PrakshiGoyal10 commented on code in PR #69998:
URL: https://github.com/apache/airflow/pull/69998#discussion_r3804655605
##########
providers/databricks/src/airflow/providers/databricks/plugins/databricks_workflow.py:
##########
@@ -515,6 +547,266 @@ def get_link( # type: ignore[override] # Signature
intentionally kept this way
return url_for("RepairDatabricksTasks.repair", **query_params)
+# Airflow-3 repair backend. Flask-AppBuilder was dropped in Airflow 3, so the
repair
+# action is re-implemented as a FastAPI sub-application mounted on the API
server, and the
+# repair links (below) build URLs that point at it.
+REPAIR_URL_PREFIX = "/databricks/workflow/repair"
+
+
+def _build_repair_url(
+ dag_id: str,
+ run_id: str,
+ launch_task_id: str,
+ *,
+ repair_all: bool = False,
+ task_id: str | None = None,
+) -> str:
+ """
+ Build the URL to the Airflow-3 FastAPI repair confirmation page for a
workflow run.
+
+ The URL carries only Airflow identifiers: the run's launch ``task_id``
(from which the
+ endpoint reads the trusted ``WorkflowRunMetadata`` XCom) and, for a
single-task repair, the
+ target ``task_id``. The Databricks connection, run id, and task keys are
never placed in the
+ link — the endpoint derives them server-side, so the request cannot point
the repair at an
+ arbitrary connection or Databricks run.
+ """
+ from urllib.parse import urlencode
+
+ query: dict[str, Any] = {"launch_task_id": launch_task_id}
+ if repair_all:
+ query["repair_all"] = "true"
+ if task_id:
+ query["task_id"] = task_id
+
+ base_url = conf.get("api", "base_url", fallback="").rstrip("/")
+ return (
+ f"{base_url}{REPAIR_URL_PREFIX}/{quote(dag_id,
safe='')}/{quote(run_id, safe='')}?{urlencode(query)}"
+ )
+
+
+def _get_launch_task_id_v3(operator: BaseOperator, ti_key: TaskInstanceKey) ->
str | None:
+ """
+ Resolve the ``task_id`` of the workflow's launch task for an extra-link
render.
+
+ The link only needs to name the launch task; the repair endpoint reads
that task's trusted
+ ``WorkflowRunMetadata`` XCom server-side. Returns ``None`` when the
operator is not part of a
+ Databricks workflow task group (so the link is not rendered).
+ """
+ task_group = operator.task_group
+ if not task_group:
+ return None
+ if ".launch" in ti_key.task_id:
+ return ti_key.task_id
+ return get_launch_task_id(task_group)
+
+
+if AIRFLOW_V_3_1_PLUS:
+ from fastapi import Depends, FastAPI, HTTPException, Request
+ from fastapi.responses import HTMLResponse, RedirectResponse
+ from markupsafe import escape
+
+ from airflow.api_fastapi.auth.managers.base_auth_manager import
COOKIE_NAME_JWT_TOKEN
+ from airflow.api_fastapi.auth.managers.models.resource_details import
DagAccessEntity, DagDetails
+ from airflow.api_fastapi.core_api.security import resolve_user_from_token
+
+ repair_app = FastAPI(
+ title="Databricks Workflow Repair",
+ description="Repair failed tasks of a Databricks workflow run from
Airflow.",
+ )
+
+ async def _resolve_request_user(request: Request):
+ """Authenticate via the bearer header (UI XHR) or the ``_token``
cookie (link navigation)."""
+ token = None
+ auth_header = request.headers.get("Authorization", "")
+ if auth_header.lower().startswith("bearer "):
+ token = auth_header.split(" ", 1)[1]
+ if not token:
+ token = request.cookies.get(COOKIE_NAME_JWT_TOKEN)
+ # resolve_user_from_token raises HTTP 401 for a missing/invalid token.
+ return await resolve_user_from_token(token)
+
+ async def _require_dag_run_edit(dag_id: str, request: Request):
+ from airflow.api_fastapi.app import get_auth_manager
+
+ user = await _resolve_request_user(request)
+ authorized = get_auth_manager().is_authorized_dag(
+ method="PUT",
+ access_entity=DagAccessEntity.RUN,
+ details=DagDetails(id=dag_id),
+ user=user,
+ )
+ if not authorized:
+ raise HTTPException(status_code=403, detail="Not authorized to
repair runs of this Dag.")
+ return user
+
+ def _task_id_to_key(dag_id: str, task_id: str, task_key_map: dict[str,
str]) -> str:
+ """
+ Resolve a task's Databricks ``task_key`` from the launch task's
trusted key map.
+
+ An explicit ``databricks_task_key`` does not survive Dag
serialization, so the serialized
+ task can't be trusted to reproduce it. The launch task captured the
real keys from the live
+ operators into ``task_key_map``. Runs launched before that map existed
fall back to the
+ operator's default derivation, ``md5(dag_id__task_id)`` — correct for
any task that did not
+ set an explicit key (the common case).
