Bugfix/connector deletion lockout (#2901)

* first cut at deletion hardening

* clean up logging

* remove commented code

Author: rkuo-danswer

backend/danswer/background/celery/apps/primary.py

@@ -17,6 +17,7 @@
 from danswer.background.celery.celery_redis import RedisConnectorDeletion
 from danswer.background.celery.celery_redis import RedisConnectorIndexing
 from danswer.background.celery.celery_redis import RedisConnectorPruning
+from danswer.background.celery.celery_redis import RedisConnectorStop
 from danswer.background.celery.celery_redis import RedisDocumentSet
 from danswer.background.celery.celery_redis import RedisUserGroup
 from danswer.background.celery.celery_utils import celery_is_worker_primary
@@ -161,6 +162,9 @@ def on_worker_init(sender: Any, **kwargs: Any) -> None:
         for key in r.scan_iter(RedisConnectorIndexing.FENCE_PREFIX + "*"):
             r.delete(key)
 
+    for key in r.scan_iter(RedisConnectorStop.FENCE_PREFIX + "*"):
+        r.delete(key)
+
 
 # @worker_process_init.connect
 # def on_worker_process_init(sender: Any, **kwargs: Any) -> None:

backend/danswer/background/celery/celery_redis.py

@@ -313,6 +313,8 @@ def generate_tasks(
         lock: redis.lock.Lock,
         tenant_id: str | None,
     ) -> int | None:
+        """Returns None if the cc_pair doesn't exist.
+        Otherwise, returns an int with the number of generated tasks."""
         last_lock_time = time.monotonic()
 
         async_results = []
@@ -540,6 +542,29 @@ def is_indexing(self, redis_client: Redis) -> bool:
         return False
 
 
+class RedisConnectorStop(RedisObjectHelper):
+    """Used to signal any running tasks for a connector to stop. We should refactor
+    connector related redis helpers into a single class.
+    """
+
+    PREFIX = "connectorstop"
+    FENCE_PREFIX = PREFIX + "_fence"  # a fence for the entire indexing process
+    TASKSET_PREFIX = PREFIX + "_taskset"  # stores a list of prune tasks id's
+
+    def __init__(self, id: int) -> None:
+        super().__init__(str(id))
+
+    def generate_tasks(
+        self,
+        celery_app: Celery,
+        db_session: Session,
+        redis_client: Redis,
+        lock: redis.lock.Lock | None,
+        tenant_id: str | None,
+    ) -> int | None:
+        return None
+
+
 def celery_get_queue_length(queue: str, r: Redis) -> int:
     """This is a redis specific way to get the length of a celery queue.
     It is priority aware and knows how to count across the multiple redis lists

backend/danswer/background/celery/celery_utils.py

@@ -1,11 +1,11 @@
-from collections.abc import Callable
 from datetime import datetime
 from datetime import timezone
 from typing import Any
 
 from sqlalchemy.orm import Session
 
 from danswer.background.celery.celery_redis import RedisConnectorDeletion
+from danswer.background.indexing.run_indexing import RunIndexingCallbackInterface
 from danswer.configs.app_configs import MAX_PRUNING_DOCUMENT_RETRIEVAL_PER_MINUTE
 from danswer.connectors.cross_connector_utils.rate_limit_wrapper import (
     rate_limit_builder,
@@ -79,7 +79,7 @@ def document_batch_to_ids(
 
 def extract_ids_from_runnable_connector(
     runnable_connector: BaseConnector,
-    progress_callback: Callable[[int], None] | None = None,
+    callback: RunIndexingCallbackInterface | None = None,
 ) -> set[str]:
     """
     If the PruneConnector hasnt been implemented for the given connector, just pull
@@ -110,8 +110,10 @@ def extract_ids_from_runnable_connector(
             max_calls=MAX_PRUNING_DOCUMENT_RETRIEVAL_PER_MINUTE, period=60
         )(document_batch_to_ids)
     for doc_batch in doc_batch_generator:
-        if progress_callback:
-            progress_callback(len(doc_batch))
+        if callback:
+            if callback.should_stop():
+                raise RuntimeError("Stop signal received")
+            callback.progress(len(doc_batch))
         all_connector_doc_ids.update(doc_batch_processing_func(doc_batch))
 
     return all_connector_doc_ids

backend/danswer/background/celery/tasks/connector_deletion/tasks.py

@@ -1,3 +1,6 @@
+from datetime import datetime
+from datetime import timezone
+
 import redis
 from celery import Celery
 from celery import shared_task
@@ -8,16 +11,28 @@
 
 from danswer.background.celery.apps.app_base import task_logger
 from danswer.background.celery.celery_redis import RedisConnectorDeletion
+from danswer.background.celery.celery_redis import RedisConnectorIndexing
+from danswer.background.celery.celery_redis import RedisConnectorPruning
+from danswer.background.celery.celery_redis import RedisConnectorStop
+from danswer.background.celery.tasks.shared.RedisConnectorDeletionFenceData import (
+    RedisConnectorDeletionFenceData,
+)
 from danswer.configs.app_configs import JOB_TIMEOUT
 from danswer.configs.constants import CELERY_VESPA_SYNC_BEAT_LOCK_TIMEOUT
 from danswer.configs.constants import DanswerRedisLocks
 from danswer.db.connector_credential_pair import get_connector_credential_pair_from_id
 from danswer.db.connector_credential_pair import get_connector_credential_pairs
 from danswer.db.engine import get_session_with_tenant
 from danswer.db.enums import ConnectorCredentialPairStatus
+from danswer.db.search_settings import get_all_search_settings
 from danswer.redis.redis_pool import get_redis_client
 
