MRG: Merge pull request #541 from octue/fix/improve-crash-diagnostics

Fix and improve crash diagnostics

docs/source/testing_services.rst

@@ -172,7 +172,7 @@ To emulate your children in tests, patch the :mod:`Child <octue.resources.child.
         app_src=app_directory_path,
         twine=os.path.join(app_directory_path, "twine.json"),
         children=children,
-        service_id="you/your-service:latest",
+        service_id="your-org/your-service:latest",
     )
 
     emulated_children = [
@@ -212,7 +212,8 @@ change correspondingly (or at all). So, it's up to you to define a set of realis
   the child's twine - this is because the twine is only available to the real child. This is ok - you're testing your
   service, not the child.
 
-You can create test fixtures manually or by recording messages from a real child to a JSON file. To record messages:
+You can create test fixtures manually or by using the ``Child.received_messages`` property after questioning a real
+child.
 
 .. code-block:: python
 
@@ -225,20 +226,13 @@ You can create test fixtures manually or by recording messages from a real child
         backend={"name": "GCPPubSubBackend", "project_name": "my-project"},
     )
 
-    result = child.ask(
-        input_values=[1, 2, 3, 4],
-        record_messages_to="child_messages.json",
-    )
-
-    with open("child_messages.json") as f:
-        child_messages = json.load(f)
+    result = child.ask(input_values=[1, 2, 3, 4])
 
-    child_messages
+    child.received_messages
     >>> [
             {
                 'type': 'delivery_acknowledgement',
                 'delivery_time': '2022-08-16 11:49:57.244263',
-                'message_number': 0
             },
             {
                 'type': 'log_record',
@@ -248,14 +242,11 @@ You can create test fixtures manually or by recording messages from a real child
                     'levelname': 'INFO',
                     ...
                 },
-                'analysis_id': '0ce8386d-564d-47fa-9d11-3b728f557bfe',
-                'message_number': 1
             },
             {
                 'type': 'result',
                 'output_values': {"some": "results"},
                 'output_manifest': None,
-                'message_number': 2
             }
         ]
 
@@ -266,7 +257,9 @@ You can then feed these into a child emulator to emulate one possible response o
     from octue.cloud.emulators import ChildEmulator
 
 
-    child_emulator = ChildEmulator(messages=child_messages)
+    child_emulator = ChildEmulator(messages=child.received_messages)
 
     child_emulator.ask(input_values=[1, 2, 3, 4])
     >>> {"some": "results"}
+
+You can also create test fixtures from :ref:`downloaded service crash diagnostics <test_fixtures_from_crash_diagnostics>`.

docs/source/troubleshooting_services.rst

@@ -2,40 +2,27 @@
 Troubleshooting services
 ========================
 
-Allowing crash diagnostics
-==========================
-A parent can give a child permission to save the following data to the cloud in the event the child fails while
-processing a question:
+Crash diagnostics
+=================
+Services save the following data to the cloud if they crash while processing a question:
 
 - Input values
 - Input manifest and datasets
 - Child configuration values
 - Child configuration manifest and datasets
-- Messages sent from the child to the parent
+- Inputs to and messages received in answer to each question the service asked its children (if it has any). These are
+  stored in the order the questions were asked.
 
-The parent can give permission on a question-by-question basis by setting ``allow_save_diagnostics_data_on_crash=True``
-in :mod:`Child.ask <octue.resources.child.Child.ask>`. For example:
+.. important::
 
-.. code-block:: python
-
-    child = Child(
-        id="my-organisation/my-service:latest",
-        backend={"name": "GCPPubSubBackend", "project_name": "my-project"},
-    )
-
-    answer = child.ask(
-        input_values={"height": 32, "width": 3},
-        allow_save_diagnostics_data_on_crash=True,
-    )
+    For this feature to be enabled, the child must have the ``crash_diagnostics_cloud_path`` field in its service
+    configuration (:ref:`octue.yaml <octue_yaml>` file) set to a Google Cloud Storage path.
 
-For crash diagnostics to be saved, the child must have the ``crash_diagnostics_cloud_path`` field in its service
-configuration (:ref:`octue.yaml <octue_yaml>` file) set to a Google Cloud Storage path.
 
