zhuqi-lucas commented on code in PR #186:
URL: https://github.com/apache/datafusion-site/pull/186#discussion_r3522646576
##########
content/blog/2026-07-05-sort-pushdown.md:
##########
@@ -0,0 +1,839 @@
+---
+layout: post
+title: Sort Pushdown in DataFusion: Skip Sorts, Skip Decode, Skip I/O
+date: 2026-07-05
+author: Qi Zhu
+categories: [performance]
+---
+
+<!--
+{% comment %}
+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.
+{% endcomment %}
+-->
+
+[TOC]
+
+*Qi Zhu, [Massive](https://www.massive.com/)*
+
+Many [Apache Parquet] datasets are already sorted on disk. Time-series
+files are usually written in ingestion-time order. Event logs are sharded
+and sorted by event id. Partitioned tables come with a natural ordering
+implied by the partition key. The information about that ordering is
+sitting right there in the file metadata.
+
+[Apache Parquet]: https://parquet.apache.org/
+
+[Apache DataFusion] has always been able to skip the sort in the
+*trivial* case: when the catalog declares an ordering (`WITH ORDER`
+or parquet `sorting_columns`) **and** the on-disk file listing
+already matches that order, the `EnsureRequirements` rule sees that
+the scan's `output_ordering` already satisfies the request and
+**removes the redundant `SortExec`**. The hard cases were
Review Comment:
Reframed as 'everything else — the messier real-world cases' with a bulleted
list.
##########
content/blog/2026-07-05-sort-pushdown.md:
##########
@@ -0,0 +1,839 @@
+---
+layout: post
+title: Sort Pushdown in DataFusion: Skip Sorts, Skip Decode, Skip I/O
+date: 2026-07-05
+author: Qi Zhu
+categories: [performance]
+---
+
+<!--
+{% comment %}
+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.
+{% endcomment %}
+-->
+
+[TOC]
+
+*Qi Zhu, [Massive](https://www.massive.com/)*
+
+Many [Apache Parquet] datasets are already sorted on disk. Time-series
+files are usually written in ingestion-time order. Event logs are sharded
+and sorted by event id. Partitioned tables come with a natural ordering
+implied by the partition key. The information about that ordering is
+sitting right there in the file metadata.
+
+[Apache Parquet]: https://parquet.apache.org/
+
+[Apache DataFusion] has always been able to skip the sort in the
+*trivial* case: when the catalog declares an ordering (`WITH ORDER`
+or parquet `sorting_columns`) **and** the on-disk file listing
+already matches that order, the `EnsureRequirements` rule sees that
+the scan's `output_ordering` already satisfies the request and
+**removes the redundant `SortExec`**. The hard cases were
+everything just slightly looser — files in the wrong on-disk order,
+declared but overlapping ranges, no declaration at all,
+`ORDER BY ... DESC` on ASC-sorted data — which all paid a full
+external sort across the entire scan. CPU wasted. Memory wasted.
+Streaming defeated.
+
+[Apache DataFusion]: https://datafusion.apache.org/
+
+This post walks through the **sort pushdown** series of work — a
Review Comment:
Restructured the opening as **problem → techniques → results**, and cut the
linear phase-by-phase framing.
##########
content/blog/2026-07-05-sort-pushdown.md:
##########
@@ -0,0 +1,839 @@
+---
+layout: post
+title: Sort Pushdown in DataFusion: Skip Sorts, Skip Decode, Skip I/O
+date: 2026-07-05
+author: Qi Zhu
+categories: [performance]
+---
+
+<!--
+{% comment %}
+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.
+{% endcomment %}
+-->
+
+[TOC]
+
+*Qi Zhu, [Massive](https://www.massive.com/)*
+
+Many [Apache Parquet] datasets are already sorted on disk. Time-series
+files are usually written in ingestion-time order. Event logs are sharded
+and sorted by event id. Partitioned tables come with a natural ordering
+implied by the partition key. The information about that ordering is
+sitting right there in the file metadata.
