Convert `nth_value` builtIn function to User Defined Window Function by buraksenn · Pull Request #13201 · apache/datafusion

buraksenn · 2024-10-31T13:26:06Z

Which issue does this PR close?

Closes#12649

Rationale for this change

Context: #8709

What changes are included in this PR?

cleanup BuiltinWindowFunctions
move builtin nth_value function to udwf

Are these changes tested?

yes

Are there any user-facing changes?

no

alamb · 2024-10-31T19:21:20Z

THis is so exciting. FYI @jonathanc-n and @Omega359

alamb · 2024-10-31T19:24:18Z

It got pretty out of hand because of nth_value and BuiltinWindowFunction references across the repo so I wanted to open a draft PR to get feedback.

I personally think it would be fine to leave BuiltInWindowFunction initially while we port over the function, and them remove it as a follow on PR

Perhaps you can leave a stub in like

enumBuiltInWindowFunction{// Never created, will be removed in a follow on PRStub};

Then we can focus this PR on making sure that nth_value is working as a user defined function and then move on in the next PR to mechanically ripping out the old BuiltInWindowFunction

buraksenn · 2024-11-01T11:08:29Z

It got pretty out of hand because of nth_value and BuiltinWindowFunction references across the repo so I wanted to open a draft PR to get feedback.
I personally think it would be fine to leave BuiltInWindowFunction initially while we port over the function, and them remove it as a follow on PR
Perhaps you can leave a stub in like
enumBuiltInWindowFunction{// Never created, will be removed in a follow on PRStub};
Then we can focus this PR on making sure that nth_value is working as a user defined function and then move on in the next PR to mechanically ripping out the old BuiltInWindowFunction

Thanks @alamb will continue with what you've said

buraksenn · 2024-11-03T20:30:20Z

Wanted to update here. I think I'm almost finished but probably encountered a side effect. This query fails in slt file:

External error: query failed: DataFusion error: Arrow error: Invalid argument error: It is not possible to concatenate arrays of different data types.
[SQL] select
c9,
sum(c5) over (order by c9) as sum1,
avg(c5) over (order by c9) as avg1,
count(c5) over (order by c9) as count1,
max(c5) over (order by c9) as max1,
min(c5) over (order by c9) as min1,
first_value(c5) over (order by c9) as fv1,
last_value(c5) over (order by c9) as lv1,
nth_value(c5, 2) over (order by c9) as nv1
from aggregate_test_100
order by c9
limit 5
at test_files/window.slt:97
Error: Execution("1 failures")
error: test failed, to rerun pass `-p datafusion-sqllogictest --test sqllogictests`

I hope to fix this and make this ready tomorrow

jcsherin · 2024-11-04T06:51:31Z

External error: query failed: DataFusion error: Arrow error: Invalid argument error: It is not possible to concatenate arrays of different data types.

In the built-in (older) version the output field is defined like:

fnfield(&self) -> Result<Field>{let nullable = true;Ok(Field::new(&self.name,self.data_type.clone(), nullable))}

In the current code, the data type of the field is hard-coded as DataType::UInt64:

fnfield(&self,field_args:WindowUDFFieldArgs) -> Result<Field>{let nullable = true;Ok(Field::new(field_args.name(),DataType::UInt64, nullable))}

To fix this use field_args to get the data type of the input expression rather than hard-coding DataType::UInt64.

buraksenn · 2024-11-04T09:33:04Z

External error: query failed: DataFusion error: Arrow error: Invalid argument error: It is not possible to concatenate arrays of different data types.
In the built-in (older) version the output field is defined like:
fnfield(&self) -> Result<Field>{let nullable = true;Ok(Field::new(&self.name,self.data_type.clone(), nullable))}
In the current code, the data type of the field is hard-coded as DataType::UInt64:
fnfield(&self,field_args:WindowUDFFieldArgs) -> Result<Field>{let nullable = true;Ok(Field::new(field_args.name(),DataType::UInt64, nullable))}
To fix this use field_args to get the data type of the input expression rather than hard-coding DataType::UInt64.

