summaryrefslogtreecommitdiffstats
diff options
context:
space:
mode:
-rw-r--r--roles/openshift_health_checker/callback_plugins/zz_failure_summary.py291
-rw-r--r--roles/openshift_health_checker/test/conftest.py1
-rw-r--r--roles/openshift_health_checker/test/zz_failure_summary_test.py70
3 files changed, 243 insertions, 119 deletions
diff --git a/roles/openshift_health_checker/callback_plugins/zz_failure_summary.py b/roles/openshift_health_checker/callback_plugins/zz_failure_summary.py
index 9f9fe123a..46cc3b577 100644
--- a/roles/openshift_health_checker/callback_plugins/zz_failure_summary.py
+++ b/roles/openshift_health_checker/callback_plugins/zz_failure_summary.py
@@ -1,161 +1,214 @@
-"""
-Ansible callback plugin to give a nicely formatted summary of failures.
-"""
+"""Ansible callback plugin to print a nicely formatted summary of failures.
-# Reason: In several locations below we disable pylint protected-access
-# for Ansible objects that do not give us any public way
-# to access the full details we need to report check failures.
-# Status: disabled permanently or until Ansible object has a public API.
-# This does leave the code more likely to be broken by future Ansible changes.
+The file / module name is prefixed with `zz_` to make this plugin be loaded last
+by Ansible, thus making its output the last thing that users see.
+"""
-from pprint import pformat
+from collections import defaultdict
from ansible.plugins.callback import CallbackBase
from ansible import constants as C
from ansible.utils.color import stringc
+FAILED_NO_MSG = u'Failed without returning a message.'
+
+
class CallbackModule(CallbackBase):
- """
- This callback plugin stores task results and summarizes failures.
- The file name is prefixed with `zz_` to make this plugin be loaded last by
- Ansible, thus making its output the last thing that users see.
- """
+ """This callback plugin stores task results and summarizes failures."""
CALLBACK_VERSION = 2.0
CALLBACK_TYPE = 'aggregate'
CALLBACK_NAME = 'failure_summary'
CALLBACK_NEEDS_WHITELIST = False
- _playbook_file = None
def __init__(self):
super(CallbackModule, self).__init__()
self.__failures = []
+ self.__playbook_file = ''
def v2_playbook_on_start(self, playbook):
super(CallbackModule, self).v2_playbook_on_start(playbook)
- # re: playbook attrs see top comment # pylint: disable=protected-access
- self._playbook_file = playbook._file_name
+ # pylint: disable=protected-access; Ansible gives us no public API to
+ # get the file name of the current playbook from a callback plugin.
+ self.__playbook_file = playbook._file_name
def v2_runner_on_failed(self, result, ignore_errors=False):
super(CallbackModule, self).v2_runner_on_failed(result, ignore_errors)
if not ignore_errors:
- self.__failures.append(dict(result=result, ignore_errors=ignore_errors))
+ self.__failures.append(result)
def v2_playbook_on_stats(self, stats):
super(CallbackModule, self).v2_playbook_on_stats(stats)
if self.__failures:
- self._print_failure_details(self.__failures)
-
- def _print_failure_details(self, failures):
- """Print a summary of failed tasks or checks."""
- self._display.display(u'\nFailure summary:\n')
-
- width = len(str(len(failures)))
- initial_indent_format = u' {{:>{width}}}. '.format(width=width)
- initial_indent_len = len(initial_indent_format.format(0))
- subsequent_indent = u' ' * initial_indent_len
- subsequent_extra_indent = u' ' * (initial_indent_len + 10)
-
- for i, failure in enumerate(failures, 1):
- entries = _format_failure(failure)
- self._display.display(u'\n{}{}'.format(initial_indent_format.format(i), entries[0]))
- for entry in entries[1:]:
- entry = entry.replace(u'\n', u'\n' + subsequent_extra_indent)
- indented = u'{}{}'.format(subsequent_indent, entry)
- self._display.display(indented)
-
- failed_checks = set()
- playbook_context = None
- # re: result attrs see top comment # pylint: disable=protected-access
- for failure in failures:
- # Get context from check task result since callback plugins cannot access task vars.