+
+[Apache Parquet]: https://parquet.apache.org/
+
+[Apache DataFusion] has always been able to skip the sort in the
+*trivial* case: when the catalog declares an ordering (`WITH ORDER`
+or parquet `sorting_columns`) **and** the on-disk file listing
+already matches that order, the `EnsureRequirements` rule sees that
+the scan's `output_ordering` already satisfies the request and
+**removes the redundant `SortExec`**. The hard cases were
+everything just slightly looser — files in the wrong on-disk order,
+declared but overlapping ranges, no declaration at all,
+`ORDER BY ... DESC` on ASC-sorted data — which all paid a full
+external sort across the entire scan. CPU wasted. Memory wasted.
+Streaming defeated.
+
+[Apache DataFusion]: https://datafusion.apache.org/
+
+This post walks through the **sort pushdown** series of work — a
+multi-release effort spanning DataFusion **v52 through v55+** — that
+closed that gap on three layers at once:
+
+1. **`Exact` path** — delete the `SortExec` entirely when statistics
+ prove the scan is already ordered. Headline: **2.1×–49×** on the
+ `sort_pushdown` benchmark suite.
+2. **`Inexact` path** — keep the `SortExec`, but read the
+ most-promising data first so `TopK`'s dynamic filter tightens fast
Review Comment:
Linked from the abstract and multiple in-body mentions of `TopK`'s dynamic
filter.
##########
content/blog/2026-07-05-sort-pushdown.md:
##########
@@ -0,0 +1,839 @@
+---
+layout: post
+title: Sort Pushdown in DataFusion: Skip Sorts, Skip Decode, Skip I/O
+date: 2026-07-05
+author: Qi Zhu
+categories: [performance]
+---
+
+<!--
+{% comment %}
+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.
+{% endcomment %}
+-->
+
+[TOC]
+
+*Qi Zhu, [Massive](https://www.massive.com/)*
+
+Many [Apache Parquet] datasets are already sorted on disk. Time-series
+files are usually written in ingestion-time order. Event logs are sharded
+and sorted by event id. Partitioned tables come with a natural ordering
+implied by the partition key. The information about that ordering is
+sitting right there in the file metadata.
+
+[Apache Parquet]: https://parquet.apache.org/
+
+[Apache DataFusion] has always been able to skip the sort in the
+*trivial* case: when the catalog declares an ordering (`WITH ORDER`
+or parquet `sorting_columns`) **and** the on-disk file listing
+already matches that order, the `EnsureRequirements` rule sees that
+the scan's `output_ordering` already satisfies the request and
+**removes the redundant `SortExec`**. The hard cases were
+everything just slightly looser — files in the wrong on-disk order,
+declared but overlapping ranges, no declaration at all,
+`ORDER BY ... DESC` on ASC-sorted data — which all paid a full
+external sort across the entire scan. CPU wasted. Memory wasted.
+Streaming defeated.
+
+[Apache DataFusion]: https://datafusion.apache.org/
+
+This post walks through the **sort pushdown** series of work — a
+multi-release effort spanning DataFusion **v52 through v55+** — that
+closed that gap on three layers at once:
+
+1. **`Exact` path** — delete the `SortExec` entirely when statistics
+ prove the scan is already ordered. Headline: **2.1×–49×** on the
+ `sort_pushdown` benchmark suite.
+2. **`Inexact` path** — keep the `SortExec`, but read the
+ most-promising data first so `TopK`'s dynamic filter tightens fast
+ and stale row groups get pruned by statistics.
+3. **Runtime row-group pruning** ([#22450]) — the latest piece, which
+ re-checks the dynamic filter at every row-group boundary and
+ physically removes pruned row groups from the open `RecordBatch`
+ stream. Headline: **5 of 11** `topk_tpch` queries run **3–4× faster**
Review Comment:
Trimmed to `2×–49× faster on ASC-LIMIT queries` — dropped the suite name.
