diff --git a/synapse/metrics/metric.py b/synapse/metrics/metric.py
index e8c15a60e..4df5ebfda 100644
--- a/synapse/metrics/metric.py
+++ b/synapse/metrics/metric.py
@@ -79,9 +79,13 @@ class CallbackMetric(BaseMetric):
         self.callback = callback
 
     def render(self):
-        # TODO(paul): work out something we can do with keys and vectors
-        return ["%s %d" % (self.name, self.callback())]
+        value = self.callback()
 
+        if self.is_scalar():
+            return ["%s %d" % (self.name, value)]
+
+        return ["%s{%s} %d" % (self.name, self._render_key(k), value[k])
+                for k in sorted(value.keys())]
 
 class CacheMetric(object):
     """A combination of two CounterMetrics, one to count cache hits and one to
diff --git a/tests/metrics/test_metric.py b/tests/metrics/test_metric.py
index 32fd178ed..b7facb858 100644
--- a/tests/metrics/test_metric.py
+++ b/tests/metrics/test_metric.py
@@ -65,7 +65,7 @@ class CounterMetricTestCase(unittest.TestCase):
 
 class CallbackMetricTestCase(unittest.TestCase):
 
-    def test_callback(self):
+    def test_scalar(self):
         d = dict()
 
         metric = CallbackMetric("size", lambda: len(d))
@@ -80,6 +80,22 @@ class CallbackMetricTestCase(unittest.TestCase):
             "size 1",
         ])
 
+    def test_vector(self):
+        vals = dict()
+
+        metric = CallbackMetric("values", lambda: vals, keys=["type"])
+
+        self.assertEquals(metric.render(), [])
+
+        # Keys have to be tuples, even if they're 1-element
+        vals[("foo",)] = 1
+        vals[("bar",)] = 2
+
+        self.assertEquals(metric.render(), [
+            "values{type=bar} 2",
+            "values{type=foo} 1",
+        ])
+
 
 class CacheMetricTestCase(unittest.TestCase):