Fix stack-use-after-scope

src/arena_allocator.hpp

@@ -9,6 +9,7 @@
 #include <iostream>
 #include <limits>
 #include <new>
+#include <span>
 #include <type_traits>
 #include <typeinfo>
 #include <utility>
@@ -301,38 +302,87 @@ public:
                                         new_size * sizeof(T), alignof(T)));
   }
 
+  /**
+   * @brief Smart pointer for arena-allocated objects with non-trivial
+   * destructors.
+   *
+   * ArenaAllocator::Ptr calls the destructor but does not free memory (assumes
+   * arena allocation). This provides RAII semantics for objects that need
+   * cleanup without the overhead of individual memory deallocation.
+   *
+   * @tparam T The type of object being managed
+   */
+  template <typename T> struct Ptr {
+    Ptr() noexcept : ptr_(nullptr) {}
+
+    explicit Ptr(T *ptr) noexcept : ptr_(ptr) {}
+
+    Ptr(const Ptr &) = delete;
+    Ptr &operator=(const Ptr &) = delete;
+
+    Ptr(Ptr &&other) noexcept : ptr_(other.ptr_) { other.ptr_ = nullptr; }
+
+    Ptr &operator=(Ptr &&other) noexcept {
+      if (this != &other) {
+        reset();
+        ptr_ = other.ptr_;
+        other.ptr_ = nullptr;
+      }
+      return *this;
+    }
+
+    ~Ptr() { reset(); }
+
+    T *operator->() const noexcept { return ptr_; }
+    T &operator*() const noexcept { return *ptr_; }
+
+    T *get() const noexcept { return ptr_; }
+
+    explicit operator bool() const noexcept { return ptr_ != nullptr; }
+
+    T *release() noexcept {
+      T *result = ptr_;
+      ptr_ = nullptr;
+      return result;
+    }
+
+    void reset(T *new_ptr = nullptr) noexcept {
+      if (ptr_) {
+        ptr_->~T();
+      }
+      ptr_ = new_ptr;
+    }
+
+  private:
+    T *ptr_;
+  };
+
   /**
    * @brief Construct an object of type T in the arena using placement new.
    *
-   * This is a convenience method that combines allocation with in-place
-   * construction. It properly handles alignment requirements for type T.
+   * This method returns different types based on whether T is trivially
+   * destructible:
+   * - For trivially destructible types: returns T* (raw pointer)
+   * - For non-trivially destructible types: returns ArenaAllocator::Ptr<T>
+   * (smart pointer that calls destructor)
    *
-   * @tparam T The type of object to construct (must be trivially destructible)
+   * @tparam T The type of object to construct
    * @tparam Args Types of constructor arguments
    * @param args Arguments to forward to T's constructor
-   * @return Pointer to the constructed object
+   * @return T* for trivially destructible types, ArenaAllocator::Ptr<T>
+   * otherwise
    * @note Prints error to stderr and calls std::abort() if memory allocation
    * fails
-   *
-   * ## Type Requirements:
-   * T must be trivially destructible (std::is_trivially_destructible_v<T>).
-   * This prevents subtle bugs since destructors are never called for objects
-   * constructed in the arena.
-   *
-   *
-   * ## Note:
-   * Objects constructed this way cannot be individually destroyed.
-   * Their destructors will NOT be called automatically - hence the requirement
-   * for trivially destructible types.
    */
-  template <typename T, typename... Args> T *construct(Args &&...args) {
-    static_assert(
-        std::is_trivially_destructible_v<T>,
-        "ArenaAllocator::construct requires trivially destructible types. "
-        "Objects constructed in the arena will not have their destructors "
-        "called.");
+  template <typename T, typename... Args> auto construct(Args &&...args) {
     void *ptr = allocate_raw(sizeof(T), alignof(T));
-    return new (ptr) T(std::forward<Args>(args)...);
+    T *obj = new (ptr) T(std::forward<Args>(args)...);
+
+    if constexpr (std::is_trivially_destructible_v<T>) {
+      return obj;
+    } else {
+      return Ptr<T>(obj);
+    }
   }
 
   /**
@@ -639,6 +689,12 @@ template <typename T> struct ArenaVector {
 
   void clear() { size_ = 0; }
 
+  // Implicit conversion to std::span
+  operator std::span<T>() { return std::span<T>(data_, size_); }
+  operator std::span<const T>() const {
+    return std::span<const T>(data_, size_);
+  }
+
   // Iterator support for range-based for loops
   T *begin() { return data_; }
   const T *begin() const { return data_; }

src/metric.cpp

@@ -11,7 +11,9 @@
 #include <cstdlib>
 #include <cstring>
 #include <functional>
+#include <map>
 #include <mutex>
+#include <set>
 #include <string>
 #include <thread>
 #include <type_traits>
@@ -168,6 +170,21 @@ struct LabelsKey {
     }
     return true;
   }
+
+  bool operator<(const LabelsKey &other) const {
+    if (labels.size() != other.labels.size()) {
+      return labels.size() < other.labels.size();
+    }
+    for (size_t i = 0; i < labels.size(); ++i) {
+      if (labels[i].first != other.labels[i].first) {
+        return labels[i].first < other.labels[i].first;
+      }
+      if (labels[i].second != other.labels[i].second) {
+        return labels[i].second < other.labels[i].second;
+      }
+    }
+    return false; // They are equal
+  }
 };
 
 } // namespace metric
@@ -214,9 +231,8 @@ template <> struct Family<Counter>::State {
       global_accumulated_values;
 
   // Callback-based metrics (global, not per-thread)
-  std::unordered_map<
-      LabelsKey, MetricCallback<Counter>, std::hash<LabelsKey>,
-      std::equal_to<LabelsKey>,
+  std::map<
+      LabelsKey, MetricCallback<Counter>, std::less<LabelsKey>,
       ArenaStlAllocator<std::pair<const LabelsKey, MetricCallback<Counter>>>>
       callbacks;
 
