SERVER-70395 Change spilling for SBE HashAggStage to use a more efficient algorithm

The new algorithm spills the entire hash table to a
RecordStore whenever the memory budget is exceeded. Once all
the input is consumed, it switches to a streaming approach,
merging the partial aggregates recovered from disk.

1 files changed, 4 insertions, 2 deletions

diff --git a/jstests/noPassthrough/spill_to_disk_secondary_read.js b/jstests/noPassthrough/spill_to_disk_secondary_read.js
index 217ce510eef..8ab9178b4c5 100644
--- a/jstests/noPassthrough/spill_to_disk_secondary_read.js
+++ b/jstests/noPassthrough/spill_to_disk_secondary_read.js
@@ -91,8 +91,10 @@ const readColl = secondary.getDB("test").foo;
         assert(hashAggGroup.hasOwnProperty("usedDisk"), hashAggGroup);
         assert(hashAggGroup.usedDisk, hashAggGroup);
         assert.eq(hashAggGroup.spilledRecords, expectedSpilledRecords, hashAggGroup);
-        assert.gte(hashAggGroup.spilledBytesApprox, expectedSpilledBytesAtLeast, hashAggGroup);
-        assert.gt(hashAggGroup.spilledRecordEstimatedStorageSize, 0, hashAggGroup);
+        // We expect each record to be individually spilled, so the number of spill events and the
+        // number of spilled records should be equal.
+        assert.eq(hashAggGroup.numSpills, hashAggGroup.spilledRecords, hashAggGroup);
+        assert.gt(hashAggGroup.spilledDataStorageSize, expectedSpilledBytesAtLeast, hashAggGroup);
     } finally {
         assert.commandWorked(secondary.adminCommand({
             setParameter: 1,