diff --git a/velox/common/base/Scratch.h b/velox/common/base/Scratch.h
new file mode 100644
index 000000000000..a2110e01aaac
--- /dev/null
+++ b/velox/common/base/Scratch.h
@@ -0,0 +1,159 @@
+/*
+ * Copyright (c) Facebook, Inc. and its affiliates.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#pragma once
+
+#include "velox/common/base/RawVector.h"
+
+/// A utility for reusable scoped temporary scratch areas.
+namespace facebook::velox {
+
+/// A collection of temporary reusable scratch vectors. The vectors are accessed
+/// via the ScratchPtr scoped lease. The vectors are padded so that their last
+/// element can be written at full SIMD width, as with raw_vector.
+class Scratch {
+ public:
+  using Item = raw_vector<char>;
+
+  Scratch() = default;
+  Scratch(const Scratch& other) = delete;
+
+  ~Scratch() {
+    reserve(0);
+    ::free(items_);
+    items_ = nullptr;
+    capacity_ = 0;
+    fill_ = 0;
+  }
+  void operator=(const Scratch& other) = delete;
+
+  /// Returns the next reusable scratch vector or makes a new one.
+  Item get() {
+    if (fill_ == 0) {
+      return Item();
+    }
+    auto temp = std::move(items_[fill_ - 1]);
+    --fill_;
+    retainedSize_ -= temp.capacity();
+    return temp;
+  }
+
+  void release(Item&& item) {
+    retainedSize_ += item.capacity();
+    if (fill_ == capacity_) {
+      reserve(std::max(16, 2 * capacity_));
+    }
+    items_[fill_++] = std::move(item);
+  }
+
+  void trim() {
+    reserve(0);
+    retainedSize_ = 0;
+  }
+
+  size_t retainedSize() {
+    return retainedSize_;
+  }
+
+ private:
+  void reserve(int32_t newCapacity) {
+    VELOX_CHECK_LE(fill_, capacity_);
+    // Delete the items above the new capacity.
+    for (auto i = newCapacity; i < fill_; ++i) {
+      std::destroy_at(&items_[i]);
+    }
+    if (newCapacity > capacity_) {
+      Item* newItems =
+          reinterpret_cast<Item*>(::malloc(sizeof(Item) * newCapacity));
+      if (fill_ > 0) {
+        memcpy(newItems, items_, fill_ * sizeof(Item));
+      }
+      memset(newItems + fill_, 0, (newCapacity - fill_) * sizeof(Item));
+      free(items_);
+      items_ = newItems;
+      capacity_ = newCapacity;
+    }
+    fill_ = std::min(fill_, newCapacity);
+  }
+
+  Item* items_{nullptr};
+  int32_t fill_{0};
+  int32_t capacity_{0};
+  // The total size held. If too large from outlier use cases, 'this' should be
+  // trimmed.
+  int64_t retainedSize_{0};
+};
+
+/// A scoped lease for a scratch area of T. For scratch areas <=
+/// 'inlineSize' the scratch area is inlined, typically on stack, and
+/// no allocation will ever take place. The inline storage is padded
+/// with a trailer of simd::kPadding bytes to allow writing at full
+/// SIMD width at the end of the area.
+template <typename T, int32_t inlineSize = 0>
+class ScratchPtr {
+ public:
+  ScratchPtr(Scratch& scratch) : scratch_(&scratch) {}
+
+  ScratchPtr(const ScratchPtr& other) = delete;
+  ScratchPtr(ScratchPtr&& other) = delete;
+
+  inline ~ScratchPtr() {
+    if (data_.data()) {
+      scratch_->release(std::move(data_));
+    }
+  }
+
+  void operator=(ScratchPtr&& other) = delete;
+  void operator=(const ScratchPtr& other) = delete;
+
+  /// Returns a writable pointer to at least 'size' uninitialized
+  /// elements of T. The last element is followed by simd::kPadding
+  /// bytes to allow a full width SIMD store for any element. This may
+  /// be called once per lifetime.
+  T* get(int32_t size) {
+    VELOX_CHECK_NULL(ptr_);
+    size_ = size;
+    if (size <= inlineSize) {
+      ptr_ = inline_;
+      return ptr_;
+    }
+    data_ = scratch_->get();
+    data_.resize(size * sizeof(T));
+    ptr_ = reinterpret_cast<T*>(data_.data());
+    return ptr_;
+  }
+
+  /// Returns the pointer returned by a previous get(int32_t).