@@ -238,9 +254,7 @@ template <> struct Family<Gauge>::State {
       instances;
 
   // Callback-based metrics
-  std::unordered_map<
-      LabelsKey, MetricCallback<Gauge>, std::hash<LabelsKey>,
-      std::equal_to<LabelsKey>,
+  std::map<LabelsKey, MetricCallback<Gauge>, std::less<LabelsKey>,
            ArenaStlAllocator<std::pair<const LabelsKey, MetricCallback<Gauge>>>>
       callbacks;
 
@@ -321,9 +335,8 @@ struct Metric {
 
   // Function-local statics to avoid static initialization order fiasco
   static auto &get_counter_families() {
-    using FamilyMap = std::unordered_map<
-        std::string_view, Family<Counter>::State *, std::hash<std::string_view>,
-        std::equal_to<std::string_view>,
+    using FamilyMap = std::map<
+        std::string_view, Family<Counter>::State *, std::less<std::string_view>,
         ArenaStlAllocator<
             std::pair<const std::string_view, Family<Counter>::State *>>>;
     static FamilyMap *counterFamilies = new FamilyMap(
@@ -334,9 +347,8 @@ struct Metric {
   }
 
   static auto &get_gauge_families() {
-    using FamilyMap = std::unordered_map<
-        std::string_view, Family<Gauge>::State *, std::hash<std::string_view>,
-        std::equal_to<std::string_view>,
+    using FamilyMap = std::map<
+        std::string_view, Family<Gauge>::State *, std::less<std::string_view>,
         ArenaStlAllocator<
             std::pair<const std::string_view, Family<Gauge>::State *>>>;
     static FamilyMap *gaugeFamilies = new FamilyMap(
@@ -347,11 +359,11 @@ struct Metric {
   }
 
   static auto &get_histogram_families() {
-    using FamilyMap = std::unordered_map<
-        std::string_view, Family<Histogram>::State *,
-        std::hash<std::string_view>, std::equal_to<std::string_view>,
-        ArenaStlAllocator<
-            std::pair<const std::string_view, Family<Histogram>::State *>>>;
+    using FamilyMap =
+        std::map<std::string_view, Family<Histogram>::State *,
+                 std::less<std::string_view>,
+                 ArenaStlAllocator<std::pair<const std::string_view,
+                                             Family<Histogram>::State *>>>;
     static FamilyMap *histogramFamilies = new FamilyMap(
         ArenaStlAllocator<
             std::pair<const std::string_view, Family<Histogram>::State *>>(
@@ -573,6 +585,152 @@ struct Metric {
     result.p = ptr;
     return result;
   }
+
+  // Pre-computed data structures with resolved pointers to eliminate hash
+  // lookups
+  struct CounterLabelData {
+    LabelsKey labels_key;
+    std::vector<Counter::State *> thread_states; // Pre-resolved pointers
+    Counter::State *global_state; // Pre-resolved global state pointer
+
+    CounterLabelData(const LabelsKey &key)
+        : labels_key(key), global_state(nullptr) {}
+  };
+
+  struct GaugeLabelData {
+    LabelsKey labels_key;
+    Gauge::State *instance_state; // Direct pointer to gauge instance
+
+    GaugeLabelData(const LabelsKey &key)
+        : labels_key(key), instance_state(nullptr) {}
+  };
+
+  struct HistogramLabelData {
+    LabelsKey labels_key;
+    std::vector<Histogram::State *> thread_states; // Pre-resolved pointers
+    Histogram::State *global_state; // Pre-resolved global state pointer
+    size_t bucket_count;            // Cache bucket count from family
+
+    HistogramLabelData(const LabelsKey &key)
+        : labels_key(key), global_state(nullptr), bucket_count(0) {}
+  };
+
+  // Pre-computed data for each family type, built once and reused
+  struct LabelSets {
+    std::vector<std::vector<CounterLabelData>> counter_data;
+    std::vector<std::vector<GaugeLabelData>> gauge_data;
+    std::vector<std::vector<HistogramLabelData>> histogram_data;
+  };
+
+  // Build label sets once for reuse in both phases
+  static LabelSets build_label_sets(ArenaAllocator &arena) {
+    LabelSets label_sets;
+
+    // Build counter data with pre-resolved pointers
+    for (const auto &[name, family] : Metric::get_counter_families()) {
+      // Collect all unique labels first
+      std::set<LabelsKey, std::less<LabelsKey>, ArenaStlAllocator<LabelsKey>>
+          all_labels{ArenaStlAllocator<LabelsKey>(&arena)};
+
+      for (const auto &[thread_id, per_thread] : family->per_thread_state) {
+        for (const auto &[labels_key, instance] : per_thread.instances) {
+          all_labels.insert(labels_key);
+        }
+      }
+      for (const auto &[labels_key, global_state] :
+           family->global_accumulated_values) {
+        if (global_state) {
+          all_labels.insert(labels_key);
+        }
+      }
+
+      // Pre-resolve all pointers for each label set
+      std::vector<CounterLabelData> family_data;
+      for (const auto &labels_key : all_labels) {
+        CounterLabelData data(labels_key);
+
+        // Pre-resolve thread-local state pointers
+        for (const auto &[thread_id, per_thread] : family->per_thread_state) {
+          auto it = per_thread.instances.find(labels_key);
+          if (it != per_thread.instances.end()) {
+            data.thread_states.push_back(it->second);
+          }
+        }
+
+        // Pre-resolve global accumulated state pointer
+        auto global_it = family->global_accumulated_values.find(labels_key);
+        data.global_state =
+            (global_it != family->global_accumulated_values.end() &&
+             global_it->second)
+                ? global_it->second
+                : nullptr;
+
+        family_data.push_back(std::move(data));
+      }
+      label_sets.counter_data.push_back(std::move(family_data));
+    }
+
+    // Build gauge data with pre-resolved pointers
+    for (const auto &[name, family] : Metric::get_gauge_families()) {
+      std::vector<GaugeLabelData> family_data;
+
+      // Gauges iterate directly over instances
+      for (const auto &[labels_key, instance] : family->instances) {
+        GaugeLabelData data(labels_key);
+        data.instance_state = instance;
+        family_data.push_back(std::move(data));
+      }
+
+      label_sets.gauge_data.push_back(std::move(family_data));
+    }
+
+    // Build histogram data with pre-resolved pointers
+    for (const auto &[name, family] : Metric::get_histogram_families()) {
+      // Collect all unique labels first
+      std::set<LabelsKey, std::less<LabelsKey>, ArenaStlAllocator<LabelsKey>>
+          all_labels{ArenaStlAllocator<LabelsKey>(&arena)};
+
+      for (const auto &[thread_id, per_thread] : family->per_thread_state) {
+        for (const auto &[labels_key, instance] : per_thread.instances) {
+          all_labels.insert(labels_key);
+        }
+      }
+      for (const auto &[labels_key, global_state] :
+           family->global_accumulated_values) {
+        if (global_state) {
+          all_labels.insert(labels_key);
+        }
+      }
+
+      // Pre-resolve all pointers for each label set
+      std::vector<HistogramLabelData> family_data;
+      for (const auto &labels_key : all_labels) {
+        HistogramLabelData data(labels_key);
+        data.bucket_count = family->buckets.size(); // Cache bucket count
+
+        // Pre-resolve thread-local state pointers
+        for (const auto &[thread_id, per_thread] : family->per_thread_state) {
+          auto it = per_thread.instances.find(labels_key);
+          if (it != per_thread.instances.end()) {
+            data.thread_states.push_back(it->second);
+          }
+        }
+
+        // Pre-resolve global accumulated state pointer
+        auto global_it = family->global_accumulated_values.find(labels_key);
+        data.global_state =
+            (global_it != family->global_accumulated_values.end() &&
+             global_it->second)
+                ? global_it->second
+                : nullptr;
+
+        family_data.push_back(std::move(data));
+      }
+      label_sets.histogram_data.push_back(std::move(family_data));
+    }
+
+    return label_sets;
+  }
 };
 
