Add throttling counter in gcsio and refactor retrying (#32428)

* Add retry instance that records throttling metric.

* Use retry with throttling counters by default. Add pipeline option.

* Fix lint

* Fix broken tests.

* Retrieve a more accurate throttling time from the caller frame.

* Apply yapf and linter

* Refactoring copy and delete

- Remove extra retries for copy, delete, _gcs_object.
- Remove the use of client.batch() as the function has no built-in
  retry.

* Fix a typo and apply yapf

* Use counter instead of counters in pipeline option.

Additionally, the variable name for the new retry object is changed.

Add a new pipeline option to enable the use of blob generation to
mitigate race conditions (at the expense of more http requests)

* Parameterize existing tests for the new pipeline options.

* Apply yapf

* Fix a typo.

* Revert the change of copy_batch and delete_batch and add warning in their docstring.

* Fix lint

* Minor change according to code review.

* Restore the previous tox.ini that got accidentally changed.

sdks/python/apache_beam/io/gcp/gcsio.py

@@ -43,10 +43,10 @@
 
 from apache_beam import version as beam_version
 from apache_beam.internal.gcp import auth
+from apache_beam.io.gcp import gcsio_retry
 from apache_beam.metrics.metric import Metrics
 from apache_beam.options.pipeline_options import GoogleCloudOptions
 from apache_beam.options.pipeline_options import PipelineOptions
-from apache_beam.utils import retry
 from apache_beam.utils.annotations import deprecated
 
 __all__ = ['GcsIO', 'create_storage_client']
@@ -155,6 +155,9 @@ def __init__(self, storage_client=None, pipeline_options=None):
     self.client = storage_client
     self._rewrite_cb = None
     self.bucket_to_project_number = {}
+    self._storage_client_retry = gcsio_retry.get_retry(pipeline_options)
+    self._use_blob_generation = getattr(
+        google_cloud_options, 'enable_gcsio_blob_generation', False)
 
   def get_project_number(self, bucket):
     if bucket not in self.bucket_to_project_number:
@@ -167,7 +170,8 @@ def get_project_number(self, bucket):
   def get_bucket(self, bucket_name, **kwargs):
     """Returns an object bucket from its name, or None if it does not exist."""
     try:
-      return self.client.lookup_bucket(bucket_name, **kwargs)
+      return self.client.lookup_bucket(
+          bucket_name, retry=self._storage_client_retry, **kwargs)
     except NotFound:
       return None
 
@@ -188,7 +192,7 @@ def create_bucket(
           bucket_or_name=bucket,
           project=project,
           location=location,
-      )
+          retry=self._storage_client_retry)
       if kms_key:
         bucket.default_kms_key_name(kms_key)
         bucket.patch()
@@ -224,33 +228,43 @@ def open(
       return BeamBlobReader(
           blob,
           chunk_size=read_buffer_size,
-          enable_read_bucket_metric=self.enable_read_bucket_metric)
+          enable_read_bucket_metric=self.enable_read_bucket_metric,
+          retry=self._storage_client_retry)
     elif mode == 'w' or mode == 'wb':
       blob = bucket.blob(blob_name)
       return BeamBlobWriter(
           blob,
           mime_type,
-          enable_write_bucket_metric=self.enable_write_bucket_metric)
+          enable_write_bucket_metric=self.enable_write_bucket_metric,
+          retry=self._storage_client_retry)
     else:
       raise ValueError('Invalid file open mode: %s.' % mode)
 
-  @retry.with_exponential_backoff(
-      retry_filter=retry.retry_on_server_errors_and_timeout_filter)
   def delete(self, path):
     """Deletes the object at the given GCS path.
 
