Backlog/v12 agent connection

installer/templates/front-end.go

@@ -70,46 +70,51 @@ server {
         proxy_request_buffering off;
     }
 
-    # The agent's persistent gRPC streams to agent-manager. nginx's
-    # grpc_*_timeout of 0 means "no inactivity timeout" (removing the
-    # directive would silently fall back to a 60s default), so an idle
-    # AgentStream / PingService / CollectorService is never torn down by the
-    # proxy. Liveness is carried by the gRPC keepalive (30s ping / 10s
-    # timeout on both ends) plus grpc_socket_keepalive, which still reaps a
-    # genuinely dead peer.
+    # The agent's persistent gRPC streams to agent-manager. nginx has no
+    # "no timeout" for upstream read/send (0 means "fire immediately", which
+    # breaks every request), so use a large inactivity backstop of 24h.
+    # gRPC keepalive PINGs are HTTP/2 control frames - nginx only ACKs them
+    # and they do NOT reset these timers; only DATA does. Every timer below
+    # must therefore be large for streams legitimately silent between calls.
     location /agent.AgentService/ {
         grpc_pass grpcs://$utmstack_agent_manager_grpc;
-        grpc_read_timeout 0;
-        grpc_send_timeout 0;
-		client_body_timeout 1h;
+        grpc_read_timeout 86400;
+        grpc_send_timeout 86400;
+        client_body_timeout 86400;
         grpc_socket_keepalive on;
     }
 
     location /agent.PanelService/ {
         grpc_pass grpcs://$utmstack_agent_manager_grpc;
-        grpc_read_timeout 0;
-        grpc_send_timeout 0;
+        grpc_read_timeout 86400;
+        grpc_send_timeout 86400;
+        client_body_timeout 86400;
         grpc_socket_keepalive on;
     }
 
     location /agent.CollectorService/ {
         grpc_pass grpcs://$utmstack_agent_manager_grpc;
-        grpc_read_timeout 0;
-        grpc_send_timeout 0;
+        grpc_read_timeout 86400;
+        grpc_send_timeout 86400;
+        client_body_timeout 86400;
         grpc_socket_keepalive on;
     }
 
     # log-input's ingest, whose service lives in the SDK's "plugins" package.
+    # log-input's ingest: also a persistent collector stream, so the same
+    # 24h inactivity backstop (a quiet host can go >60s without a log line).
     location /plugins.Integration/ {
         grpc_pass grpcs://$utmstack_log_input_grpc;
-        grpc_read_timeout 900;
-        grpc_send_timeout 900;
+        grpc_read_timeout 86400;
+        grpc_send_timeout 86400;
+        client_body_timeout 86400;
     }
 
     location /agent.PingService/ {
         grpc_pass grpcs://$utmstack_agent_manager_grpc;
-        grpc_read_timeout 0;
-        grpc_send_timeout 0;
+        grpc_read_timeout 86400;
+        grpc_send_timeout 86400;
+        client_body_timeout 86400;
         grpc_socket_keepalive on;
     }
 

installer/templates/proxy.go

@@ -27,17 +27,27 @@ server {
     location /agent. {
         grpc_pass grpc://127.0.0.1:10001;
         grpc_set_header x-shared-key $shared_key;
-        grpc_read_timeout 900;
-        grpc_send_timeout 900;
+        # Long-lived agent streams: nginx's gRPC keepalive PINGs are HTTP/2
+        # control frames and do NOT reset these timers; only DATA does.
+        # An AgentStream is pure DATA between commands, so the 60s-default
+        # client_body_timeout (request body inactivity) is what evicted idle
+        # agents at exactly 60s. Use a 24h inactivity backstop; liveness is
+        # the app's job (gRPC keepalive + TCP keepalive).
+        grpc_read_timeout 86400;
+        grpc_send_timeout 86400;
+        client_body_timeout 86400;
     }
 
     # log-input's ingest. Separate only because its proto package differs; the
     # router behind decides which of the two it is.
     location /plugins. {
         grpc_pass grpc://127.0.0.1:10001;
         grpc_set_header x-shared-key $shared_key;
-        grpc_read_timeout 900;
-        grpc_send_timeout 900;
+        # Same long-lived-stream treatment: a quiet collector can easily go
+        # more than 60s without a single log line on the stream.
+        grpc_read_timeout 86400;
+        grpc_send_timeout 86400;
+        client_body_timeout 86400;
     }
 
     location / {
@@ -105,17 +115,27 @@ server {
     location /agent. {
         grpc_pass grpc://127.0.0.1:10001;
         grpc_set_header x-shared-key $shared_key;
-        grpc_read_timeout 900;
-        grpc_send_timeout 900;
+        # Long-lived agent streams: nginx's gRPC keepalive PINGs are HTTP/2
+        # control frames and do NOT reset these timers; only DATA does.
+        # An AgentStream is pure DATA between commands, so the 60s-default
+        # client_body_timeout (request body inactivity) is what evicted idle
+        # agents at exactly 60s. Use a 24h inactivity backstop; liveness is
+        # the app's job (gRPC keepalive + TCP keepalive).
+        grpc_read_timeout 86400;
+        grpc_send_timeout 86400;
+        client_body_timeout 86400;
     }
 
     # log-input's ingest. Separate only because its proto package differs; the
     # router behind decides which of the two it is.
     location /plugins. {
         grpc_pass grpc://127.0.0.1:10001;
         grpc_set_header x-shared-key $shared_key;
-        grpc_read_timeout 900;
-        grpc_send_timeout 900;
+        # Same long-lived-stream treatment: a quiet collector can easily go
+        # more than 60s without a single log line on the stream.
+        grpc_read_timeout 86400;
+        grpc_send_timeout 86400;
+        client_body_timeout 86400;
     }
 
     location / {