 Accessing crash diagnostics
 ===========================
-In the event of a child crash, the child will upload the crash diagnostics and send the cloud path to them to the
-parent as a log message. A user with credentials to access this path can use the ``octue`` CLI to retrieve the crash
-diagnostics data:
+In the event of a crash, the service will upload the crash diagnostics and send the upload path to the parent as a log
+message. A user with credentials to access this path can use the ``octue`` CLI to retrieve the crash diagnostics data:
 
 .. code-block:: shell
 
@@ -59,4 +46,82 @@ More information on the command:
       --local-path DIRECTORY  The path to a directory to store the directory of
                               diagnostics data in. Defaults to the current working
                               directory.
+      --download-datasets     If provided, download any datasets from the crash
+                              diagnostics and update their paths in their
+                              manifests to the new local paths.
       -h, --help              Show this message and exit.
+
+.. _test_fixtures_from_crash_diagnostics:
+
+Creating test fixtures from crash diagnostics
+=============================================
+You can create test fixtures directly from crash diagnostics, allowing you to recreate the exact conditions that caused
+your service to fail.
+
+.. code-block:: python
+
+    from unittest.mock import patch
+
+    from octue import Runner
+    from octue.utils.testing import load_test_fixture_from_crash_diagnostics
+
+
+     (
+         configuration_values,
+         configuration_manifest,
+         input_values,
+         input_manifest,
+         child_emulators,
+     ) = load_test_fixture_from_crash_diagnostics(path="path/to/downloaded/crash/diagnostics")
+
+    # You can explicitly specify your children here as shown or
+    # read the same information in from your app configuration file.
+    children = [
+        {
+            "key": "my_child",
+            "id": "octue/my-child-service:latest",
+            "backend": {
+                "name": "GCPPubSubBackend",
+                "project_name": "my-project",
+            }
+        },
+        {
+            "key": "another_child",
+            "id": "octue/another-child-service:latest",
+            "backend": {
+                "name": "GCPPubSubBackend",
+                "project_name": "my-project",
+            }
+        }
+    ]
+
+    runner = Runner(
+        app_src="path/to/directory_containing_app",
+        twine=os.path.join(app_directory_path, "twine.json"),
+        children=children,
+        configuration_values=configuration_values,
+        configuration_manifest=configuration_manifest,
+        service_id="your-org/your-service:latest",
+    )
+
+    with patch("octue.runner.Child", side_effect=child_emulators):
+        analysis = runner.run(input_values=input_values, input_manifest=input_manifest)
+
+
+Disabling crash diagnostics
+===========================
+When asking a question to a child, parents can disable crash diagnostics upload in the child on a question-by-question
+basis by setting ``allow_save_diagnostics_data_on_crash`` to ``False`` in :mod:`Child.ask <octue.resources.child.Child.ask>`.
+For example:
+
+.. code-block:: python
+
+    child = Child(
+        id="my-organisation/my-service:latest",
+        backend={"name": "GCPPubSubBackend", "project_name": "my-project"},
+    )
+
+    answer = child.ask(
+        input_values={"height": 32, "width": 3},
+        allow_save_diagnostics_data_on_crash=False,
+    )

