feat(benchmarks): update database handling and add new benchmark scripts

caddysnake.c

@@ -89,6 +89,8 @@ MapKeyVal *MapKeyVal_new(size_t capacity) {
   new_map->capacity = capacity;
   new_map->keys = malloc(sizeof(char *) * capacity);
   new_map->values = malloc(sizeof(char *) * capacity);
+  new_map->keysLen = malloc(sizeof(int) * capacity);
+  new_map->valuesLen = malloc(sizeof(int) * capacity);
   return new_map;
 }
 
@@ -108,6 +110,15 @@ static void MapKeyVal_append(MapKeyVal *map, char *key, char *value) {
   map->length++;
 }
 
+static void MapKeyVal_append_v2(MapKeyVal *map, char *key, int keyLen,
+                                char *value, int valueLen) {
+  map->keys[map->length] = key;
+  map->keysLen[map->length] = keyLen;
+  map->values[map->length] = value;
+  map->valuesLen[map->length] = valueLen;
+  map->length++;
+}
+
 typedef struct {
   PyObject_HEAD WsgiApp *app;
   int64_t request_id;
@@ -605,6 +616,7 @@ void AsgiEvent_set(AsgiEvent *self, const char *body, size_t body_len,
       Py_DECREF(self->request_body);
     }
     self->request_body = PyBytes_FromStringAndSize(body, body_len);
+    free(body);
   }
   self->more_body = more_body;
   PyObject *set_fn = NULL;
@@ -725,7 +737,16 @@ static void AsgiEvent_websocket_disconnect(AsgiEvent *self, PyObject *data) {
 
 static PyObject *AsgiEvent_receive_start(AsgiEvent *self, PyObject *args) {
   PyObject *result = Py_False;
-  if (asgi_receive_start(self->request_id, self) == 1) {
+
+  size_t cbuf_size = 1 << 13;
+  char *cbuf = malloc(cbuf_size);
+
+  PyThreadState *_save = PyEval_SaveThread();
+  uint8_t receive_result =
+      asgi_receive_start(self->request_id, self, cbuf, cbuf_size);
+  PyEval_RestoreThread(_save);
+
+  if (receive_result == 1) {
     Py_INCREF(self->event_ts_receive);
     result = self->event_ts_receive;
   }
@@ -823,7 +844,8 @@ static void AsgiEvent_response_start(AsgiEvent *self, PyObject *data) {
   MapKeyVal *http_headers = MapKeyVal_new(headers_count);
 
   PyObject *key, *value, *item;
-  size_t len = 0;
+  size_t keyLen = 0;
+  size_t valueLen = 0;
   while ((item = PyIter_Next(iterator))) {
     // if (!PyTuple_Check(item) || PyTuple_Size(item) != 2) {
     //   PyErr_SetString(PyExc_RuntimeError,
@@ -837,8 +859,9 @@ static void AsgiEvent_response_start(AsgiEvent *self, PyObject *data) {
     // }
     key = PyTuple_GetItem(item, 0);
     value = PyTuple_GetItem(item, 1);
-    MapKeyVal_append(http_headers, copy_pybytes(key, &len),
-                     copy_pybytes(value, &len));
+    char *keyBuf = copy_pybytes(key, &keyLen);
+    char *valueBuf = copy_pybytes(value, &valueLen);
+    MapKeyVal_append_v2(http_headers, keyBuf, keyLen, valueBuf, valueLen);
     Py_DECREF(item);
   }
   Py_DECREF(iterator);
@@ -856,7 +879,10 @@ static void AsgiEvent_write_body(AsgiEvent *self, PyObject *data) {
   PyObject *pybody = PyDict_GetItemString(data, "body");
   size_t body_len = 0;
   char *body = copy_pybytes(pybody, &body_len);
+
+  PyThreadState *_save = PyEval_SaveThread();
   asgi_send_response(self->request_id, body, body_len, send_more_body, self);
+  PyEval_RestoreThread(_save);
 }
 
 static void AsgiEvent_websocket_accept(AsgiEvent *self, PyObject *data) {

caddysnake.go

@@ -15,6 +15,7 @@
 	"net/url"
 	"os"
 	"path/filepath"
+	"reflect"
 	"regexp"
 	"strconv"
 	"strings"
@@ -29,12 +30,16 @@
 	"github.com/caddyserver/caddy/v2/modules/caddyhttp"
 	"github.com/gorilla/websocket"
 	"go.uber.org/zap"
+
+	// debug on Linux
+	_ "github.com/ianlancetaylor/cgosymbolizer"
 )
 
