0a6137218a
Draining hints may occur in one of the two scenarios: * a node leaves the cluster and the local node drains all of the hints saved for that node, * the local node is being decommissioned. Draining may take some time and the hint manager won't stop until it finishes. It's not a problem when decommissioning a node, especially because we want the cluster to retain the data stored in the hints. However, it may become a problem when the local node started draining hints saved for another node and now it's being shut down. There are two reasons for that: * Generally, in situations like that, we'd like to be able to shut down nodes as fast as possible. The data stored in the hints won't disappear from the cluster yet since we can restart the local node. * Draining hints may introduce flakiness in tests. Replaying hints doesn't have the highest priority and it's reflected in the scheduling groups we use as well as the explicitly enforced throughput. If there are a large number of hints to be replayed, it might affect our tests. It's already happened, see: scylladb/scylladb#21949. To solve those problems, we change the semantics of draining. It will behave as before when the local node is being decommissioned. However, when the local node is only being stopped, we will immediately cancel all ongoing draining processes and stop the hint manager. To amend for that, when we start a node and it initializes a hint endpoint manager corresponding to a node that's already left the cluster, we will begin the draining process of that endpoint manager right away. That should ensure all data is retained, while possibly speeding up the shutdown process. There's a small trade-off to it, though. If we stop a node, we can then remove it. It won't have a chance to replay hints it might've before these changes, but that's an edge case. We expect this commit to bring more benefit than harm. We also provide tests verifying that the implementation works as intended. Fixes scylladb/scylladb#21949 Closes scylladb/scylladb#22811
324 lines
14 KiB
Python
324 lines
14 KiB
Python
#
|
|
# Copyright (C) 2024-present ScyllaDB
|
|
#
|
|
# SPDX-License-Identifier: LicenseRef-ScyllaDB-Source-Available-1.0
|
|
#
|
|
import asyncio
|
|
import pytest
|
|
import time
|
|
import logging
|
|
import requests
|
|
import re
|
|
|
|
from cassandra.cluster import ConnectionException, NoHostAvailable # type: ignore
|
|
from cassandra.query import SimpleStatement, ConsistencyLevel
|
|
|
|
from test.pylib.internal_types import ServerInfo
|
|
from test.pylib.manager_client import ManagerClient
|
|
from test.pylib.rest_client import inject_error
|
|
from test.pylib.util import wait_for
|
|
|
|
from test.cluster.conftest import skip_mode
|
|
from test.cluster.util import get_topology_coordinator, find_server_by_host_id, new_test_keyspace
|
|
|
|
|
|
logger = logging.getLogger(__name__)
|
|
|
|
def get_hint_manager_metric(server: ServerInfo, metric_name: str) -> int:
|
|
result = 0
|
|
metrics = requests.get(f"http://{server.ip_addr}:9180/metrics").text
|
|
pattern = re.compile(f"^scylla_hints_manager_{metric_name}")
|
|
for metric in metrics.split('\n'):
|
|
if pattern.match(metric) is not None:
|
|
result += int(float(metric.split()[1]))
|
|
return result
|
|
|
|
# Creates a sync point for ALL hosts.
|
|
def create_sync_point(node: ServerInfo) -> str:
|
|
return requests.post(f"http://{node.ip_addr}:10000/hinted_handoff/sync_point/").json()
|
|
|
|
def await_sync_point(node: ServerInfo, sync_point: str, timeout: int) -> bool:
|
|
params = {
|
|
"id": sync_point,
|
|
"timeout": str(timeout)
|
|
}
|
|
|
|
response = requests.get(f"http://{node.ip_addr}:10000/hinted_handoff/sync_point", params=params).json()
|
|
match response:
|
|
case "IN_PROGRESS":
|
|
return False
|
|
case "DONE":
|
|
return True
|
|
case _:
|
|
pytest.fail(f"Unexpected response from the server: {response}")
|
|
|
|
# Write with RF=1 and CL=ANY to a dead node should write hints and succeed
|
|
@pytest.mark.asyncio
|
|
async def test_write_cl_any_to_dead_node_generates_hints(manager: ManagerClient):
|
|
node_count = 2
|
|
servers = await manager.servers_add(node_count)
|
|
|
|
cql = manager.get_cql()
|
|
async with new_test_keyspace(manager, "WITH replication = {'class': 'NetworkTopologyStrategy', 'replication_factor': 1}") as ks:
|
|
table = f"{ks}.t"
|
|
await cql.run_async(f"CREATE TABLE {table} (pk int primary key, v int)")
|
|
|
|
await manager.server_stop_gracefully(servers[1].server_id)
|
|
|
|
def get_hints_written_count(server):
|
|
return get_hint_manager_metric(server, "written")
|
|
|
|
hints_before = get_hints_written_count(servers[0])
