diff --git a/graphite.go b/graphite.go
index 264bb106..3ee4ed21 100644
--- a/graphite.go
+++ b/graphite.go
@@ -31,6 +31,7 @@ package nebula
 
 import (
 	"bufio"
+	"context"
 	"fmt"
 	"net"
 	"strconv"
@@ -51,17 +52,29 @@ type graphiteConfigExport struct {
 	Percentiles   []float64        // Percentiles to export from timers and histograms
 }
 
+// graphiteTimeout bounds one submission, the connect and the write together. An export is a few KB, a host that can't
+// take that in this long isn't going to.
+const graphiteTimeout = 30 * time.Second
+
 // graphiteOnce performs a single submission to Graphite, returning a non-nil
-// error on failed connections.
-func graphiteOnce(c graphiteConfigExport) error {
+// error if it could not connect or write all of it within timeout. It gives up
+// as soon as ctx is done.
+func graphiteOnce(ctx context.Context, c graphiteConfigExport, timeout time.Duration) error {
 	now := time.Now().Unix()
 	du := float64(c.DurationUnit)
 	flushSeconds := float64(c.FlushInterval) / float64(time.Second)
-	conn, err := net.DialTCP("tcp", nil, c.Addr)
+	deadline := time.Now().Add(timeout)
+	d := net.Dialer{Deadline: deadline}
+	conn, err := d.DialContext(ctx, "tcp", c.Addr.String())
 	if err != nil {
 		return err
 	}
 	defer conn.Close()
+	if err := conn.SetDeadline(deadline); err != nil {
+		return err
+	}
+	stop := context.AfterFunc(ctx, func() { _ = conn.Close() })
+	defer stop()
 	w := bufio.NewWriter(conn)
 	c.Registry.Each(func(name string, i any) {
 		switch metric := i.(type) {
@@ -111,7 +124,7 @@ func graphiteOnce(c graphiteConfigExport) error {
 			fmt.Fprintf(w, "%s.%s.fifteen-minute %.2f %d\n", c.Prefix, name, t.Rate15(), now)
 			fmt.Fprintf(w, "%s.%s.mean-rate %.2f %d\n", c.Prefix, name, t.RateMean(), now)
 		}
-		w.Flush()
 	})
-	return nil
+	// A failed write sticks, so this reports the first one
+	return w.Flush()
 }
diff --git a/graphite_test.go b/graphite_test.go
new file mode 100644
index 00000000..0a951da5
--- /dev/null
+++ b/graphite_test.go
@@ -0,0 +1,315 @@
+package nebula
+
+import (
+	"bytes"
+	"context"
+	"errors"
+	"fmt"
+	"io"
+	"net"
+	"strings"
+	"sync"
+	"testing"
+	"time"
+
+	"github.com/rcrowley/go-metrics"
+	"github.com/slackhq/nebula/test"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+// graphiteHost accepts connections and hands each one to the test
+type graphiteHost struct {
+	ln    net.Listener
+	conns chan net.Conn
+
+	mu     sync.Mutex
+	closed bool
+	open   []net.Conn
+}
+
+// newGraphiteHost accepts connections for the test. A non-zero window shrinks each connection's receive buffer, so a
+// sender that is not read from blocks after a little whatever the host's TCP tuning. Only for tests that need a stall,
+// a small window makes reading very slow on Linux.
+func newGraphiteHost(t *testing.T, window int) *graphiteHost {
+	t.Helper()
+	ln, err := net.Listen("tcp", "127.0.0.1:0")
+	require.NoError(t, err)
+	h := &graphiteHost{ln: ln, conns: make(chan net.Conn, 8)}
+	done := make(chan struct{})
+	t.Cleanup(func() {
+		_ = ln.Close()
+		h.mu.Lock()
+		h.closed = true
+		for _, c := range h.open {
+			_ = c.Close()
+		}
+		h.mu.Unlock()
+		close(done)
+	})
+	go func() {
+		for {
+			c, err := ln.Accept()
+			if err != nil {
+				return
+			}
+			h.mu.Lock()
+			if h.closed {
+				h.mu.Unlock()
+				_ = c.Close()
+				return
+			}
+			h.open = append(h.open, c)
+			h.mu.Unlock()
+			if tc, ok := c.(*net.TCPConn); ok && window > 0 {
+				_ = tc.SetReadBuffer(window)
+			}
+			select {
+			case h.conns <- c:
+			case <-done:
+				return
+			}
+		}
+	}()
+	return h
+}
+
+func (h *graphiteHost) addr(t *testing.T) *net.TCPAddr {
+	a, err := net.ResolveTCPAddr("tcp", h.ln.Addr().String())
+	require.NoError(t, err)
+	return a
+}
+
+// readExport reads what one send wrote, bounded so a send that never finishes fails the test rather than hanging it
+func readExport(t *testing.T, c net.Conn) []byte {
+	t.Helper()
+	require.NoError(t, c.SetReadDeadline(time.Now().Add(10*time.Second)))
+	b, err := io.ReadAll(c)
+	require.NoError(t, err)
+	return b
+}
+
+func (h *graphiteHost) next(t *testing.T, wait time.Duration) net.Conn {
+	t.Helper()
+	select {
+	case c := <-h.conns:
+		return c
+	case <-time.After(wait):
+		t.Fatal("no connection")
+		return nil
+	}
+}
+
+// lockedBuffer is a log destination a test can read while the code under test writes to it from another goroutine
+type lockedBuffer struct {
+	mu  sync.Mutex
+	buf bytes.Buffer
+}
+
+func (b *lockedBuffer) Write(p []byte) (int, error) {
+	b.mu.Lock()
+	defer b.mu.Unlock()
+	return b.buf.Write(p)
+}
+
+func (b *lockedBuffer) String() string {
+	b.mu.Lock()
+	defer b.mu.Unlock()
+	return b.buf.String()
+}
+
+func testGraphiteConfig(addr *net.TCPAddr, r metrics.Registry) graphiteConfigExport {
+	return graphiteConfigExport{Addr: addr, Registry: r, FlushInterval: time.Second, DurationUnit: time.Nanosecond, Prefix: "nebula"}
+}
+
+// Enough metrics that one export overflows the socket buffers on any platform, so a send to a host that stops
+// reading blocks
+func bigGraphiteRegistry(r metrics.Registry, t *testing.T) {
+	for i := range 200000 {
+		name := fmt.Sprintf("big.gauge.%d", i)
+		metrics.GetOrRegisterGauge(name, r).Update(int64(i))
+		t.Cleanup(func() { r.Unregister(name) })
+	}
+}
+
+func TestGraphiteOnce_sends(t *testing.T) {
+	r := metrics.NewRegistry()
+	metrics.GetOrRegisterGauge("g", r).Update(7)
+	metrics.GetOrRegisterCounter("c", r).Inc(3)
+	h := newGraphiteHost(t, 0)
+
+	require.NoError(t, graphiteOnce(t.Context(), testGraphiteConfig(h.addr(t), r), time.Minute))
+	b := readExport(t, h.next(t, 5*time.Second))
+	assert.Contains(t, string(b), "nebula.g.value 7 ")
+	assert.Contains(t, string(b), "nebula.c.count 3 ")
+	assert.Contains(t, string(b), "nebula.c.count_ps 3.00 ")
+}
+
+// A host that accepts and never reads is given up on, and the failed write is reported
+func TestGraphiteOnce_stalledHostTimesOut(t *testing.T) {
+	r := metrics.NewRegistry()
+	bigGraphiteRegistry(r, t)
+	h := newGraphiteHost(t, 4096)