Thanks, @jcsherin that was the fix. I've fixed that issue but encountered another one. I return Error from partition evaluator but I think it is not honored.

External error: query is expected to fail, but actually succeed:
[SQL] SELECT NTH_VALUE('+Inf'::Double, v1) OVER (PARTITION BY v1) FROM t1;
at test_files/window.slt:4895
Error: Execution("1 failures")
error: test failed, to rerun pass `-p datafusion-sqllogictest --test sqllogictests`
Caused by:
process didn't exit successfully: `/Users/buraksen/d/datafusion/target/debug/deps/sqllogictests-267639e9b963ccc2 window` (exit status: 1)

But it should not succeed since:

 let n =
match get_scalar_value_from_args(partition_evaluator_args.input_exprs(), 1)?
.map(get_signed_integer)
{
Some(Ok(n)) => {
if partition_evaluator_args.is_reversed() {
-n
} else {
n
}
}
_ => {
return exec_err!(
"Expected a signed integer literal for the second argument of nth_value"
)
}
};

jcsherin · 2024-11-04T18:29:05Z

TL;DR

For invalid input expressions, built-in window functions fail early when converting logical plan to physical plan. But user-defined window functions will complete planning, and fail only during physical execution.

Validation of input expressions in user-defined window runs only during physical execution.

In this case is it not better for udwf to fail early when converting to physical plan?

A possible solution is to update WindowUDFImpl API so that we parse input expressions when creating the window expressions (similar to built-in window functions). And hook it here so argument parsing happens earlier:

datafusion/datafusion/physical-plan/src/windows/mod.rs

Lines 158 to 164 in b61b2fc

fncreate_udwf_window_expr(
fun:&Arc<WindowUDF>,
args:&[Arc<dynPhysicalExpr>],
input_schema:&Schema,
name:String,
ignore_nulls:bool,
) -> Result<Arc<dynBuiltInWindowFunctionExpr>>{

Edge Case: Empty Table

DataFusion CLI v42.2.0> CREATE TABLE t1(v1 BIGINT);
0 row(s) fetched.
Elapsed 0.020 seconds.

There are currently no rows in t1 so an early return happens in WindowAggStream::compute_aggregates.

datafusion/datafusion/physical-plan/src/windows/window_agg_exec.rs

Lines 319 to 321 in 89e96b4

if batch.num_rows() == 0{
returnOk(RecordBatch::new_empty(Arc::clone(&self.schema)));
}

The nth_value window function is never executed when this query is run on the empty table t1.

>SELECT NTH_VALUE('+Inf'::Double, v1) OVER (PARTITION BY v1) FROM t1;
+-------------------------------------------------------------------------------------------------------------+
| nth_value(Utf8("+Inf"),t1.v1) PARTITION BY [t1.v1] ROWS BETWEEN UNBOUNDED PRECEDING AND UNBOUNDED FOLLOWING |
+-------------------------------------------------------------------------------------------------------------++-------------------------------------------------------------------------------------------------------------+0 row(s) fetched.
Elapsed 0.018 seconds.

After we insert a few values into t1 the nth_value aggregation will be computed. Only then do we see the expected error message,

>insert into t1 values (123), (456);
+-------+
| count |
+-------+
| 2 |
+-------+1 row(s) fetched.
Elapsed 0.007 seconds.
>SELECT NTH_VALUE('+Inf'::Double, v1) OVER (PARTITION BY v1) FROM t1;
This feature is not implemented: There is only support Literal types for field at idx: 1in Window Function

Planning divergence between built-in & user-defined window functions

In main branch where nth_value is a BuiltInWindowFunction the argument parsing fails early when mapping logical to physical plan. In explain output there is no physical plan.

