dragonfly/tools/cluster_mgr.py
mkaruza 6cddc4385e
fix(cluster_family): Cancel slot migration from incoming node on OOM
If applying command on incoming node will result in OOM (we overflow
max_memory_limit) we are closing migration and switch state to FATAL.

Signed-off-by: mkaruza <mario@dragonflydb.io>
2025-05-07 20:40:14 +02:00

631 lines
20 KiB
Python
Executable file

#!/usr/bin/env python3
import argparse
from argparse import RawTextHelpFormatter
import json
import math
from typing import Iterable, List
import redis
import subprocess
import time
"""
To install: pip install -r requirements.txt
"""
def die_with_err(err):
print("!!!", err)
exit(-1)
class Node:
def __init__(self, host, port):
self.id = ""
self.host = host
self.port = port
def update_id(node):
node.id = send_command(node, ["cluster", "myid"])
print(f"- ID {node.id}")
def __repr__(self):
return f"{self.host}:{self.port}/{self.id}"
def to_dict(self):
return {"id": self.id, "ip": self.host, "port": self.port}
class Master(Node):
def __init__(self, host, port):
Node.__init__(self, host, port)
self.replicas = []
def start_node(node, dragonfly_bin, threads):
f = open(f"/tmp/dfly.cluster.node.{node.port}.log", "w")
print(f"- Log file for node {node.port}: {f.name}")
subprocess.Popen(
[
f"{dragonfly_bin}",
f"--port={node.port}",
"--cluster_mode=yes",
f"--proactor_threads={threads}",
"--dbfilename=",
f"--logtostderr",
"--proactor_affinity_mode=off",
"--omit_basic_usage",
],
stderr=f,
)
def send_command(node, command, print_errors=True):
client = redis.Redis(decode_responses=True, host=node.host, port=node.port)
for i in range(0, 5):
try:
result = client.execute_command(*command)
return result
except Exception as e:
if print_errors:
print(e)
time.sleep(0.1 * i)
finally:
client.close()
if print_errors:
print(f"Unable to run command {command} against {node.host}:{node.port} after 5 attempts!")
return Exception()
class SlotRange:
def __init__(self, start, end):
assert start <= end
self.start = start
self.end = end
def to_dict(self):
return {"start": self.start, "end": self.end}
@classmethod
def from_dict(cls, d):
return cls(d["start"], d["end"])
def __repr__(self):
return f"({self.start}-{self.end})"
def merge(self, other: "SlotRange"):
if self.end + 1 == other.start:
self.end = other.end
return True
elif other.end + 1 == self.start:
self.start = other.start
return True
return False
def contains(self, slot_id):
return self.start <= slot_id <= self.end
def remove(self, slot_id):
assert self.contains(slot_id)
if self.start < self.end:
if slot_id == self.start:
return None, SlotRange(self.start + 1, self.end)
elif slot_id == self.end:
return SlotRange(self.start, self.end - 1), None
elif self.start < slot_id < self.end:
return SlotRange(self.start, slot_id - 1), SlotRange(slot_id + 1, self.end)
return None, None
# Custom JSON encoder to handle SlotRange objects
class ClusterConfigEncoder(json.JSONEncoder):
def default(self, obj):
if isinstance(obj, SlotRange) or isinstance(obj, Node):
return obj.to_dict()
return super().default(obj)
def build_config_from_list(masters: List[Master]):
total_slots = 16384
slots_per_node = math.floor(total_slots / len(masters))
config = []
for i, master in enumerate(masters):
slot_range = SlotRange(i * slots_per_node, (i + 1) * slots_per_node - 1)
c = {
"slot_ranges": [slot_range],
"master": master,
"replicas": master.replicas,
}
config.append(c)
# Adjust the last slot range to include any remaining slots
config[-1]["slot_ranges"][-1].end += total_slots % len(masters)
return config
def get_nodes_from_config(config):
nodes = []
for shard in config:
nodes.append(shard["master"])
for replica in shard["replicas"]:
nodes.append(replica)
for node in nodes:
node.update_id()
return nodes
def push_config(config):
def push_to_node(node, config):
# Use the custom encoder to convert SlotRange objects during serialization
config_str = json.dumps(config, indent=2, cls=ClusterConfigEncoder)
response = send_command(node, ["dflycluster", "config", config_str])
print(f"- Push to {node.port}: {response}")
for node in get_nodes_from_config(config):
push_to_node(node, config)
def create_locally(args):
print(f"Setting up a Dragonfly cluster:")
print(f"- Master nodes: {args.num_masters}")
print(f"- Ports: {args.first_port}...{args.first_port + args.num_masters - 1}")
print(f"- Replicas for each master: {args.replicas_per_master}")
print()
next_port = args.first_port
masters = []
for i in range(args.num_masters):
master = Master("127.0.0.1", next_port)
next_port += 1
for j in range(args.replicas_per_master):
replica = Node("127.0.0.1", next_port)
master.replicas.append(replica)
next_port += 1
masters.append(master)
nodes = []
for master in masters:
nodes.append(master)
for replica in master.replicas:
nodes.append(replica)
print("Starting nodes...")
