"""Standard retry behavior. This contains the default standard retry behavior. It provides consistent behavior with other AWS SDKs. The key base classes uses for retries: * ``BaseRetryableChecker`` - Use to check a specific condition that indicates a retry should happen. This can include things like max attempts, HTTP status code checks, error code checks etc. * ``RetryBackoff`` - Use to determine how long we should backoff until we retry a request. This is the class that will implement delay such as exponential backoff. * ``RetryPolicy`` - Main class that determines if a retry should happen. It can combine data from a various BaseRetryableCheckers to make a final call as to whether or not a retry should happen. It then uses a ``BaseRetryBackoff`` to determine how long to delay. * ``RetryHandler`` - The bridge between botocore's event system used by endpoint.py to manage retries and the interfaces defined in this module. This allows us to define an API that has minimal coupling to the event based API used by botocore. """ import random import logging from botocore.exceptions import ConnectionError, HTTPClientError from botocore.exceptions import ReadTimeoutError, ConnectTimeoutError from botocore.retries import quota from botocore.retries import special from botocore.retries.base import BaseRetryBackoff, BaseRetryableChecker DEFAULT_MAX_ATTEMPTS = 3 logger = logging.getLogger(__name__) def register_retry_handler(client, max_attempts=DEFAULT_MAX_ATTEMPTS): retry_quota = RetryQuotaChecker(quota.RetryQuota()) service_id = client.meta.service_model.service_id service_event_name = service_id.hyphenize() client.meta.events.register('after-call.%s' % service_event_name, retry_quota.release_retry_quota) handler = RetryHandler( retry_policy=RetryPolicy( retry_checker=StandardRetryConditions(max_attempts=max_attempts), retry_backoff=ExponentialBackoff(), ), retry_event_adapter=RetryEventAdapter(), retry_quota=retry_quota, ) unique_id = 'retry-config-%s' % service_event_name client.meta.events.register( 'needs-retry.%s' % service_event_name, handler.needs_retry, unique_id=unique_id ) return handler class RetryHandler(object): """Bridge between botocore's event system and this module. This class is intended to be hooked to botocore's event system as an event handler. """ def __init__(self, retry_policy, retry_event_adapter, retry_quota): self._retry_policy = retry_policy self._retry_event_adapter = retry_event_adapter self._retry_quota = retry_quota def needs_retry(self, **kwargs): """Connect as a handler to the needs-retry event.""" retry_delay = None context = self._retry_event_adapter.create_retry_context(**kwargs) if self._retry_policy.should_retry(context): # Before we can retry we need to ensure we have sufficient # capacity in our retry quota. if self._retry_quota.acquire_retry_quota(context): retry_delay = self._retry_policy.compute_retry_delay(context) logger.debug("Retry needed, retrying request after " "delay of: %s", retry_delay) else: logger.debug("Retry needed but retry quota reached, " "not retrying request.") else: logger.debug("Not retrying request.") self._retry_event_adapter.adapt_retry_response_from_context( context) return retry_delay class RetryEventAdapter(object): """Adapter to existing retry interface used in the endpoints layer. This existing interface for determining if a retry needs to happen is event based and used in ``botocore.endpoint``. The interface has grown organically over the years and could use some cleanup. This adapter converts that interface into the interface used by the new retry strategies. """ def create_retry_context(self, **kwargs): """Create context based on needs-retry kwargs.""" response = kwargs['response'] if response is None: # If response is None it means that an exception was raised # because we never received a response from the service. This # could be something like a ConnectionError we get from our # http layer. http_response = None parsed_response = None else: http_response, parsed_response = response # This provides isolation between the kwargs emitted in the # needs-retry event, and what this module uses to check for # retries. context = RetryContext( attempt_number=kwargs['attempts'], operation_model=kwargs['operation'], http_response=http_response, parsed_response=parsed_response, caught_exception=kwargs['caught_exception'], request_context=kwargs['request_dict']['context'], ) return context def adapt_retry_response_from_context(self, context): """Modify response back to user back from context.""" # This will mutate attributes that are returned back to the end # user. We do it this way so that all the various retry classes # don't mutate any input parameters from the needs-retry event. metadata = context.get_retry_metadata() if context.parsed_response is not None: context.parsed_response.setdefault( 'ResponseMetadata', {}).update(metadata) # Implementation note: this is meant to encapsulate all the misc stuff # that gets sent in the needs-retry event. This is mapped so that params # are more clear and explicit. class RetryContext(object): """Normalize a response that we use to check if a retry should occur. This class smoothes over the different types of responses we may get from a service including: * A modeled error response from the service that contains a service code and error message. * A raw HTTP response that doesn't contain service protocol specific error keys. * An exception received