|
|
|
|
# Some of the inserts will be targeted to the dead node.
|
|
# The coordinator doesn't have live targets to send the write to, but it should write a hint.
|
|
for i in range(100):
|
|
await cql.run_async(SimpleStatement(f"INSERT INTO {table} (pk, v) VALUES ({i}, {i+1})", consistency_level=ConsistencyLevel.ANY))
|
|
|
|
# Verify hints are written
|
|
hints_after = get_hints_written_count(servers[0])
|
|
assert hints_after > hints_before
|
|
|
|
# For dropping the keyspace
|
|
await manager.server_start(servers[1].server_id)
|
|
|
|
@pytest.mark.asyncio
|
|
async def test_limited_concurrency_of_writes(manager: ManagerClient):
|
|
"""
|
|
We want to verify that Scylla correctly limits the concurrency of writing hints to disk.
|
|
To do that, we leverage error injections decreasing the threshold when hints should start
|
|
being rejected, and we expect to receive an exception indicating that a node is overloaded.
|
|
"""
|
|
node1 = await manager.server_add(config={
|
|
"error_injections_at_startup": ["decrease_max_size_of_hints_in_progress"]
|
|
})
|
|
node2 = await manager.server_add()
|
|
|
|
cql = manager.get_cql()
|
|
async with new_test_keyspace(manager, "WITH replication = {'class': 'SimpleStrategy', 'replication_factor': 2}") as ks:
|
|
table = f"{ks}.t"
|
|
await cql.run_async(f"CREATE TABLE {table} (pk int primary key, v int)")
|
|
|
|
await manager.server_stop_gracefully(node2.server_id)
|
|
|
|
async with inject_error(manager.api, node1.ip_addr, "slow_down_writing_hints"):
|
|
try:
|
|
for i in range(100):
|
|
await cql.run_async(SimpleStatement(f"INSERT INTO {table} (pk, v) VALUES ({i}, {i})", consistency_level=ConsistencyLevel.ONE))
|
|
pytest.fail("The coordinator node has not been overloaded, which indiciates that the concurrency of writing hints is NOT limited")
|
|
except NoHostAvailable as e:
|
|
for _, err in e.errors.items():
|
|
assert err.summary == "Coordinator node overloaded" and re.match(r"Too many in flight hints: \d+", err.message)
|
|
|
|
# For dropping the keyspace
|
|
await manager.server_start(node2.server_id)
|
|
|
|
@pytest.mark.asyncio
|
|
async def test_sync_point(manager: ManagerClient):
|
|
"""
|
|
We want to verify that the sync point API is compliant with its design.
|
|
This test concerns one particular aspect of it: Scylla should create a sync point
|
|
for ALL nodes if the parameter `target_hosts` of a request is empty, not just
|
|
live nodes.
|
|
"""
|
|
node_count = 3
|
|
[node1, node2, node3] = await manager.servers_add(node_count)
|
|
|
|
cql = manager.get_cql()
|
|
async with new_test_keyspace(manager, "WITH replication = {'class': 'SimpleStrategy', 'replication_factor': 3}") as ks:
|
|
table = f"{ks}.t"
|
|
await cql.run_async(f"CREATE TABLE {table} (pk int primary key, v int)")
|
|
|
|
await manager.server_stop_gracefully(node2.server_id)
|
|
await manager.server_stop_gracefully(node3.server_id)
|
|
|
|
await manager.server_not_sees_other_server(node1.ip_addr, node2.ip_addr)
|
|
await manager.server_not_sees_other_server(node1.ip_addr, node3.ip_addr)
|
|
|
|
mutation_count = 5
|
|
for primary_key in range(mutation_count):
|
|
await cql.run_async(SimpleStatement(f"INSERT INTO {table} (pk, v) VALUES ({primary_key}, {primary_key})", consistency_level=ConsistencyLevel.ONE))
