Add test for network failover

Signed-off-by: a.y.volkov <a.y.volkov@yadro.com>

pytest_tests/helpers/iptables_helper.py

@@ -0,0 +1,16 @@
+from ssh_helper import HostClient
+
+
+class IpTablesHelper:
+
+    @staticmethod
+    def drop_input_traffic_to_port(client: HostClient, ports: list[str]):
+        for port in ports:
+            cmd_output = client.exec(cmd=f'iptables -A INPUT -p tcp --dport {port} -j DROP')
+            assert cmd_output.rc == 0
+
+    @staticmethod
+    def restore_input_traffic_to_port(client: HostClient, ports: list[str]):
+        for port in ports:
+            cmd_output = client.exec(cmd=f'iptables -D INPUT -p tcp --dport {port} -j DROP')
+            assert cmd_output.rc == 0

pytest_tests/helpers/ssh_helper.py

@@ -66,15 +66,19 @@ class HostClient:
     TIMEOUT_RESTORE_CONNECTION = 10, 24
 
     def __init__(self, ip: str, login: str, password: Optional[str] = None,
-                 private_key_path: Optional[str] = None, init_ssh_client=True) -> None:
+                 private_key_path: Optional[str] = None, private_key_passphrase: Optional[str] = None,
+                 init_ssh_client=True) -> None:
         self.ip = ip
         self.login = login
         self.password = password
         self.private_key_path = private_key_path
+        self.private_key_passphrase = private_key_passphrase
         if init_ssh_client:
             self.create_connection(self.SSH_CONNECTION_ATTEMPTS)
 
     def exec(self, cmd: str, verify=True, timeout=90) -> SSHCommand:
+        if self.login != 'root':
+            cmd = f'sudo {cmd}'
         cmd_result = self._inner_exec(cmd, timeout)
         if verify:
             assert cmd_result.rc == 0, f'Non zero rc from command: "{cmd}"'
@@ -110,6 +114,15 @@ class HostClient:
         self.login, self.password = keep_user, keep_password
         self.create_connection()
 
+    @contextmanager
+    def create_ssh_connection(self) -> 'SSHClient':
+        if not self.ssh_client:
+            self.create_connection()
+        try:
+            yield self.ssh_client
+        finally:
+            self.drop()
+
     @allure.step('Restore connection')
     def restore_ssh_connection(self):
         retry_time, retry_count = self.TIMEOUT_RESTORE_CONNECTION
@@ -151,13 +164,13 @@ class HostClient:
             try:
                 if self.private_key_path:
                     logging.info(
-                        f"Trying to connect to host {self.ip} using SSH key "
+                        f"Trying to connect to host {self.ip} as {self.login} using SSH key "
                         f"{self.private_key_path} (attempt {attempt})"
                     )
                     self.ssh_client.connect(
                         hostname=self.ip,
                         username=self.login,
-                        pkey=RSAKey.from_private_key_file(self.private_key_path, self.password),
+                        pkey=RSAKey.from_private_key_file(self.private_key_path, self.private_key_passphrase),
                         timeout=self.CONNECTION_TIMEOUT
                     )
                 else:
@@ -175,7 +188,7 @@ class HostClient:
 
             except AuthenticationException as auth_err:
                 logging.error(f'Host: {self.ip}. {auth_err}')
-
+                self.drop()
                 raise auth_err
 
             except (
@@ -186,6 +199,7 @@ class HostClient:
                     OSError
             ) as ssh_err:
                 exc_err = ssh_err
+                self.drop()
                 logging.error(f'Host: {self.ip}, connection error. {exc_err}')
 
         raise HostIsNotAvailable(self.ip, exc_err)
@@ -197,8 +211,6 @@ class HostClient:
     def _inner_exec(self, cmd: str, timeout: int) -> SSHCommand:
         if not self.ssh_client:
             self.create_connection()
-        if self.login != "root":
-            cmd = f"sudo {cmd}"
         for _ in range(self.SSH_CONNECTION_ATTEMPTS):
             try:
                 _, stdout, stderr = self.ssh_client.exec_command(cmd, timeout=timeout)

pytest_tests/pytest.ini

@@ -18,4 +18,6 @@ markers =
     long: long tests (with long execution time)
     node_mgmt: neofs control commands
     acl: tests for basic and extended ACL
-    failover: tests for system recovery after a failure
+    failover: tests for system recovery after a failure
+    failover_panic: tests for system recovery after panic reboot of a node
+    failover_net: tests for network failure

pytest_tests/testsuites/conftest.py

@@ -23,6 +23,13 @@ deco.keyword = robot_keyword_adapter
 logger = logging.getLogger('NeoLogger')
 
