add window expression stream, delegated window aggregation to aggregate functions, and implement `row_number` #375

jimexist · 2021-05-21T11:45:17Z

Basic aggregation based window function
Streaming results from window functions
Basic Structure for row number

This partly fix-es #298

codecov-commenter · 2021-05-22T10:42:25Z

Codecov Report

Merging #375 (ad3d9c1) into master (3593d1f) will increase coverage by 0.49%.
The diff coverage is 77.02%.

@@            Coverage Diff             @@
##           master     #375      +/-   ##
==========================================
+ Coverage   74.85%   75.34%   +0.49%     
==========================================
  Files         146      147       +1     
  Lines       24565    24782     +217     
==========================================
+ Hits        18387    18673     +286     
+ Misses       6178     6109      -69

Impacted Files	Coverage Δ
datafusion/src/physical_plan/expressions/mod.rs	`71.42% <ø> (ø)`
datafusion/src/physical_plan/hash_aggregate.rs	`85.21% <ø> (ø)`
datafusion/src/physical_plan/sort.rs	`91.26% <ø> (ø)`
datafusion/src/physical_plan/window_functions.rs	`85.71% <57.89%> (-3.01%)`	⬇️
datafusion/src/physical_plan/mod.rs	`78.70% <61.90%> (-4.06%)`	⬇️
datafusion/src/physical_plan/windows.rs	`73.78% <76.57%> (+73.78%)`	⬆️
...fusion/src/physical_plan/expressions/row_number.rs	`81.25% <81.25%> (ø)`
datafusion/src/execution/context.rs	`92.08% <100.00%> (+0.03%)`	⬆️
datafusion/src/physical_plan/planner.rs	`80.45% <100.00%> (+3.94%)`	⬆️
datafusion/tests/sql.rs	`99.89% <100.00%> (+<0.01%)`	⬆️
... and 11 more

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 3593d1f...ad3d9c1. Read the comment docs.

alamb · 2021-05-23T10:49:07Z

Thank @jimexist -- I plan to review this carefully tomorrow

datafusion/src/physical_plan/windows.rs

datafusion/src/physical_plan/expressions/row_number.rs

datafusion/src/physical_plan/mod.rs

alamb · 2021-05-24T18:42:32Z

datafusion/src/physical_plan/mod.rs

+pub trait WindowAccumulator: Send + Sync + Debug {
+    /// scans the accumulator's state from a vector of scalars, similar to Accumulator it also
+    /// optionally generates values.
+    fn scan(&mut self, values: &[ScalarValue]) -> Result<Option<ScalarValue>>;


It would help me if you could explain where the "window" for the window function appears in this trait. I assume you already have a design in mind, so I figured I would ask here

I am thinking about a query like the following

select sum(value) OVER (ROWS 5 PRECEDING) FROM ....

I think in this case, you end up with 10 aggregate values from 10 different windows, in the the following manner:

1 2 3 4 5 6 7 8 9 input window 1 ─ window 2 ─── window 3 ───── window 4 ─────── window 5 ───────── window 6 ───────── window 7 ───────── window 8 ───────── window 9 ─────────

I would have expected the WindowAccumulator to have functions like

/// Add a new row to the current window fn new_row_in_window(ScalarValue); /// remove a row from the current window fn remove_row_from_window(ScalarValue); /// The current value of this function for the given window fn current_value(ScalarValue);

Or possibly something like

evaluate(window: &[ArrayRef]) -> ScalarValue

@alamb good question!

The window word in this trait is purely indicating the fact that window functions will use this. it can be of a better name but...

for a design, there are two complications:

multiple window functions, each having different window frames, can be scanning batches at the same time, so i'd want each to create its own window accumulator, remembering to push_back, and remove front, on their own; this trait would not put that into the API, it just scans

specifically for window that peeks ahead, because batches arrive in async stream, it is not feasible to peek, so my plan is to allow them to optionally execute a one time shift upon finishing up (e.g. lead is just producing the same value in situ, but with a one time shift at the end)

Due to 1 and 2, a best possible state vector for window accumulator is possibly VecDequeue. And the name scan is because accumulators iterate the list and either optionally produce one value at a time (max with order by), or exactly one accumulated value upon finish scanning (max with empty over), but not both

I see -- looking at how these functions are used for nth_value here helped me get a better sense for how these traits work

I think as you go through the implementation, adding some additional details to help future implementers of this trait. For example, when it is ok to return values from scan() or scan_batch() and what the expected total number of rows produced?

However having several examples of of implemented window functions I think will help too so no need to change anything more at this time.

alamb · 2021-05-24T18:43:32Z

datafusion/src/physical_plan/windows.rs

+
+#[cfg(test)]
+mod tests {
+    // use super::*;


I think tests would be good to have

i try to showcase the design here, so if the interface is generally okay with reviewers, i can go ahead and write down tests.

also there's integration tests in sql.rs

jimexist · 2021-05-25T05:13:38Z

local test:

> select c1, row_number() over () from test limit 5;
+----+--------------+
| c1 | ROW_NUMBER() |
+----+--------------+
| c  | 1            |
| d  | 2            |
| b  | 3            |
| a  | 4            |
| b  | 5            |
+----+--------------+
5 rows in set. Query took 0 seconds.

alamb

I personally think this is a great start. The structure is clear and I think the follow on PRs will make it even better. Thank you so much @jimexist

I am still a bit confused about what the requirements are while implementing a WindowAccumulator but I think they will become clearer as we add more window aggregate support.

