Stop re-implementing the log upload in backfill_events.py

**Impact:** `tools/scripts/backfill_events.py`, a manually run script
**Risk:** low

## What
`upload_log` now POSTs to torchci's `/api/log-uploader/backfill` route instead of
downloading the log, gzipping it, putting it in S3, and pinging the classifier
itself. Needs `LOG_UPLOADER_BOT_KEY`, and honours `HUD_URL` for testing against a
preview deployment.

Also drops the S3 raw-event archive write from `process_event`, leaving only the
DynamoDB write.

ghstack-source-id: 009c4f5
Pull-Request: #8596

Author: huydhn

tools/scripts/backfill_events.py

@@ -1,77 +1,70 @@
 #!/usr/bin/env python3
 
-import gzip
 import json
 import os
 from typing import Any
-from urllib.request import urlopen
+from urllib.error import HTTPError
+from urllib.request import Request, urlopen
 from warnings import warn
 
 import boto3
 from octokit import Octokit
 
 
-S3 = boto3.resource("s3")
-BUCKET_NAME = "ossci-raw-job-status"
-BUCKET = S3.Bucket(BUCKET_NAME)
-
 DYNAMO = boto3.resource("dynamodb")
 
-
-def json_dumps(body: Any) -> str:
-    # This logic is copied from github-status-test lambda function
-    return json.dumps(body, sort_keys=True, indent=4, separators=(",", ": "))
+# torchci's authenticated backfill route, which hands the job to the
+# gha-log-uploader lambda. That lambda owns log downloading now; re-implementing
+# it here is how this script and github-status-test used to drift apart.
+HUD_URL = os.environ.get("HUD_URL", "https://hud.pytorch.org")
+LOG_UPLOADER_BOT_KEY = os.environ.get("LOG_UPLOADER_BOT_KEY", "")
 
 
-def upload_log(
-    client: Octokit, owner: str, repo: str, job_id: int, conclusion: str
-) -> None:
-    # This logic is copied from github-status-test lambda function
-    log = client.actions.download_job_logs_for_workflow_run(
-        owner=owner, repo=repo, job_id=job_id
-    ).json
+def upload_log(owner: str, repo: str, job_id: int, conclusion: str) -> None:
+    if not LOG_UPLOADER_BOT_KEY:
+        warn("LOG_UPLOADER_BOT_KEY is not set, skipping the log upload...")
+        return
 
-    log_path = f"log/{job_id}"
-    if repo != "pytorch":
-        log_path = f"log/{owner}/{repo}/{job_id}"
+    print(f"..Requesting a log upload for {owner}/{repo} job {job_id}")
+    request = Request(
+        f"{HUD_URL}/api/log-uploader/backfill",
+        method="POST",
+        data=json.dumps(
+            {
+                "repo": f"{owner}/{repo}",
+                "job_id": job_id,
+                "conclusion": conclusion,
+            }
+        ).encode(),
+        headers={
+            "Content-Type": "application/json",
+            "Authorization": LOG_UPLOADER_BOT_KEY,
+        },
+    )
 
-    print(f"..Uploading log to {log_path}")
     try:
-        # This needs to be in try catch because GitHub doesn't keep log older than 60 days I think
-        S3.Object(BUCKET_NAME, log_path).put(
-            Body=gzip.compress(log.encode(encoding="UTF-8")),
-            ContentType="text/plain",
-            ContentEncoding="gzip",
-            Metadata={"conclusion": conclusion},
-        )
-
-        # Invoke log classifier
-        urlopen(
-            f"https://vwg52br27lx5oymv4ouejwf4re0akoeg.lambda-url.us-east-1.on.aws/?job_id={job_id}&repo={owner}/{repo}"
-        )
-    except Exception as error:
+        # GitHub drops logs after around 60 days, so an old job simply has
+        # nothing to fetch. That is a warning, not a reason to stop the backfill.
+        with urlopen(request, timeout=30) as response:
+            if response.status != 200:
+                warn(f"Log upload for job {job_id} returned {response.status}")
+    except (HTTPError, OSError) as error:
         warn(
-            f"Failed to upload {log} for job {job_id} from repo {owner}/{repo}: "
-            + f"{error}, skipping..."
+            f"Failed to request a log upload for job {job_id} from repo "
+            f"{owner}/{repo}: {error}, skipping..."
         )
 
 
 def process_event(owner: str, repo: str, event: str, body: Any) -> None:
-    # This logic is copied from github-status-test lambda function
-    if repo == "pytorch":
-        repo_prefix = ""
-    else:
-        repo_prefix = f"{owner}/{repo}/"
-
+    # Only DynamoDB, which is what clickhouse-replicator-dynamo reads. The raw
+    # event archive github-status-test used to write to S3 was never read by
+    # anything and is not reproduced here.
     if "id" not in body:
         warn(f"Missing ID in {body}, skipping...")
         return
 
     id = body["id"]
-    print(f"{event}/{repo_prefix}{id}")
-    S3.Object(BUCKET_NAME, f"{event}/{repo_prefix}{id}").put(
-        Body=json_dumps(body), ContentType="application/json"
-    )
+    print(f"{event} {owner}/{repo}/{id}")
 
     dynamodb_table = ""
     if event == "workflow_run":
@@ -131,7 +124,7 @@ def process_workflow_run(
             conclusion = workflow_job["conclusion"]
 
             process_event(owner, repo, "workflow_job", workflow_job)
-            upload_log(client, owner, repo, job_id, conclusion)
+            upload_log(owner, repo, job_id, conclusion)
 
         if not count or count >= total_count:
             # Finish processing all events