while attempting to retrieve a response. This could be a ConnectionError we receive from our HTTP layer which could represent that we weren't able to receive a response from the service. This class guarantees that at least one of the above attributes will be non None. This class is meant to provide a read-only view into the properties associated with a possible retryable response. None of the properties are meant to be modified directly. """ def __init__(self, attempt_number, operation_model=None, parsed_response=None, http_response=None, caught_exception=None, request_context=None): # 1-based attempt number. self.attempt_number = attempt_number self.operation_model = operation_model # This is the parsed response dictionary we get from parsing # the HTTP response from the service. self.parsed_response = parsed_response # This is an instance of botocore.awsrequest.AWSResponse. self.http_response = http_response # This is a subclass of Exception that will be non None if # an exception was raised when retrying to retrieve a response. self.caught_exception = caught_exception # This is the request context dictionary that's added to the # request dict. This is used to story any additional state # about the request. We use this for storing retry quota # capacity. if request_context is None: request_context = {} self.request_context = request_context self._retry_metadata = {} # These are misc helper methods to avoid duplication in the various # checkers. def get_error_code(self): """Check if there was a parsed response with an error code. If we could not find any error codes, ``None`` is returned. """ if self.parsed_response is None: return return self.parsed_response.get('Error', {}).get('Code') def add_retry_metadata(self, **kwargs): """Add key/value pairs to the retry metadata. This allows any objects during the retry process to add metadata about any checks/validations that happened. This gets added to the response metadata in the retry handler. """ self._retry_metadata.update(**kwargs) def get_retry_metadata(self): return self._retry_metadata.copy() class RetryPolicy(object): def __init__(self, retry_checker, retry_backoff): self._retry_checker = retry_checker self._retry_backoff = retry_backoff def should_retry(self, context): return self._retry_checker.is_retryable(context) def compute_retry_delay(self, context): return self._retry_backoff.delay_amount(context) class ExponentialBackoff(BaseRetryBackoff): _BASE = 2 _MAX_BACKOFF = 20 def __init__(self, max_backoff=20, random=random.random): self._base = self._BASE self._max_backoff = max_backoff self._random = random def delay_amount(self, context): """Calculates delay based on exponential backoff. This class implements truncated binary exponential backoff with jitter:: t_i = min(rand(0, 1) * 2 ** attempt, MAX_BACKOFF) where ``i`` is the request attempt (0 based). """ # The context.attempt_number is a 1-based value, but we have # to calculate the delay based on i based a 0-based value. We # want the first delay to just be ``rand(0, 1)``. return min( self._random() * (self._base ** (context.attempt_number - 1)), self._max_backoff ) class MaxAttemptsChecker(BaseRetryableChecker): def __init__(self, max_attempts): self._max_attempts = max_attempts def is_retryable(self, context): under_max_attempts = context.attempt_number < self._max_attempts if not under_max_attempts: logger.debug("Max attempts of %s reached.", self._max_attempts) context.add_retry_metadata(MaxAttemptsReached=True) return under_max_attempts class TransientRetryableChecker(BaseRetryableChecker): _TRANSIENT_ERROR_CODES = [ 'RequestTimeout', 'RequestTimeoutException', 'PriorRequestNotComplete', ] _TRANSIENT_STATUS_CODES = [500, 502, 503, 504] _TRANSIENT_EXCEPTION_CLS = ( ConnectionError, HTTPClientError, ) def __init__(self, transient_error_codes=None, transient_status_codes=None, transient_exception_cls=None): if transient_error_codes is None: transient_error_codes = self._TRANSIENT_ERROR_CODES[:] if transient_status_codes is None: transient_status_codes = self._TRANSIENT_STATUS_CODES[:] if transient_exception_cls is None: transient_exception_cls = self._TRANSIENT_EXCEPTION_CLS self._transient_error_codes = transient_error_codes self._transient_status_codes = transient_status_codes self._transient_exception_cls = transient_exception_cls def is_retryable(self, context): if context.get_error_code() in self._transient_error_codes: return True if context.http_response is not None: if context.http_response.status_code in \ self._transient_status_codes: return True if context.caught_exception is not None: return isinstance(context.caught_exception, self._transient_exception_cls) return False class ThrottledRetryableChecker(BaseRetryableChecker): # This is the union of all error codes we've seen that represent # a throttled error. _THROTTLED_ERROR_CODES = [ 'Throttling', 'ThrottlingException', 'ThrottledException', 'RequestThrottledException', 'TooManyRequestsException', 'ProvisionedThroughputExceededException', 'TransactionInProgressException', 'RequestLimitExceeded', 'BandwidthLimitExceeded', 'LimitExceededException', 'RequestThrottled', 'SlowDown', 'PriorRequestNotComplete', 'EC2ThrottledException', ] def __init__(self, throttled_error_codes=None): if throttled_error_codes is None: throttled_error_codes = self._THROTTLED_ERROR_CODES[:] self._throttled_error_codes = throttled_error_codes def is_retryable(self, context): # Only the error code from