> EXPLAIN SELECT NTH_VALUE('+Inf'::Double, v1) OVER (PARTITION BY v1) FROM t1;
+--------------+----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
| plan_type | plan |
+--------------+----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
| logical_plan | Projection: NTH_VALUE(Utf8("+Inf"),t1.v1) PARTITION BY [t1.v1] ROWS BETWEEN UNBOUNDED PRECEDING AND UNBOUNDED FOLLOWING |
| | WindowAggr: windowExpr=[[NTH_VALUE(Float64(inf), t1.v1) PARTITION BY [t1.v1] ROWS BETWEEN UNBOUNDED PRECEDING AND UNBOUNDED FOLLOWING AS NTH_VALUE(Utf8("+Inf"),t1.v1) PARTITION BY [t1.v1] ROWS BETWEEN UNBOUNDED PRECEDING AND UNBOUNDED FOLLOWING]] |
| | TableScan: t1 projection=[v1] |
+--------------+----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+1 row(s) fetched.
Elapsed 0.009 seconds.

But this is not the case for user-defined window functions. In this branch we instead see that a complete plan is built and failure is happening only when the query executes,

> EXPLAIN SELECT NTH_VALUE('+Inf'::Double, v1) OVER (PARTITION BY v1) FROM t1;
+---------------+-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
| plan_type | plan |
+---------------+-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
| logical_plan | Projection: nth_value(Utf8("+Inf"),t1.v1) PARTITION BY [t1.v1] ROWS BETWEEN UNBOUNDED PRECEDING AND UNBOUNDED FOLLOWING |
| | WindowAggr: windowExpr=[[nth_value(Float64(inf), t1.v1) PARTITION BY [t1.v1] ROWS BETWEEN UNBOUNDED PRECEDING AND UNBOUNDED FOLLOWING AS nth_value(Utf8("+Inf"),t1.v1) PARTITION BY [t1.v1] ROWS BETWEEN UNBOUNDED PRECEDING AND UNBOUNDED FOLLOWING]] |
| | TableScan: t1 projection=[v1] |
| physical_plan | ProjectionExec: expr=[nth_value(Utf8("+Inf"),t1.v1) PARTITION BY [t1.v1] ROWS BETWEEN UNBOUNDED PRECEDING AND UNBOUNDED FOLLOWING@1as nth_value(Utf8("+Inf"),t1.v1) PARTITION BY [t1.v1] ROWS BETWEEN UNBOUNDED PRECEDING AND UNBOUNDED FOLLOWING] |
| | WindowAggExec: wdw=[nth_value(Utf8("+Inf"),t1.v1) PARTITION BY [t1.v1] ROWS BETWEEN UNBOUNDED PRECEDING AND UNBOUNDED FOLLOWING: Ok(Field { name: "nth_value(Utf8(\"+Inf\"),t1.v1) PARTITION BY [t1.v1] ROWS BETWEEN UNBOUNDED PRECEDING AND UNBOUNDED FOLLOWING", data_type: Float64, nullable: true, dict_id: 0, dict_is_ordered: false, metadata: {} }), frame: WindowFrame { units: Rows, start_bound: Preceding(UInt64(NULL)), end_bound: Following(UInt64(NULL)), is_causal: false }] |
| | SortExec: expr=[v1@0ASC NULLS LAST], preserve_partitioning=[false] |
| | MemoryExec: partitions=1, partition_sizes=[0] |
| | |
+---------------+-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+2 row(s) fetched.
Elapsed 0.019 seconds.

buraksenn · 2024-11-04T20:17:34Z

@jcsherin thanks for the very detailed explanation. In this case, I think it would be better to update WindowUDFImpl in a followup PR for enhancement right? I can skip this test case in the scope of this PR. Correct me if I'm wrong please

jcsherin · 2024-11-05T03:28:42Z

I think it would be better to update WindowUDFImpl in a followup PR for enhancement right? I can skip this test case in the scope of this PR. Correct me if I'm wrong please

Sure, we can improve the API in another PR.

