Avoid duplicated runs during W&B publication (#484)

Co-authored-by: Evgeny Pavlov <[email protected]>

tracking/translations_parser/cli/taskcluster_group.py

@@ -13,6 +13,7 @@
 from collections import defaultdict
 from pathlib import Path
 
+import wandb
 import yaml
 
 import taskcluster
@@ -265,6 +266,16 @@ def publish_task_group(group_id: str) -> None:
         )
 
     # Group and publish remaining metrics tasks via the logs publication
+    if (
+        len(
+            wandb.Api().runs(
+                project_name, filters={"display_name": "group_logs", "group": group_name}
+            )
+        )
+        > 0
+    ):
+        logger.warning("Skipping group_logs fake run publication as it already exists")
+        return
     with tempfile.TemporaryDirectory() as temp_dir:
         logs_folder = Path(temp_dir) / "logs"
         eval_folder = logs_folder / project_name / group_name / "eval"

tracking/translations_parser/publishers.py

@@ -101,6 +101,21 @@ def open(self, parser) -> None:
         self.parser = parser
         config = parser.config
         config.update(self.extra_kwargs.pop("config", {}))
+
+        # Ensure no W&B run already exists
+        if name := self.extra_kwargs.get("name"):
+            existing_runs = list(
+                wandb.Api().runs(
+                    self.project,
+                    filters={"display_name": name, "group": self.extra_kwargs.get("group")},
+                )
+            )
+            if len(existing_runs) > 0:
+                logger.warning(
+                    f"This run already exists on W&B: {existing_runs}. No data will be published."
+                )
+                return
+
         # Start a W&B run
         try:
             self.wandb = wandb.init(