fix(vanilla-io_uring): lazy json-comp gz cache (parity + drop pointless boot precompute)

frameworks/vanilla-io_uring/main.v

@@ -60,7 +60,8 @@ mut:
 	// json-comp cache: the gzipped response for a given (count, m) is fully
 	// deterministic and gzip dominates the cost, so compress once and reuse.
 	// Key = (count << 32) | m. The benchmark hits only a few pairs, so it's tiny.
-	gz_cache map[u64][]u8 // json-comp: precomputed at boot, READ-ONLY during serving (lock-free reads)
+	gz_cache map[u64][]u8
+	gz_mu    &sync.RwMutex = unsafe { nil }
 }
 
 struct CrudCreate {
@@ -385,39 +386,45 @@ fn (sh &Shared) write_json_response(mut out []u8, count int, m i64) {
 	ws(mut out, '}')
 }
 
-// gz_response builds the COMPLETE gzipped /json response for (count, m). Pure
-// function of the shared dataset (deterministic), so it is precomputed once at
-// boot (see main) and also serves the cold path for an unexpected param.
-fn gz_response(sh &Shared, count int, m i64) []u8 {
+// write_json_gzip is the json-comp path. The gzipped response for a given
+// (count, m) is fully deterministic and gzip CPU dominates the cost, so we cache
+// the COMPLETE response bytes and just append the cached copy on a hit — no
+// rebuild, no recompress. Compressing once instead of per-request is the whole
+// win for json-comp (the profile is compression-bound, not allocation-bound).
+//
+// Kept LAZY, not precomputed at boot: precomputing the full (count x m) grid would
+// compress ~800 keys vs the ~8 the profile sends. On the `-gc none` epoll sibling
+// that leaked ~135 MiB via gzip.compress scratch (vlang/v#27606); this io_uring
+// binary is default-GC so it would not leak, but lazy keeps both entries identical
+// and the shared lock is not a bottleneck (the json-comp@16384 collapse is thread
+// oversubscription from the co-hosted json-tls listener, enghitalo/vanilla#89).
+fn (mut sh Shared) write_json_gzip(mut out []u8, count int, m i64) {
+	key := (u64(u32(count)) << 32) | u64(u32(m))
+	sh.gz_mu.@rlock()
+	cached := sh.gz_cache[key] or { []u8{} }
+	sh.gz_mu.runlock()
+	if cached.len > 0 {
+		out << cached
+		return
+	}
 	body := sh.json_body(count, m)
-	gz := gzip.compress(body.bytes()) or { return []u8{} }
+	gz := gzip.compress(body.bytes()) or {
+		write_resp(mut out, 'application/json', body)
+		return
+	}
 	mut resp := []u8{cap: gz.len + 128}
 	ws(mut resp,
 		'HTTP/1.1 200 OK\r\nServer: vanilla\r\nContent-Encoding: gzip\r\nContent-Type: application/json\r\nContent-Length: ')
 	wi(mut resp, i64(gz.len))
 	ws(mut resp, '\r\nConnection: keep-alive\r\n\r\n')
 	unsafe { resp.push_many(gz.data, gz.len) }
-	return resp
-}
-
-// write_json_gzip is the json-comp path. The gzipped response per (count, m) is
-// deterministic, so the cache is fully precomputed at boot and is READ-ONLY during
-// serving — the hot path is a lock-free map read (no per-request shared RwMutex,
-// which contended across all workers and collapsed io_uring json-comp @16384).
-fn (sh &Shared) write_json_gzip(mut out []u8, count int, m i64) {
-	key := (u64(u32(count)) << 32) | u64(u32(m))
-	if cached := sh.gz_cache[key] {
-		out << cached
-		return
-	}
-	// Param outside the precomputed grid (not sent by the benchmark): build and
-	// send WITHOUT caching, so the hot path stays write-free and lock-free.
-	resp := gz_response(sh, count, m)
-	if resp.len > 0 {
-		out << resp
-	} else {
-		write_resp(mut out, 'application/json', sh.json_body(count, m))
+	// Store it (bounded so a flood of distinct m values can't grow it without limit).
+	sh.gz_mu.@lock()
+	if sh.gz_cache.len < 1024 {
+		sh.gz_cache[key] = resp
 	}
+	sh.gz_mu.unlock()
+	out << resp
 }
 
 // json_body builds just the /json body string (used for the gzip path).
@@ -762,20 +769,7 @@ fn main() {
 		cache:    map[int]string{}
 		cache_mu: sync.new_rwmutex()
 		gz_cache: map[u64][]u8{}
-	}
-
-	// Precompute the json-comp (gzip) responses at boot so write_json_gzip reads the
-	// cache LOCK-FREE during serving (no per-request shared RwMutex — that contended
-	// across workers and collapsed io_uring json-comp @16384 conns). The profile's
-	// params are bounded; cover count 1..min(dataset,64) x m 1..16.
-	gz_cap_count := if dataset.len < 64 { dataset.len } else { 64 }
-	for c in 1 .. gz_cap_count + 1 {
-		for mm in 1 .. 17 {
-			r := gz_response(sh, c, i64(mm))
-			if r.len > 0 {
-				sh.gz_cache[(u64(u32(c)) << 32) | u64(u32(mm))] = r
-			}
-		}
+		gz_mu:    sync.new_rwmutex()
 	}
 