for node in nodes:
start_node(node, args.dragonfly_bin, args.threads)
print()
time.sleep(0.5)
if args.replicas_per_master > 0:
print("Configuring replication...")
for master in masters:
for replica in master.replicas:
response = send_command(replica, ["replicaof", master.host, master.port])
print(f"- {replica.port} replicating {master.port}: {response}")
print()
print(f"Getting IDs...")
for n in nodes:
n.update_id()
print()
config = build_config_from_list(masters)
print(f"Pushing config:\n{config}\n")
push_config(config)
print()
def config_single_remote(args):
print(
f"Configuring remote Dragonfly {args.target_host}:{args.target_port} to be a single-server cluster"
)
master = Master(args.target_host, args.target_port)
master.update_id()
test = send_command(master, ["get", "x"], print_errors=False)
if type(test) is not Exception:
die_with_err("Node either not found or already configured")
config = build_config_from_list([master])
print(f"Pushing config:\n{config}\n")
push_config(config)
print()
def build_config_from_existing(args):
def list_to_dict(l):
return {l[i]: l[i + 1] for i in range(0, len(l), 2)}
def build_node(node_list):
d = list_to_dict(node_list)
node = Node(d["endpoint"], d["port"])
node.id = d["id"]
return node
def build_slots(slot_list):
slots = []
for i in range(0, len(slot_list), 2):
slots.append(SlotRange(slot_list[i], slot_list[i + 1]))
return slots
client = redis.Redis(decode_responses=True, host=args.target_host, port=args.target_port)
existing = client.execute_command("cluster", "shards")
config = []
for shard_list in existing:
shard = list_to_dict(shard_list)
config.append(
{
"slot_ranges": build_slots(shard["slots"]),
"master": build_node(shard["nodes"][0]),
"replicas": [build_node(replica) for replica in shard["nodes"][1::]],
}
)
client.close()
return config
def find_master(config, host, port, die_if_not_found=True):
new_owner = None
for shard in config:
if shard["master"].host == host and shard["master"].port == port:
new_owner = shard
break
if new_owner == None and die_if_not_found:
die_with_err(f"Can't find master (hint: use flag --target_host / --target_port).")
return new_owner
def find_replica(config, host, port):
for shard in config:
for replica in shard["replicas"]:
if replica.host == host and replica.port == port:
return replica, shard
die_with_err("Can't find target node")
def attach(args):
print(f"Attaching remote Dragonfly {args.attach_host}:{args.attach_port} to cluster")
if args.attach_as_replica:
newcomer = Node(args.attach_host, args.attach_port)
replica_resp = send_command(newcomer, ["info", "replication"])
if replica_resp["role"] != "slave":
die_with_err("Node is not in replica mode")
if (
replica_resp["master_host"] != args.target_host
or replica_resp["master_port"] != args.target_port
):
die_with_err("Node is not a replica of target")
newcomer.update_id()
config = build_config_from_existing(args)
master_node = find_master(config, args.target_host, args.target_port)
master_node["replicas"].append(newcomer)
print(f"Pushing config:\n{config}\n")
push_config(config)
else:
newcomer = Master(args.attach_host, args.attach_port)
replica_resp = send_command(newcomer, ["info", "replication"])
if replica_resp["role"] != "master":
die_with_err("Node is not in master mode")
newcomer.update_id()
newcomer_config = build_config_from_list([newcomer])
newcomer_config[0]["slot_ranges"] = []
config = build_config_from_existing(args)
print(f"Pushing config:\n{config}\n")
push_config([*config, newcomer_config[0]])
print()
def detach(args):
print(f"Detaching remote Dragonfly {args.target_host}:{args.target_port} from cluster")
print(
"Important: detached node will not receive a new config! This means that the detached node will still 'think' that it belongs to the cluster"
)
config = build_config_from_existing(args)
node = find_master(config, args.target_host, args.target_port, die_if_not_found=False)
if node == None:
replica, master = find_replica(config, args.target_host, args.target_port)
master["replicas"].remove(replica)
else:
if len(node["slot_ranges"]) != 0:
die_with_err("Can't detach a master with assigned slots")
if len(node["replicas"]) != 0:
die_with_err("Can't detach a master with replicas")
config = [m for m in config if m != node]
push_config(config)
def takeover(args):
