Adez017 commented on code in PR #66:
URL: https://github.com/apache/datafusion-site/pull/66#discussion_r2047186316
##########
content/blog/2025-04-04-datafusion-userdefined-window-functions.md:
##########
@@ -0,0 +1,339 @@
+---
+layout: post
+title: User defined Window Functions in DataFusion
+date: 2025-04-04
+author: Aditya Singh Rathore
+categories: [tutorial]
+---
+
+<!--
+{% 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 %}
+-->
+
+## Introduction
+Window functions are a powerful feature in SQL, allowing for complex
analytical computations over a subset of data. However, efficiently
implementing them, especially sliding windows, can be quite challenging. With
DataFusion's recent support for user-defined window functions , developers now
have more flexibility and performance improvements at their disposal.
+
+In this post, we'll explore:
+
+- What window functions are and why they matter
+
+- Understanding sliding windows
+
+- The challenges of computing window aggregates efficiently
+
+- How DataFusion optimizes these computations
+
+- Alternative approaches and their trade-offs
+
+## Understanding Window Functions in SQL
+
+Imagine you're analyzing sales data and want insights without losing the finer
details. This is where **window functions** come into play. Unlike **GROUP
BY**, which condenses data, window functions let you retain each row while
performing calculations over a defined **range** —like having a moving lens
over your dataset.
+
+Picture a business tracking daily sales. They need a running total to
understand cumulative revenue trends without collapsing individual
transactions. SQL makes this easy:
+```sql
+SELECT id, value, SUM(value) OVER (ORDER BY id) AS running_total
+FROM sales;
+```
+```text
+example:
++------------+--------+-------------------------------+
+| Date | Sales | Rows Considered |
++------------+--------+-------------------------------+
+| Jan 01 | 100 | [100] |
+| Jan 02 | 120 | [100, 120] |
+| Jan 03 | 130 | [100, 120, 130] |
+| Jan 04 | 150 | [100, 120, 130, 150] |
+| Jan 05 | 160 | [100, 120, 130, 150, 160] |
+| Jan 06 | 180 | [100, 120, 130, 150, 160, 180]|
+| Jan 07 | 170 | [100, ..., 170] (7 days) |
+| Jan 08 | 175 | [120, ..., 175] |
++------------+--------+-------------------------------+
+A row-by-row representation of how a 7-day moving average includes the
previous 6 days and the current one.
+```
+
+
+This helps in analytical queries where we need cumulative sums, moving
averages, or ranking without losing individual records.
+
+
+## User Defined Window Functions
+Built-in window functions serve many use cases, but sometimes custom logic is
needed—for example:
+
+- Calculating moving averages with complex conditions
+
+- Implementing a custom ranking strategy
+
+- Tracking non-standard cumulative logic
+
+Thus, **User-Defined Window Functions (UDWFs)** allow developers to define
their own behavior using a combination of SQL and *bit of logic*.
+
+Writing a user defined window function is slightly more complex than an
aggregate function due
+to the variety of ways that window functions are called. I recommend reviewing
the
+[online
documentation](https://datafusion.apache.org/library-user-guide/adding-udfs.html#registering-a-window-udf)
+for a description of which functions need to be implemented.
+
+## Understaing Sliding Window
+
+Sliding windows define a **moving range** of data over which aggregations are
computed. Unlike simple cumulative functions, these windows are dynamically
updated as new data arrives.
+
+For instance, if we want a 7-day moving average of sales:
+
+```sql
+SELECT date, sales,
+ AVG(sales) OVER (ORDER BY date ROWS BETWEEN 6 PRECEDING AND CURRENT
ROW) AS moving_avg
+FROM sales;
+```
+Here, each row’s result is computed based on the last 7 days, making it
computationally intensive as data grows.
+
+## Why Computing Sliding Windows Is Hard
+
+Imagine you’re at a café, and the barista is preparing coffee orders. If they
made each cup from scratch without using pre-prepared ingredients, the process
would be painfully slow. This is exactly the problem with naïve sliding window
computations.
+
+Computing sliding windows efficiently is tricky because:
+
+- **High Computation Costs:** Just like making coffee from scratch for each
customer, recalculating aggregates for every row is expensive.
