Do not use squashing for materialized views (fixes excessive memory usage)

src/Processors/Transforms/buildPushingToViewsChain.cpp

@@ -297,7 +297,11 @@ Chain buildPushingToViewsChain(
                     insert_columns.emplace_back(column.name);
             }
 
-            InterpreterInsertQuery interpreter(nullptr, insert_context, false, false, false);
+            /// Create interpreter w/o squashing
+            /// since we have squashing (SquashingChunksTransform) for each block.
+            InterpreterInsertQuery interpreter(nullptr, insert_context,
+                /* allow_materialized_= */ false,
+                /* no_squash_= */ true);
             out = interpreter.buildChain(inner_table, inner_metadata_snapshot, insert_columns, view_thread_status, view_counter_ms);
             out.addStorageHolder(dependent_table);
             out.addStorageHolder(inner_table);

tests/queries/0_stateless/02223_mv_aggregator_leak.sql

@@ -0,0 +1,18 @@
+drop table if exists mv_02223;
+drop table if exists in_02223;
+drop table if exists out_02223;
+
+create table in_02223 (key UInt64) engine=Null();
+create table out_02223 (keys AggregateFunction(uniqExact, String)) engine=Null();
+
+create materialized view mv_02223 to out_02223 as select uniqExactState(toString(key)) as keys from in_02223 group by intDiv(key, 1000);
+
+-- Here SET is used (over SETTINGS in INSERT) to test it on 19.x
+-- since the issue was introduced in 19.11.12.69
+-- https://github.com/ClickHouse/ClickHouse/pull/3796
+set max_memory_usage=200000000;
+insert into in_02223 select * from numbers(10000000);
+
+drop table mv_02223;
+drop table out_02223;
+drop table in_02223;