diff --git a/HISTORY.md b/HISTORY.md
index 2055e6d74..baf61150a 100644
--- a/HISTORY.md
+++ b/HISTORY.md
@@ -4,6 +4,7 @@
 * Fixed a bug where manual flush would block forever even though flush options had wait=false.
 * Fixed a bug where RocksDB could corrupt DBs with `avoid_flush_during_recovery == true` by removing valid WALs, leading to `Status::Corruption` with message like "SST file is ahead of WALs" when attempting to reopen.
 * Fixed a bug in async_io path where incorrect length of data is read by FilePrefetchBuffer if data is consumed from two populated buffers and request for more data is sent.
+* Fixed a CompactionFilter bug. Compaction filter used to use `Delete` to remove keys, even if the keys should be removed with `SingleDelete`. Mixing `Delete` and `SingleDelete` may cause undefined behavior.
 
 ### New Features
 * DB::GetLiveFilesStorageInfo is ready for production use.
@@ -12,6 +13,7 @@
 ### Public API changes
 * Add rollback_deletion_type_callback to TransactionDBOptions so that write-prepared transactions know whether to issue a Delete or SingleDelete to cancel a previous key written during prior prepare phase. The PR aims to prevent mixing SingleDeletes and Deletes for the same key that can lead to undefined behaviors for write-prepared transactions.
 * EXPERIMENTAL: Add new API AbortIO in file_system to abort the read requests submitted asynchronously.
+* CompactionFilter::Decision has a new value: kRemoveWithSingleDelete. If CompactionFilter returns this decision, then CompactionIterator will use `SingleDelete` to mark a key as removed.
 
 ### Bug Fixes
 * RocksDB calls FileSystem::Poll API during FilePrefetchBuffer destruction which impacts performance as it waits for read requets completion which is not needed anymore. Calling FileSystem::AbortIO to abort those requests instead fixes that performance issue.
diff --git a/db/compaction/compaction_iterator.cc b/db/compaction/compaction_iterator.cc
index 9f5bc4bc8..be7b94997 100644
--- a/db/compaction/compaction_iterator.cc
+++ b/db/compaction/compaction_iterator.cc
@@ -307,6 +307,14 @@ bool CompactionIterator::InvokeFilterIfNeeded(bool* need_skip,
     // no value associated with delete
     value_.clear();
     iter_stats_.num_record_drop_user++;
+  } else if (filter == CompactionFilter::Decision::kRemoveWithSingleDelete) {
+    // convert the current key to a single delete; key_ is pointing into
+    // current_key_ at this point, so updating current_key_ updates key()
+    ikey_.type = kTypeSingleDeletion;
+    current_key_.UpdateInternalKey(ikey_.sequence, kTypeSingleDeletion);
+    // no value associated with single delete
+    value_.clear();
+    iter_stats_.num_record_drop_user++;
   } else if (filter == CompactionFilter::Decision::kChangeValue) {
     if (ikey_.type == kTypeBlobIndex) {
       // value transfer from blob file to inlined data
diff --git a/db/db_compaction_filter_test.cc b/db/db_compaction_filter_test.cc
index abf8b6c8f..5a3f27d12 100644
--- a/db/db_compaction_filter_test.cc
+++ b/db/db_compaction_filter_test.cc
@@ -968,6 +968,71 @@ TEST_F(DBTestCompactionFilter, IgnoreSnapshotsFalseRecovery) {
   ASSERT_TRUE(TryReopen(options).IsNotSupported());
 }
 
