use deepdiff to get differences between gpu and cpu results

pyproject.toml

@@ -67,6 +67,7 @@ dev = [
     "types-mock",
     "types-PyYAML",
     "types-aiofiles",
+    "deepdiff",
 ]
 
 [project.scripts]

src/dodal/devices/zocalo/zocalo_results.py

@@ -10,6 +10,7 @@
 import workflows.recipe
 import workflows.transport
 from bluesky.protocols import Descriptor, Triggerable
+from deepdiff import DeepDiff
 from numpy.typing import NDArray
 from ophyd_async.core import HintedSignal, StandardReadable, soft_signal_r_and_setter
 from ophyd_async.core.async_status import AsyncStatus
@@ -60,11 +61,21 @@ def get_dict_differences(
     dict1: dict, dict1_source: str, dict2: dict, dict2_source: str
 ) -> str:
     differences_str = ""
-    for key in dict1.keys():
-        if not dict1[key] == dict2[key]:
-            differences_str += f"Results differed in {key}: {dict1_source} contains {dict1[key]} while {dict2_source} contains {dict2[key]} \n"
-    if differences_str:
-        LOGGER.warning(differences_str)
+    diff = DeepDiff(dict1, dict2, math_epsilon=1e-5, ignore_numeric_type_changes=True)
+    # diff is None if there are no differences, otherwise a dictionary with various keys depending
+    # on the type of difference detected. Eg 'values_changed', dictionary_item_added', 'iterable_item_added'
+
+    if diff:
+        differences_str += (
+            f"Differences found between {dict1_source} and {dict2_source} results:\n"
+        )
+
+    # Format the difference more nicely in a string
+    for key, value in diff.items():
+        if key == "values_changed":
+            differences_str += f"   {key}: {str(value).replace('old_value', dict1_source).replace('new_value', dict2_source)}\n"
+        else:
+            differences_str += f"  {key}: {value}\n"
     return differences_str
 
 

tests/devices/unit_tests/test_zocalo_results.py

@@ -357,10 +357,48 @@ async def test_if_cpu_results_arrive_before_gpu_then_warn(
     )
 
 
+@pytest.mark.parametrize(
+    "dict1,dict2,output",
+    [
+        (
+            {"ispyb_ids": {"gpu": True}, "results": [{"test": 0}]},
+            {"ispyb_ids": {}, "results": [{"test": 1}]},
+            "Differences found between GPU and CPU results:\n   values_changed: {\"root['test']\": {'CPU': 1, 'GPU': 0}}\n",
+        ),
+        (
+            {
+                "ispyb_ids": {"gpu": True},
+                "results": [{"test": [[1, 2 + 1e-6, 3], [1, 2, 3]]}],
+            },
+            {"ispyb_ids": {}, "results": [{"test": [[1, 2, 3], [1, 2, 3]]}]},
+            None,
+        ),
+        (
+            {
+                "ispyb_ids": {"gpu": True},
+                "results": [
+                    {"test": [[1, 2 + 1e-6, 3], [1, 2, 3]], "extra_key": "test"}
+                ],
+            },
+            {"ispyb_ids": {}, "results": [{"test": [[1, 2, 3], [1, 2, 3]]}]},
+            "Differences found between GPU and CPU results:\n  dictionary_item_removed: SetOrdered([\"root['extra_key']\"])\n",
+        ),
+        (
+            {
+                "ispyb_ids": {"gpu": False},
+                "results": [
+                    {"test": [[1, 2 + 1e-6, 3], [1, 2, 3]], "extra_key": "test"}
+                ],
+            },
+            {"ispyb_ids": {}, "results": [{"test": [[1, 3, 3], [1, 2, 3]]}]},
+            "Differences found between CPU and GPU results:\n  dictionary_item_removed: SetOrdered([\"root['extra_key']\"])\n   values_changed: {\"root['test'][0][1]\": {'GPU': 3, 'CPU': 2.000001}}\n",
+        ),
+    ],
+)
 @patch("dodal.devices.zocalo.zocalo_results.LOGGER")
 @patch("dodal.devices.zocalo.zocalo_results._get_zocalo_connection", autospec=True)
 async def test_warning_if_results_are_different(
-    mock_connection, mock_logger, RE: RunEngine
+    mock_connection, mock_logger, RE: RunEngine, dict1, dict2, output
 ):
     zocalo_results = ZocaloResults(
         name="zocalo", zocalo_environment="dev_artemis", use_cpu_and_gpu_zocalo=True
@@ -369,14 +407,14 @@ async def test_warning_if_results_are_different(
     await zocalo_results.stage()
     zocalo_results._raw_results_received.get = MagicMock(
         side_effect=[
-            {"ispyb_ids": {}, "results": [{"test": 0}]},
-            {"ispyb_ids": {}, "results": [{"test": 1}]},
+            dict1,
+            dict2,
         ]
     )
     RE(bps.trigger(zocalo_results, wait=False))
     mock_logger.warning.assert_called_with(
-        "Results differed in test: CPU contains 0 while GPU contains 1 \n"
-    )
+        output
+    ) if output else mock_logger.warning.assert_not_called()
 
 
 @patch("dodal.devices.zocalo.zocalo_results._get_zocalo_connection", autospec=True)