##########
content/blog/2026-07-05-sort-pushdown.md:
##########
@@ -0,0 +1,839 @@
+---
+layout: post
+title: Sort Pushdown in DataFusion: Skip Sorts, Skip Decode, Skip I/O
+date: 2026-07-05
+author: Qi Zhu
+categories: [performance]
+---
+
+<!--
+{% comment %}
+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.
+{% endcomment %}
+-->
+
+[TOC]
+
+*Qi Zhu, [Massive](https://www.massive.com/)*
+
+Many [Apache Parquet] datasets are already sorted on disk. Time-series
+files are usually written in ingestion-time order. Event logs are sharded
+and sorted by event id. Partitioned tables come with a natural ordering
+implied by the partition key. The information about that ordering is
+sitting right there in the file metadata.
+
+[Apache Parquet]: https://parquet.apache.org/
+
+[Apache DataFusion] has always been able to skip the sort in the
+*trivial* case: when the catalog declares an ordering (`WITH ORDER`
+or parquet `sorting_columns`) **and** the on-disk file listing
+already matches that order, the `EnsureRequirements` rule sees that
+the scan's `output_ordering` already satisfies the request and
+**removes the redundant `SortExec`**. The hard cases were
+everything just slightly looser — files in the wrong on-disk order,
+declared but overlapping ranges, no declaration at all,
+`ORDER BY ... DESC` on ASC-sorted data — which all paid a full
+external sort across the entire scan. CPU wasted. Memory wasted.
+Streaming defeated.
+
+[Apache DataFusion]: https://datafusion.apache.org/
+
+This post walks through the **sort pushdown** series of work — a
+multi-release effort spanning DataFusion **v52 through v55+** — that
+closed that gap on three layers at once:
+
+1. **`Exact` path** — delete the `SortExec` entirely when statistics
+ prove the scan is already ordered. Headline: **2.1×–49×** on the
+ `sort_pushdown` benchmark suite.
+2. **`Inexact` path** — keep the `SortExec`, but read the
+ most-promising data first so `TopK`'s dynamic filter tightens fast
+ and stale row groups get pruned by statistics.
+3. **Runtime row-group pruning** ([#22450]) — the latest piece, which
+ re-checks the dynamic filter at every row-group boundary and
+ physically removes pruned row groups from the open `RecordBatch`
+ stream. Headline: **5 of 11** `topk_tpch` queries run **3–4× faster**
+ with zero regressions; total benchmark time drops by **−44%**.
+
+[#22450]: https://github.com/apache/datafusion/pull/22450
+[#20839]: https://github.com/apache/datafusion/pull/20839
Review Comment:
Yes — that's [#21956](https://github.com/apache/datafusion/pull/21956)
(`Runtime reorder for TopK convergence`), listed under Landed PRs. Called it
out more explicitly in the Inexact section.
##########
content/blog/2026-07-05-sort-pushdown.md:
##########
@@ -0,0 +1,839 @@
+---
+layout: post
+title: Sort Pushdown in DataFusion: Skip Sorts, Skip Decode, Skip I/O
+date: 2026-07-05
+author: Qi Zhu
+categories: [performance]
+---
+
+<!--
+{% comment %}
+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.
+{% endcomment %}
+-->
+
+[TOC]
+
+*Qi Zhu, [Massive](https://www.massive.com/)*
+
+Many [Apache Parquet] datasets are already sorted on disk. Time-series
+files are usually written in ingestion-time order. Event logs are sharded
+and sorted by event id. Partitioned tables come with a natural ordering
+implied by the partition key. The information about that ordering is
+sitting right there in the file metadata.
+
+[Apache Parquet]: https://parquet.apache.org/
+
+[Apache DataFusion] has always been able to skip the sort in the
+*trivial* case: when the catalog declares an ordering (`WITH ORDER`
+or parquet `sorting_columns`) **and** the on-disk file listing
+already matches that order, the `EnsureRequirements` rule sees that
+the scan's `output_ordering` already satisfies the request and
+**removes the redundant `SortExec`**. The hard cases were
+everything just slightly looser — files in the wrong on-disk order,
+declared but overlapping ranges, no declaration at all,
+`ORDER BY ... DESC` on ASC-sorted data — which all paid a full
+external sort across the entire scan. CPU wasted. Memory wasted.