+	done := make(chan error, 1)
+	cfg := testGraphiteConfig(h.addr(t), r)
+	go func() { done <- graphiteOnce(t.Context(), cfg, 300*time.Millisecond) }()
+	h.next(t, 5*time.Second) // accepted, never read
+
+	select {
+	case err := <-done:
+		require.Error(t, err)
+	case <-time.After(10 * time.Second):
+		t.Fatal("the send did not give up on a host that stopped reading")
+	}
+}
+
+// The timeout covers the whole send, a host that keeps reading but too slowly is still given up on
+func TestGraphiteOnce_timeoutCoversTheWholeSend(t *testing.T) {
+	r := metrics.NewRegistry()
+	bigGraphiteRegistry(r, t)
+	h := newGraphiteHost(t, 0)
+	done := make(chan error, 1)
+	start := time.Now()
+	cfg := testGraphiteConfig(h.addr(t), r)
+	go func() { done <- graphiteOnce(t.Context(), cfg, 500*time.Millisecond) }()
+	c := h.next(t, 5*time.Second)
+	stopReading := make(chan struct{})
+	defer close(stopReading)
+	// 32KB every 10ms keeps every write well inside the timeout, the whole export takes seconds
+	go func() {
+		buf := make([]byte, 32<<10)
+		for {
+			select {
+			case <-stopReading:
+				return
+			case <-time.After(10 * time.Millisecond):
+			}
+			if _, err := c.Read(buf); err != nil {
+				return
+			}
+		}
+	}()
+
+	select {
+	case err := <-done:
+		require.Error(t, err)
+		assert.Less(t, time.Since(start), 5*time.Second)
+	case <-time.After(10 * time.Second):
+		t.Fatal("a host that kept reading slowly held the send past its timeout")
+	}
+}
+
+func TestGraphiteOnce_cancelledCtxDoesNotConnect(t *testing.T) {
+	h := newGraphiteHost(t, 0)
+	ctx, cancel := context.WithCancel(t.Context())
+	cancel()
+	require.Error(t, graphiteOnce(ctx, testGraphiteConfig(h.addr(t), metrics.NewRegistry()), time.Minute))
+	select {
+	case <-h.conns:
+		t.Fatal("connected on a cancelled ctx")
+	case <-time.After(200 * time.Millisecond):
+	}
+}
+
+// A stop or reload abandons a stuck send rather than waiting out its timeout
+func TestGraphiteOnce_cancelAbandonsSend(t *testing.T) {
+	r := metrics.NewRegistry()
+	bigGraphiteRegistry(r, t)
+	h := newGraphiteHost(t, 4096)
+	ctx, cancel := context.WithCancel(t.Context())
+	done := make(chan error, 1)
+	cfg := testGraphiteConfig(h.addr(t), r)
+	go func() { done <- graphiteOnce(ctx, cfg, time.Minute) }()
+	h.next(t, 5*time.Second)
+	select {
+	case err := <-done:
+		t.Fatalf("the send should be stuck on a host that is not reading, it ended with %v", err)
+	case <-time.After(200 * time.Millisecond):
+	}
+
+	cancel()
+	select {
+	case err := <-done:
+		require.Error(t, err)
+	case <-time.After(5 * time.Second):
+		t.Fatal("the send is still stuck after its ctx was cancelled")
+	}
+}
+
+// buildRuntime hands graphite the configured prefix and interval
+func TestStatsServer_graphiteExportUsesConfig(t *testing.T) {
+	const name = "graphite.wiring.test"
+	metrics.GetOrRegisterCounter(name, nil).Inc(8)
+	t.Cleanup(func() { metrics.DefaultRegistry.Unregister(name) })
+	h := newGraphiteHost(t, 0)
+	s, c := newTestStatsServer(t)
+	setStatsConfig(c, map[string]any{"type": "graphite", "interval": "4s", "prefix": "pfx", "host": h.ln.Addr().String()})
+	cfg, err := loadStatsConfig(c)
+	require.NoError(t, err)
+
+	s.runMu.Lock()
+	fns, _ := s.buildRuntime(t.Context(), cfg)
+	s.runMu.Unlock()
+	for _, fn := range fns {
+		fn()
+	}
+	b := readExport(t, h.next(t, 5*time.Second))
+	assert.Contains(t, string(b), "pfx."+name+".count 8 ")
+	assert.Contains(t, string(b), "pfx."+name+".count_ps 2.00 ")
+}
+
+func TestStatsServer_graphiteFailureIsLogged(t *testing.T) {
+	logs := &lockedBuffer{}
+	s, c := newTestStatsServer(t)
+	s.l = test.NewLoggerWithOutput(logs)
+	setStatsConfig(c, map[string]any{"type": "graphite", "interval": "1h", "host": "127.0.0.1:" + freeTCPPort(t)})
+	cfg, err := loadStatsConfig(c)
+	require.NoError(t, err)
+
+	s.runMu.Lock()
+	fns, _ := s.buildRuntime(t.Context(), cfg)
+	s.runMu.Unlock()
+	for _, fn := range fns {
+		fn()
+	}
+	assert.Contains(t, logs.String(), "Graphite export failed")
+}
+
+// Stop abandons the running export, so it goes through Start rather than handing buildRuntime a ctx, and a stop is
+// not logged as a failure
+func TestStatsServer_stopAbandonsGraphiteSend(t *testing.T) {
+	bigGraphiteRegistry(metrics.DefaultRegistry, t)
+	h := newGraphiteHost(t, 4096)
+	logs := &lockedBuffer{}
+	s, c := newTestStatsServer(t)
+	s.l = test.NewLoggerWithOutput(logs)
+	setStatsConfig(c, map[string]any{"type": "graphite", "interval": "50ms", "host": h.ln.Addr().String()})
+	require.NoError(t, s.reload(c, true))
+
+	started := make(chan struct{})
+	go func() {
+		s.Start()
+		close(started)
+	}()
+	stuck := h.next(t, 30*time.Second)
+	s.Stop()
+	select {
+	case <-started:
+	case <-time.After(5 * time.Second):
+		t.Fatal("Start did not return after Stop")
+	}
+
+	// A closed sender resets the connection once it sees this, one still sending just buffers it. Either way the
+	// read below doesn't have to drain the export through the small window to find out
+	_, err := stuck.Write([]byte("x"))
+	require.NoError(t, err)
+	require.NoError(t, stuck.SetReadDeadline(time.Now().Add(10*time.Second)))
+	n, err := io.Copy(io.Discard, stuck)
+	var ne net.Error
+	if errors.As(err, &ne) && ne.Timeout() {
+		t.Fatalf("the send kept the connection after Stop, read %d bytes", n)
+	}
+	assert.Less(t, n, int64(1<<20), "the send kept going after Stop")
+	// The capture loop logs on its own goroutine after Start has returned, give it the chance
+	assert.Never(t, func() bool { return strings.Contains(logs.String(), "Graphite export failed") }, 500*time.Millisecond, 10*time.Millisecond, "a stop is not a failure")
+}
diff --git a/stats.go b/stats.go
index a82bbc2b..c6895eef 100644
--- a/stats.go
+++ b/stats.go
@@ -153,8 +153,8 @@ func (s *statsServer) Start() {
 		return
 	}
 	cfg := *s.runCfg
-	captureFns, listener := s.buildRuntime(cfg)
 	runCtx, cancel := context.WithCancel(s.ctx)
+	captureFns, listener := s.buildRuntime(runCtx, cfg)
 	rt := &statsRuntime{cancel: cancel, listener: listener}
 	s.run = rt
 	s.runMu.Unlock()
@@ -237,7 +237,8 @@ func (s *statsServer) Stop() {
 