+ """
+ mapped = task_key_map.get(task_id)
+ if mapped:
+ return mapped
+ import hashlib
+
+ return hashlib.md5(f"{dag_id}__{task_id}".encode()).hexdigest()
+
+ def _read_launch_metadata(dag_id: str, run_id: str, launch_task_id: str,
session) -> Any:
+ """
+ Read the launch task's trusted ``WorkflowRunMetadata`` XCom (conn_id,
job_id, run_id).
+
+ The Databricks connection and run id come from here — never from the
request — so a crafted
+ link cannot redirect the repair at an arbitrary connection or
Databricks run.
+ """
+ from airflow.models.xcom import XComModel
+ from airflow.providers.databricks.operators.databricks_workflow import
WorkflowRunMetadata
+
+ result = session.scalars(
+ XComModel.get_many(
+ run_id=run_id,
+ key="return_value",
+ task_ids=launch_task_id,
+ dag_ids=dag_id,
+ limit=1,
+ )
+ ).first()
+ if result is None:
+ raise HTTPException(status_code=404, detail="Databricks workflow
run metadata not found.")
+ return WorkflowRunMetadata(**XComModel.deserialize_value(result))
+
+ def _clear_repaired_and_downstream(
+ dag, run_id: str, task_ids: list[str], session, logger: logging.Logger
+ ) -> None:
+ """
+ Clear the repaired tasks' instances and their downstream instances for
this run.
+
+ Runs inside the API server (the DB-facing component), so clearing the
repaired tasks plus
+ their downstream lets the upstream-failed dependents resume
deterministically when the
+ repaired Databricks sub-runs succeed — without clearing the whole Dag.
+ """
+ from sqlalchemy import select
+
+ from airflow.models.taskinstance import clear_task_instances
+
+ target_task_ids: set[str] = set(task_ids)
+ for task_id in task_ids:
+
target_task_ids.update(dag.get_task(task_id).get_flat_relative_ids(upstream=False))
+
+ dr = session.scalars(select(DagRun).where(DagRun.dag_id == dag.dag_id,
DagRun.run_id == run_id)).one()
+ tis_to_clear = [ti for ti in dr.get_task_instances(session=session) if
ti.task_id in target_task_ids]
+ logger.info("Clearing %s task instances after Databricks repair",
len(tis_to_clear))
+ clear_task_instances(tis_to_clear, session)
+
+ def _repair_confirmation_page(dag_id: str, run_id: str, action: str,
summary: str) -> HTMLResponse:
+ """Render the read-only confirmation page whose form issues the
state-changing POST."""
+ return HTMLResponse(
+ "<!doctype html><html><head><title>Repair Databricks
workflow</title></head><body>"
+ "<h2>Repair Databricks workflow tasks</h2>"
+ f"<p>Dag <b>{escape(dag_id)}</b>, run <b>{escape(run_id)}</b>.</p>"
+ f"<p>{escape(summary)}</p>"
+ f'<form method="post" action="{escape(action)}">'
+ '<button type="submit">Repair</button></form>'
+ "</body></html>"
+ )
+
+ @repair_app.get("/{dag_id}/{run_id}")
+ def repair_databricks_workflow_confirm(
+ dag_id: str,
+ run_id: str,
+ request: Request,
+ launch_task_id: str,
+ task_id: str | None = None,
+ repair_all: bool = False,
+ _user=Depends(_require_dag_run_edit),
+ ):
+ """Render a read-only confirmation page; the repair itself happens on
the POST below."""
+ run_id = unquote(run_id)
+ summary = (
+ "This will repair all failed tasks of the run and resume their
downstream tasks."
+ if repair_all
+ else f"This will repair task '{task_id}' and resume its downstream
tasks."
+ )
+ # Same-site relative action; SameSite=Lax on the auth cookie means a
cross-site POST cannot
+ # carry it, so moving the mutation to POST is what protects it from
CSRF.
+ action = f"{request.url.path}?{request.url.query}"
+ return _repair_confirmation_page(dag_id, run_id, action, summary)
Review Comment:
Thanks — addressed the scanner flag on the repair redirect.
The redirect and the confirmation-page form action were assembled from the
request's own path and query. The values were percent-encoded into a fixed
same-site `/dags/{dag_id}/runs/{run_id}` path and could not actually leave the
origin, but CodeQL traced request input into a redirect target and kept the
open-redirect alert open regardless.
Pushed a change so no request-controlled string reaches either target:
- The redirect is now built from the `DagRun`'s persisted `dag_id` /
`run_id`, read from the database, rather than the request path params.
- The confirmation form action is rebuilt from the already-validated
identifiers instead of echoing the request URL.
- Fetching the run also lets the endpoint return a `404` for an unknown run
and reuse that row when clearing task instances.
That breaks the taint flow the scanner was following, so the finding should
clear on the next analysis. All provider unit tests pass locally.
---
Drafted-by: Claude Code (Opus 4.8); reviewed by @PrakshiGoyal10
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