 Counter::Counter() = default;
@@ -911,9 +1069,146 @@ bool is_valid_label_value(std::string_view value) {
   return simdutf::validate_utf8(value.data(), value.size());
 }
 
+union MetricValue {
+  double as_double;
+  uint64_t as_uint64;
+};
+
+// Phase 1: Compute all metric values in deterministic order
+static ArenaVector<MetricValue>
+compute_metric_values(ArenaAllocator &arena,
+                      const Metric::LabelSets &label_sets) {
+  ArenaVector<MetricValue> values(&arena);
+
+  // Compute counter values - ITERATION ORDER MUST MATCH FORMAT PHASE
+  size_t counter_family_idx = 0;
+  for (const auto &[name, family] : Metric::get_counter_families()) {
+    // Callback values
+    for (const auto &[labels_key, callback] : family->callbacks) {
+      auto value = callback();
+      values.push_back({.as_double = value});
+    }
+
+    // Use pre-computed data with resolved pointers - no hash lookups!
+    const auto &family_data = label_sets.counter_data[counter_family_idx++];
+    for (const auto &data : family_data) {
+      double total_value = 0.0;
+
+      // Sum thread-local values using pre-resolved pointers
+      for (auto *state_ptr : data.thread_states) {
+        // Atomic read to match atomic store in Counter::inc()
+        double value;
+        __atomic_load(&state_ptr->value, &value, __ATOMIC_RELAXED);
+        total_value += value;
+      }
+
+      // Add global accumulated value using pre-resolved pointer
+      if (data.global_state) {
+        total_value += data.global_state->value;
+      }
+
+      values.push_back({.as_double = total_value});
+    }
+  }
+
+  // Compute gauge values - ITERATION ORDER MUST MATCH FORMAT PHASE
+  size_t gauge_family_idx = 0;
+  for (const auto &[name, family] : Metric::get_gauge_families()) {
+    // Callback values
+    for (const auto &[labels_key, callback] : family->callbacks) {
+      auto value = callback();
+      values.push_back({.as_double = value});
+    }
+
+    // Use pre-computed data with resolved pointers - no hash lookups!
+    const auto &family_data = label_sets.gauge_data[gauge_family_idx++];
+    for (const auto &data : family_data) {
+      auto value = std::bit_cast<double>(
+          data.instance_state->value.load(std::memory_order_relaxed));
+      values.push_back({.as_double = value});
+    }
+  }
+
+  // Compute histogram values - ITERATION ORDER MUST MATCH FORMAT PHASE
+  size_t histogram_family_idx = 0;
+  for (const auto &family_pair : Metric::get_histogram_families()) {
+    // Use pre-computed data with resolved pointers - no hash lookups!
+    const auto &family_data = label_sets.histogram_data[histogram_family_idx++];
+
+    for (const auto &data : family_data) {
+      size_t bucket_count = data.bucket_count; // Use cached bucket count
+
+      ArenaVector<uint64_t> total_counts(&arena);
+      for (size_t i = 0; i < bucket_count; ++i) {
+        total_counts.push_back(0);
+      }
+      double total_sum = 0.0;
+      uint64_t total_observations = 0;
+
+      // Sum thread-local values using pre-resolved pointers
+      for (auto *instance : data.thread_states) {
+        // Extract data under lock - minimize critical section
+        uint64_t *counts_snapshot = arena.allocate<uint64_t>(bucket_count);
+        double sum_snapshot;
+        uint64_t observations_snapshot;
+
+        {
+          std::lock_guard<std::mutex> lock(instance->mutex);
+          for (size_t i = 0; i < instance->counts.size(); ++i) {
+            counts_snapshot[i] = instance->counts[i];
+          }
+          sum_snapshot = instance->sum;
+          observations_snapshot = instance->observations;
+        }
+
+        // Add to totals
+        for (size_t i = 0; i < bucket_count; ++i) {
+          total_counts[i] += counts_snapshot[i];
+        }
+        total_sum += sum_snapshot;
+        total_observations += observations_snapshot;
+      }
+
+      // Add global accumulated value using pre-resolved pointer
+      if (data.global_state) {
+        auto *global_state = data.global_state;
+        for (size_t i = 0; i < global_state->counts.size(); ++i) {
+          total_counts[i] += global_state->counts[i];
+        }
+        total_sum += global_state->sum;
+        total_observations += global_state->observations;
+      }
+
+      // Store histogram values
+      // Store explicit bucket counts
+      for (size_t i = 0; i < total_counts.size(); ++i) {
+        values.push_back({.as_uint64 = total_counts[i]});
+      }
+      // Store +Inf bucket (total observations)
+      values.push_back({.as_uint64 = total_observations});
+      // Store sum
+      values.push_back({.as_double = total_sum});
+      // Store count
+      values.push_back({.as_uint64 = total_observations});
+    }
+  }
+
+  return values;
+}
+
+// Phase 2: Format metrics using pre-computed values
 std::span<std::string_view> render(ArenaAllocator &arena) {
+  // Hold lock throughout both phases to prevent registry changes
   std::unique_lock<std::mutex> _{Metric::mutex};
 