octue/app_loading.py

@@ -0,0 +1,63 @@
+import importlib
+import logging
+import os
+import sys
+
+
+logger = logging.getLogger(__name__)
+
+
+class AppFrom:
+    """A context manager that imports the module "app" from a file named "app.py" in the given directory on entry (by
+    making a temporary addition to the system path) and unloads it (by deleting it from `sys.modules`) on exit. It will
+    issue a warning if an existing module called "app" is already loaded. Usage example:
+
+    ```python3
+    with AppFrom('/path/to/dir') as app:
+        Runner().run(app)
+    ```
+
+    :param str app_path: path to directory containing module named "app.py".
+    :return None:
+    """
+
+    def __init__(self, app_path="."):
+        self.app_path = os.path.abspath(os.path.normpath(app_path))
+        logger.debug("Initialising AppFrom context at app_path %s", self.app_path)
+        self.app_module = None
+
+    def __enter__(self):
+        # Warn on an app present on the system path
+        if "app" in sys.modules.keys():
+            logger.warning(
+                "Module 'app' already on system path. Using 'AppFrom' context will yield unexpected results. Avoid "
+                "using 'app' as a python module, except for your main entrypoint."
+            )
+
+        # Insert the present directory first on the system path.
+        sys.path.insert(0, self.app_path)
+
+        # Import the app from the present directory.
+        self.app_module = importlib.import_module("app")
+
+        # Immediately clean up the entry to the system path (don't use "remove" because if the user has it in their
+        # path, this'll be an unexpected side effect, and don't do it in cleanup in case the called code inserts a path)
+        sys.path.pop(0)
+        logger.debug("Imported app at app_path and cleaned up temporary modification to sys.path %s", self.app_path)
+        return self
+
+    def __exit__(self, exc_type, exc_value, traceback):
+        """Unload the imported module.
+
+        :return None:
+        """
+        try:
+            del sys.modules["app"]
+            logger.debug("Deleted app from sys.modules")
+
+        except KeyError:
+            context_manager_name = type(self).__name__
+            logger.warning(
+                f"Module 'app' was already removed from the system path prior to exiting the {context_manager_name} "
+                f"context manager. Using the {context_manager_name} context may yield unexpected results."
+            )

octue/cli.py

@@ -20,7 +20,7 @@
 from octue.definitions import MANIFEST_FILENAME, VALUES_FILENAME
 from octue.exceptions import ServiceAlreadyExists
 from octue.log_handlers import apply_log_handler, get_remote_handler
-from octue.resources import service_backends
+from octue.resources import Manifest, service_backends
 from octue.runner import Runner
 from octue.utils.encoders import OctueJSONEncoder
 from twined import Twine
@@ -289,25 +289,61 @@ def start(service_config, revision_tag, timeout, no_rm):
 @click.option(
     "--local-path",
     type=click.Path(file_okay=False),
-    default=None,
+    default=".",
     help="The path to a directory to store the directory of diagnostics data in. Defaults to the current working "
     "directory.",
 )
-def get_crash_diagnostics(cloud_path, local_path):
+@click.option(
+    "--download-datasets",
+    is_flag=True,
+    help="If provided, download any datasets from the crash diagnostics and update their paths in the configuration and "
+    "input manifests to the new local paths.",
+)
+def get_crash_diagnostics(cloud_path, local_path, download_datasets):
     """Download crash diagnostics for an analysis from the given directory in Google Cloud Storage. The cloud path
     should end in the analysis ID.
 
     CLOUD_PATH: The path to the directory in Google Cloud Storage containing the diagnostics data.
     """
     analysis_id = storage.path.split(cloud_path)[-1]
-    local_path = os.path.join((local_path or "."), analysis_id)
+    local_path = os.path.join(local_path, analysis_id)
+
+    if download_datasets:
+        filter = None
+    else:
+        filter = lambda blob: any(
+            (
+                blob.name.endswith(f"configuration_{VALUES_FILENAME}"),
+                blob.name.endswith(f"configuration_{MANIFEST_FILENAME}"),
+                blob.name.endswith(f"input_{VALUES_FILENAME}"),
+                blob.name.endswith(f"input_{MANIFEST_FILENAME}"),
+                blob.name.endswith("questions.json"),
+            )
+        )
 
     GoogleCloudStorageClient().download_all_files(
         local_path=local_path,
         cloud_path=cloud_path,
+        filter=filter,
         recursive=True,
     )
 
+    # Update the manifests with the local paths of the datasets.
+    if download_datasets:
+        for manifest_type in ("configuration_manifest", "input_manifest"):
+            manifest_path = os.path.join(local_path, manifest_type + ".json")
+
+            if not os.path.exists(manifest_path):
+                continue
+
+            manifest = Manifest.from_file(manifest_path)
+
+            manifest.update_dataset_paths(
+                path_generator=lambda dataset: os.path.join(local_path, f"{manifest_type}_datasets", dataset.name)
+            )
+
+            manifest.to_file(manifest_path)
+
     logger.info("Downloaded crash diagnostics from %r to %r.", cloud_path, local_path)