|
|
|
|
# Mutations need to be applied to hinted handoff's commitlog before we create the sync point.
|
|
# Otherwise, the sync point will correspond to no hints at all.
|
|
|
|
# We need to wrap the function in an async function to make `wait_for` be able to use it below.
|
|
async def check_no_hints_in_progress_node1() -> bool:
|
|
return get_hint_manager_metric(node1, "size_of_hints_in_progress") == 0
|
|
|
|
deadline = time.time() + 30
|
|
await wait_for(check_no_hints_in_progress_node1, deadline)
|
|
|
|
sync_point1 = create_sync_point(node1)
|
|
|
|
await manager.server_start(node2.server_id)
|
|
await manager.server_sees_other_server(node1.ip_addr, node2.ip_addr)
|
|
|
|
assert not await_sync_point(node1, sync_point1, 30)
|
|
|
|
await manager.server_start(node3.server_id)
|
|
await manager.server_sees_other_server(node1.ip_addr, node3.ip_addr)
|
|
|
|
assert await_sync_point(node1, sync_point1, 30)
|
|
|
|
|
|
@pytest.mark.asyncio
|
|
@skip_mode('release', "error injections aren't enabled in release mode")
|
|
async def test_hints_consistency_during_decommission(manager: ManagerClient):
|
|
"""
|
|
This test reproduces the failure observed in scylladb/scylla-dtest#4582
|
|
in a more reliable way than the test_hintedhandoff_decom dtest.
|
|
|
|
We want to make sure that data stored in hints will not get lost if hints replay
|
|
happens in parallel to streaming during decommission.
|
|
|
|
The test is vnodes-specific.
|
|
"""
|
|
(server1, server2, server3) = await manager.servers_add(3, config={
|
|
"error_injections_at_startup": ["decrease_hints_flush_period"]
|
|
})
|
|
cql = manager.cql
|
|
|
|
logger.info("Creatting a keyspace with RF=1 and a table")
|
|
async with new_test_keyspace(manager, "WITH replication = {'class': 'NetworkTopologyStrategy', 'replication_factor': 1} AND tablets = { 'enabled': false }") as ks:
|
|
table = f"{ks}.t"
|
|
await cql.run_async(f"CREATE TABLE {table} (pk int primary key, v int)")
|
|
|
|
logger.info("Stopping node 3")
|
|
await manager.server_stop_gracefully(server3.server_id)
|
|
await manager.others_not_see_server(server3.ip_addr)
|
|
|
|
# Write 100 rows with CL=ANY. Some of the rows will only be stored as hints because of RF=1
|
|
logger.info("Writing 100 rows with CL=ANY")
|
|
for i in range(100):
|
|
await cql.run_async(SimpleStatement(f"INSERT INTO {table} (pk, v) VALUES ({i}, {i + 1})", consistency_level=ConsistencyLevel.ANY))
|
|
|
|
# Temporarily pause hints replay, we will unpause it after decommission starts and streaming is done,
|
|
# but before switching to writing to new nodes
|
|
logger.info("Pause hints replay on nodes 1 and 2")
|
|
for srv in (server1, server2):
|
|
await manager.api.enable_injection(srv.ip_addr, "hinted_handoff_pause_hint_replay", one_shot=False)
|
|
|
|
# Start the node
|
|
logger.info("Start node 3")
|
|
await manager.server_start(server3.server_id)
|
|
await manager.servers_see_each_other([server1, server2, server3])
|
|
|
|
# Record the current position of hints so that we can wait for them later
|
|
sync_points = [create_sync_point(srv) for srv in (server1, server2)]
|
|
|
|
async with asyncio.TaskGroup() as tg:
|
|
coord = await get_topology_coordinator(manager)
|
|
coord_srv = await find_server_by_host_id(manager, [server1, server2, server3], coord)
|
|
|
|
# Make sure topology coordinator will pause right after streaming
|
|
logger.info("Enabling injection on the topology coordinator that will tell it to pause streaming")
|
|
await manager.api.enable_injection(coord_srv.ip_addr, "topology_coordinator_pause_after_streaming", one_shot=False)
|
|
coord_log = await manager.server_open_log(coord_srv.server_id)
|
|
coord_mark = await coord_log.mark()
|
|
|
|
# Start decommission - it will get stuck on error injection so do it in the background
|
|
logger.info("Starting decommission in the background")
|
|
decommission_result = tg.create_task(manager.decommission_node(server3.server_id))
|
|
|
|
# Wait until streaming ends
|
|
logger.info("Wait until decomission finishes streaming")
|
|
await coord_log.wait_for(f'decommissioning: streaming completed for node', from_mark=coord_mark)
|
|
|
|
# Now, unpause hints and let them be replayed
|
|
logger.info("Unpause hints replay on nodes 1 and 2")
|
|
for srv in (server1, server2):
|
|
await manager.api.disable_injection(srv.ip_addr, "hinted_handoff_pause_hint_replay")
|
|
|
|
logger.info("Wait until hints are replayed from nodes 1 and 2")
|
|
await asyncio.gather(*(asyncio.to_thread(await_sync_point, srv, pt, timeout=30) for srv, pt in zip((server1, server2), sync_points)))
|
|
|
|
# Unpause streaming and let decommission finish
|
|
logger.info("Unpause streaming")
|
|
await manager.api.disable_injection(coord_srv.ip_addr, "topology_coordinator_pause_after_streaming")
|
|
|
|
logger.info("Wait until decomission finishes")
|
|
await decommission_result
|
|
|
|
# Verify that no data has been lost - if the hints replay only sent the hints to the original destination (server3),
|
|
# then they will be only present on server3 which already left the cluster
|
|
logger.info("Verify that no data stored in hints have been lost")
|
|
for i in range(100):
|
|
assert list(await cql.run_async(f"SELECT v FROM {table} WHERE pk = {i}")) == [(i + 1,)]
|
|
|
|
@pytest.mark.asyncio
|
|
async def test_draining_hints(manager: ManagerClient):
|
|
"""
|
|
This test verifies that all hints are drained when a node is being decommissioned.