 
+@pytest.fixture(scope='session')
+def free_storage_check():
+    if os.getenv('FREE_STORAGE', default='False').lower() not in ('true', '1'):
+        pytest.skip('Test only works on SberCloud infrastructure')
+    yield
+
+
 @pytest.fixture(scope='session', autouse=True)
 @allure.title('Check binary versions')
 def check_binary_versions(request):
@@ -81,6 +88,7 @@ def init_wallet_with_address(prepare_tmp_dir):
 def prepare_wallet_and_deposit(init_wallet_with_address):
     wallet, addr, _ = init_wallet_with_address
     logger.info(f'Init wallet: {wallet},\naddr: {addr}')
+    allure.attach.file(wallet, os.path.basename(wallet), allure.attachment_type.JSON)
 
     if not FREE_STORAGE:
         deposit = 30

pytest_tests/testsuites/failovers/failover_utils.py

@@ -0,0 +1,49 @@
+import logging
+from time import sleep
+from typing import Optional
+
+import allure
+
+from common import NEOFS_NETMAP_DICT
+from python_keywords.node_management import node_healthcheck
+from storage_policy import get_nodes_with_object
+
+logger = logging.getLogger('NeoLogger')
+
+
+@allure.step('Wait for object replication')
+def wait_object_replication_on_nodes(wallet: str, cid: str, oid: str, expected_copies: int,
+                                     excluded_nodes: Optional[list[str]] = None) -> list[str]:
+    excluded_nodes = excluded_nodes or []
+    sleep_interval, attempts = 10, 18
+    nodes = []
+    for __attempt in range(attempts):
+        nodes = get_nodes_with_object(wallet, cid, oid, skip_nodes=excluded_nodes)
+        if len(nodes) == expected_copies:
+            return nodes
+        sleep(sleep_interval)
+    raise AssertionError(f'Expected {expected_copies} copies of object, but found {len(nodes)}. '
+                         f'Waiting time {sleep_interval * attempts}')
+
+
+@allure.step('Wait for storage node returned to cluster')
+def wait_all_storage_node_returned():
+    sleep_interval, attempts = 10, 12
+    for __attempt in range(attempts):
+        if is_all_storage_node_returned():
+            return
+        sleep(sleep_interval)
+    raise AssertionError('Storage node(s) is broken')
+
+
+def is_all_storage_node_returned() -> bool:
+    with allure.step('Run health check for all storage nodes'):
+        for node_name in NEOFS_NETMAP_DICT.keys():
+            try:
+                health_check = node_healthcheck(node_name)
+            except Exception as err:
+                logger.warning(f'Node healthcheck fails with error {err}')
+                return False
+            if health_check.health_status != 'READY' or health_check.network_status != 'ONLINE':
+                return False
+    return True