+TEST_F(DBTestCompactionFilter, DropKeyWithSingleDelete) {
+  Options options = GetDefaultOptions();
+  options.create_if_missing = true;
+
+  Reopen(options);
+
+  ASSERT_OK(Put("a", "v0"));
+  ASSERT_OK(Put("b", "v0"));
+  const Snapshot* snapshot = db_->GetSnapshot();
+
+  ASSERT_OK(SingleDelete("b"));
+  ASSERT_OK(Flush());
+
+  {
+    CompactRangeOptions cro;
+    cro.change_level = true;
+    cro.target_level = options.num_levels - 1;
+    ASSERT_OK(db_->CompactRange(cro, nullptr, nullptr));
+  }
+
+  db_->ReleaseSnapshot(snapshot);
+  Close();
+
+  class DeleteFilterV2 : public CompactionFilter {
+   public:
+    Decision FilterV2(int /*level*/, const Slice& key, ValueType /*value_type*/,
+                      const Slice& /*existing_value*/,
+                      std::string* /*new_value*/,
+                      std::string* /*skip_until*/) const override {
+      if (key.starts_with("b")) {
+        return Decision::kRemoveWithSingleDelete;
+      }
+      return Decision::kRemove;
+    }
+
+    const char* Name() const override { return "DeleteFilterV2"; }
+  } delete_filter_v2;
+
+  options.compaction_filter = &delete_filter_v2;
+  options.level0_file_num_compaction_trigger = 2;
+  Reopen(options);
+
+  ASSERT_OK(Put("b", "v1"));
+  ASSERT_OK(Put("x", "v1"));
+  ASSERT_OK(Flush());
+
+  ASSERT_OK(Put("r", "v1"));
+  ASSERT_OK(Put("z", "v1"));
+  ASSERT_OK(Flush());
+
+  ASSERT_OK(dbfull()->TEST_WaitForCompact());
+
+  Close();
+
+  options.compaction_filter = nullptr;
+  Reopen(options);
+  ASSERT_OK(SingleDelete("b"));
+  ASSERT_OK(Flush());
+  {
+    CompactRangeOptions cro;
+    cro.bottommost_level_compaction = BottommostLevelCompaction::kForce;
+    ASSERT_OK(db_->CompactRange(cro, nullptr, nullptr));
+  }
+}
+
 }  // namespace ROCKSDB_NAMESPACE
 
 int main(int argc, char** argv) {
diff --git a/db_stress_tool/db_stress_compaction_filter.h b/db_stress_tool/db_stress_compaction_filter.h
index f55062e0a..c967622db 100644
--- a/db_stress_tool/db_stress_compaction_filter.h
+++ b/db_stress_tool/db_stress_compaction_filter.h
@@ -46,11 +46,13 @@ class DbStressCompactionFilter : public CompactionFilter {
     // Reaching here means we acquired the lock.
 
     bool key_exists = state_->Exists(cf_id_, key_num);
+    const bool allow_overwrite = state_->AllowsOverwrite(key_num);
 
     key_mutex->Unlock();
 
     if (!key_exists) {
-      return Decision::kRemove;
+      return allow_overwrite ? Decision::kRemove
+                             : Decision::kRemoveWithSingleDelete;
     }
     return Decision::kKeep;
   }
diff --git a/db_stress_tool/db_stress_driver.cc b/db_stress_tool/db_stress_driver.cc
index 5cc5ca8dc..a56454abe 100644
--- a/db_stress_tool/db_stress_driver.cc
+++ b/db_stress_tool/db_stress_driver.cc
@@ -58,8 +58,8 @@ void ThreadBody(void* v) {
 
 bool RunStressTest(StressTest* stress) {
   SystemClock* clock = db_stress_env->GetSystemClock().get();
-  stress->InitDb();
   SharedState shared(db_stress_env, stress);
+  stress->InitDb(&shared);
   stress->FinishInitDb(&shared);
 
 #ifndef NDEBUG
diff --git a/db_stress_tool/db_stress_shared_state.cc b/db_stress_tool/db_stress_shared_state.cc
index bc1363504..078cc2fef 100644
--- a/db_stress_tool/db_stress_shared_state.cc
+++ b/db_stress_tool/db_stress_shared_state.cc
@@ -12,8 +12,6 @@
 #include "db_stress_tool/db_stress_shared_state.h"
 
 namespace ROCKSDB_NAMESPACE {
-const uint32_t SharedState::UNKNOWN_SENTINEL = 0xfffffffe;
-const uint32_t SharedState::DELETION_SENTINEL = 0xffffffff;
 #if defined(ROCKSDB_SUPPORT_THREAD_LOCAL)
 #if defined(OS_SOLARIS)
 __thread bool SharedState::ignore_read_error;
diff --git a/db_stress_tool/db_stress_shared_state.h b/db_stress_tool/db_stress_shared_state.h
index e7eb4aade..b05144548 100644
--- a/db_stress_tool/db_stress_shared_state.h
+++ b/db_stress_tool/db_stress_shared_state.h
@@ -45,9 +45,9 @@ class SharedState {
  public:
   // indicates a key may have any value (or not be present) as an operation on
   // it is incomplete.
-  static const uint32_t UNKNOWN_SENTINEL;
+  static constexpr uint32_t UNKNOWN_SENTINEL = 0xfffffffe;
   // indicates a key should definitely be deleted
-  static const uint32_t DELETION_SENTINEL;
+  static constexpr uint32_t DELETION_SENTINEL = 0xffffffff;
 