 	// ── json-tls profile: /json over HTTPS on :8081 via the epoll + kTLS backend ──

site/data/api-16-1024.json

@@ -1507,28 +1507,28 @@
   {
     "framework": "vanilla-io_uring",
     "language": "V",
-    "rps": 29136,
-    "avg_latency": "33.91ms",
-    "p99_latency": "270.00ms",
-    "cpu": "1499.7%",
-    "memory": "2.3GiB",
+    "rps": 28450,
+    "avg_latency": "34.72ms",
+    "p99_latency": "268.30ms",
+    "cpu": "1459.0%",
+    "memory": "2.1GiB",
     "connections": 1024,
     "threads": 64,
     "duration": "5s",
     "pipeline": 1,
-    "bandwidth": "146.61MB/s",
-    "input_bw": "1.64MB/s",
-    "reconnects": 87278,
-    "status_2xx": 437050,
+    "bandwidth": "143.12MB/s",
+    "input_bw": "1.60MB/s",
+    "reconnects": 85218,
+    "status_2xx": 426754,
     "status_3xx": 0,
     "status_4xx": 0,
     "status_5xx": 0,
-    "tpl_baseline": 163773,
-    "tpl_json": 163787,
+    "tpl_baseline": 160037,
+    "tpl_json": 160077,
     "tpl_db": 0,
     "tpl_upload": 0,
     "tpl_static": 0,
-    "tpl_async_db": 109490
+    "tpl_async_db": 106640
   },
   {
     "framework": "workerman",

site/data/api-4-256.json

@@ -1507,28 +1507,28 @@
   {
     "framework": "vanilla-io_uring",
     "language": "V",
-    "rps": 28454,
-    "avg_latency": "7.83ms",
-    "p99_latency": "54.30ms",
-    "cpu": "360.5%",
-    "memory": "2.1GiB",
+    "rps": 28434,
+    "avg_latency": "7.65ms",
+    "p99_latency": "57.70ms",
+    "cpu": "356.8%",
+    "memory": "2.0GiB",
     "connections": 256,
     "threads": 64,
     "duration": "5s",
     "pipeline": 1,
-    "bandwidth": "143.08MB/s",
+    "bandwidth": "143.02MB/s",
     "input_bw": "1.60MB/s",
-    "reconnects": 85343,
-    "status_2xx": 426819,
+    "reconnects": 85290,
+    "status_2xx": 426516,
     "status_3xx": 0,
     "status_4xx": 0,
     "status_5xx": 0,
-    "tpl_baseline": 160147,
-    "tpl_json": 159929,
+    "tpl_baseline": 159969,
+    "tpl_json": 159971,
     "tpl_db": 0,
     "tpl_upload": 0,
     "tpl_static": 0,
-    "tpl_async_db": 106741
+    "tpl_async_db": 106576
   },
   {
     "framework": "workerman",

site/data/async-db-1024.json

@@ -1259,19 +1259,19 @@
   {
     "framework": "vanilla-io_uring",
     "language": "V",
-    "rps": 10645,
-    "avg_latency": "95.83ms",
-    "p99_latency": "405.90ms",
-    "cpu": "4906.4%",
-    "memory": "1.9GiB",
+    "rps": 11025,
+    "avg_latency": "92.40ms",
+    "p99_latency": "385.40ms",
+    "cpu": "5086.3%",
+    "memory": "1.8GiB",
     "connections": 1024,
     "threads": 64,
     "duration": "5s",
     "pipeline": 1,
-    "bandwidth": "40.86MB/s",
-    "input_bw": "727.69KB/s",
-    "reconnects": 3851,
-    "status_2xx": 106453,
+    "bandwidth": "42.50MB/s",
+    "input_bw": "753.66KB/s",
+    "reconnects": 3988,
+    "status_2xx": 110258,
     "status_3xx": 0,
     "status_4xx": 0,
     "status_5xx": 0

site/data/baseline-4096.json

@@ -1751,19 +1751,19 @@
   {
     "framework": "vanilla-io_uring",
     "language": "V",
-    "rps": 3815248,
-    "avg_latency": "1.07ms",
-    "p99_latency": "6.97ms",
-    "cpu": "5989.3%",
+    "rps": 3655431,
+    "avg_latency": "1.12ms",
+    "p99_latency": "7.14ms",
+    "cpu": "5411.7%",
     "memory": "1.5GiB",
     "connections": 4096,
     "threads": 64,
     "duration": "5s",
     "pipeline": 1,
-    "bandwidth": "389.16MB/s",
-    "input_bw": "294.72MB/s",
+    "bandwidth": "372.79MB/s",
+    "input_bw": "282.37MB/s",
     "reconnects": 0,
-    "status_2xx": 19076241,
+    "status_2xx": 18277158,
     "status_3xx": 0,
     "status_4xx": 0,
     "status_5xx": 0