 
+class TaskDependencyError(RuntimeError):
+    """Raised to the caller to indicate dependent tasks are running that would interfere
+    with connector deletion."""
+
+
 @shared_task(
     name="check_for_connector_deletion_task",
     soft_time_limit=JOB_TIMEOUT,
@@ -37,17 +52,30 @@ def check_for_connector_deletion_task(self: Task, *, tenant_id: str | None) -> N
         if not lock_beat.acquire(blocking=False):
             return
 
+        # collect cc_pair_ids
         cc_pair_ids: list[int] = []
         with get_session_with_tenant(tenant_id) as db_session:
             cc_pairs = get_connector_credential_pairs(db_session)
             for cc_pair in cc_pairs:
                 cc_pair_ids.append(cc_pair.id)
 
+        # try running cleanup on the cc_pair_ids
         for cc_pair_id in cc_pair_ids:
             with get_session_with_tenant(tenant_id) as db_session:
-                try_generate_document_cc_pair_cleanup_tasks(
-                    self.app, cc_pair_id, db_session, r, lock_beat, tenant_id
-                )
+                rcs = RedisConnectorStop(cc_pair_id)
+                try:
+                    try_generate_document_cc_pair_cleanup_tasks(
+                        self.app, cc_pair_id, db_session, r, lock_beat, tenant_id
+                    )
+                except TaskDependencyError as e:
+                    # this means we wanted to start deleting but dependent tasks were running
+                    # Leave a stop signal to clear indexing and pruning tasks more quickly
+                    task_logger.info(str(e))
+                    r.set(rcs.fence_key, cc_pair_id)
+                else:
+                    # clear the stop signal if it exists ... no longer needed
+                    r.delete(rcs.fence_key)
+
     except SoftTimeLimitExceeded:
         task_logger.info(
             "Soft time limit exceeded, task is being terminated gracefully."
@@ -70,6 +98,10 @@ def try_generate_document_cc_pair_cleanup_tasks(
     """Returns an int if syncing is needed. The int represents the number of sync tasks generated.
     Note that syncing can still be required even if the number of sync tasks generated is zero.
     Returns None if no syncing is required.
+
+    Will raise TaskDependencyError if dependent tasks such as indexing and pruning are
+    still running. In our case, the caller reacts by setting a stop signal in Redis to
+    exit those tasks as quickly as possible.
     """
 
     lock_beat.reacquire()
@@ -90,28 +122,63 @@ def try_generate_document_cc_pair_cleanup_tasks(
     if cc_pair.status != ConnectorCredentialPairStatus.DELETING:
         return None
 
-    # add tasks to celery and build up the task set to monitor in redis
-    r.delete(rcd.taskset_key)
-
-    # Add all documents that need to be updated into the queue
-    task_logger.info(
-        f"RedisConnectorDeletion.generate_tasks starting. cc_pair_id={cc_pair.id}"
+    # set a basic fence to start
+    fence_value = RedisConnectorDeletionFenceData(
+        num_tasks=None,
+        submitted=datetime.now(timezone.utc),
     )
-    tasks_generated = rcd.generate_tasks(app, db_session, r, lock_beat, tenant_id)
-    if tasks_generated is None:
+    r.set(rcd.fence_key, fence_value.model_dump_json())
+
+    try:
+        # do not proceed if connector indexing or connector pruning are running
+        search_settings_list = get_all_search_settings(db_session)
+        for search_settings in search_settings_list:
+            rci = RedisConnectorIndexing(cc_pair_id, search_settings.id)
+            if r.get(rci.fence_key):
+                raise TaskDependencyError(
+                    f"Connector deletion - Delayed (indexing in progress): "
+                    f"cc_pair={cc_pair_id} "
+                    f"search_settings={search_settings.id}"
+                )
+
+        rcp = RedisConnectorPruning(cc_pair_id)
+        if r.get(rcp.fence_key):
+            raise TaskDependencyError(
+                f"Connector deletion - Delayed (pruning in progress): "
+                f"cc_pair={cc_pair_id}"
+            )
+
+        # add tasks to celery and build up the task set to monitor in redis
+        r.delete(rcd.taskset_key)
+
+        # Add all documents that need to be updated into the queue
+        task_logger.info(
+            f"RedisConnectorDeletion.generate_tasks starting. cc_pair={cc_pair_id}"
+        )
+        tasks_generated = rcd.generate_tasks(app, db_session, r, lock_beat, tenant_id)
+        if tasks_generated is None:
+            raise ValueError("RedisConnectorDeletion.generate_tasks returned None")
+    except TaskDependencyError:
+        r.delete(rcd.fence_key)
+        raise
+    except Exception:
+        task_logger.exception("Unexpected exception")
+        r.delete(rcd.fence_key)
         return None
+    else:
+        # Currently we are allowing the sync to proceed with 0 tasks.
+        # It's possible for sets/groups to be generated initially with no entries
+        # and they still need to be marked as up to date.
+        # if tasks_generated == 0:
+        #     return 0
 
-    # Currently we are allowing the sync to proceed with 0 tasks.
-    # It's possible for sets/groups to be generated initially with no entries
-    # and they still need to be marked as up to date.
-    # if tasks_generated == 0:
-    #     return 0
+        task_logger.info(
+            f"RedisConnectorDeletion.generate_tasks finished. "
+            f"cc_pair={cc_pair_id} tasks_generated={tasks_generated}"
+        )
 
-    task_logger.info(
-        f"RedisConnectorDeletion.generate_tasks finished. "
-        f"cc_pair_id={cc_pair.id} tasks_generated={tasks_generated}"
-    )
+        # set this only after all tasks have been added
+        fence_value.num_tasks = tasks_generated
+        r.set(rcd.fence_key, fence_value.model_dump_json())
 
-    # set this only after all tasks have been added
-    r.set(rcd.fence_key, tasks_generated)
     return tasks_generated