+  // Build label sets once for both phases
+  Metric::LabelSets label_sets = Metric::build_label_sets(arena);
+
+  // Phase 1: Compute all metric values
+  ArenaVector<MetricValue> metric_values =
+      compute_metric_values(arena, label_sets);
+  const MetricValue *next_value = metric_values.data();
+
   ArenaVector<std::string_view> output(&arena);
 
   auto format_labels =
@@ -972,7 +1267,8 @@ std::span<std::string_view> render(ArenaAllocator &arena) {
     return std::string_view(buf, p - buf);
   };
 
-  // Render counters
+  // Format counters - ITERATION ORDER MUST MATCH COMPUTE PHASE
+  size_t counter_family_idx = 0;
   for (const auto &[name, family] : Metric::get_counter_families()) {
     output.push_back(format(arena, "# HELP %.*s %.*s\n",
                             static_cast<int>(name.length()), name.data(),
@@ -983,8 +1279,10 @@ std::span<std::string_view> render(ArenaAllocator &arena) {
 
     ArenaVector<std::pair<std::string_view, std::string_view>> labels_sv(
         &arena);
+
+    // Format callback values
     for (const auto &[labels_key, callback] : family->callbacks) {
-      auto value = callback();
+      auto value = next_value++->as_double;
       labels_sv.clear();
       for (size_t i = 0; i < labels_key.labels.size(); ++i) {
         labels_sv.push_back(labels_key.labels[i]);
@@ -996,36 +1294,15 @@ std::span<std::string_view> render(ArenaAllocator &arena) {
                               value));
     }
 
-    // Aggregate all counter values (thread-local + global accumulated)
-    std::unordered_map<LabelsKey, double, std::hash<LabelsKey>,
-                       std::equal_to<LabelsKey>,
-                       ArenaStlAllocator<std::pair<const LabelsKey, double>>>
-        aggregated_values{
-            ArenaStlAllocator<std::pair<const LabelsKey, double>>(&arena)};
+    // Use pre-computed data (same as compute phase)
+    const auto &family_data = label_sets.counter_data[counter_family_idx++];
 
-    // First, add thread-local values
-    for (const auto &[thread_id, per_thread] : family->per_thread_state) {
-      for (const auto &[labels_key, instance] : per_thread.instances) {
-        // Atomic read to match atomic store in Counter::inc()
-        double value;
-        __atomic_load(&instance->value, &value, __ATOMIC_RELAXED);
-        aggregated_values[labels_key] += value;
-      }
-    }
-
-    // Then, add globally accumulated values from destroyed threads
-    for (const auto &[labels_key, global_state] :
-         family->global_accumulated_values) {
-      if (global_state) {
-        aggregated_values[labels_key] += global_state->value;
-      }
-    }
-
-    // Render aggregated counter values
-    for (const auto &[labels_key, total_value] : aggregated_values) {
+    // Format counter values using pre-computed values
+    for (const auto &data : family_data) {
+      auto total_value = next_value++->as_double;
       labels_sv.clear();
-      for (size_t i = 0; i < labels_key.labels.size(); ++i) {
-        labels_sv.push_back(labels_key.labels[i]);
+      for (size_t i = 0; i < data.labels_key.labels.size(); ++i) {
+        labels_sv.push_back(data.labels_key.labels[i]);
       }
       auto labels = format_labels(labels_sv);
       output.push_back(format(arena, "%.*s%.*s %.17g\n",
@@ -1035,7 +1312,8 @@ std::span<std::string_view> render(ArenaAllocator &arena) {
     }
   }
 
-  // Render gauges
+  // Format gauges - ITERATION ORDER MUST MATCH COMPUTE PHASE
+  size_t gauge_family_idx = 0;
   for (const auto &[name, family] : Metric::get_gauge_families()) {
     output.push_back(format(arena, "# HELP %.*s %.*s\n",
                             static_cast<int>(name.length()), name.data(),
@@ -1046,8 +1324,10 @@ std::span<std::string_view> render(ArenaAllocator &arena) {
 
     ArenaVector<std::pair<std::string_view, std::string_view>> labels_sv(
         &arena);
+
+    // Format callback values
     for (const auto &[labels_key, callback] : family->callbacks) {
-      auto value = callback();
+      auto value = next_value++->as_double;
       labels_sv.clear();
       for (size_t i = 0; i < labels_key.labels.size(); ++i) {
         labels_sv.push_back(labels_key.labels[i]);
@@ -1059,12 +1339,13 @@ std::span<std::string_view> render(ArenaAllocator &arena) {
                               value));
     }
 