+Streaming defeated.
+
+[Apache DataFusion]: https://datafusion.apache.org/
+
+This post walks through the **sort pushdown** series of work — a
+multi-release effort spanning DataFusion **v52 through v55+** — that
+closed that gap on three layers at once:
+
+1. **`Exact` path** — delete the `SortExec` entirely when statistics
+ prove the scan is already ordered. Headline: **2.1×–49×** on the
+ `sort_pushdown` benchmark suite.
+2. **`Inexact` path** — keep the `SortExec`, but read the
+ most-promising data first so `TopK`'s dynamic filter tightens fast
+ and stale row groups get pruned by statistics.
+3. **Runtime row-group pruning** ([#22450]) — the latest piece, which
+ re-checks the dynamic filter at every row-group boundary and
+ physically removes pruned row groups from the open `RecordBatch`
+ stream. Headline: **5 of 11** `topk_tpch` queries run **3–4× faster**
+ with zero regressions; total benchmark time drops by **−44%**.
+
+[#22450]: https://github.com/apache/datafusion/pull/22450
+[#20839]: https://github.com/apache/datafusion/pull/20839
+
+Together these compose into a **three-layer pruning stack** —
Review Comment:
Removed.
##########
content/blog/2026-07-05-sort-pushdown.md:
##########
@@ -0,0 +1,839 @@
+---
+layout: post
+title: Sort Pushdown in DataFusion: Skip Sorts, Skip Decode, Skip I/O
+date: 2026-07-05
+author: Qi Zhu
+categories: [performance]
+---
+
+<!--
+{% comment %}
+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.
+{% endcomment %}
+-->
+
+[TOC]
+
+*Qi Zhu, [Massive](https://www.massive.com/)*
+
+Many [Apache Parquet] datasets are already sorted on disk. Time-series
+files are usually written in ingestion-time order. Event logs are sharded
+and sorted by event id. Partitioned tables come with a natural ordering
+implied by the partition key. The information about that ordering is
+sitting right there in the file metadata.
+
+[Apache Parquet]: https://parquet.apache.org/
+
+[Apache DataFusion] has always been able to skip the sort in the
+*trivial* case: when the catalog declares an ordering (`WITH ORDER`
+or parquet `sorting_columns`) **and** the on-disk file listing
+already matches that order, the `EnsureRequirements` rule sees that
+the scan's `output_ordering` already satisfies the request and
+**removes the redundant `SortExec`**. The hard cases were
+everything just slightly looser — files in the wrong on-disk order,
+declared but overlapping ranges, no declaration at all,
+`ORDER BY ... DESC` on ASC-sorted data — which all paid a full
+external sort across the entire scan. CPU wasted. Memory wasted.
+Streaming defeated.
+
+[Apache DataFusion]: https://datafusion.apache.org/
+
+This post walks through the **sort pushdown** series of work — a
+multi-release effort spanning DataFusion **v52 through v55+** — that
+closed that gap on three layers at once:
+
+1. **`Exact` path** — delete the `SortExec` entirely when statistics
+ prove the scan is already ordered. Headline: **2.1×–49×** on the
+ `sort_pushdown` benchmark suite.
+2. **`Inexact` path** — keep the `SortExec`, but read the
+ most-promising data first so `TopK`'s dynamic filter tightens fast
+ and stale row groups get pruned by statistics.
+3. **Runtime row-group pruning** ([#22450]) — the latest piece, which
+ re-checks the dynamic filter at every row-group boundary and
+ physically removes pruned row groups from the open `RecordBatch`
+ stream. Headline: **5 of 11** `topk_tpch` queries run **3–4× faster**
+ with zero regressions; total benchmark time drops by **−44%**.