- # NOTE: thus context is not known unless checks run. Failures prior to checks running
- # don't have playbook_context in the results. But we only use it now when checks fail.
- playbook_context = playbook_context or failure['result']._result.get('playbook_context')
- failed_checks.update(
- name
- for name, result in failure['result']._result.get('checks', {}).items()
- if result.get('failed')
- )
- if failed_checks:
- self._print_check_failure_summary(failed_checks, playbook_context)
-
- def _print_check_failure_summary(self, failed_checks, context):
- checks = ','.join(sorted(failed_checks))
- # The purpose of specifying context is to vary the output depending on what the user was
- # expecting to happen (based on which playbook they ran). The only use currently is to
- # vary the message depending on whether the user was deliberately running checks or was
- # trying to install/upgrade and checks are just included. Other use cases may arise.
- summary = ( # default to explaining what checks are in the first place
- '\n'
- 'The execution of "{playbook}"\n'
- 'includes checks designed to fail early if the requirements\n'
- 'of the playbook are not met. One or more of these checks\n'
- 'failed. To disregard these results, you may choose to\n'
- 'disable failing checks by setting an Ansible variable:\n\n'
- ' openshift_disable_check={checks}\n\n'
- 'Failing check names are shown in the failure details above.\n'
- 'Some checks may be configurable by variables if your requirements\n'
- 'are different from the defaults; consult check documentation.\n'
- 'Variables can be set in the inventory or passed on the\n'
- 'command line using the -e flag to ansible-playbook.\n\n'
- ).format(playbook=self._playbook_file, checks=checks)
- if context in ['pre-install', 'health', 'adhoc']:
- summary = ( # user was expecting to run checks, less explanation needed
- '\n'
- 'You may choose to configure or disable failing checks by\n'
- 'setting Ansible variables. To disable those above:\n\n'
- ' openshift_disable_check={checks}\n\n'
- 'Consult check documentation for configurable variables.\n'
- 'Variables can be set in the inventory or passed on the\n'
- 'command line using the -e flag to ansible-playbook.\n\n'
- ).format(checks=checks)
- self._display.display(summary)
-
-
-# re: result attrs see top comment # pylint: disable=protected-access
-def _format_failure(failure):
+ self._display.display(failure_summary(self.__failures, self.__playbook_file))
+
+
+def failure_summary(failures, playbook):
+ """Return a summary of failed tasks, including details on health checks."""
+ if not failures:
+ return u''
+
+ # NOTE: because we don't have access to task_vars from callback plugins, we
+ # store the playbook context in the task result when the
+ # openshift_health_check action plugin is used, and we use this context to
+ # customize the error message.
+ # pylint: disable=protected-access; Ansible gives us no sufficient public
+ # API on TaskResult objects.
+ context = next((
+ context for context in
+ (failure._result.get('playbook_context') for failure in failures)
+ if context
+ ), None)
+
+ failures = [failure_to_dict(failure) for failure in failures]
+ failures = deduplicate_failures(failures)
+
+ summary = [u'', u'', u'Failure summary:', u'']
+
+ width = len(str(len(failures)))
+ initial_indent_format = u' {{:>{width}}}. '.format(width=width)
+ initial_indent_len = len(initial_indent_format.format(0))
+ subsequent_indent = u' ' * initial_indent_len
+ subsequent_extra_indent = u' ' * (initial_indent_len + 10)
+
+ for i, failure in enumerate(failures, 1):
+ entries = format_failure(failure)
+ summary.append(u'\n{}{}'.format(initial_indent_format.format(i), entries[0]))
+ for entry in entries[1:]:
+ entry = entry.replace(u'\n', u'\n' + subsequent_extra_indent)
+ indented = u'{}{}'.format(subsequent_indent, entry)
+ summary.append(indented)
+
+ failed_checks = set()
+ for failure in failures:
+ failed_checks.update(name for name, message in failure['checks'])
+ if failed_checks:
+ summary.append(check_failure_footer(failed_checks, context, playbook))
+
+ return u'\n'.join(summary)
+
+
+def failure_to_dict(failed_task_result):
+ """Extract information out of a failed TaskResult into a dict.
