X-Git-Url: https://git.arvados.org/arvados.git/blobdiff_plain/0f8dc3d824f03b82b8db9f61bbcc0592b62b998f..0561bd0c3c07257fd58ded6c7cfa5feeae97af57:/services/nodemanager/arvnodeman/computenode/driver/__init__.py diff --git a/services/nodemanager/arvnodeman/computenode/driver/__init__.py b/services/nodemanager/arvnodeman/computenode/driver/__init__.py index 9bdcc5f7a1..8f881b04e3 100644 --- a/services/nodemanager/arvnodeman/computenode/driver/__init__.py +++ b/services/nodemanager/arvnodeman/computenode/driver/__init__.py @@ -1,4 +1,7 @@ #!/usr/bin/env python +# Copyright (C) The Arvados Authors. All rights reserved. +# +# SPDX-License-Identifier: AGPL-3.0 from __future__ import absolute_import, print_function @@ -8,7 +11,7 @@ from operator import attrgetter import libcloud.common.types as cloud_types from libcloud.compute.base import NodeDriver, NodeAuthSSHKey -from ...config import NETWORK_ERRORS +from ...config import CLOUD_ERRORS from .. import RetryMixin class BaseComputeNodeDriver(RetryMixin): @@ -24,7 +27,7 @@ class BaseComputeNodeDriver(RetryMixin): Subclasses must implement arvados_create_kwargs, sync_node, node_fqdn, and node_start_time. """ - CLOUD_ERRORS = NETWORK_ERRORS + (cloud_types.LibcloudError,) + @RetryMixin._retry() def _create_driver(self, driver_class, **auth_kwargs): @@ -79,33 +82,60 @@ class BaseComputeNodeDriver(RetryMixin): key = NodeAuthSSHKey(ssh_file.read()) return 'auth', key - def search_for(self, term, list_method, key=attrgetter('id'), **kwargs): + def search_for_now(self, term, list_method, key=attrgetter('id'), **kwargs): """Return one matching item from a list of cloud objects. Raises ValueError if the number of matching objects is not exactly 1. Arguments: * term: The value that identifies a matching item. - * list_method: A string that names the method to call on this - instance's libcloud driver for a list of objects. + * list_method: A string that names the method to call for a + list of objects. * key: A function that accepts a cloud object and returns a value search for a `term` match on each item. Returns the object's 'id' attribute by default. """ + try: + list_func = getattr(self, list_method) + except AttributeError: + list_func = getattr(self.real, list_method) + items = list_func(**kwargs) + results = [item for item in items if key(item) == term] + count = len(results) + if count != 1: + raise ValueError("{} returned {} results for {!r}".format( + list_method, count, term)) + return results[0] + + def search_for(self, term, list_method, key=attrgetter('id'), **kwargs): + """Return one cached matching item from a list of cloud objects. + + See search_for_now() for details of arguments and exceptions. + This method caches results, so it's good to find static cloud objects + like node sizes, regions, etc. + """ cache_key = (list_method, term) if cache_key not in self.SEARCH_CACHE: - items = getattr(self.real, list_method)(**kwargs) - results = [item for item in items - if key(item) == term] - count = len(results) - if count != 1: - raise ValueError("{} returned {} results for '{}'".format( - list_method, count, term)) - self.SEARCH_CACHE[cache_key] = results[0] + self.SEARCH_CACHE[cache_key] = self.search_for_now( + term, list_method, key, **kwargs) return self.SEARCH_CACHE[cache_key] - def list_nodes(self): - return self.real.list_nodes(**self.list_kwargs) + def list_nodes(self, **kwargs): + l = self.list_kwargs.copy() + l.update(kwargs) + return self.real.list_nodes(**l) + + def create_cloud_name(self, arvados_node): + """Return a cloud node name for the given Arvados node record. + + Subclasses must override this method. It should return a string + that can be used as the name for a newly-created cloud node, + based on identifying information in the Arvados node record. + + Arguments: + * arvados_node: This Arvados node record to seed the new cloud node. + """ + raise NotImplementedError("BaseComputeNodeDriver.create_cloud_name") def arvados_create_kwargs(self, size, arvados_node): """Return dynamic keyword arguments for create_node. @@ -141,19 +171,17 @@ class BaseComputeNodeDriver(RetryMixin): kwargs.update(self.arvados_create_kwargs(size, arvados_node)) kwargs['size'] = size return self.real.create_node(**kwargs) - except self.CLOUD_ERRORS: + except CLOUD_ERRORS as create_error: # Workaround for bug #6702: sometimes the create node request # succeeds but times out and raises an exception instead of # returning a result. If this happens, we get stuck in a retry # loop forever because subsequent create_node attempts will fail # due to node name collision. So check if the node we intended to # create shows up in the cloud node list and return it if found. - node = self.search_for(kwargs['name'], 'list_nodes', self._name_key) - if node: - return node - else: - # something else went wrong, re-raise the exception - raise + try: + return self.search_for_now(kwargs['name'], 'list_nodes', self._name_key) + except ValueError: + raise create_error def post_create_node(self, cloud_node): # ComputeNodeSetupActor calls this method after the cloud node is @@ -180,13 +208,22 @@ class BaseComputeNodeDriver(RetryMixin): # seconds since the epoch UTC. raise NotImplementedError("BaseComputeNodeDriver.node_start_time") - @classmethod - def is_cloud_exception(cls, exception): - # libcloud compute drivers typically raise bare Exceptions to - # represent API errors. Return True for any exception that is - # exactly an Exception, or a better-known higher-level exception. - return (isinstance(exception, cls.CLOUD_ERRORS) or - type(exception) is Exception) + def destroy_node(self, cloud_node): + try: + return self.real.destroy_node(cloud_node) + except CLOUD_ERRORS as destroy_error: + # Sometimes the destroy node request succeeds but times out and + # raises an exception instead of returning success. If this + # happens, we get a noisy stack trace. Check if the node is still + # on the node list. If it is gone, we can declare victory. + try: + self.search_for_now(cloud_node.id, 'list_nodes') + except ValueError: + # If we catch ValueError, that means search_for_now didn't find + # it, which means destroy_node actually succeeded. + return True + # The node is still on the list. Re-raise. + raise # Now that we've defined all our own methods, delegate generic, public # attributes of libcloud drivers that we haven't defined ourselves. @@ -196,6 +233,11 @@ class BaseComputeNodeDriver(RetryMixin): lambda self, value: setattr(self.real, attr_name, value), doc=getattr(getattr(NodeDriver, attr_name), '__doc__', None)) + # node id + @classmethod + def node_id(cls): + raise NotImplementedError("BaseComputeNodeDriver.node_id") + _locals = locals() for _attr_name in dir(NodeDriver): if (not _attr_name.startswith('_')) and (_attr_name not in _locals):