|
|
"""
|
|
|
|
s1, s2, _ = await manager.servers_add(3)
|
|
cql = manager.get_cql()
|
|
|
|
await manager.api.set_logger_level(s1.ip_addr, "hints_manager", "trace")
|
|
|
|
await cql.run_async("CREATE KEYSPACE ks WITH REPLICATION = {'class': 'SimpleStrategy', 'replication_factor': 3}")
|
|
await cql.run_async("CREATE TABLE ks.t (pk int PRIMARY KEY, v int)")
|
|
|
|
await manager.server_stop_gracefully(s2.server_id)
|
|
|
|
# Generate hints towards s2 on s1 with probability 1 - ((#nodes - 1) / #nodes)^1000 ~= 1.
|
|
for i in range(1000):
|
|
await cql.run_async(SimpleStatement(f"INSERT INTO ks.t (pk, v) VALUES ({i}, {i + 1})", consistency_level=ConsistencyLevel.ANY))
|
|
|
|
sync_point = create_sync_point(s1)
|
|
await manager.server_start(s2.server_id)
|
|
|
|
|
|
async def wait():
|
|
assert await_sync_point(s1, sync_point, 60)
|
|
|
|
async with asyncio.TaskGroup() as tg:
|
|
_ = tg.create_task(manager.decommission_node(s1.server_id, timeout=60))
|
|
_ = tg.create_task(wait())
|
|
|
|
@pytest.mark.asyncio
|
|
@skip_mode("release", "error injections are not supported in release mode")
|
|
async def test_canceling_hint_draining(manager: ManagerClient):
|
|
"""
|
|
This test verifies that draining hints is canceled as soon as we issue a shutdown,
|
|
but it's resumed after starting the node again.
|
|
"""
|
|
|
|
s1, s2, _ = await manager.servers_add(3)
|
|
cql = manager.get_cql()
|
|
|
|
host_id2 = await manager.get_host_id(s2.server_id)
|
|
|
|
await manager.api.set_logger_level(s1.ip_addr, "hints_manager", "trace")
|
|
|
|
await cql.run_async("CREATE KEYSPACE ks WITH REPLICATION = {'class': 'SimpleStrategy', 'replication_factor': 3}")
|
|
await cql.run_async("CREATE TABLE ks.t (pk int PRIMARY KEY, v int)")
|
|
|
|
await manager.server_stop_gracefully(s2.server_id)
|
|
|
|
# Generate hints towards s2 on s1 with probability 1 - ((#nodes - 1) / #nodes)^1000 ~= 1.
|
|
for i in range(1000):
|
|
await cql.run_async(SimpleStatement(f"INSERT INTO ks.t (pk, v) VALUES ({i}, {i + 1})", consistency_level=ConsistencyLevel.ANY))
|
|
|
|
sync_point = create_sync_point(s1)
|
|
|
|
await manager.api.enable_injection(s1.ip_addr, "hinted_handoff_pause_hint_replay", False, {})
|
|
await manager.remove_node(s1.server_id, s2.server_id)
|
|
await manager.server_stop_gracefully(s1.server_id)
|
|
|
|
s1_log = await manager.server_open_log(s1.server_id)
|
|
s1_mark = await s1_log.mark()
|
|
|
|
await manager.server_update_cmdline(s1.server_id, ["--logger-log-level", "hints_manager=trace"])
|
|
await manager.server_start(s1.server_id)
|
|
|
|
s1_log = await manager.server_open_log(s1.server_id)
|
|
|
|
# Make sure the node still knows about the decommissioned node and does start draining for it.
|
|
await s1_log.wait_for(f"Draining starts for {host_id2}", s1_mark)
|
|
|
|
# Make sure draining finishes successfully.
|
|
assert await_sync_point(s1, sync_point, 60)
|
|
await s1_log.wait_for(f"Removed hint directory for {host_id2}")
|