pytest_tests/testsuites/failovers/test_failover_network.py

@@ -0,0 +1,118 @@
+import logging
+from random import choices
+from time import sleep
+
+import allure
+import pytest
+
+from common import STORAGE_NODE_SSH_PRIVATE_KEY_PATH, STORAGE_NODE_SSH_USER, STORAGE_NODE_SSH_PASSWORD, \
+    NEOFS_NETMAP_DICT
+from iptables_helper import IpTablesHelper
+from python_keywords.container import create_container
+from python_keywords.neofs_verbs import get_object, put_object
+from python_keywords.utility_keywords import generate_file, get_file_hash
+from ssh_helper import HostClient
+from wellknown_acl import PUBLIC_ACL
+from .failover_utils import wait_all_storage_node_returned, wait_object_replication_on_nodes
+
+logger = logging.getLogger('NeoLogger')
+STORAGE_NODE_COMMUNICATION_PORT = '8080'
+STORAGE_NODE_COMMUNICATION_PORT_TLS = '8082'
+PORTS_TO_BLOCK = [STORAGE_NODE_COMMUNICATION_PORT, STORAGE_NODE_COMMUNICATION_PORT_TLS]
+blocked_hosts = []
+
+
+@pytest.fixture(autouse=True)
+@allure.step('Install iptables if needed')
+def install_iptables_if_needed():
+    check_command = 'iptables --version'
+    install_command = 'apt-get --yes install iptables'
+    for node_config in NEOFS_NETMAP_DICT.values():
+        host = node_config.get('rpc').split(':')[0]
+        client = HostClient(ip=host, login=STORAGE_NODE_SSH_USER,
+                            password=STORAGE_NODE_SSH_PASSWORD,
+                            private_key_path=STORAGE_NODE_SSH_PRIVATE_KEY_PATH)
+        with client.create_ssh_connection():
+            try:
+                client.exec(check_command)
+            except AssertionError as err:
+                logger.info(f'Command {check_command} fails with error {err}')
+                client.exec(install_command)
+                client.exec(check_command)
+
+
+@pytest.fixture(autouse=True)
+@allure.step('Restore network')
+def restore_network():
+    yield
+
+    not_empty = len(blocked_hosts) != 0
+    for host in list(blocked_hosts):
+        with allure.step(f'Start storage node {host}'):
+            client = HostClient(ip=host, login=STORAGE_NODE_SSH_USER,
+                                password=STORAGE_NODE_SSH_PASSWORD,
+                                private_key_path=STORAGE_NODE_SSH_PRIVATE_KEY_PATH)
+            with client.create_ssh_connection():
+                IpTablesHelper.restore_input_traffic_to_port(client, PORTS_TO_BLOCK)
+        blocked_hosts.remove(host)
+    if not_empty:
+        wait_all_storage_node_returned()
+
+
+@allure.title('Block Storage node traffic')
+@pytest.mark.failover
+@pytest.mark.failover_net
+def test_block_storage_node_traffic(prepare_wallet_and_deposit, free_storage_check):
+    """
+    Block storage nodes traffic using iptables and wait for replication for objects.
+    """
+    wallet = prepare_wallet_and_deposit
+    placement_rule = 'REP 2 IN X CBF 2 SELECT 2 FROM * AS X'
+    excluded_nodes = []
+    wakeup_node_timeout = 10  # timeout to let nodes detect that traffic has blocked
+    nodes_to_block_count = 2
+
+    source_file_path = generate_file()
+    cid = create_container(wallet, rule=placement_rule, basic_acl=PUBLIC_ACL)
+    oid = put_object(wallet, source_file_path, cid)
+    nodes = wait_object_replication_on_nodes(wallet, cid, oid, 2)
+
+    logger.info(f'Nodes are {nodes}')
+    random_nodes = [(node, node.split(':')[0]) for node in nodes]
+    if nodes_to_block_count > len(nodes):
+        random_nodes = [(node, node.split(':')[0]) for node in choices(nodes, k=2)]
+
+    for random_node, random_node_ip in random_nodes:
+        client = HostClient(ip=random_node_ip, login=STORAGE_NODE_SSH_USER,
+                            password=STORAGE_NODE_SSH_PASSWORD,
+                            private_key_path=STORAGE_NODE_SSH_PRIVATE_KEY_PATH)
+
+        with allure.step(f'Block incoming traffic for node {random_node} on port {PORTS_TO_BLOCK}'):
+            with client.create_ssh_connection():
+                IpTablesHelper.drop_input_traffic_to_port(client, PORTS_TO_BLOCK)
+            blocked_hosts.append(random_node_ip)
+            excluded_nodes.append(random_node)
+            sleep(wakeup_node_timeout)
+
+        new_nodes = wait_object_replication_on_nodes(wallet, cid, oid, 2, excluded_nodes=excluded_nodes)
+
+        assert random_node not in new_nodes
+
+        got_file_path = get_object(wallet, cid, oid, endpoint=new_nodes[0])
+        assert get_file_hash(source_file_path) == get_file_hash(got_file_path)
+
+    for random_node, random_node_ip in random_nodes:
+        client = HostClient(ip=random_node_ip, login=STORAGE_NODE_SSH_USER,
+                            password=STORAGE_NODE_SSH_PASSWORD,
+                            private_key_path=STORAGE_NODE_SSH_PRIVATE_KEY_PATH)
+
+        with allure.step(f'Unblock incoming traffic for node {random_node} on port {PORTS_TO_BLOCK}'):
+            with client.create_ssh_connection():
+                IpTablesHelper.restore_input_traffic_to_port(client, PORTS_TO_BLOCK)
+            blocked_hosts.remove(random_node_ip)
+        sleep(wakeup_node_timeout)
+
+    new_nodes = wait_object_replication_on_nodes(wallet, cid, oid, 2)
+
+    got_file_path = get_object(wallet, cid, oid, endpoint=new_nodes[0])
+    assert get_file_hash(source_file_path) == get_file_hash(got_file_path)

pytest_tests/testsuites/failovers/test_failover_storage.py

@@ -1,91 +1,54 @@
 import logging
 import os
-from time import sleep
 
 import allure
 import pytest
-from common import NEOFS_NETMAP_DICT
+
+from common import STORAGE_NODE_SSH_PRIVATE_KEY_PATH, STORAGE_NODE_SSH_USER, STORAGE_NODE_SSH_PASSWORD
 from python_keywords.container import create_container
 from python_keywords.neofs_verbs import get_object, put_object
-from python_keywords.node_management import node_healthcheck
 from python_keywords.utility_keywords import generate_file, get_file_hash
 from sbercloud_helper import SberCloud
 from ssh_helper import HostClient, HostIsNotAvailable
-from storage_policy import get_nodes_with_object
 from wellknown_acl import PUBLIC_ACL
+from .failover_utils import wait_all_storage_node_returned, wait_object_replication_on_nodes
 
 logger = logging.getLogger('NeoLogger')
 stopped_hosts = []
 