     Args:
       path: GCS file path pattern in the form gs://<bucket>/<name>.
     """
     bucket_name, blob_name = parse_gcs_path(path)
+    bucket = self.client.bucket(bucket_name)
+    if self._use_blob_generation:
+      # blob can be None if not found
+      blob = bucket.get_blob(blob_name, retry=self._storage_client_retry)
+      generation = getattr(blob, "generation", None)
+    else:
+      generation = None
     try:
-      bucket = self.client.bucket(bucket_name)
-      bucket.delete_blob(blob_name)
+      bucket.delete_blob(
+          blob_name,
+          if_generation_match=generation,
+          retry=self._storage_client_retry)
     except NotFound:
       return
 
   def delete_batch(self, paths):
     """Deletes the objects at the given GCS paths.
+    Warning: any exception during batch delete will NOT be retried.
 
     Args:
       paths: List of GCS file path patterns or Dict with GCS file path patterns
@@ -287,8 +301,6 @@ def delete_batch(self, paths):
 
     return final_results
 
-  @retry.with_exponential_backoff(
-      retry_filter=retry.retry_on_server_errors_and_timeout_filter)
   def copy(self, src, dest):
     """Copies the given GCS object from src to dest.
 
@@ -297,19 +309,32 @@ def copy(self, src, dest):
       dest: GCS file path pattern in the form gs://<bucket>/<name>.
 
     Raises:
-      TimeoutError: on timeout.
+      Any exceptions during copying
     """
     src_bucket_name, src_blob_name = parse_gcs_path(src)
     dest_bucket_name, dest_blob_name= parse_gcs_path(dest, object_optional=True)
     src_bucket = self.client.bucket(src_bucket_name)
-    src_blob = src_bucket.blob(src_blob_name)
+    if self._use_blob_generation:
+      src_blob = src_bucket.get_blob(src_blob_name)
+      if src_blob is None:
+        raise NotFound("source blob %s not found during copying" % src)
+      src_generation = src_blob.generation
+    else:
+      src_blob = src_bucket.blob(src_blob_name)
+      src_generation = None
     dest_bucket = self.client.bucket(dest_bucket_name)
     if not dest_blob_name:
       dest_blob_name = None
-    src_bucket.copy_blob(src_blob, dest_bucket, new_name=dest_blob_name)
+    src_bucket.copy_blob(
+        src_blob,
+        dest_bucket,
+        new_name=dest_blob_name,
+        source_generation=src_generation,
+        retry=self._storage_client_retry)
 
   def copy_batch(self, src_dest_pairs):
     """Copies the given GCS objects from src to dest.
+    Warning: any exception during batch copy will NOT be retried.
 
     Args:
       src_dest_pairs: list of (src, dest) tuples of gs://<bucket>/<name> files
@@ -450,8 +475,6 @@ def _status(self, path):
       file_status['size'] = gcs_object.size
     return file_status
 
-  @retry.with_exponential_backoff(
-      retry_filter=retry.retry_on_server_errors_and_timeout_filter)
   def _gcs_object(self, path):
     """Returns a gcs object for the given path
 
@@ -462,7 +485,7 @@ def _gcs_object(self, path):
     """
     bucket_name, blob_name = parse_gcs_path(path)
     bucket = self.client.bucket(bucket_name)
-    blob = bucket.get_blob(blob_name)
+    blob = bucket.get_blob(blob_name, retry=self._storage_client_retry)
     if blob:
       return blob
     else:
@@ -510,7 +533,8 @@ def list_files(self, path, with_metadata=False):
     else:
       _LOGGER.debug("Starting the size estimation of the input")
     bucket = self.client.bucket(bucket_name)
-    response = self.client.list_blobs(bucket, prefix=prefix)
+    response = self.client.list_blobs(
+        bucket, prefix=prefix, retry=self._storage_client_retry)
     for item in response:
       file_name = 'gs://%s/%s' % (item.bucket.name, item.name)
       if file_name not in file_info:
@@ -546,8 +570,7 @@ def _updated_to_seconds(updated):
   def is_soft_delete_enabled(self, gcs_path):
     try:
       bucket_name, _ = parse_gcs_path(gcs_path)
-      # set retry timeout to 5 seconds when checking soft delete policy
-      bucket = self.get_bucket(bucket_name, retry=DEFAULT_RETRY.with_timeout(5))
+      bucket = self.get_bucket(bucket_name)
       if (bucket.soft_delete_policy is not None and
           bucket.soft_delete_policy.retention_duration_seconds > 0):
         return True
@@ -563,8 +586,9 @@ def __init__(
       self,
       blob,
       chunk_size=DEFAULT_READ_BUFFER_SIZE,
-      enable_read_bucket_metric=False):
-    super().__init__(blob, chunk_size=chunk_size)
+      enable_read_bucket_metric=False,
+      retry=DEFAULT_RETRY):
+    super().__init__(blob, chunk_size=chunk_size, retry=retry)
     self.enable_read_bucket_metric = enable_read_bucket_metric
     self.mode = "r"
 