+
+[#22450]: https://github.com/apache/datafusion/pull/22450
+[#20839]: https://github.com/apache/datafusion/pull/20839
+
+Together these compose into a **three-layer pruning stack** —
+file-level, row-group-level, row-level — all driven by the same
+`TopK` dynamic filter. **None of the runtime parts of this post
+would be possible without TopK's dynamic filter pushdown** — the
+primitive that lets `TopK`'s threshold reach the parquet scanner
+mid-execution. The earlier [Dynamic Filters][dyn-filters-blog]
+post covers that primitive in detail and is recommended background
+for what follows.
+
+The page-level reverse primitive we are adding upstream in
Review Comment:
Moved to Future Directions only. Intro no longer mentions it.
##########
content/blog/2026-07-05-sort-pushdown.md:
##########
@@ -0,0 +1,839 @@
+---
+layout: post
+title: Sort Pushdown in DataFusion: Skip Sorts, Skip Decode, Skip I/O
+date: 2026-07-05
+author: Qi Zhu
+categories: [performance]
+---
+
+<!--
+{% comment %}
+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.
+{% endcomment %}
+-->
+
+[TOC]
+
+*Qi Zhu, [Massive](https://www.massive.com/)*
+
+Many [Apache Parquet] datasets are already sorted on disk. Time-series
+files are usually written in ingestion-time order. Event logs are sharded
+and sorted by event id. Partitioned tables come with a natural ordering
+implied by the partition key. The information about that ordering is
+sitting right there in the file metadata.
+
+[Apache Parquet]: https://parquet.apache.org/
+
+[Apache DataFusion] has always been able to skip the sort in the
+*trivial* case: when the catalog declares an ordering (`WITH ORDER`
+or parquet `sorting_columns`) **and** the on-disk file listing
+already matches that order, the `EnsureRequirements` rule sees that
+the scan's `output_ordering` already satisfies the request and
+**removes the redundant `SortExec`**. The hard cases were
+everything just slightly looser — files in the wrong on-disk order,
+declared but overlapping ranges, no declaration at all,
+`ORDER BY ... DESC` on ASC-sorted data — which all paid a full
+external sort across the entire scan. CPU wasted. Memory wasted.
+Streaming defeated.
+
+[Apache DataFusion]: https://datafusion.apache.org/
+
+This post walks through the **sort pushdown** series of work — a
+multi-release effort spanning DataFusion **v52 through v55+** — that
+closed that gap on three layers at once:
+
+1. **`Exact` path** — delete the `SortExec` entirely when statistics
+ prove the scan is already ordered. Headline: **2.1×–49×** on the
+ `sort_pushdown` benchmark suite.
+2. **`Inexact` path** — keep the `SortExec`, but read the
+ most-promising data first so `TopK`'s dynamic filter tightens fast
+ and stale row groups get pruned by statistics.
+3. **Runtime row-group pruning** ([#22450]) — the latest piece, which
+ re-checks the dynamic filter at every row-group boundary and
+ physically removes pruned row groups from the open `RecordBatch`
+ stream. Headline: **5 of 11** `topk_tpch` queries run **3–4× faster**
+ with zero regressions; total benchmark time drops by **−44%**.
+
+[#22450]: https://github.com/apache/datafusion/pull/22450
+[#20839]: https://github.com/apache/datafusion/pull/20839
+
+Together these compose into a **three-layer pruning stack** —
+file-level, row-group-level, row-level — all driven by the same
+`TopK` dynamic filter. **None of the runtime parts of this post
+would be possible without TopK's dynamic filter pushdown** — the
+primitive that lets `TopK`'s threshold reach the parquet scanner
+mid-execution. The earlier [Dynamic Filters][dyn-filters-blog]
+post covers that primitive in detail and is recommended background
+for what follows.
+
+The page-level reverse primitive we are adding upstream in
+[arrow-rs] will push the `DESC` gains further still.
+
+[arrow-rs]: https://github.com/apache/arrow-rs
+
+## TL;DR
Review Comment:
Removed the TL;DR block entirely; the 3-sentence abstract at the top now
covers it.
--
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]