site/data/baseline-512.json

@@ -1751,19 +1751,19 @@
   {
     "framework": "vanilla-io_uring",
     "language": "V",
-    "rps": 3326849,
-    "avg_latency": "153us",
-    "p99_latency": "1.41ms",
-    "cpu": "5576.7%",
+    "rps": 3275759,
+    "avg_latency": "155us",
+    "p99_latency": "1.43ms",
+    "cpu": "6403.6%",
     "memory": "1.5GiB",
     "connections": 512,
     "threads": 64,
     "duration": "5s",
     "pipeline": 1,
-    "bandwidth": "339.35MB/s",
-    "input_bw": "256.99MB/s",
+    "bandwidth": "334.17MB/s",
+    "input_bw": "253.04MB/s",
     "reconnects": 0,
-    "status_2xx": 16634248,
+    "status_2xx": 16378798,
     "status_3xx": 0,
     "status_4xx": 0,
     "status_5xx": 0

site/data/crud-4096.json

@@ -522,19 +522,19 @@
   {
     "framework": "vanilla-io_uring",
     "language": "V",
-    "rps": 221360,
-    "avg_latency": "18.47ms",
-    "p99_latency": "177.90ms",
-    "cpu": "1445.6%",
-    "memory": "1.7GiB",
+    "rps": 231065,
+    "avg_latency": "17.69ms",
+    "p99_latency": "178.50ms",
+    "cpu": "1458.0%",
+    "memory": "1.9GiB",
     "connections": 4096,
     "threads": 64,
     "duration": "5s",
     "pipeline": 1,
-    "bandwidth": "69.78MB/s",
-    "input_bw": "19.00MB/s",
-    "reconnects": 14598,
-    "status_2xx": 3320411,
+    "bandwidth": "73.10MB/s",
+    "input_bw": "19.83MB/s",
+    "reconnects": 15393,
+    "status_2xx": 3465984,
     "status_3xx": 0,
     "status_4xx": 0,
     "status_5xx": 0

site/data/fortunes-1024.json

@@ -192,18 +192,18 @@
   {
     "framework": "vanilla-io_uring",
     "language": "V",
-    "rps": 70,
-    "avg_latency": "2.27s",
-    "p99_latency": "5.00s",
-    "cpu": "1055.2%",
-    "memory": "1.4GiB",
+    "rps": 37,
+    "avg_latency": "2.96s",
+    "p99_latency": "4.99s",
+    "cpu": "966.9%",
+    "memory": "1.7GiB",
     "connections": 1024,
     "threads": 64,
     "duration": "5s",
     "pipeline": 1,
-    "bandwidth": "1.66MB/s",
+    "bandwidth": "902.86KB/s",
     "reconnects": 0,
-    "status_2xx": 350,
+    "status_2xx": 186,
     "status_3xx": 0,
     "status_4xx": 0,
     "status_5xx": 0

site/data/json-4096.json

@@ -1479,19 +1479,19 @@
   {
     "framework": "vanilla-io_uring",
     "language": "V",
-    "rps": 2407630,
+    "rps": 2415906,
     "avg_latency": "881us",
-    "p99_latency": "7.41ms",
-    "cpu": "6373.4%",
+    "p99_latency": "7.34ms",
+    "cpu": "6379.9%",
     "memory": "1.4GiB",
     "connections": 4096,
     "threads": 64,
     "duration": "5s",
     "pipeline": 1,
-    "bandwidth": "8.15GB/s",
-    "input_bw": "114.80MB/s",
-    "reconnects": 481458,
-    "status_2xx": 12038150,
+    "bandwidth": "8.18GB/s",
+    "input_bw": "115.20MB/s",
+    "reconnects": 483317,
+    "status_2xx": 12079530,
     "status_3xx": 0,
     "status_4xx": 0,
     "status_5xx": 0

site/data/json-comp-16384.json

@@ -1222,19 +1222,19 @@
   {
     "framework": "vanilla-io_uring",
     "language": "V",
-    "rps": 159300,
-    "avg_latency": "43.02ms",
-    "p99_latency": "1.48s",
-    "cpu": "437.8%",
+    "rps": 155961,
+    "avg_latency": "47.80ms",
+    "p99_latency": "1.79s",
+    "cpu": "484.9%",
     "memory": "2.0GiB",
     "connections": 16384,
     "threads": 64,
     "duration": "5s",
     "pipeline": 1,
-    "bandwidth": "260.34MB/s",
-    "input_bw": "11.85MB/s",
-    "reconnects": 30427,
-    "status_2xx": 796504,
+    "bandwidth": "254.75MB/s",
+    "input_bw": "11.60MB/s",
+    "reconnects": 29439,
+    "status_2xx": 779805,
     "status_3xx": 0,
     "status_4xx": 0,
     "status_5xx": 0