@@ -585,13 +609,14 @@ def __init__(
       content_type,
       chunk_size=16 * 1024 * 1024,
       ignore_flush=True,
-      enable_write_bucket_metric=False):
+      enable_write_bucket_metric=False,
+      retry=DEFAULT_RETRY):
     super().__init__(
         blob,
         content_type=content_type,
         chunk_size=chunk_size,
         ignore_flush=ignore_flush,
-        retry=DEFAULT_RETRY)
+        retry=retry)
     self.mode = "w"
     self.enable_write_bucket_metric = enable_write_bucket_metric
 

sdks/python/apache_beam/io/gcp/gcsio_integration_test.py

@@ -34,6 +34,7 @@
 
 import mock
 import pytest
+from parameterized import parameterized_class
 
 from apache_beam.io.filesystems import FileSystems
 from apache_beam.options.pipeline_options import GoogleCloudOptions
@@ -51,6 +52,9 @@
 
 
 @unittest.skipIf(gcsio is None, 'GCP dependencies are not installed')
+@parameterized_class(
+    ('no_gcsio_throttling_counter', 'enable_gcsio_blob_generation'),
+    [(False, False), (False, True), (True, False), (True, True)])
 class GcsIOIntegrationTest(unittest.TestCase):
 
   INPUT_FILE = 'gs://dataflow-samples/shakespeare/kinglear.txt'
@@ -67,7 +71,6 @@ def setUp(self):
     self.gcs_tempdir = (
         self.test_pipeline.get_option('temp_location') + '/gcs_it-' +
         str(uuid.uuid4()))
-    self.gcsio = gcsio.GcsIO()
 
   def tearDown(self):
     FileSystems.delete([self.gcs_tempdir + '/'])
@@ -92,14 +95,47 @@ def _verify_copy(self, src, dest, dest_kms_key_name=None):
 
   @pytest.mark.it_postcommit
   def test_copy(self):
+    self.gcsio = gcsio.GcsIO(
+        pipeline_options={
+            "no_gcsio_throttling_counter": self.no_gcsio_throttling_counter,
+            "enable_gcsio_blob_generation": self.enable_gcsio_blob_generation
+        })
+    src = self.INPUT_FILE
+    dest = self.gcs_tempdir + '/test_copy'
+
+    self.gcsio.copy(src, dest)
+    self._verify_copy(src, dest)
+
+    unknown_src = self.test_pipeline.get_option('temp_location') + \
+        '/gcs_it-' + str(uuid.uuid4())
+    with self.assertRaises(NotFound):
+      self.gcsio.copy(unknown_src, dest)
+
+  @pytest.mark.it_postcommit
+  def test_copy_and_delete(self):
+    self.gcsio = gcsio.GcsIO(
+        pipeline_options={
+            "no_gcsio_throttling_counter": self.no_gcsio_throttling_counter,
+            "enable_gcsio_blob_generation": self.enable_gcsio_blob_generation
+        })
     src = self.INPUT_FILE
     dest = self.gcs_tempdir + '/test_copy'
 
     self.gcsio.copy(src, dest)
     self._verify_copy(src, dest)
 