+
+ The intent is to transform a TaskResult object into something easier to
+ manipulate. TaskResult is ansible.executor.task_result.TaskResult.
+ """
+ # pylint: disable=protected-access; Ansible gives us no sufficient public
+ # API on TaskResult objects.
+ _result = failed_task_result._result
+ return {
+ 'host': failed_task_result._host.get_name(),
+ 'play': play_name(failed_task_result._task),
+ 'task': failed_task_result.task_name,
+ 'msg': _result.get('msg', FAILED_NO_MSG),
+ 'checks': tuple(
+ (name, result.get('msg', FAILED_NO_MSG))
+ for name, result in sorted(_result.get('checks', {}).items())
+ if result.get('failed')
+ ),
+ }
+
+
+def play_name(obj):
+ """Given a task or block, return the name of its parent play.
+
+ This is loosely inspired by ansible.playbook.base.Base.dump_me.
+ """
+ # pylint: disable=protected-access; Ansible gives us no sufficient public
+ # API to implement this.
+ if not obj:
+ return ''
+ if hasattr(obj, '_play'):
+ return obj._play.get_name()
+ return play_name(getattr(obj, '_parent'))
+
+
+def deduplicate_failures(failures):
+ """Group together similar failures from different hosts.
+
+ Returns a new list of failures such that identical failures from different
+ hosts are grouped together in a single entry. The relative order of failures
+ is preserved.
+ """
+ groups = defaultdict(list)
+ for failure in failures:
+ group_key = tuple(sorted((key, value) for key, value in failure.items() if key != 'host'))
+ groups[group_key].append(failure)
+ result = []
+ for failure in failures:
+ group_key = tuple(sorted((key, value) for key, value in failure.items() if key != 'host'))
+ if group_key not in groups:
+ continue
+ failure['host'] = tuple(sorted(g_failure['host'] for g_failure in groups.pop(group_key)))
+ result.append(failure)
+ return result
+
+
+def format_failure(failure):
"""Return a list of pretty-formatted text entries describing a failure, including
relevant information about it. Expect that the list of text entries will be joined
by a newline separator when output to the user."""
- result = failure['result']
- host = result._host.get_name()
- play = _get_play(result._task)
- if play:
- play = play.get_name()
- task = result._task.get_name()
- msg = result._result.get('msg', u'???')
+ host = u', '.join(failure['host'])
+ play = failure['play']
+ task = failure['task']
+ msg = failure['msg']
+ checks = failure['checks']
fields = (
- (u'Host', host),
+ (u'Hosts', host),
(u'Play', play),
(u'Task', task),
(u'Message', stringc(msg, C.COLOR_ERROR)),
)
- if 'checks' in result._result:
- fields += ((u'Details', _format_failed_checks(result._result['checks'])),)
+ if checks:
+ fields += ((u'Details', format_failed_checks(checks)),)
row_format = '{:10}{}'
return [row_format.format(header + u':', body) for header, body in fields]
-def _format_failed_checks(checks):
+def format_failed_checks(checks):
"""Return pretty-formatted text describing checks that failed."""
- failed_check_msgs = []
- for check, body in checks.items():
- if body.get('failed', False): # only show the failed checks
- msg = body.get('msg', u"Failed without returning a message")
- failed_check_msgs.append('check "%s":\n%s' % (check, msg))
- if failed_check_msgs:
- return stringc("\n\n".join(failed_check_msgs), C.COLOR_ERROR)
- else: # something failed but no checks will admit to it, so dump everything
- return stringc(pformat(checks), C.COLOR_ERROR)
-
-
-# This is inspired by ansible.playbook.base.Base.dump_me.
-# re: play/task/block attrs see top comment # pylint: disable=protected-access
-def _get_play(obj):
- """Given a task or block, recursively try to find its parent play."""
- if hasattr(obj, '_play'):
- return obj._play
- if getattr(obj, '_parent'):
- return _get_play(obj._parent)
+ messages = []
+ for name, message in checks:
+ messages.append(u'check "{}":\n{}'.format(name, message))
+ return stringc(u'\n\n'.join(messages), C.COLOR_ERROR)
+
+
+def check_failure_footer(failed_checks, context, playbook):
+ """Return a textual explanation about checks depending on context.
