test: add a global dimension to test failure metric

The current test failure metrics have high cardinality because they
include the test name. This makes it hard to visualize them in
CloudWatch since we run into aggregation limits.

It would be nice to have one global failure rate, and some views
per-kernel and per-CPU type.

This adds 10 new metrics = 6 per-CPU + 3 per-host kernel + 1 global
metric so it is not too costly.

Signed-off-by: Pablo Barbáchano <pablob@amazon.com>

tests/conftest.py

@@ -117,17 +117,24 @@ def record_props(request, record_property):
 def pytest_runtest_logreport(report):
     """Send general test metrics to CloudWatch"""
     if report.when == "call":
-        dimensions = {
-            "test": report.nodeid,
-            "instance": global_props.instance,
-            "cpu_model": global_props.cpu_model,
-            "host_kernel": "linux-" + global_props.host_linux_version,
-        }
+        METRICS.set_dimensions(
+            {
+                "test": report.nodeid,
+                "instance": global_props.instance,
+                "cpu_model": global_props.cpu_model,
+                "host_kernel": "linux-" + global_props.host_linux_version,
+            },
+            # per host kernel
+            {"host_kernel": "linux-" + global_props.host_linux_version},
+            # per CPU
+            {"cpu_model": global_props.cpu_model},
+            # and global
+            {},
+        )
         METRICS.set_property("result", report.outcome)
         METRICS.set_property("location", report.location)
         for prop_name, prop_val in report.user_properties:
             METRICS.set_property(prop_name, prop_val)
-        METRICS.set_dimensions(dimensions)
         METRICS.put_metric(
             "duration",
             report.duration,