+    self.gcsio.delete(dest)
+
+    # no exception if we delete an nonexistent file.
+    self.gcsio.delete(dest)
+
   @pytest.mark.it_postcommit
   def test_batch_copy_and_delete(self):
+    self.gcsio = gcsio.GcsIO(
+        pipeline_options={
+            "no_gcsio_throttling_counter": self.no_gcsio_throttling_counter,
+            "enable_gcsio_blob_generation": self.enable_gcsio_blob_generation
+        })
     num_copies = 10
     srcs = [self.INPUT_FILE] * num_copies
     dests = [
@@ -152,6 +188,7 @@ def test_batch_copy_and_delete(self):
   @mock.patch('apache_beam.io.gcp.gcsio.default_gcs_bucket_name')
   @unittest.skipIf(NotFound is None, 'GCP dependencies are not installed')
   def test_create_default_bucket(self, mock_default_gcs_bucket_name):
+    self.gcsio = gcsio.GcsIO()
     google_cloud_options = self.test_pipeline.options.view_as(
         GoogleCloudOptions)
     # overwrite kms option here, because get_or_create_default_gcs_bucket()

sdks/python/apache_beam/io/gcp/gcsio_retry.py

@@ -0,0 +1,71 @@
+#
+# Licensed to the Apache Software Foundation (ASF) under one or more
+# contributor license agreements.  See the NOTICE file distributed with
+# this work for additional information regarding copyright ownership.
+# The ASF licenses this file to You under the Apache License, Version 2.0
+# (the "License"); you may not use this file except in compliance with
+# the License.  You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+
+"""
+Throttling Handler for GCSIO
+"""
+
+import inspect
+import logging
+import math
+
+from google.api_core import exceptions as api_exceptions
+from google.api_core import retry
+from google.cloud.storage.retry import DEFAULT_RETRY
+from google.cloud.storage.retry import _should_retry  # pylint: disable=protected-access
+
+from apache_beam.metrics.metric import Metrics
+from apache_beam.options.pipeline_options import GoogleCloudOptions
+
+_LOGGER = logging.getLogger(__name__)
+
+__all__ = ['DEFAULT_RETRY_WITH_THROTTLING_COUNTER']
+
+
+class ThrottlingHandler(object):
+  _THROTTLED_SECS = Metrics.counter('gcsio', "cumulativeThrottlingSeconds")
+
+  def __call__(self, exc):
+    if isinstance(exc, api_exceptions.TooManyRequests):
+      _LOGGER.debug('Caught GCS quota error (%s), retrying.', exc.reason)
+      # TODO: revisit the logic here when gcs client library supports error
+      # callbacks
+      frame = inspect.currentframe()
+      if frame is None:
+        _LOGGER.warning('cannot inspect the current stack frame')
+        return
+
+      prev_frame = frame.f_back
+      if prev_frame is None:
+        _LOGGER.warning('cannot inspect the caller stack frame')
+        return
+
+      # next_sleep is one of the arguments in the caller
+      # i.e. _retry_error_helper() in google/api_core/retry/retry_base.py
+      sleep_seconds = prev_frame.f_locals.get("next_sleep", 0)
+      ThrottlingHandler._THROTTLED_SECS.inc(math.ceil(sleep_seconds))
+
+
+DEFAULT_RETRY_WITH_THROTTLING_COUNTER = retry.Retry(
+    predicate=_should_retry, on_error=ThrottlingHandler())
+
+
+def get_retry(pipeline_options):
+  if pipeline_options.view_as(GoogleCloudOptions).no_gcsio_throttling_counter:
+    return DEFAULT_RETRY
+  else:
+    return DEFAULT_RETRY_WITH_THROTTLING_COUNTER