 
-@pytest.fixture(scope='session')
-def free_storage_check():
-    if os.getenv('FREE_STORAGE', default='False').lower() not in ('true', '1'):
-        pytest.skip('Test only works on SberCloud infrastructure')
-    yield
-
-
 @pytest.fixture(scope='session')
 def sbercloud_client():
     with allure.step('Connect to SberCloud'):
         try:
             yield SberCloud(f'{os.getcwd()}/configuration/sbercloud.yaml')
-        except Exception:
-            pytest.fail('SberCloud infrastructure not available')
+        except Exception as err:
+            pytest.fail(f'SberCloud infrastructure not available. Error\n{err}')
 
 
 @pytest.fixture(autouse=True)
+@allure.step('Return all storage nodes')
 def return_all_storage_nodes_fixture(sbercloud_client):
     yield
     return_all_storage_nodes(sbercloud_client)
 
 
 def panic_reboot_host(ip: str = None):
-    ssh = HostClient(ip=ip, login="root", private_key_path=f"{os.getcwd()}/configuration/id_rsa")
-    ssh.exec('echo 1 > /proc/sys/kernel/sysrq')
+    ssh = HostClient(ip=ip, login=STORAGE_NODE_SSH_USER,
+                     password=STORAGE_NODE_SSH_PASSWORD,
+                     private_key_path=STORAGE_NODE_SSH_PRIVATE_KEY_PATH)
+    ssh.exec('sudo sh -c "echo 1 > /proc/sys/kernel/sysrq"')
     with pytest.raises(HostIsNotAvailable):
-        ssh.exec('echo b > /proc/sysrq-trigger', timeout=1)
+        ssh.exec('sudo sh -c "echo b > /proc/sysrq-trigger"', timeout=1)
 
 
 def return_all_storage_nodes(sbercloud_client: SberCloud):
-    for host in stopped_hosts:
+    for host in list(stopped_hosts):
         with allure.step(f'Start storage node {host}'):
             sbercloud_client.start_node(node_ip=host.split(':')[-2])
+        stopped_hosts.remove(host)
+
     wait_all_storage_node_returned()
-    stopped_hosts.clear()
-
-
-def is_all_storage_node_returned() -> bool:
-    with allure.step('Run health check for all storage nodes'):
-        for node_name in NEOFS_NETMAP_DICT.keys():
-            try:
-                health_check = node_healthcheck(node_name)
-            except (AssertionError, HostIsNotAvailable, TimeoutError):
-                return False
-            if health_check.health_status != 'READY' or health_check.network_status != 'ONLINE':
-                return False
-    return True
-
-
-def wait_all_storage_node_returned():
-    sleep_interval, attempts = 10, 12
-    for __attempt in range(attempts):
-        if is_all_storage_node_returned():
-            return
-        sleep(sleep_interval)
-    raise AssertionError('Storage node(s) is broken')
-
-
-def wait_object_replication(wallet, cid, oid, expected_copies: int, excluded_nodes: [str] = None) -> [str]:
-    excluded_nodes = excluded_nodes or []
-    sleep_interval, attempts = 10, 12
-    nodes = []
-    for __attempt in range(attempts):
-        nodes = [node for node in get_nodes_with_object(wallet, cid, oid) if node not in excluded_nodes]
-        if len(nodes) == expected_copies:
-            return nodes
-        sleep(sleep_interval)
-    raise AssertionError(f'Expected {expected_copies} copies of object, but found {len(nodes)} ')
 