I think we should merge this one in. cc @Dandandan @andygrove @jorgecarleitao

alamb · 2021-05-25T16:48:20Z

datafusion/src/physical_plan/mod.rs

+pub trait WindowAccumulator: Send + Sync + Debug {
+    /// scans the accumulator's state from a vector of scalars, similar to Accumulator it also
+    /// optionally generates values.
+    fn scan(&mut self, values: &[ScalarValue]) -> Result<Option<ScalarValue>>;


I see -- looking at how these functions are used for nth_value here helped me get a better sense for how these traits work

I think as you go through the implementation, adding some additional details to help future implementers of this trait. For example, when it is ok to return values from scan() or scan_batch() and what the expected total number of rows produced?

However having several examples of of implemented window functions I think will help too so no need to change anything more at this time.

alamb · 2021-05-25T16:50:04Z

datafusion/tests/sql.rs

+    min(c3) over () \
+    from aggregate_test_100 limit 5";
+    let actual = execute(&mut ctx, sql).await;
+    let expected = vec![


datafusion/src/physical_plan/windows.rs

datafusion/src/execution/context.rs

datafusion/src/physical_plan/expressions/row_number.rs

Dandandan · 2021-05-25T17:32:42Z

datafusion/src/physical_plan/windows.rs

+
+    for i in 0..(schema.fields().len() - window_expr.len()) {
+        let col = concat(
+            &original_batches


I guess this is only needed for the empty over clause right, as in other cases we probably only want to access the rows.
Also for functions like RowNumber I think we could be streaming/emit batches instead of "caching" them as we don't need to scan all the rows in order to compute the result. Both good in this PR or an issue / follow up work!

yes I am aware of the potential changes on how rows are scanned but plan to leave this as is because down the road (when adding partition by and order by) things could change a lot.

I'd like this to gradually fledge out.

datafusion/src/physical_plan/expressions/row_number.rs

Dandandan · 2021-05-25T17:41:53Z

datafusion/tests/sql.rs

+    count(c3) over (), \
+    max(c3) over (), \
+    min(c3) over () \
+    from aggregate_test_100 limit 5";


I think an explicit order by is required (or sorting of results) to make this deterministic.

Now the ordering is "just" a implementation detail of the batches getting merged / concatenated into 1 batch and keeping the same ordering.

will add sort by c1

Dandandan · 2021-05-25T17:46:03Z

datafusion/src/physical_plan/mod.rs

+        &mut self,
+        num_rows: usize,
+        values: &[ArrayRef],
+    ) -> Result<Option<ArrayRef>> {


I don't get completely when this scan_batch is being used instead of the scan_batch of the implementations? Or can this be removed by now?

it's a catch-all implementations. with more window functions implemented it'll be clearer and by that time we can feel free to remove this.

Dandandan

I think this is a good start / continuation - thanks @jimexist for this amazing contribution!

I think this is good to go - also feel free addressing some of the comments within this PR if you want.