Here is a workaround that fixes the failing test:

// In datafusion/physical-plan/src/windows/mod.rsfncreate_udwf_window_expr(fun:&Arc<WindowUDF>,args:&[Arc<dynPhysicalExpr>],input_schema:&Schema,name:String,ignore_nulls:bool,) -> Result<Arc<dynBuiltInWindowFunctionExpr>>{// need to get the types into an owned vec for some reasonlet input_types:Vec<_> = args
.iter().map(|arg| arg.data_type(input_schema)).collect::<Result<_>>()?;let udwf_expr =
Arc::new(WindowUDFExpr{fun:Arc::clone(fun),args: args.to_vec(),
input_types,
name,is_reversed:false,
ignore_nulls,});/// Early validation of input expressions////// We create a partition evaluator because in the user-defined window/// implementation this is where code for parsing input expressions/// exist. The benefits are:/// - If any of the input expressions are invalid we catch them early/// in the planning phase, rather than during execution./// - Maintains compatibility with built-in (now removed) window/// functions validation behavior./// - Predictable and reliable error handling.////// See discussion here:/// https://github.com/apache/datafusion/pull/13201#issuecomment-2454209975let _ = udwf_expr.create_evaluator()?;Ok(udwf_expr)}

I verified that this works in your branch.

DataFusion CLI v42.2.0> CREATE TABLE t1(v1 BIGINT);
0 row(s) fetched.
Elapsed 0.019 seconds.
> EXPLAIN SELECT NTH_VALUE('+Inf'::Double, v1) OVER (PARTITION BY v1) FROM t1;
+--------------+----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
| plan_type | plan |
+--------------+----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
| logical_plan | Projection: nth_value(Utf8("+Inf"),t1.v1) PARTITION BY [t1.v1] ROWS BETWEEN UNBOUNDED PRECEDING AND UNBOUNDED FOLLOWING |
| | WindowAggr: windowExpr=[[nth_value(Float64(inf), t1.v1) PARTITION BY [t1.v1] ROWS BETWEEN UNBOUNDED PRECEDING AND UNBOUNDED FOLLOWING AS nth_value(Utf8("+Inf"),t1.v1) PARTITION BY [t1.v1] ROWS BETWEEN UNBOUNDED PRECEDING AND UNBOUNDED FOLLOWING]] |
| | TableScan: t1 projection=[v1] |
+--------------+----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+1 row(s) fetched.
Elapsed 0.018 seconds.
>SELECT NTH_VALUE('+Inf'::Double, v1) OVER (PARTITION BY v1) FROM t1;
This feature is not implemented: There is only support Literal types for field at idx: 1in Window Function

This workaround may not be ideal, but at least we do not have to skip this test. Also please feel free to update the code/comments as you see fit.

jcsherin · 2024-11-11T12:23:50Z

+
+/// Create an expression to represent the `nth_value` window function
+///
+pub fn nth_value(arg: datafusion_expr::Expr, n: Option<i64>) -> datafusion_expr::Expr {


The type of n is i64, not Option<i64>.
See the rust docs: https://docs.rs/datafusion/latest/datafusion/logical_expr/window_function/fn.nth_value.html

Also add a roundtrip logical plan test for this API here:
datafusion/datafusion/proto/tests/cases/roundtrip_logical_plan.rs
Line 943 in ac79ef3
cume_dist(),

Fixed and added test

jcsherin · 2024-11-11T13:55:55Z

@buraksenn Tremendous effort 🙌. These changes look good to me.

Co-authored-by: Sherin Jacob <jacob@protoship.io>

jcsherin · 2024-11-11T15:23:07Z

@buraksenn and @berkaysynnada Thanks!

@alamb This PR is ready.

alamb · 2024-11-13T06:56:14Z

Awesome -- thank you so much. I will review this PR hopefully later today

alamb

Thank you so much @buraksenn , @jcsherin -- it is just so beautiful to see this PR now after all the work. It is basically perfect from my perspective 🏆

alamb · 2024-11-13T15:31:36Z

 [dev-dependencies]
 criterion = { version = "0.5", features = ["async_futures"] }
 datafusion-functions-aggregate = { workspace = true }
+datafusion-functions-window = { workspace = true }


some day I hope we can remove these dependencies (so we can make testing physical-plan faster, but not a part of this PR

alamb · 2024-11-13T15:32:32Z

+ // We create a partition evaluator because in the user-defined window
+ // implementation this is where code for parsing input expressions
+ // exist. The benefits are:
+ // - If any of the input expressions are invalid we catch them early


