From c238ce9737f7e59d3aeb9ebf354aa877d09deb36 Mon Sep 17 00:00:00 2001
From: Benjamin Wang <benjamin.ahrtr@gmail.com>
Date: Mon, 20 Jan 2025 18:25:42 +0000
Subject: [PATCH] add comment to clarify the etcd shutting down workflow

Signed-off-by: Benjamin Wang <benjamin.ahrtr@gmail.com>
---
 server/embed/etcd.go | 38 +++++++++++++++++++++++++++++++++++---
 1 file changed, 35 insertions(+), 3 deletions(-)

diff --git a/server/embed/etcd.go b/server/embed/etcd.go
index 5f8f7e33d960..9b4bccdaba72 100644
--- a/server/embed/etcd.go
+++ b/server/embed/etcd.go
@@ -388,6 +388,24 @@ func (e *Etcd) Config() Config {
 // Close gracefully shuts down all servers/listeners.
 // Client requests will be terminated with request timeout.
 // After timeout, enforce remaning requests be closed immediately.
+//
+// The rough workflow to shut down etcd:
+//  1. close the `stopc` channel, so that all error handlers (child
+//     goroutines) won't send back any errors anymore;
+//  2. close all client and metrics listeners, so that etcd server
+//     stops receiving any new connection immediately;
+//  3. stop the http and grpc servers gracefully, within request timeout;
+//  4. call the cancel function to close the gateway context, so that
+//     all gateway connections are closed.
+//  5. stop etcd server gracefully, and ensure the main raft loop
+//     goroutine is stopped;
+//  6. stop all peer listeners, so that it stops receives peer connections
+//     and messages (wait up to 1-second);
+//  7. wait for all child goroutines (i.e. client handlers, peer handlers
+//     and metrics handlers) to exit;
+//  8. close the `errc` channel to release the resource. Note that it's only
+//     safe to close the `errc` after step 6 above is done, otherwise the
+//     child goroutines may send errors back to already closed `errc` channel.
 func (e *Etcd) Close() {
 	fields := []zap.Field{
 		zap.String("name", e.cfg.Name),
@@ -407,10 +425,14 @@ func (e *Etcd) Close() {
 		lg.Sync()
 	}()
 
+	//  1. close the `stopc` channel, so that all error handlers (child
+	//     goroutines) won't send back any errors anymore;
 	e.closeOnce.Do(func() {
 		close(e.stopc)
 	})
 
+	//  2. close all client and metrics listeners, so that etcd server
+	//     stops receiving any new connection immediately;
 	for i := range e.Clients {
 		if e.Clients[i] != nil {
 			e.Clients[i].Close()
@@ -421,7 +443,7 @@ func (e *Etcd) Close() {
 		e.metricsListeners[i].Close()
 	}
 
-	// close client requests with request timeout
+	//  3. stop the http and grpc servers gracefully, within request timeout;
 	timeout := 2 * time.Second
 	if e.Server != nil {
 		timeout = e.Server.Cfg.ReqTimeout()
@@ -434,6 +456,8 @@ func (e *Etcd) Close() {
 		}
 	}
 
+	//  4. call the cancel function to close the gateway context, so that
+	//     all gateway connections are closed.
 	for _, sctx := range e.sctxs {
 		sctx.cancel()
 	}
@@ -443,12 +467,14 @@ func (e *Etcd) Close() {
 		e.tracingExporterShutdown()
 	}
 
-	// close rafthttp transports
+	//  5. stop etcd server gracefully, and ensure the main raft loop
+	//     goroutine is stopped;
 	if e.Server != nil {
 		e.Server.Stop()
 	}
 
-	// close all idle connections in peer handler (wait up to 1-second)
+	//  6. stop all peer listeners, so that it stops receives peer connections
+	//     and messages (wait up to 1-second);
 	for i := range e.Peers {
 		if e.Peers[i] != nil && e.Peers[i].close != nil {
 			ctx, cancel := context.WithTimeout(context.Background(), time.Second)
@@ -457,7 +483,13 @@ func (e *Etcd) Close() {
 		}
 	}
 	if e.errc != nil {
+		//  7. wait for all child goroutines (i.e. client handlers, peer handlers
+		//     and metrics handlers) to exit;
 		e.wg.Wait()
+
+		//  8. close the `errc` channel to release the resource. Note that it's only
+		//     safe to close the `errc` after step 6 above is done, otherwise the
+		//     child goroutines may send errors back to already closed `errc` channel.
 		close(e.errc)
 	}
 }