-    for (const auto &[labels_key, instance] : family->instances) {
-      auto value = std::bit_cast<double>(
-          instance->value.load(std::memory_order_relaxed));
+    // Use pre-computed data (same as compute phase)
+    const auto &family_data = label_sets.gauge_data[gauge_family_idx++];
+    for (const auto &data : family_data) {
+      auto value = next_value++->as_double;
       labels_sv.clear();
-      for (size_t i = 0; i < labels_key.labels.size(); ++i) {
-        labels_sv.push_back(labels_key.labels[i]);
+      for (size_t i = 0; i < data.labels_key.labels.size(); ++i) {
+        labels_sv.push_back(data.labels_key.labels[i]);
       }
       auto labels = format_labels(labels_sv);
       output.push_back(format(arena, "%.*s%.*s %.17g\n",
@@ -1074,7 +1355,8 @@ std::span<std::string_view> render(ArenaAllocator &arena) {
     }
   }
 
-  // Render histograms
+  // Format histograms - ITERATION ORDER MUST MATCH COMPUTE PHASE
+  size_t histogram_family_idx = 0;
   for (const auto &[name, family] : Metric::get_histogram_families()) {
     output.push_back(format(arena, "# HELP %.*s %.*s\n",
                             static_cast<int>(name.length()), name.data(),
@@ -1083,166 +1365,67 @@ std::span<std::string_view> render(ArenaAllocator &arena) {
     output.push_back(format(arena, "# TYPE %.*s histogram\n",
                             static_cast<int>(name.length()), name.data()));
 
-    // Aggregate all histogram values (thread-local + global accumulated)
-    // Use a simpler structure to avoid tuple constructor issues
-    struct AggregatedHistogram {
-      ArenaVector<double> thresholds;
-      ArenaVector<uint64_t> counts;
-      double sum;
-      uint64_t observations;
-
-      AggregatedHistogram(ArenaAllocator &arena)
-          : thresholds(&arena), counts(&arena), sum(0.0), observations(0) {}
-    };
-    std::unordered_map<
-        LabelsKey, AggregatedHistogram *, std::hash<LabelsKey>,
-        std::equal_to<LabelsKey>,
-        ArenaStlAllocator<std::pair<const LabelsKey, AggregatedHistogram *>>>
-        aggregated_histograms{ArenaStlAllocator<
-            std::pair<const LabelsKey, AggregatedHistogram *>>(&arena)};
+    // Use pre-computed data (same as compute phase)
+    const auto &family_data = label_sets.histogram_data[histogram_family_idx++];
 
     ArenaVector<std::pair<std::string_view, std::string_view>> bucket_labels_sv(
         &arena);
 
-    // First, collect thread-local histogram data
-    for (const auto &[thread_id, per_thread] : family->per_thread_state) {
-      for (const auto &[labels_key, instance] : per_thread.instances) {
-        // Extract data under lock - minimize critical section
-        // Note: thresholds and counts sizes never change after histogram
-        // creation
-        ArenaVector<double> thresholds_snapshot(&arena);
-        ArenaVector<uint64_t> counts_snapshot(&arena);
-        double sum_snapshot;
-        uint64_t observations_snapshot;
+    // Format histogram data using pre-computed values
+    for (const auto &data : family_data) {
+      // Get bucket count from pre-computed data
+      size_t bucket_count = data.bucket_count;
 
-        // Copy data with minimal critical section
-        {
-          std::lock_guard<std::mutex> lock(instance->mutex);
-          // Copy thresholds
-          for (size_t i = 0; i < instance->thresholds.size(); ++i) {
-            thresholds_snapshot.push_back(instance->thresholds[i]);
-          }
-          // Copy counts
-          for (size_t i = 0; i < instance->counts.size(); ++i) {
-            counts_snapshot.push_back(instance->counts[i]);
-          }
-          sum_snapshot = instance->sum;
-          observations_snapshot = instance->observations;
-        }
-
-        // Initialize or aggregate into aggregated_histograms
-        auto it = aggregated_histograms.find(labels_key);
-        if (it == aggregated_histograms.end()) {
-          // Create new entry
-          auto *agg_hist = new (arena.allocate_raw(
-              sizeof(AggregatedHistogram), alignof(AggregatedHistogram)))
-              AggregatedHistogram(arena);
-          for (size_t i = 0; i < thresholds_snapshot.size(); ++i) {
-            agg_hist->thresholds.push_back(thresholds_snapshot[i]);
-          }
-          for (size_t i = 0; i < counts_snapshot.size(); ++i) {
-            agg_hist->counts.push_back(counts_snapshot[i]);
-          }
-          agg_hist->sum = sum_snapshot;
-          agg_hist->observations = observations_snapshot;
-          aggregated_histograms[labels_key] = agg_hist;
-        } else {
-          // Aggregate with existing entry
-          auto *agg_hist = it->second;
-          // Aggregate counts
-          for (size_t i = 0; i < counts_snapshot.size(); ++i) {
-            agg_hist->counts[i] += counts_snapshot[i];
-          }
-          agg_hist->sum += sum_snapshot;
-          agg_hist->observations += observations_snapshot;
-        }
-      }
-    }
-
-    // Then, add globally accumulated values from destroyed threads
-    for (const auto &[labels_key, global_state] :
-         family->global_accumulated_values) {
-      if (global_state) {
-        auto it = aggregated_histograms.find(labels_key);
-        if (it == aggregated_histograms.end()) {
-          // Create new entry from global state
-          auto *agg_hist = new (arena.allocate_raw(
-              sizeof(AggregatedHistogram), alignof(AggregatedHistogram)))
-              AggregatedHistogram(arena);
-          for (size_t i = 0; i < global_state->thresholds.size(); ++i) {
-            agg_hist->thresholds.push_back(global_state->thresholds[i]);
-          }
-          for (size_t i = 0; i < global_state->counts.size(); ++i) {
-            agg_hist->counts.push_back(global_state->counts[i]);
-          }
-          agg_hist->sum = global_state->sum;
-          agg_hist->observations = global_state->observations;
-          aggregated_histograms[labels_key] = agg_hist;
-        } else {
-          // Add global accumulated values to existing entry
-          auto *agg_hist = it->second;
-          for (size_t i = 0; i < global_state->counts.size(); ++i) {
-            agg_hist->counts[i] += global_state->counts[i];
-          }
-          agg_hist->sum += global_state->sum;
-          agg_hist->observations += global_state->observations;
-        }
-      }
-    }
-
-    // Render aggregated histogram data
-    for (const auto &[labels_key, agg_hist] : aggregated_histograms) {
-
-      // Render explicit bucket counts
-      for (size_t i = 0; i < agg_hist->thresholds.size(); ++i) {
+      // Format explicit bucket counts
+      for (size_t i = 0; i < bucket_count; ++i) {
+        auto count = next_value++->as_uint64;
         bucket_labels_sv.clear();
-        for (size_t j = 0; j < labels_key.labels.size(); ++j) {
-          bucket_labels_sv.push_back(labels_key.labels[j]);
+        for (size_t j = 0; j < data.labels_key.labels.size(); ++j) {
+          bucket_labels_sv.push_back(data.labels_key.labels[j]);
         }
-
         bucket_labels_sv.push_back(
-            {"le", static_format(arena, agg_hist->thresholds[i])});
+            {"le", static_format(arena, family->buckets[i])});
         auto labels = format_labels(bucket_labels_sv);
         output.push_back(format(
             arena, "%.*s_bucket%.*s %llu\n", static_cast<int>(name.length()),
             name.data(), static_cast<int>(labels.length()), labels.data(),
-            static_cast<unsigned long long>(agg_hist->counts[i])));
+            static_cast<unsigned long long>(count)));
       }
 