+  T* get() const {
+    VELOX_DCHECK_NOT_NULL(ptr_);
+    return ptr_;
+  }
+
+  /// Returns the size of the previous get(int32_t).
+  int32_t size() const {
+    return size_;
+  }
+
+ private:
+  Scratch* scratch_{nullptr};
+  raw_vector<char> data_;
+  T* ptr_{nullptr};
+  int32_t size_{0};
+  T inline_[inlineSize];
+  char padding_[inlineSize == 0 ? 0 : simd::kPadding];
+};
+
+} // namespace facebook::velox
diff --git a/velox/common/base/tests/CMakeLists.txt b/velox/common/base/tests/CMakeLists.txt
index ff97bbe63711..5b305719e1b4 100644
--- a/velox/common/base/tests/CMakeLists.txt
+++ b/velox/common/base/tests/CMakeLists.txt
@@ -22,6 +22,7 @@ add_executable(
   FsTest.cpp
   RangeTest.cpp
   RawVectorTest.cpp
+  ScratchTest.cpp
   RuntimeMetricsTest.cpp
   ScopedLockTest.cpp
   SemaphoreTest.cpp
diff --git a/velox/common/base/tests/ScratchTest.cpp b/velox/common/base/tests/ScratchTest.cpp
new file mode 100644
index 000000000000..36bdd2da2f81
--- /dev/null
+++ b/velox/common/base/tests/ScratchTest.cpp
@@ -0,0 +1,74 @@
+/*
+ * Copyright (c) Facebook, Inc. and its affiliates.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "velox/common/base/Scratch.h"
+
+#include <gtest/gtest.h>
+
+using namespace facebook::velox;
+
+TEST(ScratchTest, basic) {
+  Scratch scratch;
+  {
+    ScratchPtr<int32_t> ints(scratch);
+    ScratchPtr<int64_t> longs(scratch);
+    auto tempInts = ints.get(1000);
+    auto tempLongs = longs.get(2000);
+    std::fill(tempInts, tempInts + 1000, -1);
+    std::fill(tempLongs, tempLongs + 2000, -1);
+    EXPECT_EQ(0, scratch.retainedSize());
+  }
+  EXPECT_EQ(20352, scratch.retainedSize());
+  {
+    ScratchPtr<int32_t> ints(scratch);
+    ScratchPtr<int64_t> longs(scratch);
+    auto tempLongs = longs.get(2000);
+    auto tempInts = ints.get(1000);
+    std::fill(tempInts, tempInts + 1000, -1);
+    std::fill(tempInts, tempInts + 2000, -1);
+    EXPECT_EQ(0, scratch.retainedSize());
+  }
+  // The scratch vectors were acquired in a different order, so the smaller got
+  // resized to the larger size.
+  EXPECT_EQ(32640, scratch.retainedSize());
+  scratch.trim();
+  EXPECT_EQ(0, scratch.retainedSize());
+  {
+    ScratchPtr<int32_t, 10> ints(scratch);
+    // The size is the inline size, nothing gets returned to 'scratch'.
+    auto temp = ints.get(10);
+    temp[0] = 1;
+  }
+  EXPECT_EQ(0, scratch.retainedSize());
+}
+
+TEST(ScratchTest, large) {
+  constexpr int32_t kSize = 100;
+  Scratch scratch;
+  std::vector<std::unique_ptr<ScratchPtr<int32_t>>> pointers;
+  for (auto i = 0; i < kSize; ++i) {
+    pointers.push_back(std::make_unique<ScratchPtr<int32_t>>(scratch));
+    pointers.back()->get(1000);
+  }
+  pointers.clear();
+  // 100 times 1000 bytes returned.
+  EXPECT_LT(100'000, scratch.retainedSize());
+  for (auto i = 0; i < kSize; ++i) {
+    pointers.push_back(std::make_unique<ScratchPtr<int32_t>>(scratch));
+    pointers.back()->get(1000);
+  }
+  EXPECT_EQ(0, scratch.retainedSize());
+}