 // buildRuntime produces the capture functions and, for prometheus, an un-served
 // http.Server from cfg. cfg has already been validated by loadStatsConfig.
-func (s *statsServer) buildRuntime(cfg statsConfig) ([]func(), *http.Server) {
+// A graphite send in progress is abandoned when ctx is done.
+func (s *statsServer) buildRuntime(ctx context.Context, cfg statsConfig) ([]func(), *http.Server) {
 	// rcrowley/go-metrics guards these registrations with a private sync.Once,
 	// so subsequent reloads are no-ops.
 	metrics.RegisterDebugGCStats(metrics.DefaultRegistry)
@@ -262,7 +263,7 @@ func (s *statsServer) buildRuntime(cfg statsConfig) ([]func(), *http.Server) {
 			Percentiles:   []float64{0.5, 0.75, 0.95, 0.99, 0.999},
 		}
 		captureFns = append(captureFns, func() {
-			if err := graphiteOnce(gcfg); err != nil {
+			if err := graphiteOnce(ctx, gcfg, graphiteTimeout); err != nil && ctx.Err() == nil {
 				s.l.Error("Graphite export failed", "error", err)
 			}
 		})
diff --git a/stats_test.go b/stats_test.go
index 9d1aae08..58bfe417 100644
--- a/stats_test.go
+++ b/stats_test.go
@@ -2,7 +2,6 @@ package nebula
 
