feat: add cron task that runs the minimal training pipeline nightly

.cron.yml

@@ -0,0 +1,13 @@
+jobs:
+  # Run the minimal pipeline once a day to support integration testing before
+  # worker image changes are made (see https://bugzilla.mozilla.org/show_bug.cgi?id=1937882).
+  # These runs will pick up cached tasks, so most of the time this will simply
+  # end up running `all-pipeline`.
+  - name: run-pipeline
+    job:
+      type: decision-task
+      # we don't use treeherder...but this is a required field
+      treeherder-symbol: pipeline
+      target-tasks-method: train-target-tasks
+    when:
+      - {hour: 0, minute: 0}

taskcluster/translations_taskgraph/actions/train.py

@@ -1,6 +1,7 @@
 # This Source Code Form is subject to the terms of the Mozilla Public
 # License, v. 2.0. If a copy of the MPL was not distributed with this
 # file, You can obtain one at http://mozilla.org/MPL/2.0/.
+import collections.abc
 import json
 import logging
 
@@ -58,6 +59,15 @@ def validate_pretrained_models(params):
             )
 
 
+def deep_update(d, u):
+    for k, v in u.items():
+        if isinstance(v, collections.abc.Mapping):
+            d[k] = deep_update(d.get(k, {}), v)
+        else:
+            d[k] = v
+    return d
+
+
 @register_callback_action(
     name="train",
     title="Train",
@@ -397,6 +407,11 @@ def train_action(parameters, graph_config, input, task_group_id, task_id):
     # etc.
 
     parameters = dict(parameters)
+    # Although we provide defaults to everything in the `schema` when
+    # registering the action in the decorator above, these are not passed
+    # along when the action runs. To make sure we have the proper defaults
+    # we use the defaults with the input provided overlaid on top of them.
+    input = deep_update(defaults, input)
 
     start_stage = input.pop("start-stage", None)
     if start_stage:

taskcluster/translations_taskgraph/parameters.py

@@ -2,11 +2,13 @@
 # License, v. 2.0. If a copy of the MPL was not distributed with this
 # file, You can obtain one at http://mozilla.org/MPL/2.0/.
 
+import logging
 from pathlib import Path
 from taskgraph.parameters import extend_parameters_schema
 from voluptuous import Extra, Optional, Required
 import yaml
 
+logger = logging.getLogger(__name__)
 
 # By default, provide a very minimal config for CI that runs very quickly. This allows
 # the pipeline to be validated in CI. The production training configs should override
@@ -103,3 +105,10 @@ def deep_setdefault(dict_, defaults):
 def get_decision_parameters(graph_config, parameters):
     parameters.setdefault("training_config", {})
     deep_setdefault(parameters, get_ci_training_config())
+    # We run the pipeline on a cron schedule to enable integration testing when
+    # worker images change (see https://bugzilla.mozilla.org/show_bug.cgi?id=1937882).
+    # These runs should _never_ be sent to W&B to avoid cluttering it up
+    # with data of no value.
+    if parameters["tasks_for"] == "cron" and parameters["target_tasks_method"] == "train-target-tasks":
+        logger.info("Overriding wandb-publication to be False for cron pipeline run")
+        parameters["training_config"]["wandb-publication"] = False