a parsed service response is used # to determine if the response is a throttled response. return context.get_error_code() in self._throttled_error_codes class ModeledRetryableChecker(BaseRetryableChecker): """Check if an error has been modeled as retryable.""" def __init__(self): self._error_detector = ModeledRetryErrorDetector() def is_retryable(self, context): error_code = context.get_error_code() if error_code is None: return False return self._error_detector.detect_error_type(context) is not None class ModeledRetryErrorDetector(object): """Checks whether or not an error is a modeled retryable error.""" # There are return values from the detect_error_type() method. TRANSIENT_ERROR = 'TRANSIENT_ERROR' THROTTLING_ERROR = 'THROTTLING_ERROR' # This class is lower level than ModeledRetryableChecker, which # implements BaseRetryableChecker. This object allows you to distinguish # between the various types of retryable errors. def detect_error_type(self, context): """Detect the error type associated with an error code and model. This will either return: * ``self.TRANSIENT_ERROR`` - If the error is a transient error * ``self.THROTTLING_ERROR`` - If the error is a throttling error * ``None`` - If the error is neither type of error. """ error_code = context.get_error_code() op_model = context.operation_model if op_model is None or not op_model.error_shapes: return for shape in op_model.error_shapes: if shape.metadata.get('retryable') is not None: # Check if this error code matches the shape. This can # be either by name or by a modeled error code. error_code_to_check = ( shape.metadata.get('error', {}).get('code') or shape.name ) if error_code == error_code_to_check: if shape.metadata['retryable'].get('throttling'): return self.THROTTLING_ERROR return self.TRANSIENT_ERROR class ThrottlingErrorDetector(object): def __init__(self, retry_event_adapter): self._modeled_error_detector = ModeledRetryErrorDetector() self._fixed_error_code_detector = ThrottledRetryableChecker() self._retry_event_adapter = retry_event_adapter # This expects the kwargs from needs-retry to be passed through. def is_throttling_error(self, **kwargs): context = self._retry_event_adapter.create_retry_context(**kwargs) if self._fixed_error_code_detector.is_retryable(context): return True error_type = self._modeled_error_detector.detect_error_type(context) return error_type == self._modeled_error_detector.THROTTLING_ERROR class StandardRetryConditions(BaseRetryableChecker): """Concrete class that implements the standard retry policy checks. Specifically: not max_attempts and (transient or throttled or modeled_retry) """ def __init__(self, max_attempts=DEFAULT_MAX_ATTEMPTS): # Note: This class is for convenience so you can have the # standard retry condition in a single class. self._max_attempts_checker = MaxAttemptsChecker(max_attempts) self._additional_checkers = OrRetryChecker([ TransientRetryableChecker(), ThrottledRetryableChecker(), ModeledRetryableChecker(), OrRetryChecker([ special.RetryIDPCommunicationError(), special.RetryDDBChecksumError(), ]) ]) def is_retryable(self, context): return (self._max_attempts_checker.is_retryable(context) and self._additional_checkers.is_retryable(context)) class OrRetryChecker(BaseRetryableChecker): def __init__(self, checkers): self._checkers = checkers def is_retryable(self, context): return any(checker.is_retryable(context) for checker in self._checkers) class RetryQuotaChecker(object): _RETRY_COST = 5 _NO_RETRY_INCREMENT = 1 _TIMEOUT_RETRY_REQUEST = 10 _TIMEOUT_EXCEPTIONS = (ConnectTimeoutError, ReadTimeoutError) # Implementation note: We're not making this a BaseRetryableChecker # because this isn't just a check if we can retry. This also changes # state so we have to careful when/how we call this. Making it # a BaseRetryableChecker implies you can call .is_retryable(context) # as many times as you want and not affect anything. def __init__(self, quota): self._quota = quota # This tracks the last amount self._last_amount_acquired = None def acquire_retry_quota(self, context): if self._is_timeout_error(context): capacity_amount = self._TIMEOUT_RETRY_REQUEST else: capacity_amount = self._RETRY_COST success = self._quota.acquire(capacity_amount) if success: # We add the capacity amount to the request context so we know # how much to release later. The capacity amount can vary based # on the error. context.request_context['retry_quota_capacity'] = capacity_amount return True context.add_retry_metadata(RetryQuotaReached=True) return False def _is_timeout_error(self, context): return isinstance(context.caught_exception, self._TIMEOUT_EXCEPTIONS) # This is intended to be hooked up to ``after-call``. def release_retry_quota(self, context, http_response, **kwargs): # There's three possible options. # 1. The HTTP response did not have a 2xx response. In that case we # give no quota back. # 2. The HTTP request was successful and was never retried. In # that case we give _NO_RETRY_INCREMENT back. # 3. The API call had retries, and we eventually receive an HTTP # response with a 2xx status code. In that case we give back # whatever quota was associated with the last acquisition. if http_response is None: return status_code = http_response.status_code if 200 <= status_code < 300: if 'retry_quota_capacity' not in context: self._quota.release(self._NO_RETRY_INCREMENT) else: capacity_amount = context['retry_quota_capacity'] self._quota.release(capacity_amount)