 import (
 	"context"
-	"io"
 	"log/slog"
 	"net"
 	"strconv"
@@ -110,15 +109,14 @@ func TestStatsServer_reload_initial_prometheus(t *testing.T) {
 }
 
 func TestStatsServer_Start_graphite_blocksUntilStop(t *testing.T) {
-	sink := newGraphiteSink(t)
-	defer sink.Close()
+	h := newGraphiteHost(t, 0)
 
 	s, c := newTestStatsServer(t)
 	setStatsConfig(c, map[string]any{
 		"type":     "graphite",
 		"interval": "1s",
 		"protocol": "tcp",
-		"host":     sink.Addr(),
+		"host":     h.ln.Addr().String(),
 		"prefix":   "test",
 	})
 	require.NoError(t, s.reload(c, true))
@@ -371,35 +369,6 @@ func waitForListening(t *testing.T, addr string) {
 	})
 }
 
-// graphiteSink is a minimal TCP accept-and-discard server so graphiteOnce
-// calls in tests don't spam error logs or wedge on connection refused.
-type graphiteSink struct {
-	ln net.Listener
-}
-
-func newGraphiteSink(t *testing.T) *graphiteSink {
-	t.Helper()
-	ln, err := net.Listen("tcp", "127.0.0.1:0")
-	require.NoError(t, err)
-	g := &graphiteSink{ln: ln}
-	go func() {
-		for {
-			conn, err := ln.Accept()
-			if err != nil {
-				return
-			}
-			go func(c net.Conn) {
-				_, _ = io.Copy(io.Discard, c)
-				_ = c.Close()
-			}(conn)
-		}
-	}()
-	return g
-}
-
-func (g *graphiteSink) Addr() string { return g.ln.Addr().String() }
-func (g *graphiteSink) Close()       { _ = g.ln.Close() }
-
 func freeTCPPort(t *testing.T) string {
 	t.Helper()
 	ln, err := net.Listen("tcp", "127.0.0.1:0")