+
+ The purpose of specifying context is to vary the output depending on what
+ the user was expecting to happen (based on which playbook they ran). The
+ only use currently is to vary the message depending on whether the user was
+ deliberately running checks or was trying to install/upgrade and checks are
+ just included. Other use cases may arise.
+ """
+ checks = ','.join(sorted(failed_checks))
+ summary = [u'']
+ if context in ['pre-install', 'health', 'adhoc']:
+ # User was expecting to run checks, less explanation needed.
+ summary.extend([
+ u'You may configure or disable checks by setting Ansible '
+ u'variables. To disable those above, set:',
+ u' openshift_disable_check={checks}'.format(checks=checks),
+ u'Consult check documentation for configurable variables.',
+ ])
+ else:
+ # User may not be familiar with the checks, explain what checks are in
+ # the first place.
+ summary.extend([
+ u'The execution of "{playbook}" includes checks designed to fail '
+ u'early if the requirements of the playbook are not met. One or '
+ u'more of these checks failed. To disregard these results,'
+ u'explicitly disable checks by setting an Ansible variable:'.format(playbook=playbook),
+ u' openshift_disable_check={checks}'.format(checks=checks),
+ u'Failing check names are shown in the failure details above. '
+ u'Some checks may be configurable by variables if your requirements '
+ u'are different from the defaults; consult check documentation.',
+ ])
+ summary.append(
+ u'Variables can be set in the inventory or passed on the command line '
+ u'using the -e flag to ansible-playbook.'
+ )
+ return u'\n'.join(summary)
diff --git a/roles/openshift_health_checker/test/conftest.py b/roles/openshift_health_checker/test/conftest.py
index 3cbd65507..244a1f0fa 100644
--- a/roles/openshift_health_checker/test/conftest.py
+++ b/roles/openshift_health_checker/test/conftest.py
@@ -7,5 +7,6 @@ openshift_health_checker_path = os.path.dirname(os.path.dirname(__file__))
sys.path[1:1] = [
openshift_health_checker_path,
os.path.join(openshift_health_checker_path, 'action_plugins'),
+ os.path.join(openshift_health_checker_path, 'callback_plugins'),
os.path.join(openshift_health_checker_path, 'library'),
]
diff --git a/roles/openshift_health_checker/test/zz_failure_summary_test.py b/roles/openshift_health_checker/test/zz_failure_summary_test.py
new file mode 100644
index 000000000..0fc258133
--- /dev/null
+++ b/roles/openshift_health_checker/test/zz_failure_summary_test.py
@@ -0,0 +1,70 @@
+from zz_failure_summary import deduplicate_failures
+
+import pytest
+
+
+@pytest.mark.parametrize('failures,deduplicated', [
+ (
+ [
+ {
+ 'host': 'master1',
+ 'msg': 'One or more checks failed',
+ },
+ ],
+ [
+ {
+ 'host': ('master1',),
+ 'msg': 'One or more checks failed',
+ },
+ ],
+ ),
+ (
+ [
+ {
+ 'host': 'master1',
+ 'msg': 'One or more checks failed',
+ },
+ {
+ 'host': 'node1',
+ 'msg': 'One or more checks failed',
+ },
+ ],
+ [
+ {
+ 'host': ('master1', 'node1'),
+ 'msg': 'One or more checks failed',
+ },
+ ],
+ ),
+ (
+ [
+ {
+ 'host': 'node1',
+ 'msg': 'One or more checks failed',
+ 'checks': (('test_check', 'error message'),),
+ },
+ {
+ 'host': 'master2',
+ 'msg': 'Some error happened',
+ },
+ {
+ 'host': 'master1',
+ 'msg': 'One or more checks failed',
+ 'checks': (('test_check', 'error message'),),
+ },
+ ],
+ [
+ {
+ 'host': ('master1', 'node1'),
+ 'msg': 'One or more checks failed',
+ 'checks': (('test_check', 'error message'),),
+ },
+ {
+ 'host': ('master2',),
+ 'msg': 'Some error happened',
+ },
+ ],
+ ),
+])
+def test_deduplicate_failures(failures, deduplicated):
+ assert deduplicate_failures(failures) == deduplicated