diff --git a/dbcon/joblist/tuple-bps.cpp b/dbcon/joblist/tuple-bps.cpp
index 8dd18b464..4785916a1 100644
--- a/dbcon/joblist/tuple-bps.cpp
+++ b/dbcon/joblist/tuple-bps.cpp
@@ -281,7 +281,7 @@ uint64_t TupleBPS::JoinLocalData::generateJoinResultSet(const uint32_t depth,
         {
           // Don't wait for memory, just send the data on to DL.
           RowGroup out(local_outputRG);
-          if (fe2 && tbps->runFEonPM())
+          if (fe2 && !tbps->runFEonPM())
           {
             processFE2(outputData);
             tbps->rgDataVecToDl(outputData, local_fe2Output, dlp);
diff --git a/dbcon/joblist/tuplehashjoin.cpp b/dbcon/joblist/tuplehashjoin.cpp
index 4d9959d8d..0e577beab 100644
--- a/dbcon/joblist/tuplehashjoin.cpp
+++ b/dbcon/joblist/tuplehashjoin.cpp
@@ -1892,6 +1892,19 @@ void TupleHashJoinStep::generateJoinResultSet(const vector<vector<Row::Pointer>
         // Count the memory
         if (UNLIKELY(!getMemory(l_outputRG.getMaxDataSize())))
         {
+          // MCOL-5512
+          if (fe2)
+          {
+            RowGroup l_fe2RG;
+            Row fe2InRow;
+            Row fe2OutRow;
+
+            l_fe2RG = fe2Output;
+            l_outputRG.initRow(&fe2InRow);
+            l_fe2RG.initRow(&fe2OutRow);
+
+            processFE2(l_outputRG, l_fe2RG, fe2InRow, fe2OutRow, &outputData, fe2.get());
+          }
           // Don't let the join results buffer get out of control.
           sendResult(outputData);
           outputData.clear();