-      // Render +Inf bucket using total observations count
+      // Format +Inf bucket
+      auto observations = next_value++->as_uint64;
       bucket_labels_sv.clear();
-      for (size_t j = 0; j < labels_key.labels.size(); ++j) {
-        bucket_labels_sv.push_back(labels_key.labels[j]);
+      for (size_t j = 0; j < data.labels_key.labels.size(); ++j) {
+        bucket_labels_sv.push_back(data.labels_key.labels[j]);
       }
       bucket_labels_sv.push_back({"le", "+Inf"});
       auto inf_labels = format_labels(bucket_labels_sv);
       output.push_back(format(
           arena, "%.*s_bucket%.*s %llu\n", static_cast<int>(name.length()),
           name.data(), static_cast<int>(inf_labels.length()), inf_labels.data(),
-          static_cast<unsigned long long>(agg_hist->observations)));
+          static_cast<unsigned long long>(observations)));
 
-      // Render sum
+      // Format sum
+      auto sum = next_value++->as_double;
       bucket_labels_sv.clear();
-      for (size_t j = 0; j < labels_key.labels.size(); ++j) {
-        bucket_labels_sv.push_back(labels_key.labels[j]);
+      for (size_t j = 0; j < data.labels_key.labels.size(); ++j) {
+        bucket_labels_sv.push_back(data.labels_key.labels[j]);
       }
       auto labels = format_labels(bucket_labels_sv);
-      output.push_back(format(arena, "%.*s_sum%.*s %.17g\n",
+      output.push_back(format(
+          arena, "%.*s_sum%.*s %.17g\n", static_cast<int>(name.length()),
+          name.data(), static_cast<int>(labels.length()), labels.data(), sum));
+
+      // Format count
+      auto count = next_value++->as_uint64;
+      output.push_back(format(arena, "%.*s_count%.*s %llu\n",
                               static_cast<int>(name.length()), name.data(),
                               static_cast<int>(labels.length()), labels.data(),
-                              agg_hist->sum));
-
-      // Render count
-      output.push_back(format(
-          arena, "%.*s_count%.*s %llu\n", static_cast<int>(name.length()),
-          name.data(), static_cast<int>(labels.length()), labels.data(),
-          static_cast<unsigned long long>(agg_hist->observations)));
+                              static_cast<unsigned long long>(count)));
     }
   }
 
-  auto result = arena.allocate<std::string_view>(output.size());
-  std::copy(output.data(), output.data() + output.size(), result);
-  return std::span<std::string_view>(result, output.size());
+  return output;
 }
 
 // Template specialization implementations for register_callback
@@ -1295,4 +1478,31 @@ void Family<Gauge>::register_callback(
 std::mutex Metric::mutex;
 thread_local Metric::ThreadInit Metric::thread_init;
 
+void reset_metrics_for_testing() {
+  std::lock_guard _{Metric::mutex};
+
+  // WARNING: This function assumes no metric objects are in use!
+  // Clear all family maps - this will leak the Family::State objects but
+  // that's acceptable for testing since they were allocated in the global arena
+
+  // Get references to the maps
+  auto &counter_families = Metric::get_counter_families();
+  auto &gauge_families = Metric::get_gauge_families();
+  auto &histogram_families = Metric::get_histogram_families();
+  auto &interned_labels = Metric::get_interned_labels();
+
+  // Clear all family registrations
+  counter_families.clear();
+  gauge_families.clear();
+  histogram_families.clear();
+  interned_labels.clear();
+
+  // Reset the global arena - this will invalidate all arena-allocated strings
+  // but since we're clearing everything, that's OK
+  Metric::get_global_arena().reset();
+
+  // Note: Thread-local arenas will be cleaned up by ThreadInit destructors
+  // when threads exit naturally
+}
+
 } // namespace metric

src/metric.hpp

@@ -218,6 +218,11 @@ bool is_valid_metric_name(std::string_view name);
 bool is_valid_label_key(std::string_view key);
 bool is_valid_label_value(std::string_view value);
 
