@mark1russell7/machines 0.3.10 → 0.3.12

package/README.md

@@ -95,6 +95,14 @@ A change to the version of a system changes its sources too. Each fact of the do
 
 In this repository, a snapshot of the predictions of the cells (`packages/cells/src/__snapshots__/predictions.json`) shows each prediction that a change to a document moves.
 
+### 0.3.12
+
+- **The send timeout of the client socket of Supavisor.** Supavisor 2.9.13 locks ranch 2.2.0. ranch gives each client socket a send timeout of 30 s that closes the socket (src/ranch_tcp.erl:108-109). Thus a write of a reply that waits 30 s ends the client. `supavisor-client` gives it as a source and a note, with no timer. A write waits only when the client takes no data, and on a broken path the FIN of the close waits behind the data. In run 39170773426, behind an MTU black hole, the socket closed 30.068 to 30.086 s after the fault in each of 11 runs.
+
+### 0.3.11
+
+- **The connect timer of PgBouncer can fire early.** libevent 2.1 counts a timeout from its cached time of the pass, not from the clock (event.c, `event_add_nolock` with `gettime` and `tv_cache`). Thus server_connect_timeout can fire up to one pass early, and its band has an early slack of 10 ms. In run 39160356281 the timer closed the socket at 127.998993 s, 1.007 ms before 128 s.
+
 ### 0.3.10
 
 - **Supavisor 2.9.13, in transaction mode.** Two new documents, each from the Elixir source at the tag v2.9.13:

package/build/index.d.ts

@@ -283,6 +283,11 @@ export declare const POSTGRES_SESSION: Machine;
  * At the defaults the handler writes a heartbeat to an idle client each 60 s: a ParameterStatus of
  * application_name. A failed write ends the client. client_idle_timeout does not end a session in
  * 2.9.13: its timer fires an event that no clause handles. Each client socket has SO_KEEPALIVE.
+ *
+ * Each client socket also has the send timeout of ranch: a write to the client that waits 30 s fails,
+ * and the BEAM closes the socket. This document gives it no timer. A write waits when the client takes no
+ * data: a broken path, or a client that reads nothing. On a broken path, the FIN of the close waits
+ * behind the data that the kernel did not send.
  */
 export declare const SUPAVISOR_CLIENT: Machine;
 //#endregion

package/build/index.js

@@ -7824,6 +7824,10 @@ const PGBOUNCER_SERVER = machineSchema.parse({
 			ref: `${PB} src/janitor.c:558-570 the janitor closes a server in SV_LOGIN that is older than server_connect_timeout, from its launch: 'connect timeout'`,
 			rung: "implemented"
 		},
+		{
+			ref: "run 39160356281 of mark1russell7/systems: with server_connect_timeout 128 s, the connect timer closed the socket 127.998993 s after its SYN in 1 of 11 runs, 1.007 ms before 128 s. libevent 2.1 counts a timeout from its cached time of the pass (event.c event_add_nolock, gettime with tv_cache)",
+			rung: "observed"
+		},
 		{
 			ref: `${PB} src/objects.c:1385-1393, :1246-1248 a server in a login that closes marks the pool as failing, with the reason. A login that ends clears it`,
 			rung: "implemented"
@@ -8044,6 +8048,10 @@ const PGBOUNCER_SERVER = machineSchema.parse({
 				knob: "slack",
 				unit: "ms"
 			},
+			early: {
+				knob: "slack",
+				unit: "ms"
+			},
 			arms: ["enter:connecting"],
 			cancels: [
 				"socket.connect",
@@ -8051,7 +8059,7 @@ const PGBOUNCER_SERVER = machineSchema.parse({
 				"peer.rst"
 			],
 			fires: "timer.connect-timeout",
-			note: "src/objects.c:1640-1641, src/sbuf.c:163-203: a timeout of libevent on the wait for EV_WRITE, in whole seconds. Which clock libevent reads is an assumption of this document: the monotonic one"
+			note: "src/objects.c:1640-1641, src/sbuf.c:163-203: a timeout of libevent on the wait for EV_WRITE, in whole seconds. Which clock libevent reads is an assumption of this document: the monotonic one. libevent counts the timeout from the time that it caches for the callbacks of one pass of its loop, thus from up to one pass before the connect: the timer can fire up to that pass early"
 		},
 		login_timeout: {
 			duration: {
@@ -9607,6 +9615,11 @@ const SV = "v2.9.13";
 * At the defaults the handler writes a heartbeat to an idle client each 60 s: a ParameterStatus of
 * application_name. A failed write ends the client. client_idle_timeout does not end a session in
 * 2.9.13: its timer fires an event that no clause handles. Each client socket has SO_KEEPALIVE.
+*
+* Each client socket also has the send timeout of ranch: a write to the client that waits 30 s fails,
+* and the BEAM closes the socket. This document gives it no timer. A write waits when the client takes no
+* data: a broken path, or a client that reads nothing. On a broken path, the FIN of the close waits
+* behind the data that the kernel did not send.
 */
 const SUPAVISOR_CLIENT = machineSchema.parse({
 	format: "mark1russell7.machine/1",
@@ -9633,6 +9646,14 @@ const SUPAVISOR_CLIENT = machineSchema.parse({
 			ref: `${SV} lib/supavisor/application.ex:71-83 each listener is ranch_tcp with keepalive true: each client socket has SO_KEEPALIVE`,
 			rung: "implemented"
 		},
+		{
+			ref: `ranch 2.2.0, the version of mix.lock at ${SV}: src/ranch_tcp.erl:108-109 a listener sets send_timeout 30000 and send_timeout_close true, and each accepted socket has the options of its listener. Thus a write to the client that waits 30 s fails, and the BEAM closes the socket. The server connection ignores the result of each write of a reply to the client (lib/supavisor/db_handler.ex:890-905)`,
+			rung: "implemented"
+		},
+		{
+			ref: `${SV} run 39170773426 of mark1russell7/systems, T03.defaults.59: behind an MTU black hole at a router, each of 11 runs had a reply of 2 MB in flight. The client socket of Supavisor left ESTABLISHED for FIN_WAIT1 30.068 to 30.086 s after the fault, and the client got nothing in 60 s`,
+			rung: "observed"
+		},
 		{
 			ref: `${SV} lib/supavisor/client_handler.ex:21-23, :123, :198 a state_timeout of 5_000 ms in :handshake, and 2_500 ms for a TLS handshake`,
 			rung: "implemented"
@@ -9999,7 +10020,7 @@ const SUPAVISOR_CLIENT = machineSchema.parse({
 		},
 		sets: {},
 		ends: "client.gone",
-		note: "The socket of the client. ranch sets SO_KEEPALIVE and no other keepalive option, thus the three keepalive sysctls of the namespace apply"
+		note: "The socket of the client. ranch sets SO_KEEPALIVE and no other keepalive option, thus the three keepalive sysctls of the namespace apply. It also sets a send timeout of 30 s that closes the socket. That close has no timer here: the kernel sends the FIN only after the data before it, an assumption of this document"
 	}
 });
 //#endregion