 //go:embed caddysnake.py
 var caddysnake_py string
 
 var SIZE_OF_CHAR_POINTER = unsafe.Sizeof((*C.char)(nil))
+var SIZE_OF_INT_POINTER = unsafe.Sizeof(C.int(0))
 
 // MapKeyVal wraps the same structure defined in the C layer
 type MapKeyVal struct {
@@ -85,7 +90,25 @@
 	headerValuePtr := unsafe.Pointer(uintptr(unsafe.Pointer(m.m.values)) + uintptr(pos)*SIZE_OF_CHAR_POINTER)
 	headerName := *(**C.char)(headerNamePtr)
 	headerValue := *(**C.char)(headerValuePtr)
-	return C.GoString(headerName), C.GoString(headerValue)
+
+	headerNameLenPtr := unsafe.Pointer(uintptr(unsafe.Pointer(m.m.keysLen)) + uintptr(pos)*SIZE_OF_INT_POINTER)
+	headerValueLenPtr := unsafe.Pointer(uintptr(unsafe.Pointer(m.m.valuesLen)) + uintptr(pos)*SIZE_OF_INT_POINTER)
+	headerNameLen := int(*(*C.int)(headerNameLenPtr))
+	headerValueLen := int(*(*C.int)(headerValueLenPtr))
+
+	hdr := reflect.StringHeader{
+		Data: uintptr(unsafe.Pointer(headerName)),
+		Len:  headerNameLen,
+	}
+	headerNameStr := *(*string)(unsafe.Pointer(&hdr))
+
+	hdr2 := reflect.StringHeader{
+		Data: uintptr(unsafe.Pointer(headerValue)),
+		Len:  headerValueLen,
+	}
+	headerValueStr := *(*string)(unsafe.Pointer(&hdr2))
+
+	return headerNameStr, headerValueStr
 }
 
 func (m *MapKeyVal) Len() int {
@@ -1036,21 +1059,20 @@
 	h.event = event
 }
 
-func (h *AsgiRequestHandler) readBody(event *C.AsgiEvent) {
+func (h *AsgiRequestHandler) readBody(event *C.AsgiEvent, cbuf *C.char, cbufSize C.size_t) {
 	var bodyStr *C.char
 	var bodyLen C.size_t
 	var moreBody C.uint8_t
 	if !h.completedBody {
-		buffer := make([]byte, 1<<16)
+		buffer := unsafe.Slice((*byte)(unsafe.Pointer(cbuf)), cbufSize)
 		n, err := h.r.Body.Read(buffer)
 		if err != nil && err != io.EOF {
 			h.done <- err
 			return
 		}
 		h.completedBody = (err == io.EOF)
-		buffer = append(buffer[:n], 0)
-		bodyStr = (*C.char)(unsafe.Pointer(&buffer[0]))
-		bodyLen = C.size_t(len(buffer) - 1) // -1 to remove null-terminator
+		bodyStr = cbuf
+		bodyLen = C.size_t(n)
 	}
 
 	if h.completedBody {
@@ -1064,9 +1086,9 @@
 	})
 }
 
-func (h *AsgiRequestHandler) ReceiveStart(event *C.AsgiEvent) C.uint8_t {
+func (h *AsgiRequestHandler) ReceiveStart(event *C.AsgiEvent, cbuf *C.char, cbufSize C.size_t) C.uint8_t {
 	h.operations <- AsgiOperations{op: func() {
-		h.readBody(event)
+		h.readBody(event, cbuf, cbufSize)
 	}}
 	return C.uint8_t(1)
 }
@@ -1082,6 +1104,7 @@
 	wsConn, err := upgrader.Upgrade(h.w, h.r, headers)
 	if err != nil {
 		h.websocketState = WS_DISCONNECTED
+		h.websocketConn.Close()
 		C.AsgiEvent_websocket_set_disconnected(event)
 		C.AsgiEvent_set(event, nil, 0, C.uint8_t(0), C.uint8_t(1))
 		return
@@ -1126,6 +1149,8 @@
 
 		h.w.WriteHeader(int(statusCode))
 
