yangerd: containers: widen event debounce to avoid podman contention

Author: mattiaswal

src/yangerd/internal/containermonitor/containermonitor.go

@@ -31,8 +31,15 @@ import (
 const (
 	treeKey = "infix-containers:containers"
 
-	// debounceDelay coalesces bursts of events into one re-read.
-	debounceDelay = 200 * time.Millisecond
+	// debounceDelay coalesces bursts of events into one re-read.  It is
+	// deliberately generous: container lifecycle events fire while confd
+	// is still running its own `podman` start/stop/rm operations, so
+	// re-reading too eagerly makes yangerd's `podman ps/inspect/stats`
+	// contend with confd for the libpod lock on a CPU-starved guest.
+	// Waiting for the churn to settle keeps yangerd off confd's back
+	// during config apply/reset; a couple of seconds of staleness in
+	// operational data is harmless.
+	debounceDelay = 2 * time.Second
 )
 
 // ContainerMonitor subscribes to container lifecycle events via a

src/yangerd/internal/containermonitor/containermonitor_test.go

@@ -64,7 +64,8 @@ func TestEventTriggersRefresh(t *testing.T) {
 	// A container "died" event, newline-framed as podman emits it.
 	go m.readEvents(strings.NewReader(`{"Type":"container","Status":"died","Name":"gone"}` + "\n"))
 
-	deadline := time.After(2 * time.Second)
+	// Must comfortably exceed debounceDelay, or this races the re-read.
+	deadline := time.After(debounceDelay + 3*time.Second)
 	for {
 		if tr.Get(treeKey) == nil && calls > 0 {
 			break