+// Reset all metrics state - WARNING: Only safe for testing!
+// This clears all registered families and metrics. Should only be called
+// when no metric objects are in use and no concurrent render() calls.
+void reset_metrics_for_testing();
+
 // Note: Histograms do not support callbacks due to their multi-value nature
 // (buckets + sum + count). Use static histogram metrics only.
 

tests/test_arena_allocator.cpp

@@ -598,3 +598,162 @@ TEST_CASE("format function fallback codepath") {
     CHECK(result == "Valid format: 42");
   }
 }
+
+// Test object with non-trivial destructor for ArenaAllocator::Ptr testing
+class TestObject {
+public:
+  static int destructor_count;
+  static int constructor_count;
+
+  int value;
+
+  TestObject(int v) : value(v) { constructor_count++; }
+
+  ~TestObject() { destructor_count++; }
+
+  static void reset_counters() {
+    constructor_count = 0;
+    destructor_count = 0;
+  }
+};
+
+int TestObject::destructor_count = 0;
+int TestObject::constructor_count = 0;
+
+// Test struct with trivial destructor
+struct TrivialObject {
+  int value;
+  TrivialObject(int v) : value(v) {}
+};
+
+TEST_CASE("ArenaAllocator::Ptr smart pointer functionality") {
+  TestObject::reset_counters();
+
+  SUBCASE("construct returns raw pointer for trivially destructible types") {
+    ArenaAllocator arena;
+
+    auto ptr = arena.construct<TrivialObject>(42);
+    static_assert(std::is_same_v<decltype(ptr), TrivialObject *>,
+                  "construct() should return raw pointer for trivially "
+                  "destructible types");
+    CHECK(ptr != nullptr);
+    CHECK(ptr->value == 42);
+  }
+
+  SUBCASE("construct returns ArenaAllocator::Ptr for non-trivially "
+          "destructible types") {
+    ArenaAllocator arena;
+
+    auto ptr = arena.construct<TestObject>(42);
+    static_assert(
+        std::is_same_v<decltype(ptr), ArenaAllocator::Ptr<TestObject>>,
+        "construct() should return ArenaAllocator::Ptr for non-trivially "
+        "destructible types");
+    CHECK(ptr);
+    CHECK(ptr->value == 42);
+    CHECK(TestObject::constructor_count == 1);
+    CHECK(TestObject::destructor_count == 0);
+  }
+
+  SUBCASE("ArenaAllocator::Ptr calls destructor on destruction") {
+    ArenaAllocator arena;
+
+    {
+      auto ptr = arena.construct<TestObject>(42);
+      CHECK(TestObject::constructor_count == 1);
+      CHECK(TestObject::destructor_count == 0);
+    } // ptr goes out of scope
+
+    CHECK(TestObject::destructor_count == 1);
+  }
+
+  SUBCASE("ArenaAllocator::Ptr move semantics") {
+    ArenaAllocator arena;
+
+    auto ptr1 = arena.construct<TestObject>(42);
+    CHECK(TestObject::constructor_count == 1);
+
+    auto ptr2 = std::move(ptr1);
+    CHECK(!ptr1); // ptr1 should be null after move
+    CHECK(ptr2);
+    CHECK(ptr2->value == 42);
+    CHECK(TestObject::destructor_count == 0); // No destruction yet
+
+    ptr2.reset();
+    CHECK(TestObject::destructor_count == 1); // Destructor called
+  }
+
+  SUBCASE("ArenaAllocator::Ptr access operators") {
+    ArenaAllocator arena;
+
+    auto ptr = arena.construct<TestObject>(123);
+
+    // Test operator->
+    CHECK(ptr->value == 123);
+
+    // Test operator*
+    CHECK((*ptr).value == 123);
+
+    // Test get()
+    TestObject *raw_ptr = ptr.get();
+    CHECK(raw_ptr != nullptr);
+    CHECK(raw_ptr->value == 123);
+
+    // Test bool conversion
+    CHECK(ptr);
+    CHECK(static_cast<bool>(ptr) == true);
+  }
+
+  SUBCASE("ArenaAllocator::Ptr reset functionality") {
+    ArenaAllocator arena;
+
+    auto ptr = arena.construct<TestObject>(42);
+    CHECK(TestObject::constructor_count == 1);
+    CHECK(TestObject::destructor_count == 0);
+
+    ptr.reset();
+    CHECK(!ptr);
+    CHECK(TestObject::destructor_count == 1);
+
+    // Reset with new object
+    TestObject *raw_obj = arena.construct<TestObject>(84).release();
+    ptr.reset(raw_obj);
+    CHECK(ptr);
+    CHECK(ptr->value == 84);
+    CHECK(TestObject::constructor_count == 2);
+    CHECK(TestObject::destructor_count == 1);
+  }
+
+  SUBCASE("ArenaAllocator::Ptr release functionality") {
+    ArenaAllocator arena;
+
+    auto ptr = arena.construct<TestObject>(42);
+    TestObject *raw_ptr = ptr.release();
+
+    CHECK(!ptr); // ptr should be null after release
+    CHECK(raw_ptr != nullptr);
+    CHECK(raw_ptr->value == 42);
+    CHECK(TestObject::destructor_count == 0); // No destructor called
+
+    // Manually call destructor (since we released ownership)
+    raw_ptr->~TestObject();
+    CHECK(TestObject::destructor_count == 1);
+  }
+
+  SUBCASE("ArenaAllocator::Ptr move assignment") {
+    ArenaAllocator arena;
+
+    auto ptr1 = arena.construct<TestObject>(42);
+    auto ptr2 = arena.construct<TestObject>(84);
+
+    CHECK(TestObject::constructor_count == 2);
+    CHECK(TestObject::destructor_count == 0);
+
+    ptr1 = std::move(ptr2); // Should destroy first object, move second
+
+    CHECK(!ptr2); // ptr2 should be null
+    CHECK(ptr1);
+    CHECK(ptr1->value == 84);
+    CHECK(TestObject::destructor_count == 1); // First object destroyed
+  }
+}

tests/test_metric.cpp

@@ -7,7 +7,9 @@
 #include <atomic>
 #include <chrono>
 #include <cmath>
+#include <fstream>
 #include <latch>
+#include <sstream>
 #include <thread>
 #include <vector>
 
