-in-memory batches since creating tiny batches would result in too much metadata memory consumption [...]
+| datafusion.execution.coalesce_batches | true | When set to true, record batches will be examined between each operator and small batches will be coalesced into larger batches. This is helpful when there are highly selective filters or joins that could produce tiny output batches. The target batch size is determined by the configuration setting [...]
+| datafusion.execution.collect_statistics | false | Should DataFusion collect statistics after listing files [...]
+| datafusion.execution.target_partitions | 0 | Number of partitions for query execution. Increasing partitions can increase concurrency. Defaults to the number of CPU cores on the system [...]
+| datafusion.execution.time_zone | +00:00 | The default time zone Some functions, e.g. `EXTRACT(HOUR from SOME_TIME)`, shift the underlying datetime according to this time zone, and then extract the hour [...]
+| datafusion.execution.parquet.enable_page_index | false | If true, uses parquet data page level metadata (Page Index) statistics to reduce the number of rows decoded. [...]
+| datafusion.execution.parquet.pruning | true | If true, the parquet reader attempts to skip entire row groups based on the predicate in the query and the metadata (min/max values) stored in the parquet file [...]
+| datafusion.execution.parquet.skip_metadata | true | If true, the parquet reader skip the optional embedded metadata that may be in the file Schema. This setting can help avoid schema conflicts when querying multiple parquet files with schemas containing compatible types but different metadata [...]
+| datafusion.execution.parquet.metadata_size_hint | NULL | If specified, the parquet reader will try and fetch the last `size_hint` bytes of the parquet file optimistically. If not specified, two reads are required: One read to fetch the 8-byte parquet footer and another to fetch the metadata length encoded in the footer [...]
+| datafusion.execution.parquet.pushdown_filters | false | If true, filter expressions are be applied during the parquet decoding operation to reduce the number of rows decoded [...]
+| datafusion.execution.parquet.reorder_filters | false | If true, filter expressions evaluated during the parquet decoding operation will be reordered heuristically to minimize the cost of evaluation. If false, the filters are applied in the same order as written in the query [...]
+| datafusion.execution.aggregate.scalar_update_factor | 10 | Specifies the threshold for using `ScalarValue`s to update accumulators during high-cardinality aggregations for each input batch. The aggregation is considered high-cardinality if the number of affected groups is greater than or equal to `batch_size / scalar_update_factor`. In such cases, `ScalarValue`s are utilized for updating accumulators, rather than the default batch-slice approach. This can lead to perform [...]
+| datafusion.optimizer.enable_round_robin_repartition | true | When set to true, the physical plan optimizer will try to add round robin repartitioning to increase parallelism to leverage more CPU cores [...]
+| datafusion.optimizer.filter_null_join_keys | false | When set to true, the optimizer will insert filters before a join between a nullable and non-nullable column to filter out nulls on the nullable side. This filter can add additional overhead when the file format does not fully support predicate push down. [...]
+| datafusion.optimizer.repartition_aggregations | true | Should DataFusion repartition data using the aggregate keys to execute aggregates in parallel using the provided `target_partitions` level [...]
+| datafusion.optimizer.repartition_file_min_size | 10485760 | Minimum total files size in bytes to perform file scan repartitioning. [...]
+| datafusion.optimizer.repartition_joins | true | Should DataFusion repartition data using the join keys to execute joins in parallel using the provided `target_partitions` level [...]
+| datafusion.optimizer.allow_symmetric_joins_without_pruning | true | Should DataFusion allow symmetric hash joins for unbounded data sources even when its inputs do not have any ordering or filtering If the flag is not enabled, the SymmetricHashJoin operator will be unable to prune its internal buffers, resulting in certain join types - such as Full, Left, LeftAnti, LeftSemi, Right, RightAnti, and RightSemi - being produced only at the end of the execution. This is not typical in [...]
+| datafusion.optimizer.repartition_file_scans | true | When set to true, file groups will be repartitioned to achieve maximum parallelism. Currently supported only for Parquet format in which case multiple row groups from the same file may be read concurrently. If false then each row group is read serially, though different files may be read in parallel. [...]
+| datafusion.optimizer.repartition_windows | true | Should DataFusion repartition data using the partitions keys to execute window functions in parallel using the provided `target_partitions` level [...]
+| datafusion.optimizer.repartition_sorts | true | Should DataFusion execute sorts in a per-partition fashion and merge afterwards instead of coalescing first and sorting globally. With this flag is enabled, plans in the form below `text "SortExec: [a@0 ASC]", " CoalescePartitionsExec", " RepartitionExec: partitioning=RoundRobinBatch(8), input_partitions=1", ` would turn into the plan below which performs better in multithreaded environments `text "SortPreserving [...]
+| datafusion.optimizer.skip_failed_rules | true | When set to true, the logical plan optimizer will produce warning messages if any optimization rules produce errors and then proceed to the next rule. When set to false, any rules that produce errors will cause the query to fail [...]
+| datafusion.optimizer.max_passes | 3 | Number of times that the optimizer will attempt to optimize the plan [...]
+| datafusion.optimizer.top_down_join_key_reordering | true | When set to true, the physical plan optimizer will run a top down process to reorder the join keys [...]
+| datafusion.optimizer.prefer_hash_join | true | When set to true, the physical plan optimizer will prefer HashJoin over SortMergeJoin. HashJoin can work more efficiently than SortMergeJoin but consumes more memory [...]
+| datafusion.optimizer.hash_join_single_partition_threshold | 1048576 | The maximum estimated size in bytes for one input side of a HashJoin will be collected into a single partition [...]
+| datafusion.explain.logical_plan_only | false | When set to true, the explain statement will only print logical plans [...]
+| datafusion.explain.physical_plan_only | false | When set to true, the explain statement will only print physical plans [...]
+| datafusion.sql_parser.parse_float_as_decimal | false | When set to true, SQL parser will parse float as decimal type [...]
+| datafusion.sql_parser.enable_ident_normalization | true | When set to true, SQL parser will normalize ident (convert ident to lowercase when not quoted) [...]
+| datafusion.sql_parser.dialect | generic | Configure the SQL dialect used by DataFusion's parser; supported values include: Generic, MySQL, PostgreSQL, Hive, SQLite, Snowflake, Redshift, MsSQL, ClickHouse, BigQuery, and Ansi. [...]