   // Errors when reading filter blocks are ignored, so we use a thread
   // local variable updated via sync points to keep track of errors injected
@@ -81,36 +81,9 @@ class SharedState {
         stress_test_(stress_test),
         verification_failure_(false),
         should_stop_test_(false),
-        no_overwrite_ids_(FLAGS_column_families),
+        no_overwrite_ids_(GenerateNoOverwriteIds()),
         expected_state_manager_(nullptr),
         printing_verification_results_(false) {
-    // Pick random keys in each column family that will not experience
-    // overwrite
-
-    fprintf(stdout, "Choosing random keys with no overwrite\n");
-    Random64 rnd(seed_);
-    // Start with the identity permutation. Subsequent iterations of
-    // for loop below will start with perm of previous for loop
-    int64_t* permutation = new int64_t[max_key_];
-    for (int64_t i = 0; i < max_key_; i++) {
-      permutation[i] = i;
-    }
-    // Now do the Knuth shuffle
-    int64_t num_no_overwrite_keys = (max_key_ * FLAGS_nooverwritepercent) / 100;
-    // Only need to figure out first num_no_overwrite_keys of permutation
-    no_overwrite_ids_.reserve(num_no_overwrite_keys);
-    for (int64_t i = 0; i < num_no_overwrite_keys; i++) {
-      int64_t rand_index = i + rnd.Next() % (max_key_ - i);
-      // Swap i and rand_index;
-      int64_t temp = permutation[i];
-      permutation[i] = permutation[rand_index];
-      permutation[rand_index] = temp;
-      // Fill no_overwrite_ids_ with the first num_no_overwrite_keys of
-      // permutation
-      no_overwrite_ids_.insert(permutation[i]);
-    }
-    delete[] permutation;
-
     Status status;
     // TODO: We should introduce a way to explicitly disable verification
     // during shutdown. When that is disabled and FLAGS_expected_values_dir
@@ -293,7 +266,7 @@ class SharedState {
                                                 pending);
   }
 
-  bool AllowsOverwrite(int64_t key) {
+  bool AllowsOverwrite(int64_t key) const {
     return no_overwrite_ids_.find(key) == no_overwrite_ids_.end();
   }
 
@@ -335,6 +308,36 @@ class SharedState {
     ignore_read_error = true;
   }
 
+  // Pick random keys in each column family that will not experience overwrite.
+  std::unordered_set<int64_t> GenerateNoOverwriteIds() const {
+    fprintf(stdout, "Choosing random keys with no overwrite\n");
+    // Start with the identity permutation. Subsequent iterations of
+    // for loop below will start with perm of previous for loop
+    std::vector<int64_t> permutation(max_key_);
+    for (int64_t i = 0; i < max_key_; ++i) {
+      permutation[i] = i;
+    }
+    // Now do the Knuth shuffle
+    const int64_t num_no_overwrite_keys =
+        (max_key_ * FLAGS_nooverwritepercent) / 100;
+    // Only need to figure out first num_no_overwrite_keys of permutation
+    std::unordered_set<int64_t> ret;
+    ret.reserve(num_no_overwrite_keys);
+    Random64 rnd(seed_);
+    for (int64_t i = 0; i < num_no_overwrite_keys; i++) {
+      assert(i < max_key_);
+      int64_t rand_index = i + rnd.Next() % (max_key_ - i);
+      // Swap i and rand_index;
+      int64_t temp = permutation[i];
+      permutation[i] = permutation[rand_index];
+      permutation[rand_index] = temp;
+      // Fill no_overwrite_ids_ with the first num_no_overwrite_keys of
+      // permutation
+      ret.insert(permutation[i]);
+    }
+    return ret;
+  }
+
   port::Mutex mu_;
   port::CondVar cv_;
   const uint32_t seed_;
@@ -355,7 +358,7 @@ class SharedState {
   std::atomic<bool> should_stop_test_;
 
