gitops/patroni
1
0
Fork 0
mirror of https://github.com/outbackdingo/patroni.git synced 2026-08-25 07:20:20 +00:00
Code Issues Packages Projects Releases Wiki Activity

Fix bug with switchover in synchronous_mode=quorum (#3310)

Browse Source 
When the candidate is specified we don't have to check quorum requirements.
The problem was introduced in #3278

Close #3307
This commit is contained in:
Alexander Kukushkin
2025-03-24 08:50:49 +01:00
committed by GitHub
parent a3c772dfc9
commit deb9cc6b73
2 changed files with 54 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files
patroni/ha.py
+12 -4
View File
@@ -1377,7 +1377,14 @@ class Ha(object):
quorum_votes += 1
# In case of quorum replication we need to make sure that there is enough healthy synchronous replicas!
return quorum_votes >= (self.cluster.sync.quorum if self.quorum_commit_mode_is_active() else 0)
# However, when failover candidate is set, we can ignore quorum requirements.
check_quorum = self.quorum_commit_mode_is_active() and\
not (self.cluster.failover and self.cluster.failover.candidate and not exclude_failover_candidate)
if check_quorum and quorum_votes < self.cluster.sync.quorum:
logger.info('Quorum requirement %d can not be reached', self.cluster.sync.quorum)
return False
return quorum_votes >= 0
def manual_failover_process_no_leader(self) -> Optional[bool]:
"""Handles manual failover/switchover when the old leader already stepped down.
@@ -2418,8 +2425,9 @@ class Ha(object):
return False
# Don't spend time on "nofailover" nodes checking.
# We also don't need nodes which we can't query with the api in the list.
return node.name not in exclude and \
not node.nofailover and bool(node.api_url) and \
(not failover or not failover.candidate or node.name == failover.candidate)
# And, if exclude_failover_candidate is True we want to skip node.name == failover.candidate check.
return node.name not in exclude and not node.nofailover and bool(node.api_url) and \
(exclude_failover_candidate or not failover
or not failover.candidate or node.name == failover.candidate)
return list(filter(is_eligible, self.cluster.members))
tests/test_ha.py
+42 -2
View File
@@ -68,7 +68,8 @@ def get_cluster_initialized_without_leader(leader=False, failover=None, sync=Non
'tags': {'clonefrom': True},
'scheduled_restart': {'schedule': "2100-01-01 10:53:07.560445+00:00",
'postgres_version': '99.0.0'}})
syncstate = SyncState(0 if sync else None, sync and sync[0], sync and sync[1], 0)
syncstate = SyncState(0 if sync else None, sync and sync[0],
sync and sync[1], sync[2] if sync and len(sync) > 2 else 0)
failsafe = {m.name: m.api_url for m in (m1, m2)} if failsafe else None
return get_cluster(SYSID, leader, [m1, m2], failover, syncstate, cluster_config, failsafe)
@@ -1015,7 +1016,7 @@ class TestHa(PostgresInit):
# switchover from a specific leader, but the only sync node (us, postgresql0) has nofailover tag
self.ha.cluster = get_cluster_initialized_without_leader(failover=Failover(0, 'leader', '', None),
sync=('postgresql0'))
sync=('postgresql0', None))
self.ha.patroni.nofailover = True
self.assertEqual(self.ha.run_cycle(), 'following a different leader because I am not allowed to promote')
@@ -1844,3 +1845,42 @@ class TestHa(PostgresInit):
# Test that _process_quorum_replication doesn't take longer than loop_wait
with patch('time.time', Mock(side_effect=[30, 60, 90, 120])):
self.ha.process_sync_replication()
def test_is_failover_possible(self):
self.p._major_version = 140000 # supports_multiple_sync
self.p.name = 'leader'
self.ha.fetch_node_status = get_node_status()
self.ha.cluster = get_cluster_initialized_with_leader(sync=('leader', 'foo,other', 1),
failover=Failover(0, 'leader', 'other', None))
self.ha.cluster.members.append(Member(0, 'foo', 28, {'api_url': 'http://127.0.0.1:8011/patroni'}))
# switchover when synchronous_mode = off
self.assertTrue(self.ha.is_failover_possible())
with patch.object(global_config.__class__, 'is_synchronous_mode', PropertyMock(return_value=True)):
# switchover to synchronous node when synchronous_mode = on
self.assertTrue(self.ha.is_failover_possible())
with patch.object(global_config.__class__, 'is_quorum_commit_mode', PropertyMock(return_value=True)):
# switchover to synchronous node when synchronous_mode = quorum
self.assertTrue(self.ha.is_failover_possible()) # success, despite the fact that quorum is low
# failover candidate is unhealthy, we are checking if there are other good candidates, but quorum is low
self.assertFalse(self.ha.is_failover_possible(exclude_failover_candidate=True))
# now we satisfy quorum requirements
with patch.object(SyncState, 'quorum', PropertyMock(return_value=0)):
self.assertTrue(self.ha.is_failover_possible(exclude_failover_candidate=True))
self.ha.cluster = get_cluster_initialized_with_leader(sync=('leader', 'foo,other', 1),
failover=Failover(0, '', 'foo', None))
# failover to missing node foo
self.assertFalse(self.ha.is_failover_possible())
self.ha.cluster = get_cluster_initialized_with_leader(sync=('leader', 'foo,other', 1),
failover=Failover(0, 'leader', '', None))
# switchover from leader when synchronous_mode = off
self.assertTrue(self.ha.is_failover_possible())
with patch.object(global_config.__class__, 'is_synchronous_mode', PropertyMock(return_value=True)):
# switchover from leader when synchronous_mode = on
self.assertTrue(self.ha.is_failover_possible())
with patch.object(global_config.__class__, 'is_quorum_commit_mode', PropertyMock(return_value=True)):
# switchover from leader when synchronous_mode = quorum
self.assertFalse(self.ha.is_failover_possible()) # failure, because quorum is low