 
 @allure.title('Lost and return nodes')
@@ -98,14 +61,14 @@ def test_lost_storage_node(prepare_wallet_and_deposit, sbercloud_client: SberClo
     source_file_path = generate_file()
     cid = create_container(wallet, rule=placement_rule, basic_acl=PUBLIC_ACL)
     oid = put_object(wallet, source_file_path, cid)
-    nodes = wait_object_replication(wallet, cid, oid, 2)
+    nodes = wait_object_replication_on_nodes(wallet, cid, oid, 2)
 
     new_nodes = []
     for node in nodes:
         stopped_hosts.append(node)
         with allure.step(f'Stop storage node {node}'):
             sbercloud_client.stop_node(node_ip=node.split(':')[-2], hard=hard_reboot)
-        new_nodes = wait_object_replication(wallet, cid, oid, 2, excluded_nodes=[node])
+        new_nodes = wait_object_replication_on_nodes(wallet, cid, oid, 2, excluded_nodes=[node])
 
     assert not [node for node in nodes if node in new_nodes]
     got_file_path = get_object(wallet, cid, oid, endpoint=new_nodes[0])
@@ -114,7 +77,7 @@ def test_lost_storage_node(prepare_wallet_and_deposit, sbercloud_client: SberClo
     with allure.step(f'Return storage nodes'):
         return_all_storage_nodes(sbercloud_client)
 
-    new_nodes = wait_object_replication(wallet, cid, oid, 2)
+    new_nodes = wait_object_replication_on_nodes(wallet, cid, oid, 2)
 
     got_file_path = get_object(wallet, cid, oid, endpoint=new_nodes[0])
     assert get_file_hash(source_file_path) == get_file_hash(got_file_path)
@@ -122,6 +85,7 @@ def test_lost_storage_node(prepare_wallet_and_deposit, sbercloud_client: SberClo
 
 @allure.title('Panic storage node(s)')
 @pytest.mark.parametrize('sequence', [True, False])
+@pytest.mark.failover_panic
 @pytest.mark.failover
 def test_panic_storage_node(prepare_wallet_and_deposit, free_storage_check, sequence: bool):
     wallet = prepare_wallet_and_deposit
@@ -130,18 +94,22 @@ def test_panic_storage_node(prepare_wallet_and_deposit, free_storage_check, sequ
     cid = create_container(wallet, rule=placement_rule, basic_acl=PUBLIC_ACL)
     oid = put_object(wallet, source_file_path, cid)
 
-    nodes = wait_object_replication(wallet, cid, oid, 2)
+    nodes = wait_object_replication_on_nodes(wallet, cid, oid, 2)
     allure.attach('\n'.join(nodes), 'Current nodes with object', allure.attachment_type.TEXT)
     for node in nodes:
         with allure.step(f'Hard reboot host {node} via magic SysRq option'):
             panic_reboot_host(ip=node.split(':')[-2])
             if sequence:
-                new_nodes = wait_object_replication(wallet, cid, oid, 2, excluded_nodes=[node])
+                try:
+                    new_nodes = wait_object_replication_on_nodes(wallet, cid, oid, 2, excluded_nodes=[node])
+                except AssertionError:
+                    new_nodes = wait_object_replication_on_nodes(wallet, cid, oid, 2)
+
                 allure.attach('\n'.join(new_nodes), f'Nodes with object after {node} fail',
-                                   allure.attachment_type.TEXT)
+                              allure.attachment_type.TEXT)
 
     if not sequence:
-        new_nodes = wait_object_replication(wallet, cid, oid, 2, excluded_nodes=nodes)
+        new_nodes = wait_object_replication_on_nodes(wallet, cid, oid, 2)
         allure.attach('\n'.join(new_nodes), 'Nodes with object after nodes fail', allure.attachment_type.TEXT)
 
     got_file_path = get_object(wallet, cid, oid, endpoint=new_nodes[0])

robot/resources/lib/python_keywords/storage_policy.py

@@ -5,6 +5,8 @@
     that storage policies are kept.
 """
 
+from typing import Optional
+
 from robot.api import logger
 from robot.api.deco import keyword
 
@@ -86,7 +88,7 @@ def get_complex_object_copies(wallet: str, cid: str, oid: str):
 
 
 @keyword('Get Nodes With Object')
-def get_nodes_with_object(wallet: str, cid: str, oid: str):
+def get_nodes_with_object(wallet: str, cid: str, oid: str, skip_nodes: Optional[list[str]] = None) -> list[str]:
     """
        The function returns list of nodes which store
        the given object.
@@ -95,11 +97,16 @@ def get_nodes_with_object(wallet: str, cid: str, oid: str):
                         we request the nodes
             cid (str): ID of the container which store the object
             oid (str): object ID
+            skip_nodes (list): list of nodes that should be excluded from check
        Returns:
             (list): nodes which store the object
     """
+    nodes_to_search = NEOFS_NETMAP
+    if skip_nodes:
+        nodes_to_search = [node for node in NEOFS_NETMAP if node not in skip_nodes]
+
     nodes_list = []
-    for node in NEOFS_NETMAP:
+    for node in nodes_to_search:
         try:
             res = neofs_verbs.head_object(wallet, cid, oid,
                                           endpoint=node,