[GitHub] [arrow] jorgecarleitao commented on a change in pull request #7984: ARROW-9779: [Rust] [DataFusion] Increase stability of average accumulator

[GitBox <gi...@apache.org>]

jorgecarleitao commented on a change in pull request #7984:
URL: https://github.com/apache/arrow/pull/7984#discussion_r472386086



##########
File path: rust/datafusion/src/execution/physical_plan/expressions.rs
##########
@@ -337,40 +336,86 @@ impl AggregateExpr for Avg {
     }
 }
 
-macro_rules! avg_accumulate {
+macro_rules! avg_accumulate_scalar {
     ($SELF:ident, $VALUE:expr, $ARRAY_TYPE:ident) => {{
-        match ($SELF.sum, $SELF.count) {
-            (Some(sum), Some(count)) => {
-                $SELF.sum = Some(sum + $VALUE as f64);
+        // details here: https://stackoverflow.com/a/23493727/931303
+        match ($SELF.avg, $SELF.count) {
+            (Some(avg), Some(count)) => {
                 $SELF.count = Some(count + 1);
+                let online = (avg * (count as f64) + $VALUE as f64) / (count + 1) as f64;
+                $SELF.avg = Some(online);
             }
             _ => {
-                $SELF.sum = Some($VALUE as f64);
                 $SELF.count = Some(1);
+                $SELF.avg = Some($VALUE as f64);
             }
         };
     }};
 }
+
+macro_rules! avg_accumulate_array {
+    ($SELF:ident, $ARRAY:expr, $ARRAY_TYPE:ident) => {{
+        // details here: https://stackoverflow.com/a/23493727/931303
+        let m = $ARRAY.len();
+        let sum = compute::sum($ARRAY.as_any().downcast_ref::<$ARRAY_TYPE>().unwrap());

Review comment:
       It is less likely to overflow over a single record batch than over a the set of all record batches in a single partition (over some very basic hypothesis over the statistical model behind the data).
   
   I am taking a text book authoritative argument here, as I do not have sufficient knowledge to take this debate to its details. I can see that spark [uses sum,count](https://github.com/apache/spark/blob/5264164a67df498b73facae207eda12ee133be7d/examples/src/main/java/org/apache/spark/examples/sql/JavaUserDefinedTypedAggregation.java#L66), so that is another argument (for what is worth).
   
   No specific preference here. I implemented it as an example of an aggregate UDF in one of my other PRs, and thought that it could be useful here.




----------------------------------------------------------------
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.

For queries about this service, please contact Infrastructure at:
users@infra.apache.org