💯 for these comments that explain the rationale

alamb · 2024-11-13T15:39:16Z

I also took the liberty of merging up from main to make sure we haven't hit any logical conflicts with this PR

alamb · 2024-11-13T16:10:32Z

I don't think there is any reason to wait around for this PR -- people know it is coming, so let's get this in 🚀

…pache#13201) * refactored nth_value * continue * test * proto and rustlint * fix datatype * cont * cont * apply jcsherins early validation * docs * doc * Apply suggestions from code review Co-authored-by: Sherin Jacob <jacob@protoship.io> * passes lint but does not have tests * continue * Update roundtrip_physical_plan.rs * udwf, not udaf * fix bounded but not fixed roundtrip * added * Update datafusion/sqllogictest/test_files/errors.slt Co-authored-by: Sherin Jacob <jacob@protoship.io> --------- Co-authored-by: Sherin Jacob <jacob@protoship.io> Co-authored-by: berkaysynnada <berkay.sahin@synnada.ai> Co-authored-by: Andrew Lamb <andrew@nerdnetworks.org>

…pache#13201) * refactored nth_value * continue * test * proto and rustlint * fix datatype * cont * cont * apply jcsherins early validation * docs * doc * Apply suggestions from code review Co-authored-by: Sherin Jacob <jacob@protoship.io> * passes lint but does not have tests * continue * Update roundtrip_physical_plan.rs * udwf, not udaf * fix bounded but not fixed roundtrip * added * Update datafusion/sqllogictest/test_files/errors.slt Co-authored-by: Sherin Jacob <jacob@protoship.io> --------- Co-authored-by: Sherin Jacob <jacob@protoship.io> Co-authored-by: berkaysynnada <berkay.sahin@synnada.ai> Co-authored-by: Andrew Lamb <andrew@nerdnetworks.org> (cherry picked from commit 54ab128)

…pache#13201) v44 * refactored nth_value * continue * test * proto and rustlint * fix datatype * cont * cont * apply jcsherins early validation * docs * doc * Apply suggestions from code review Co-authored-by: Sherin Jacob <jacob@protoship.io> * passes lint but does not have tests * continue * Update roundtrip_physical_plan.rs * udwf, not udaf * fix bounded but not fixed roundtrip * added * Update datafusion/sqllogictest/test_files/errors.slt Co-authored-by: Sherin Jacob <jacob@protoship.io> --------- Co-authored-by: Sherin Jacob <jacob@protoship.io> Co-authored-by: berkaysynnada <berkay.sahin@synnada.ai> Co-authored-by: Andrew Lamb <andrew@nerdnetworks.org> (cherry picked from commit 54ab128)

refactored nth_value
014306d

github-actionsBot added sql SQL Planner logical-expr Logical plan and expressions physical-expr Changes to the physical-expr crates core Core DataFusion crate sqllogictest SQL Logic Tests (.slt) substrait Changes to the substrait crate proto Related to proto crate labels Oct 31, 2024

buraksenn added 2 commits November 3, 2024 17:11

merge main
b61b2fc

continue
fda6a6f

buraksenn force-pushed the nth_value_func_ref branch from 8cf82f0 to fda6a6fCompare November 3, 2024 20:18

github-actionsBot removed the sqllogictest SQL Logic Tests (.slt) label Nov 3, 2024

test
7fb84fa

github-actionsBot added the sqllogictest SQL Logic Tests (.slt) label Nov 3, 2024

proto and rustlint
e6ef6ed

buraksenn added 3 commits November 4, 2024 12:14

fix datatype
1131df0

cont
f28163a

cont
fddbc58

buraksenn force-pushed the nth_value_func_ref branch from 0706334 to fddbc58Compare November 4, 2024 09:32

