diff --git a/CMakeLists.txt b/CMakeLists.txt
index 21f15d4..fad43d0 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -124,6 +124,7 @@ if (PHMAP_BUILD_EXAMPLES)
     add_executable(ex_basic examples/basic.cc phmap.natvis)
     add_executable(ex_bench examples/bench.cc phmap.natvis)
     add_executable(ex_emplace examples/emplace.cc phmap.natvis)
+    add_executable(ex_serialize examples/serialize.cc phmap.natvis)
     add_executable(ex_hash_std examples/hash_std.cc phmap.natvis)
     add_executable(ex_hash_value examples/hash_value.cc phmap.natvis)
     add_executable(ex_two_files examples/f1.cc examples/f2.cc phmap.natvis)
diff --git a/examples/serialize.cc b/examples/serialize.cc
new file mode 100644
index 0000000..a644672
--- /dev/null
+++ b/examples/serialize.cc
@@ -0,0 +1,82 @@
+#include <iostream>
+#include <bitset>
+#include <cinttypes>
+#include "parallel_hashmap/phmap.h"
+#include "cereal/types/unordered_map.hpp"
+#include "cereal/types/memory.hpp"
+#include "cereal/types/bitset.hpp"
+#include "cereal/archives/binary.hpp"
+#include <fstream>
+#include <random>
+#include <chrono>
+#include <functional>
+#include <cstdio>
+
+using phmap::flat_hash_map;
+using namespace std;
+template <typename T> using milliseconds = std::chrono::duration<T, std::milli>;
+
+void showtime(const char *name, std::function<void ()> doit)
+{
+    auto t1 = std::chrono::high_resolution_clock::now();
+    doit();
+    auto t2 = std::chrono::high_resolution_clock::now();
+    auto elapsed = milliseconds<double>(t2 - t1).count();
+    printf("%s: %.3fms\n", name, (int)elapsed / 1000.0f);
+}
+    
+int main()
+{
+    using MapType = phmap::parallel_flat_hash_map<bitset<42>, int>;
+    MapType table;
+    const int num_items = 100000000;
+
+    // Iterate and add keys and values 
+    // -------------------------------
+    showtime("build hash", [&table, num_items]() {
+            random_device dev;
+            mt19937 rng(dev());
+            uniform_int_distribution<mt19937::result_type> dist6(0,1);
+            table.reserve(num_items);
+            for (int i=0; i < num_items; ++i) 
+            {
+                bitset<42> bs;
+                for (int j=0; j<42; ++j) 
+                    bs[j] = dist6(rng);
+                table[bs] = i;
+            }
+        });
+
+    // cerealize and save data
+    // -----------------------
+    showtime("serialize", [&table]() {
+            ofstream os("z:\out.cereal", ios::binary);
+#if 0
+            const unsigned int length = 8192;
+            char buffer[length];
+            os.rdbuf()->pubsetbuf(buffer, length);
+#endif
+            cereal::BinaryOutputArchive archive(os);
+            archive(table.size());
+            archive(table);
+        });
+
+    MapType().swap(table);
+
+    // deserialize
+    // -----------
+    showtime("deserialize", [&table]() {
+            ifstream is("z:\out.cereal", ios::binary);
+            cereal::BinaryInputArchive archive_in(is);
+            size_t table_size;
+
+            archive_in(table_size);
+            table.reserve(table_size);
+            archive_in(table);             // deserialize from file out.cereal into variable
+        });
+
+    
+    printf("table size: %zu\n", table.size());
+
+    return 0;
+}
diff --git a/parallel_hashmap/phmap.h b/parallel_hashmap/phmap.h
index 8c844ef..748c003 100644
--- a/parallel_hashmap/phmap.h
+++ b/parallel_hashmap/phmap.h
@@ -1202,6 +1202,8 @@ public:
         // compared to destruction of the elements of the container. So we pick the
         // largest bucket_count() threshold for which iteration is still fast and
         // past that we simply deallocate the array.
+        if (empty())
+            return;
         if (capacity_ > 127) {
             destroy_slots();
         } else if (capacity_) {