print(f"Promoting Dragonfly {args.target_host}:{args.target_port} from replica to master")
print(
"Important: do not forget to send command REPLICAOF NO ONE to new master, and update "
" additional replicas if such exist"
)
print("Important: previous master will be detached from the cluster")
config = build_config_from_existing(args)
replica, master = find_replica(config, args.target_host, args.target_port)
master["replicas"].remove(replica)
master["master"] = replica
push_config(config)
def move(args):
config = build_config_from_existing(args)
new_owner = find_master(config, args.target_host, args.target_port)
def remove_slot(slot_id, from_range: SlotRange, slot_ranges: list):
slot_ranges.remove(from_range)
left, right = from_range.remove(slot_id)
if left:
slot_ranges.append(left)
if right:
slot_ranges.append(right)
def add_slot(slot, to_shard):
slot_range = SlotRange(slot, slot)
for existing_range in to_shard["slot_ranges"]:
if existing_range.merge(slot_range):
return
to_shard["slot_ranges"].append(slot_range)
def find_slot(slot, config):
for shard in config:
for slot_range in shard["slot_ranges"]:
if slot_range.contains(slot):
return shard, slot_range
return None, None
def pack(slot_ranges):
slot_objects = sorted(slot_ranges, key=lambda x: x.start)
packed = []
for slot_range in slot_objects:
if packed and packed[-1].merge(slot_range):
continue
packed.append(slot_range)
return packed
for slot in range(args.slot_start, args.slot_end + 1):
shard, slot_range = find_slot(slot, config)
if shard == None or shard == new_owner:
continue
remove_slot(slot, slot_range, shard["slot_ranges"])
add_slot(slot, new_owner)
for shard in config:
shard["slot_ranges"] = pack(shard["slot_ranges"])
# Use the custom encoder for printing the JSON
print(f"Pushing new config:\n{json.dumps(config, indent=2, cls=ClusterConfigEncoder)}\n")
push_config(config)
def migrate(args):
config = build_config_from_existing(args)
target = find_master(config, args.target_host, args.target_port)
target_node = target["master"]
target_node.update_id()
# Find source node
source = None
for node in config:
slots: Iterable[SlotRange] = node["slot_ranges"]
for slot in slots:
if slot.start <= args.slot_start and slot.end >= args.slot_end:
source = node
break
if source == None:
die_with_err("Unsupported slot range migration (currently only 1-node migration supported)")
source["migrations"] = [
{
"slot_ranges": [{"start": args.slot_start, "end": args.slot_end}],
"node_id": target_node.id,
"ip": target_node.host,
"port": target_node.port,
}
]
push_config(config)
# wait for migration finish
sync_status = []
while True:
sync_status = send_command(target_node, ["DFLYCLUSTER", "SLOT-MIGRATION-STATUS"])
if len(sync_status) == 0:
# Migration didn't start yet
continue
if len(sync_status) != 1:
die_with_err(f"Unexpected number of migrations {len(sync_status)}: {sync_status}")
if "FATAL" in sync_status[0]:
die_with_err(f"Error in migration {len(sync_status)}: {sync_status}")
if "FINISHED" in sync_status[0]:
print(f"Migration finished: {sync_status[0]}")
break
# Push new config to all nodes
print("Updating all nodes with new slots state")
move(args)
def populate(args):
config = build_config_from_existing(args)
for shard in config:
master = shard["master"]
slot_ranges = shard["slot_ranges"]
for slot_range in slot_ranges:
cmd = [
"debug",
"populate",
str(args.size),
"key",
str(args.valsize),
"SLOTS",
str(slot_range.start),
str(slot_range.end),
]
send_command(master, cmd)
def print_config(args):
config = build_config_from_existing(args)
print(json.dumps(config, indent=2, cls=ClusterConfigEncoder))
def shutdown(args):
config = build_config_from_existing(args)
for node in get_nodes_from_config(config):
send_command(node, ["shutdown"])
def main():
parser = argparse.ArgumentParser(
description="""
Dragonfly Manual Cluster Manager
This tool helps managing a Dragonfly cluster manually.
Cluster can either be local or remote:
- Starting Dragonfly instances must be done locally, binary path can be set with `--dragonfly_bin` (default: ../build-opt/dragonfly)
- Remote Dragonflies must already be started, and initialized with `--cluster_mode=yes`
Example usage:
Create a 3 node cluster locally:
./cluster_mgr.py --action=create_locally --num_masters=3
This will create 3 Dragonfly processes with ports 7001-7003.