@@ -273,11 +275,9 @@ TEST_CASE("callback-based metrics") {
   auto gauge_family = metric::create_gauge("callback_gauge", "Callback gauge");
 
   SUBCASE("counter callback") {
-    std::atomic<double> counter_value{42.0};
 
-    counter_family.register_callback(
-        {{"type", "callback"}},
-        [&counter_value]() { return counter_value.load(); });
+    counter_family.register_callback({{"type", "callback"}},
+                                     []() { return 42.0; });
 
     // Callback should be called during render
     ArenaAllocator arena;
@@ -286,11 +286,8 @@ TEST_CASE("callback-based metrics") {
   }
 
   SUBCASE("gauge callback") {
-    std::atomic<double> gauge_value{123.5};
-
-    gauge_family.register_callback({{"type", "callback"}}, [&gauge_value]() {
-      return gauge_value.load();
-    });
+    gauge_family.register_callback({{"type", "callback"}},
+                                   []() { return 123.5; });
 
     ArenaAllocator arena;
     auto output = metric::render(arena);
@@ -657,3 +654,98 @@ TEST_CASE("memory management") {
     CHECK(final_output.size() > 0);
   }
 }
+
+TEST_CASE("render output deterministic order golden test") {
+  // Clean slate - reset all metrics before this test
+  metric::reset_metrics_for_testing();
+
+  ArenaAllocator arena;
+
+  // Create a comprehensive set of metrics with deliberate ordering
+  // to test deterministic output
+
+  // Create counters with different family names and labels
+  auto z_counter_family =
+      metric::create_counter("z_last_counter", "Last counter alphabetically");
+  auto z_counter =
+      z_counter_family.create({{"method", "POST"}, {"handler", "api"}});
+  z_counter.inc(42.0);
+
+  auto a_counter_family =
+      metric::create_counter("a_first_counter", "First counter alphabetically");
+  auto a_counter1 = a_counter_family.create({{"status", "200"}});
+  auto a_counter2 = a_counter_family.create(
+      {{"method", "GET"}}); // Should come before status lexicographically
+  a_counter1.inc(100.0);
+  a_counter2.inc(200.0);
+
+  // Create gauges with different orderings
+  auto m_gauge_family = metric::create_gauge("m_middle_gauge", "Middle gauge");
+  auto m_gauge = m_gauge_family.create({{"type", "memory"}});
+  m_gauge.set(1024.0);
+
+  auto b_gauge_family = metric::create_gauge("b_second_gauge", "Second gauge");
+  auto b_gauge = b_gauge_family.create({{"region", "us-west"}});
+  b_gauge.set(256.0);
+
+  // Create histograms
+  auto x_hist_family = metric::create_histogram("x_histogram", "Test histogram",
+                                                {0.1, 0.5, 1.0});
+  auto x_hist = x_hist_family.create({{"endpoint", "/api/v1"}});
+  x_hist.observe(0.25);
+  x_hist.observe(0.75);
+
+  // Add some callbacks to test callback ordering
+  a_counter_family.register_callback({{"callback", "test"}},
+                                     []() { return 123.0; });
+  m_gauge_family.register_callback({{"callback", "dynamic"}},
+                                   []() { return 456.0; });
+
+  // Render the metrics
+  auto output = metric::render(arena);
+
+  // Concatenate all output into a single string
+  std::ostringstream oss;
+  for (const auto &line : output) {
+    oss << line;
+  }
+  std::string actual_output = oss.str();
+
+  // Define expected golden output - this represents the exact expected
+  // deterministic order
+  std::string expected_golden =
+      "# HELP a_first_counter First counter alphabetically\n"
+      "# TYPE a_first_counter counter\n"
+      "a_first_counter{callback=\"test\"} 123\n"
+      "a_first_counter{method=\"GET\"} 200\n"
+      "a_first_counter{status=\"200\"} 100\n"
+      "# HELP z_last_counter Last counter alphabetically\n"
+      "# TYPE z_last_counter counter\n"
+      "z_last_counter{handler=\"api\",method=\"POST\"} 42\n"
+      "# HELP b_second_gauge Second gauge\n"
+      "# TYPE b_second_gauge gauge\n"
+      "b_second_gauge{region=\"us-west\"} 256\n"
+      "# HELP m_middle_gauge Middle gauge\n"
+      "# TYPE m_middle_gauge gauge\n"
+      "m_middle_gauge{callback=\"dynamic\"} 456\n"
+      "m_middle_gauge{type=\"memory\"} 1024\n"
+      "# HELP x_histogram Test histogram\n"
+      "# TYPE x_histogram histogram\n"
+      "x_histogram_bucket{endpoint=\"/api/v1\",le=\"0.1\"} 0\n"
+      "x_histogram_bucket{endpoint=\"/api/v1\",le=\"0.5\"} 1\n"
+      "x_histogram_bucket{endpoint=\"/api/v1\",le=\"1.0\"} 2\n"
+      "x_histogram_bucket{endpoint=\"/api/v1\",le=\"+Inf\"} 2\n"
+      "x_histogram_sum{endpoint=\"/api/v1\"} 1\n"
+      "x_histogram_count{endpoint=\"/api/v1\"} 2\n";
+
+  // Check if output matches golden file
+  if (actual_output != expected_golden) {
+    MESSAGE("Render output does not match expected golden output.");
+    MESSAGE("This indicates the deterministic ordering has changed.");
+    MESSAGE("Expected output:\n" << expected_golden);
+    MESSAGE("Actual output:\n" << actual_output);
+    CHECK(false); // Force test failure
+  } else {
+    CHECK(true); // Test passes
+  }
+}