+		h.w.(http.Flusher).Flush()
+
 		pythonMainThread.do(func() {
 			C.AsgiEvent_set(event, nil, 0, C.uint8_t(0), C.uint8_t(1))
 		})
@@ -1135,7 +1160,7 @@
 func (h *AsgiRequestHandler) SendResponse(body *C.char, bodyLen C.size_t, moreBody C.uint8_t, event *C.AsgiEvent) {
 	h.operations <- AsgiOperations{op: func() {
 		defer C.free(unsafe.Pointer(body))
-		bodyBytes := C.GoBytes(unsafe.Pointer(body), C.int(bodyLen))
+		bodyBytes := unsafe.Slice((*byte)(unsafe.Pointer(body)), bodyLen)
 		h.accumulatedResponseSize += len(bodyBytes)
 		_, err := h.w.Write(bodyBytes)
 		if f, ok := h.w.(http.Flusher); ok {
@@ -1210,7 +1235,7 @@
 }
 
 //export asgi_receive_start
-func asgi_receive_start(requestID C.uint64_t, event *C.AsgiEvent) C.uint8_t {
+func asgi_receive_start(requestID C.uint64_t, event *C.AsgiEvent, cbuf *C.char, cbufSize C.size_t) C.uint8_t {
 	h := asgiState.GetHandler(uint64(requestID))
 	if h == nil || h.completedResponse {
 		return C.uint8_t(0)
@@ -1221,7 +1246,7 @@
 		return h.HandleWebsocket(event)
 	}
 
-	return h.ReceiveStart(event)
+	return h.ReceiveStart(event, cbuf, cbufSize)
 }
 
 //export asgi_set_headers

caddysnake.h

@@ -13,6 +13,8 @@ typedef struct {
   size_t capacity;
   char **keys;
   char **values;
+  int *keysLen;
+  int *valuesLen;
 } MapKeyVal;
 MapKeyVal *MapKeyVal_new(size_t);
 void MapKeyVal_free(MapKeyVal *map);
@@ -43,7 +45,7 @@ void AsgiEvent_websocket_set_disconnected(AsgiEvent *);
 void AsgiEvent_cleanup(AsgiEvent *);
 void AsgiApp_cleanup(AsgiApp *);
 
-extern uint8_t asgi_receive_start(uint64_t, AsgiEvent *);
+extern uint8_t asgi_receive_start(uint64_t, AsgiEvent *, char *, size_t);
 extern void asgi_send_response(uint64_t, char *, size_t, uint8_t, AsgiEvent *);
 extern void asgi_send_response_websocket(uint64_t, char *, size_t, uint8_t,
                                          AsgiEvent *);

tests/benchmarks/Caddyfile

@@ -0,0 +1,17 @@
+{
+	http_port 9080
+	admin off
+	auto_https off
+	log {
+		level info
+	}
+}
+:9080 {
+	route /* {
+		python {
+			module_asgi "main:app"
+			venv "./venv"
+			lifespan on
+		}
+	}
+}

tests/benchmarks/Dockerfile

@@ -0,0 +1,23 @@
+FROM python:3.13-bullseye
+
+# Install go version 1.24
+RUN apt-get update && \
+    apt-get install -y wget bc wrk git && \
+    wget https://go.dev/dl/go1.24.0.linux-arm64.tar.gz && \
+    tar -C /usr/local -xzf go1.24.0.linux-arm64.tar.gz && \
+    rm go1.24.0.linux-arm64.tar.gz &&\
+    pip install fastapi uvicorn hypercorn granian asyncpg pydantic
+
+ENV PATH=$PATH:/usr/local/go/bin
+ENV GODEBUG=cgocheck=0
+
+RUN go install github.com/caddyserver/xcaddy/cmd/xcaddy@latest
+
+COPY . /caddy-snake
+
+RUN cd /caddy-snake &&\
+    # git apply tests/benchmarks/caddysnake.patch &&\
+    cd /usr/local/bin &&\
+    CGO_ENABLED=1 /root/go/bin/xcaddy build --with github.com/mliezun/caddy-snake=/caddy-snake
+
+WORKDIR /caddy-snake/tests/benchmarks