Ports can be overridden with `--first_port`.
Create a 6 node cluster locally, 3 of them masters with 1 replica each:
./cluster_mgr.py --action=create_locally --num_masters=3 --replicas_per_master=1
Connect to existing cluster and print current config:
./cluster_mgr.py --action=print_config
This will connect to 127.0.0.1:6379 by default. Override with `--target_host` and `--target_port`
Configure an existing Dragonfly server to be a standalone cluster (owning all slots):
./cluster_mgr.py --action=config_single_remote
This connects to an *existing* Dragonfly server, and pushes a config telling it to own all slots.
This will connect to 127.0.0.1:6379 by default. Override with `--target_host` and `--target_port`
Attach an existing Dragonfly server to an existing cluster (owning no slots):
./cluster_mgr.py --action=attach --attach_host=HOST --attach_port=PORT
This will connect to existing cluster present at 127.0.0.1:6379 by default. Override with
`--target_host` and `--target_port`.
To attach node as a replica - use --attach_as_replica=True. In such case, the node will be a
replica of --target_host/--target_port.
To set up a new cluster - start the servers and then use
./cluster_mgr.py --action=config_single_remote ...
./cluster_mgr.py --action=attach ...
And repeat `--action=attach` for all servers.
Afterwards, distribute the slots between the servers as desired with `--action=move` or
`--action=migrate`.
To detach (remove) a node from the cluster:
./cluster_mgr.py --action=detach --target_host=X --target_port=X
Notes:
- If the node is a master, it must not have any slots assigned to it.
- The node will not be notified that it's no longer in a cluster. It's a good idea to shut it down
after detaching it from the cluster.
To take over (turn replica to master):
./cluster_mgr.py --action=takeover --target_host=X --target_port=X
Notes:
- You'll need to run REPLICAOF NO ONE on the new master
- If previous master had other replicas, you'll need to update them with REPLICAOF as well
- Previous master will be detached from cluster. It's a good idea to shut it down.
Connect to cluster and move slots 10-20 to target:
./cluster_mgr.py --action=move --slot_start=10 --slot_end=20 --target_host=X --target_port=X
WARNING: This will NOT migrate existing data, i.e. data in slots 10-20 will be erased.
Migrate slots 10-20 to target:
./cluster_mgr.py --action=migrate --slot_start=10 --slot_end=20 --target_host=X --target_port=X
Unlike --action=move above, this will migrate the data to the new owner.
Connect to cluster and shutdown all nodes:
./cluster_mgr.py --action=shutdown --target_port=X
WARNING: Be careful! This will close all Dragonfly servers connected to the cluster.
""",
formatter_class=RawTextHelpFormatter,
)
parser.add_argument(
"--action",
default="",
help="Which action to take? See `--help`",
)
parser.add_argument(
"--num_masters", type=int, default=3, help="Number of master nodes in cluster"
)
parser.add_argument(
"--replicas_per_master", type=int, default=0, help="How many replicas for each master"
)
parser.add_argument("--first_port", type=int, default=7001, help="First master's port")
parser.add_argument("--threads", type=int, default=2, help="Threads per node")
parser.add_argument(
"--slot_start", type=int, default=0, help="First slot to move / migrate (inclusive)"
)
parser.add_argument(
"--slot_end", type=int, default=100, help="Last slot to move / migrate (inclusive)"
)
parser.add_argument("--target_host", default="127.0.0.1", help="Master host/ip")
parser.add_argument("--target_port", type=int, default=6379, help="Master port")
parser.add_argument(
"--attach_host", default="127.0.0.1", help="New cluster node master host/ip"
)
parser.add_argument(
"--attach_port", type=int, default=6379, help="New cluster node master port"
)
parser.add_argument(
"--attach_as_replica", type=bool, default=False, help="Is the attached node a replica?"
)
parser.add_argument(
"--dragonfly_bin", default="../build-opt/dragonfly", help="Dragonfly binary path"
)
parser.add_argument(
"--size", type=int, default=1000000, help="Number of keys to populate in each slotrange"
)
parser.add_argument(
"--valsize", type=int, default=16, help="Value size for each key during population"
)
args = parser.parse_args()
actions = dict(
[
(f.__name__, f)
for f in [
create_locally,
shutdown,
config_single_remote,
attach,
detach,
takeover,
move,
print_config,
migrate,
populate,
]
]
)
action = actions.get(args.action.lower())
if action:
action(args)
else:
die_with_err(f'Error - unknown action "{args.action}". See --help')
if __name__ == "__main__":
main()