only read one chunk in mergetramsform when meet virtual row

src/Processors/Merges/IMergingTransform.cpp

@@ -1,3 +1,4 @@
+#include <Processors/Merges/Algorithms/MergeTreeReadInfo.h>
 #include <Processors/Merges/IMergingTransform.h>
 
 namespace DB
@@ -101,8 +102,10 @@ IProcessor::Status IMergingTransformBase::prepareInitializeInputs()
         /// setNotNeeded after reading first chunk, because in optimismtic case
         /// (e.g. with optimized 'ORDER BY primary_key LIMIT n' and small 'n')
         /// we won't have to read any chunks anymore;
-        auto chunk = input.pull(limit_hint != 0);
-        if ((limit_hint && chunk.getNumRows() < limit_hint) || always_read_till_end)
+        /// If virtual row exists, test it first, so don't read more chunks.
+        auto chunk = input.pull(true);
+        if ((limit_hint == 0 && !getVirtualRowFromChunk(chunk))
+            || (limit_hint && chunk.getNumRows() < limit_hint) || always_read_till_end)
             input.setNeeded();
 
         if (!chunk.hasRows())

tests/queries/0_stateless/03031_read_in_order_optimization.reference

@@ -2,4 +2,4 @@
 1
 2
 3
-24578
+16386

tests/queries/0_stateless/03031_read_in_order_optimization.sql

@@ -46,4 +46,24 @@ AND query not like '%system.query_log%'
 ORDER BY query_start_time DESC, read_rows DESC
 LIMIT 1;
 
+-- SELECT x
+-- FROM t
+-- ORDER BY x ASC
+-- LIMIT 4
+-- SETTINGS max_block_size = 8192,
+-- read_in_order_two_level_merge_threshold = 5, --avoid preliminary merge
+-- max_threads = 1,
+-- optimize_read_in_order = 1;
+
+-- SYSTEM FLUSH LOGS;
+
+-- -- without virtual row 16.38k, but with virtual row 24.58k, becasue read again (why?) in the non-target part after reading its virtual row and before sending the virtual row to the priority queue
+-- SELECT read_rows
+-- FROM system.query_log
+-- WHERE current_database = currentDatabase()
+-- AND query like '%SELECT x%'
+-- AND query not like '%system.query_log%'
+-- ORDER BY query_start_time DESC, read_rows DESC
+-- LIMIT 1;
+
 DROP TABLE t;