   // Keys that should not be overwritten
-  std::unordered_set<size_t> no_overwrite_ids_;
+  const std::unordered_set<int64_t> no_overwrite_ids_;
 
   std::unique_ptr<ExpectedStateManager> expected_state_manager_;
   // Cannot store `port::Mutex` directly in vector since it is not copyable
diff --git a/db_stress_tool/db_stress_test_base.cc b/db_stress_tool/db_stress_test_base.cc
index 443b825bc..37835bb21 100644
--- a/db_stress_tool/db_stress_test_base.cc
+++ b/db_stress_tool/db_stress_test_base.cc
@@ -281,12 +281,12 @@ bool StressTest::BuildOptionsTable() {
   return true;
 }
 
-void StressTest::InitDb() {
+void StressTest::InitDb(SharedState* shared) {
   uint64_t now = clock_->NowMicros();
   fprintf(stdout, "%s Initializing db_stress\n",
           clock_->TimeToString(now / 1000000).c_str());
   PrintEnv();
-  Open();
+  Open(shared);
   BuildOptionsTable();
 }
 
@@ -568,7 +568,7 @@ void StressTest::PreloadDbAndReopenAsReadOnly(int64_t number_of_keys,
     fprintf(stdout, "%s Reopening database in read-only\n",
             clock_->TimeToString(now / 1000000).c_str());
     // Reopen as read-only, can ignore all options related to updates
-    Open();
+    Open(shared);
   } else {
     fprintf(stderr, "Failed to preload db");
     exit(1);
@@ -2302,10 +2302,12 @@ void StressTest::PrintEnv() const {
   fprintf(stdout, "------------------------------------------------\n");
 }
 
-void StressTest::Open() {
+void StressTest::Open(SharedState* shared) {
   assert(db_ == nullptr);
 #ifndef ROCKSDB_LITE
   assert(txn_db_ == nullptr);
+#else
+  (void)shared;
 #endif
   if (FLAGS_options_file.empty()) {
     BlockBasedTableOptions block_based_options;
@@ -2752,7 +2754,7 @@ void StressTest::Open() {
           static_cast<size_t>(FLAGS_wp_snapshot_cache_bits);
       txn_db_options.wp_commit_cache_bits =
           static_cast<size_t>(FLAGS_wp_commit_cache_bits);
-      PrepareTxnDbOptions(txn_db_options);
+      PrepareTxnDbOptions(shared, txn_db_options);
       s = TransactionDB::Open(options_, txn_db_options, FLAGS_db,
                               cf_descriptors, &column_families_, &txn_db_);
       if (!s.ok()) {
@@ -2912,7 +2914,7 @@ void StressTest::Reopen(ThreadState* thread) {
   auto now = clock_->NowMicros();
   fprintf(stdout, "%s Reopening database for the %dth time\n",
           clock_->TimeToString(now / 1000000).c_str(), num_times_reopened_);
-  Open();
+  Open(thread->shared);
 
   if ((FLAGS_sync_fault_injection || FLAGS_disable_wal) && IsStateTracked()) {
     Status s = thread->shared->SaveAtAndAfter(db_);
diff --git a/db_stress_tool/db_stress_test_base.h b/db_stress_tool/db_stress_test_base.h
index ee37eb101..2307e1261 100644
--- a/db_stress_tool/db_stress_test_base.h
+++ b/db_stress_tool/db_stress_test_base.h
@@ -30,7 +30,7 @@ class StressTest {
 
   bool BuildOptionsTable();
 
-  void InitDb();
+  void InitDb(SharedState*);
   // The initialization work is split into two parts to avoid a circular
   // dependency with `SharedState`.
   virtual void FinishInitDb(SharedState*);
@@ -219,7 +219,7 @@ class StressTest {
 
   void PrintEnv() const;
 
-  void Open();
+  void Open(SharedState* shared);
 
   void Reopen(ThreadState* thread);
 
@@ -228,7 +228,8 @@ class StressTest {
   virtual void RegisterAdditionalListeners() {}
 
 #ifndef ROCKSDB_LITE
-  virtual void PrepareTxnDbOptions(TransactionDBOptions& /*txn_db_opts*/) {}
+  virtual void PrepareTxnDbOptions(SharedState* /*shared*/,
+                                   TransactionDBOptions& /*txn_db_opts*/) {}
 #endif
 
   std::shared_ptr<Cache> cache_;
diff --git a/db_stress_tool/multi_ops_txns_stress.cc b/db_stress_tool/multi_ops_txns_stress.cc
index 619a119cb..e29a30980 100644
--- a/db_stress_tool/multi_ops_txns_stress.cc
+++ b/db_stress_tool/multi_ops_txns_stress.cc
@@ -533,7 +533,7 @@ void MultiOpsTxnsStressTest::RegisterAdditionalListeners() {
 