tests/benchmarks/benchmark.sh

@@ -0,0 +1,50 @@
+#!/bin/bash
+
+set -euo pipefail
+
+HOST=localhost
+DURATION=20s
+THREADS=2
+CONNECTIONS=20
+RESULTS_FILE="results.txt"
+
+
+function run_benchmark() {
+    SERVER_CMD="$1"
+    SERVER_NAME="$2"
+    PORT="$3"
+    URL="http://$HOST:$PORT/pastes"
+    LOG_FILE="${SERVER_NAME}_log.txt"
+    echo "=== Starting $SERVER_NAME ===" | tee -a "$RESULTS_FILE"
+    # Start server in background
+    eval "$SERVER_CMD" > "$LOG_FILE" 2>&1 &
+    # Wait for server to be ready
+    sleep 10
+    for i in {1..30}; do
+        if curl -s "$URL" -o /dev/null; then
+            break
+        fi
+        sleep 1
+    done
+    echo "Running wrk benchmark on $SERVER_NAME..." | tee -a "$RESULTS_FILE"
+    wrk -t$THREADS -c$CONNECTIONS -d$DURATION -s "wrk_post.lua" "$URL" | tee -a "$RESULTS_FILE"
+    # Find and kill the server process by command line
+    ps aux | grep "$SERVER_CMD" | grep -v grep | awk '{print $2}' | xargs -r kill
+    echo "=== Finished $SERVER_NAME ===" | tee -a "$RESULTS_FILE"
+}
+
+# Clean previous results
+rm -f "$RESULTS_FILE"
+
+run_benchmark "./caddy run --config Caddyfile" "caddy" 9080
+run_benchmark "uvicorn main:app --host 0.0.0.0 --port 9081" "uvicorn" 9081
+run_benchmark "hypercorn main:app --bind 0.0.0.0:9082" "hypercorn" 9082
+run_benchmark "granian --interface asgi --host 0.0.0.0 --port 9083 main:app" "granian" 9083
+
+echo "=== Benchmark Summary ===" | tee -a "$RESULTS_FILE"
+grep -E 'Running wrk benchmark|Requests/sec|Latency' "$RESULTS_FILE" | tee -a "$RESULTS_FILE"
+
+# Print best performer
+BEST=$(grep 'Requests/sec' "$RESULTS_FILE" | awk '{print $2" "$3}' | sort -nr | head -1)
+BEST_SERVER=$(grep -B2 "$BEST" "$RESULTS_FILE" | head -1 | awk '{print $3}')
+echo -e "Best performer: $BEST_SERVER with $BEST requests/sec" | tee -a "$RESULTS_FILE"

tests/benchmarks/benchmark_v2.sh

@@ -0,0 +1,32 @@
+#!/bin/bash
+
+set -euo pipefail
+
+SIZE=$1
+
+HOST="[2001:bc8:1210:6e9e:dc00:ff:fe7a:770f]"
+DURATION=20s
+THREADS=2
+CONNECTIONS=20
+RESULTS_FILE="results.txt"
+
+
+function run_benchmark() {
+    SERVER_NAME="$1"
+    PORT="$2"
+    URL="http://$HOST:$PORT/pastes"
+    LOG_FILE="${SERVER_NAME}_log.txt"
+    echo "Running wrk benchmark on $SERVER_NAME..." | tee -a "$RESULTS_FILE"
+    wrk -t$THREADS -c$CONNECTIONS -d$DURATION -s "wrk_post_$SIZE.lua" "$URL" | tee -a "$RESULTS_FILE"
+}
+
+# Clean previous results
+rm -f "$RESULTS_FILE"
+
+run_benchmark "caddy" 9080
+run_benchmark "uvicorn" 9081
+run_benchmark "hypercorn" 9082
+run_benchmark "granian" 9083
+
+echo "=== Benchmark Summary ===" | tee -a "$RESULTS_FILE"
+grep -E 'Running wrk benchmark|Requests/sec|Latency' "$RESULTS_FILE" | tee -a "$RESULTS_FILE"

tests/benchmarks/caddysnake.patch

@@ -0,0 +1,29 @@
+diff --git a/caddysnake.go b/caddysnake.go
+index bf24510..63f73af 100644
+--- a/caddysnake.go
++++ b/caddysnake.go
+@@ -16,6 +16,7 @@ import (
+ 	"os"
+ 	"path/filepath"
+ 	"regexp"
++	"runtime/pprof"
+ 	"strconv"
+ 	"strings"
+ 	"sync"
+@@ -169,6 +170,8 @@ func (CaddySnake) CaddyModule() caddy.ModuleInfo {
+
+ // Provision sets up the module.
+ func (f *CaddySnake) Provision(ctx caddy.Context) error {
++	cpu_prof, _ := os.Create("cpu_profile.pprof")
++	pprof.StartCPUProfile(cpu_prof)
+ 	f.logger = ctx.Logger(f)
+ 	if f.ModuleWsgi != "" {
+ 		w, err := NewWsgi(f.ModuleWsgi, f.WorkingDir, f.VenvPath)
+@@ -200,6 +203,7 @@ func (m *CaddySnake) Validate() error {
+
+ // Cleanup frees resources uses by module
+ func (m *CaddySnake) Cleanup() error {
++	defer pprof.StopCPUProfile()
+ 	if m != nil && m.app != nil {
+ 		m.logger.Info("cleaning up module")
+ 		return m.app.Cleanup()