commit 7fb3640 Author: Jiayu Liu <jiayu.liu@airbnb.com> Date: Fri May 21 16:38:25 2021 +0800 row number done commit 1723926 Author: Jiayu Liu <jiayu.liu@airbnb.com> Date: Fri May 21 16:05:50 2021 +0800 add row number commit bf5b8a5 Author: Jiayu Liu <jiayu.liu@airbnb.com> Date: Fri May 21 15:04:49 2021 +0800 save commit d2ce852 Author: Jiayu Liu <jiayu.liu@airbnb.com> Date: Fri May 21 14:53:05 2021 +0800 add streams commit 0a861a7 Author: Jiayu Liu <jiayu.liu@airbnb.com> Date: Thu May 20 22:28:34 2021 +0800 save stream commit a9121af Author: Jiayu Liu <jiayu.liu@airbnb.com> Date: Thu May 20 22:01:51 2021 +0800 update unit test commit 2af2a27 Author: Jiayu Liu <jiayu.liu@airbnb.com> Date: Fri May 21 14:25:12 2021 +0800 fix unit test commit bb57c76 Author: Jiayu Liu <jiayu.liu@airbnb.com> Date: Fri May 21 14:23:34 2021 +0800 use upper case commit 5d96e52 Author: Jiayu Liu <jiayu.liu@airbnb.com> Date: Fri May 21 14:16:16 2021 +0800 fix unit test commit 1ecae8f Author: Jiayu Liu <jiayu.liu@airbnb.com> Date: Fri May 21 12:27:26 2021 +0800 fix unit test commit bc2271d Author: Jiayu Liu <jiayu.liu@airbnb.com> Date: Fri May 21 10:04:29 2021 +0800 fix error commit 880b94f Author: Jiayu Liu <jiayu.liu@airbnb.com> Date: Fri May 21 08:24:00 2021 +0800 fix unit test commit 4e792e1 Author: Jiayu Liu <jiayu.liu@airbnb.com> Date: Fri May 21 08:05:17 2021 +0800 fix test commit c36c04a Author: Jiayu Liu <jiayu.liu@airbnb.com> Date: Fri May 21 00:07:54 2021 +0800 add more tests commit f5e64de Author: Jiayu Liu <jiayu.liu@airbnb.com> Date: Thu May 20 23:41:36 2021 +0800 update commit a1eae86 Author: Jiayu Liu <jiayu.liu@airbnb.com> Date: Thu May 20 23:36:15 2021 +0800 enrich unit test commit 0d2a214 Author: Jiayu Liu <jiayu.liu@airbnb.com> Date: Thu May 20 23:25:43 2021 +0800 adding filter by todo commit 8b486d5 Author: Jiayu Liu <jiayu.liu@airbnb.com> Date: Thu May 20 23:17:22 2021 +0800 adding more built-in functions commit abf08cd Author: Jiayu Liu <Jimexist@users.noreply.github.com> Date: Thu May 20 22:36:27 2021 +0800 Update datafusion/src/physical_plan/window_functions.rs Co-authored-by: Andrew Lamb <andrew@nerdnetworks.org> commit 0cbca53 Author: Jiayu Liu <Jimexist@users.noreply.github.com> Date: Thu May 20 22:34:57 2021 +0800 Update datafusion/src/physical_plan/window_functions.rs Co-authored-by: Andrew Lamb <andrew@nerdnetworks.org> commit 831c069 Author: Jiayu Liu <Jimexist@users.noreply.github.com> Date: Thu May 20 22:34:04 2021 +0800 Update datafusion/src/logical_plan/builder.rs Co-authored-by: Andrew Lamb <andrew@nerdnetworks.org> commit f70c739 Author: Jiayu Liu <Jimexist@users.noreply.github.com> Date: Thu May 20 22:33:04 2021 +0800 Update datafusion/src/logical_plan/builder.rs Co-authored-by: Andrew Lamb <andrew@nerdnetworks.org> commit 3ee87aa Author: Jiayu Liu <jiayu.liu@airbnb.com> Date: Wed May 19 22:55:08 2021 +0800 fix unit test commit 5c4d92d Author: Jiayu Liu <jiayu.liu@airbnb.com> Date: Wed May 19 22:48:26 2021 +0800 fix clippy commit a0b7526 Author: Jiayu Liu <jiayu.liu@airbnb.com> Date: Wed May 19 22:46:38 2021 +0800 fix unused imports commit 1d3b076 Author: Jiayu Liu <jiayu.liu@airbnb.com> Date: Thu May 13 18:51:14 2021 +0800 add window expr

jimexist force-pushed the add-row-number branch 2 times, most recently from 90daf11 to 4ad6e4d Compare May 22, 2021 09:50

jimexist marked this pull request as ready for review May 22, 2021 10:44

jimexist changed the title ~~add streaming of window expr, row_number~~ add window expression stream, delegated window aggregation, and a basic structure for row_number May 22, 2021

jimexist mentioned this pull request May 23, 2021

Function to create ArrayRef from an iterator of ScalarValues #381

Merged

jimexist force-pushed the add-row-number branch from 1bdc43c to 1cf079c Compare May 23, 2021 11:05

Dandandan reviewed May 23, 2021

View reviewed changes

datafusion/src/physical_plan/windows.rs Outdated Show resolved Hide resolved

jimexist force-pushed the add-row-number branch from 7304d66 to 09ae29d Compare May 24, 2021 00:35

jimexist mentioned this pull request May 24, 2021

add first_value, last_value, and nth_value built-in window functions #403

Merged

jimexist force-pushed the add-row-number branch 2 times, most recently from 90a22fa to 6e4d662 Compare May 24, 2021 14:36

jimexist mentioned this pull request May 24, 2021

finish implementation of window accumulation and func row_number() and add sql.rs test #413

Closed

alamb reviewed May 24, 2021

View reviewed changes

jimexist force-pushed the add-row-number branch 2 times, most recently from 8e6c181 to b439b0e Compare May 25, 2021 01:40

jimexist changed the title ~~add window expression stream, delegated window aggregation, and a basic structure for row_number~~ add window expression stream, delegated window aggregation, and implement row_number May 25, 2021

jimexist changed the title ~~add window expression stream, delegated window aggregation, and implement row_number~~ add window expression stream, delegated window aggregation to aggregate functions, and implement row_number May 25, 2021

jimexist force-pushed the add-row-number branch from ee57340 to 23ded9f Compare May 25, 2021 05:10

jimexist force-pushed the add-row-number branch 3 times, most recently from 1b82f25 to 563192a Compare May 25, 2021 12:43

jimexist requested review from alamb and Dandandan May 25, 2021 15:56

alamb approved these changes May 25, 2021

View reviewed changes