diff --git a/src/http_handler.cpp b/src/http_handler.cpp
index b16ddc0..5a7f3a4 100644
--- a/src/http_handler.cpp
+++ b/src/http_handler.cpp
@@ -113,8 +113,12 @@ void HttpHandler::on_data_arrived(std::string_view data,
       break;
     }
 
-    // Connection state will be reset when appropriate by the server
-    // For now, we don't reset immediately since messages may still be queued
+    {
+      auto guard = ok_pipeline.push(1, true);
+      for (auto &c : guard.batch) {
+        c = std::move(conn_ptr);
+      }
+    }
   }
 }
 
diff --git a/src/http_handler.hpp b/src/http_handler.hpp
index ca5ff62..1f2555e 100644
--- a/src/http_handler.hpp
+++ b/src/http_handler.hpp
@@ -1,10 +1,14 @@
 #pragma once
 
+#include "ThreadPipeline.h"
 #include "connection.hpp"
 #include "connection_handler.hpp"
+#include "perfetto_categories.hpp"
+#include "server.hpp"
 #include <llhttp.h>
 #include <memory>
 #include <string_view>
+#include <thread>
 
 /**
  * HTTP routes supported by WeaselDB server.
@@ -57,8 +61,34 @@ struct HttpConnectionState {
  * Supports the WeaselDB REST API endpoints with enum-based routing.
  */
 class HttpHandler : public ConnectionHandler {
+  ThreadPipeline<std::unique_ptr<Connection>> ok_pipeline{10, {1}};
+  std::thread ok_thread{[this]() {
+    pthread_setname_np(pthread_self(), "stage-0");
+    for (;;) {
+      auto guard = ok_pipeline.acquire(0, 0);
+      for (auto &c : guard.batch) {
+        if (!c) {
+          return;
+        }
+        auto *state = static_cast<HttpConnectionState *>(c->user_data);
+        TRACE_EVENT("http", "pipeline thread",
+                    perfetto::Flow::Global(state->request_id));
+        Server::releaseBackToServer(std::move(c));
+      }
+    }
+  }};
+
 public:
   HttpHandler() = default;
+  ~HttpHandler() {
+    {
+      auto guard = ok_pipeline.push(1, true);
+      for (auto &c : guard.batch) {
+        c = {};
+      }
+    }
+    ok_thread.join();
+  }
 
   void on_connection_established(Connection &conn) override;
   void on_connection_closed(Connection &conn) override;
diff --git a/src/main.cpp b/src/main.cpp
index d317036..7a686f7 100644
--- a/src/main.cpp
+++ b/src/main.cpp
@@ -258,6 +258,5 @@ int main(int argc, char *argv[]) {
   server->run();
   std::cout << "Server shutdown complete." << std::endl;
 
-  g_server = nullptr;
   return 0;
 }
diff --git a/src/server.cpp b/src/server.cpp
index 099f78f..99fd688 100644
--- a/src/server.cpp
+++ b/src/server.cpp
@@ -337,7 +337,7 @@ void Server::start_io_threads(std::vector<std::thread> &threads) {
         // Process existing connections in batch
         if (batch_count > 0) {
           process_connection_batch(epollfd, {batch, (size_t)batch_count},
-                                   {batch_events, (size_t)batch_count}, false);
+                                   {batch_events, (size_t)batch_count});
         }
 
         // Only accept on listen sockets that epoll indicates are ready
@@ -372,6 +372,15 @@ void Server::start_io_threads(std::vector<std::thread> &threads) {
               perror("setsockopt SO_KEEPALIVE");
             }
 
+            // Add to epoll with no interests
+            struct epoll_event event{};
+            event.events = 0;
+            event.data.fd = fd;
+            if (epoll_ctl(epollfd, EPOLL_CTL_ADD, fd, &event) == -1) {
+              perror("epoll_ctl ADD");
+              (void)connection_registry_.remove(fd);
+            }
+
             // Transfer ownership from registry to batch processing
             size_t epoll_index = thread_id % epoll_fds_.size();
             batch[batch_count] = std::unique_ptr<Connection>(new Connection(
@@ -385,8 +394,7 @@ void Server::start_io_threads(std::vector<std::thread> &threads) {
             // Process batch if full
             if (batch_count == config_.server.event_batch_size) {
               process_connection_batch(epollfd, {batch, (size_t)batch_count},
-                                       {batch_events, (size_t)batch_count},
-                                       true);
+                                       {batch_events, (size_t)batch_count});
               batch_count = 0;
             }
           } // End inner accept loop
@@ -395,7 +403,7 @@ void Server::start_io_threads(std::vector<std::thread> &threads) {
         // Process remaining accepted connections
         if (batch_count > 0) {
           process_connection_batch(epollfd, {batch, (size_t)batch_count},
-                                   {batch_events, (size_t)batch_count}, true);
+                                   {batch_events, (size_t)batch_count});
           batch_count = 0;
         }
       }
@@ -460,7 +468,7 @@ void Server::process_connection_io(std::unique_ptr<Connection> &conn,
 
 void Server::process_connection_batch(
     int epollfd, std::span<std::unique_ptr<Connection>> batch,
-    std::span<const int> events, bool is_new) {
+    std::span<const int> events) {
   // First process I/O for each connection
   for (size_t i = 0; i < batch.size(); ++i) {
     if (batch[i]) {
@@ -487,9 +495,8 @@ void Server::process_connection_batch(
       // Put connection back in registry since handler didn't take ownership.
       // Must happen before epoll_ctl
       connection_registry_.store(fd, std::move(conn_ptr));
-      int epoll_op = is_new ? EPOLL_CTL_ADD : EPOLL_CTL_MOD;
-      if (epoll_ctl(epollfd, epoll_op, fd, &event) == -1) {
-        perror(is_new ? "epoll_ctl ADD" : "epoll_ctl MOD");
+      if (epoll_ctl(epollfd, EPOLL_CTL_MOD, fd, &event) == -1) {
+        perror("epoll_ctl MOD");
         (void)connection_registry_.remove(fd);
       }
     }
diff --git a/src/server.hpp b/src/server.hpp
index c5dd44e..268ec51 100644
--- a/src/server.hpp
+++ b/src/server.hpp
@@ -158,7 +158,7 @@ private:
   // Helper for processing a batch of connections with their events
   void process_connection_batch(int epollfd,
                                 std::span<std::unique_ptr<Connection>> batch,
-                                std::span<const int> events, bool is_new);
+                                std::span<const int> events);
 
   /**
    * Called internally to return ownership to the server.