buraksenn added 2 commits November 5, 2024 08:50

apply jcsherins early validation
c406505

docs
12c0a52

fix bounded but not fixed roundtrip
9bbeef4

buraksenn commented Nov 11, 2024
View reviewed changes

Comment threaddatafusion/proto/src/physical_plan/to_proto.rs

jcsherin reviewed Nov 11, 2024
View reviewed changes

added
9e7e5e2

jcsherin reviewed Nov 11, 2024
View reviewed changes

Comment threaddatafusion/proto/tests/cases/roundtrip_logical_plan.rs

Merge branch 'main' into nth_value_func_ref
7a8a760

jcsherin reviewed Nov 11, 2024
View reviewed changes

Comment threaddatafusion/sqllogictest/test_files/errors.slt Outdated

Update datafusion/sqllogictest/test_files/errors.slt
073d8e8
Co-authored-by: Sherin Jacob <jacob@protoship.io>

alamb mentioned this pull request Nov 13, 2024
Review Backlog and Plan - Andrew Lamb - Nov 2024 #13386
Closed

alamb changed the title ~~Convert nth_value builtIn function to UDWF~~ Convert nth_value builtIn function to User Defined Window FunctionNov 13, 2024

alamb approved these changes Nov 13, 2024
View reviewed changes

alamb added the api change Changes the API exposed to users of the crate label Nov 13, 2024

alamb changed the title ~~Convert nth_value builtIn function to User Defined Window Function~~Convert nth_value builtIn function to User Defined Window FunctionNov 13, 2024

Merge remote-tracking branch 'apache/main' into nth_value_func_ref
6ea06f7

alamb mentioned this pull request Nov 13, 2024
Remove BuiltInWindowFunction (LogicalPlans) #13393
Merged

alamb merged commit 54ab128 into apache:mainNov 13, 2024

This was referenced Nov 13, 2024
Remove old static Window Functions page #13171
Closed
Error when serializing physical window functions to proto #13401
Closed

jcsherin mentioned this pull request Nov 14, 2024
fix: serialize user-defined window functions to proto #13421
Merged

This was referenced Dec 9, 2024
Write "upgrade guide" for DataFusion 44.0.0 #13702
Closed
chore: reinstate find_df_window_func #13708
Closed

buraksenn deleted the nth_value_func_ref branch August 11, 2026 11:01

Uh oh!

Conversation

buraksenn commented Oct 31, 2024• edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Which issue does this PR close?

Rationale for this change

What changes are included in this PR?

Are these changes tested?

Are there any user-facing changes?

Uh oh!

alamb commented Oct 31, 2024

Uh oh!

alamb commented Oct 31, 2024

Uh oh!

buraksenn commented Nov 1, 2024

Uh oh!

buraksenn commented Nov 3, 2024

Uh oh!

jcsherin commented Nov 4, 2024• edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

buraksenn commented Nov 4, 2024

Uh oh!

jcsherin commented Nov 4, 2024• edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

TL;DR

Edge Case: Empty Table

Planning divergence between built-in & user-defined window functions

Uh oh!

buraksenn commented Nov 4, 2024

Uh oh!

jcsherin commented Nov 5, 2024• edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Uh oh!

jcsherinNov 11, 2024

Choose a reason for hiding this comment

Uh oh!

jcsherinNov 11, 2024

Choose a reason for hiding this comment

Uh oh!

buraksennNov 11, 2024

Choose a reason for hiding this comment

Uh oh!

Uh oh!

jcsherin commented Nov 11, 2024

Uh oh!

Uh oh!

jcsherin commented Nov 11, 2024

Uh oh!

alamb commented Nov 13, 2024

Uh oh!

alamb left a comment

Choose a reason for hiding this comment

Uh oh!

alambNov 13, 2024

Choose a reason for hiding this comment

Uh oh!

alambNov 13, 2024

Choose a reason for hiding this comment

Uh oh!

alamb commented Nov 13, 2024

Uh oh!

alamb commented Nov 13, 2024

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

buraksenn commented Oct 31, 2024•
edited
Loading

jcsherin commented Nov 4, 2024•
edited
Loading

jcsherin commented Nov 4, 2024•
edited
Loading

jcsherin commented Nov 5, 2024•
edited
Loading