diff --git a/velox/exec/HashBuild.cpp b/velox/exec/HashBuild.cpp
index 550d43de2b17..d88b13a60ebb 100644
--- a/velox/exec/HashBuild.cpp
+++ b/velox/exec/HashBuild.cpp
@@ -1094,10 +1094,24 @@ void HashBuild::reclaim(
   VELOX_CHECK(canReclaim());
   auto* driver = operatorCtx_->driver();
   VELOX_CHECK_NOT_NULL(driver);
-  VELOX_CHECK(!nonReclaimableSection_);
 
   TestValue::adjust("facebook::velox::exec::HashBuild::reclaim", this);
 
+  const auto& task = driver->task();
+  const std::vector<Operator*> operators =
+      task->findPeerOperators(operatorCtx_->driverCtx()->pipelineId, this);
+  // Worst case scenario reservation was performed in ensureTableFits() when
+  // accounting for NextRowVector for duplicated rows, i.e. we assume every
+  // single row has duplicates. That is normally not the case. So when the
+  // query is under memory pressure, the excessive (in most cases) reservations
+  // can be returned.
+  for (auto i = 0; i <= operators.size(); i++) {
+    auto* memoryPool = i == 0 ? pool() : operators[i - 1]->pool();
+    const auto oldReservedBytes = memoryPool->reservedBytes();
+    memoryPool->release();
+    stats.reclaimedBytes += (oldReservedBytes - memoryPool->reservedBytes());
+  }
+
   if (exceededMaxSpillLevelLimit_) {
     return;
   }
@@ -1120,10 +1134,7 @@ void HashBuild::reclaim(
     return;
   }
 
-  const auto& task = driver->task();
   VELOX_CHECK(task->pauseRequested());
-  const std::vector<Operator*> operators =
-      task->findPeerOperators(operatorCtx_->driverCtx()->pipelineId, this);
   for (auto* op : operators) {
     HashBuild* buildOp = dynamic_cast<HashBuild*>(op);
     VELOX_CHECK_NOT_NULL(buildOp);
diff --git a/velox/exec/Operator.cpp b/velox/exec/Operator.cpp
index 449a2aaece91..9f46464cf2b5 100644
--- a/velox/exec/Operator.cpp
+++ b/velox/exec/Operator.cpp
@@ -643,8 +643,11 @@ uint64_t Operator::MemoryReclaimer::reclaim(
       "facebook::velox::exec::Operator::MemoryReclaimer::reclaim", pool);
 
   // NOTE: we can't reclaim memory from an operator which is under
+  // non-reclaimable section, except for HashBuild operator. If it is HashBuild
+  // operator, we allow it to enter HashBuild::reclaim because there is a good
+  // chance we can release some unused reserved memory even if it's in
   // non-reclaimable section.
-  if (op_->nonReclaimableSection_) {
+  if (op_->nonReclaimableSection_ && op_->operatorType() != "HashBuild") {
     // TODO: reduce the log frequency if it is too verbose.
     ++stats.numNonReclaimableAttempts;
     RECORD_METRIC_VALUE(kMetricMemoryNonReclaimableCount);
diff --git a/velox/exec/tests/HashJoinTest.cpp b/velox/exec/tests/HashJoinTest.cpp
index 74578b37ba26..37601f0f913c 100644
--- a/velox/exec/tests/HashJoinTest.cpp
+++ b/velox/exec/tests/HashJoinTest.cpp
@@ -5925,22 +5925,29 @@ DEBUG_ONLY_TEST_F(HashJoinTest, reclaimDuringAllocation) {
     ASSERT_EQ(reclaimable, enableSpilling);
     if (enableSpilling) {
       ASSERT_GE(reclaimableBytes, 0);
+      op->reclaim(
+          folly::Random::oneIn(2) ? 0 : folly::Random::rand32(),
+          reclaimerStats_);
     } else {
       ASSERT_EQ(reclaimableBytes, 0);
+      VELOX_ASSERT_THROW(
+          op->reclaim(
+              folly::Random::oneIn(2) ? 0 : folly::Random::rand32(),
+              reclaimerStats_),
+          "");
     }
-    VELOX_ASSERT_THROW(
-        op->reclaim(
-            folly::Random::oneIn(2) ? 0 : folly::Random::rand32(),
-            reclaimerStats_),
-        "");
 
     driverWait.notify();
     Task::resume(task);
     task.reset();
 
     taskThread.join();
+    if (enableSpilling) {
+      ASSERT_GT(reclaimerStats_.reclaimedBytes, 0);
+    } else {
+      ASSERT_EQ(reclaimerStats_, memory::MemoryReclaimer::Stats{0});
+    }
   }
-  ASSERT_EQ(reclaimerStats_, memory::MemoryReclaimer::Stats{0});
 }
 
 DEBUG_ONLY_TEST_F(HashJoinTest, reclaimDuringOutputProcessing) {