 #ifndef ROCKSDB_LITE
 void MultiOpsTxnsStressTest::PrepareTxnDbOptions(
-    TransactionDBOptions& txn_db_opts) {
+    SharedState* /*shared*/, TransactionDBOptions& txn_db_opts) {
   // MultiOpsTxnStressTest uses SingleDelete to delete secondary keys, thus we
   // register this callback to let TxnDb know that when rolling back
   // a transaction, use only SingleDelete to cancel prior Put from the same
diff --git a/db_stress_tool/multi_ops_txns_stress.h b/db_stress_tool/multi_ops_txns_stress.h
index a7db1c69e..d711f2357 100644
--- a/db_stress_tool/multi_ops_txns_stress.h
+++ b/db_stress_tool/multi_ops_txns_stress.h
@@ -265,7 +265,8 @@ class MultiOpsTxnsStressTest : public StressTest {
   void RegisterAdditionalListeners() override;
 
 #ifndef ROCKSDB_LITE
-  void PrepareTxnDbOptions(TransactionDBOptions& txn_db_opts) override;
+  void PrepareTxnDbOptions(SharedState* /*shared*/,
+                           TransactionDBOptions& txn_db_opts) override;
 #endif  // !ROCKSDB_LITE
 
   Status PrimaryKeyUpdateTxn(ThreadState* thread, uint32_t old_a,
diff --git a/db_stress_tool/no_batched_ops_stress.cc b/db_stress_tool/no_batched_ops_stress.cc
index 58e32016d..75579d0ce 100644
--- a/db_stress_tool/no_batched_ops_stress.cc
+++ b/db_stress_tool/no_batched_ops_stress.cc
@@ -12,6 +12,7 @@
 #ifndef NDEBUG
 #include "utilities/fault_injection_fs.h"
 #endif // NDEBUG
+#include "rocksdb/utilities/transaction_db.h"
 
 namespace ROCKSDB_NAMESPACE {
 class NonBatchedOpsStressTest : public StressTest {
@@ -930,6 +931,21 @@ class NonBatchedOpsStressTest : public StressTest {
     }
     return true;
   }
+
+#ifndef ROCKSDB_LITE
+  void PrepareTxnDbOptions(SharedState* shared,
+                           TransactionDBOptions& txn_db_opts) override {
+    txn_db_opts.rollback_deletion_type_callback =
+        [shared](TransactionDB*, ColumnFamilyHandle*, const Slice& key) {
+          assert(shared);
+          uint64_t key_num = 0;
+          bool ok = GetIntVal(key.ToString(), &key_num);
+          assert(ok);
+          (void)ok;
+          return !shared->AllowsOverwrite(key_num);
+        };
+  }
+#endif  // ROCKSDB_LITE
 };
 
 StressTest* CreateNonBatchedOpsStressTest() {
diff --git a/include/rocksdb/compaction_filter.h b/include/rocksdb/compaction_filter.h
index bdf872fd5..29f7fae8f 100644
--- a/include/rocksdb/compaction_filter.h
+++ b/include/rocksdb/compaction_filter.h
@@ -39,6 +39,7 @@ class CompactionFilter : public Customizable {
   enum class Decision {
     kKeep,
     kRemove,
+    kRemoveWithSingleDelete,
     kChangeValue,
     kRemoveAndSkipUntil,
     kChangeBlobIndex,  // used internally by BlobDB.
diff --git a/tools/db_crashtest.py b/tools/db_crashtest.py
index 696448989..53aa95ee0 100644
--- a/tools/db_crashtest.py
+++ b/tools/db_crashtest.py
@@ -384,6 +384,8 @@ multiops_txn_default_params = {
     "flush_one_in": 1000,
     "key_spaces_path": setup_multiops_txn_key_spaces_file(),
     "rollback_one_in":  4,
+    # Re-enable once we have a compaction for MultiOpsTxnStressTest
+    "enable_compaction_filter": 0,
 }
 
 multiops_wc_txn_params = {