make up some comments

Signed-off-by: sperlingxx <lovedreamf@gmail.com>

integration_tests/src/main/python/hash_aggregate_test.py

@@ -399,6 +399,9 @@ def test_hash_reduction_pivot_without_nans(data_gen, conf):
     ('c', LongRangeGen())] for value_gen in _repeat_agg_column_for_collect_op
 ]
 
+# We wrapped sort_array functions on collect_list/collect_set because the orders of collected lists/sets are not
+# deterministic. The annotation `ignore_order` only affects on the order between rows, while with collect ops we also
+# need to guarantee the consistency of the row-wise order (the orders within each array produced by collect ops).
 @approximate_float
 @ignore_order(local=True)
 @incompat

sql-plugin/src/main/scala/com/nvidia/spark/rapids/aggregate.scala

@@ -981,6 +981,13 @@ abstract class GpuTypedImperativeSupportedAggregateExecMeta[INPUT <: SparkPlan](
     // are inconsistent with CPU buffers. Therefore, we have to fall back all Aggregate stages
     // to CPU once any of them did fallback, in order to guarantee no partial-accelerated
     // TypedImperativeAggregate function.
+    //
+    // This fallback procedure adapts AQE. As what GpuExchanges do, it leverages the
+    // `gpuSupportedTag` to store the information about whether instances are GPU-supported
+    // or not, which is produced by the side effect of `willNotWorkOnGpu`. When AQE is on,
+    // during the preparation stage, there will be a run of GpuOverrides on the entire plan to
+    // trigger these side effects if necessary, before AQE splits the entire query into several
+    // query stages.
     GpuTypedImperativeSupportedAggregateExecMeta.checkAndFallbackEntirely(this)
   }