+
+- **Data Shuffling:** In large distributed systems, data must often be
shuffled between nodes, causing delays—like passing orders between multiple
baristas who don’t communicate efficiently.
+
+- **State Management:** Keeping track of past computations is like remembering
previous orders without writing them down—error-prone and inefficient.
+
+Many traditional query engines struggle to optimize these computations
effectively, leading to sluggish performance.
+
+## How DataFusion Evaluates Window Functions Quickly
+In the world of big data, every millisecond counts. Imagine you’re analyzing
stock market data, tracking sensor readings from millions of IoT devices, or
crunching through massive customer logs—speed matters. This is where
[DataFusion](https://datafusion.apache.org/) shines, making window function
computations blazing fast. Let’s break down how it achieves this remarkable
performance.
+
+DataFusion now supports [user-defined window aggregates
(UDWAs)](https://datafusion.apache.org/library-user-guide/adding-udfs.html),
meaning you can bring your own aggregation logic and use it within a window
function.
+
+For example, we will declare a user defined window function that computes a
moving average.
+```sql
+use datafusion::arrow::{array::{ArrayRef, Float64Array, AsArray},
datatypes::Float64Type};
+use datafusion::logical_expr::{PartitionEvaluator};
+use datafusion::common::ScalarValue;
+use datafusion::error::Result;
+/// This implements the lowest level evaluation for a window function
+///
+/// It handles calculating the value of the window function for each
+/// distinct values of `PARTITION BY`
+#[derive(Clone, Debug)]
+struct MyPartitionEvaluator {}
+
+impl MyPartitionEvaluator {
+ fn new() -> Self {
+ Self {}
+ }
+}
+
+/// Different evaluation methods are called depending on the various
+/// settings of WindowUDF. This example uses the simplest and most
+/// general, `evaluate`. See `PartitionEvaluator` for the other more
+/// advanced uses.
+impl PartitionEvaluator for MyPartitionEvaluator {
+ /// Tell DataFusion the window function varies based on the value
+ /// of the window frame.
+ fn uses_window_frame(&self) -> bool {
+ true
+ }
+
+ /// This function is called once per input row.
+ ///
+ /// `range`specifies which indexes of `values` should be
+ /// considered for the calculation.
+ ///
+ /// Note this is the SLOWEST, but simplest, way to evaluate a
+ /// window function. It is much faster to implement
+ /// evaluate_all or evaluate_all_with_rank, if possible
+ fn evaluate(
+ &mut self,
+ values: &[ArrayRef],
+ range: &std::ops::Range<usize>,
+ ) -> Result<ScalarValue> {
+ // Again, the input argument is an array of floating
+ // point numbers to calculate a moving average
+ let arr: &Float64Array =
values[0].as_ref().as_primitive::<Float64Type>();
+
+ let range_len = range.end - range.start;
+
+ // our smoothing function will average all the values in the
+ let output = if range_len > 0 {
+ let sum: f64 =
arr.values().iter().skip(range.start).take(range_len).sum();
+ Some(sum / range_len as f64)
+ } else {
+ None
+ };
+
+ Ok(ScalarValue::Float64(output))
+ }
+}
+
+/// Create a `PartitionEvaluator` to evaluate this function on a new
+/// partition.
+fn make_partition_evaluator() -> Result<Box<dyn PartitionEvaluator>> {
+ Ok(Box::new(MyPartitionEvaluator::new()))
+}
+```
+### Registering a Window UDF
+To register a Window UDF, you need to wrap the function implementation in a
`WindowUDF` struct and then register it with the `SessionContext`. DataFusion
provides the `create_udwf` helper functions to make this easier. There is a
lower level API with more functionality but is more complex, that is documented
in
[advanced_udwf.rs](https://github.com/apache/datafusion/blob/main/datafusion-examples/examples/advanced_udwf.rs).
Review Comment:
done
--
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: github-unsubscr...@datafusion.apache.org
For queries about this service, please contact Infrastructure at:
us...@infra.apache.org
---------------------------------------------------------------------
To unsubscribe, e-mail: github-unsubscr...@datafusion.apache.org
For additional commands, e-mail: github-h...@datafusion.apache.org
