Files
qdrant/tests/consensus_tests/test_shard_stream_transfer.py
Tim Visée c63f9eef29 Fix stream records transfer data race, losing pending updates (#8103)
* Add integration test to assert all queued updates are also transferred

* Plunge update queue in stream records transfers

* Migrate existing plunger usages to new plunge helper

* Skip test if not compiled with staging flag

* Only send delay operation when staging feature is enabled

* Reformat

* Fix review remarks
2026-02-11 17:18:44 +01:00

481 lines
18 KiB
Python

import multiprocessing
import pathlib
import random
from time import sleep
from .fixtures import upsert_random_points, create_collection
from .utils import *
N_PEERS = 3
N_SHARDS = 3
N_REPLICA = 1
COLLECTION_NAME = "test_collection"
def update_points_in_loop(peer_url, collection_name, offset=0, throttle=False, duration=None):
start = time.time()
limit = 3
while True:
upsert_random_points(peer_url, limit, collection_name, offset=offset)
offset += limit
if throttle:
sleep(0.1)
if duration is not None and (time.time() - start) > duration:
break
def run_update_points_in_background(peer_url, collection_name, num_points=None, num_cities=None, shard_key=None, init_offset=0, throttle=False, duration=None):
p = multiprocessing.Process(target=update_points_in_loop, args=(peer_url, collection_name, num_points, num_cities, shard_key, init_offset, throttle, duration))
p.start()
return p
def check_data_consistency(data):
assert(len(data) > 1)
for i in range(len(data) - 1):
j = i + 1
data_i = data[i]
data_j = data[j]
if data_i != data_j:
ids_i = set(x["id"] for x in data_i["points"])
ids_j = set(x["id"] for x in data_j["points"])
diff = ids_i - ids_j
if len(diff) < 100:
print(f"Diff between {i} and {j}: {diff}")
else:
print(f"Diff len between {i} and {j}: {len(diff)}")
assert False, "Data on all nodes should be consistent"
# Transfer shards from one node to another
#
# Simply does the most basic transfer: no concurrent updates during the
# transfer.
#
# Test that data on the both sides is consistent
def test_shard_stream_transfer(tmp_path: pathlib.Path):
assert_project_root()
# seed port to reuse the same port for the restarted nodes
peer_api_uris, peer_dirs, bootstrap_uri = start_cluster(tmp_path, N_PEERS, 20000)
create_collection(peer_api_uris[0], shard_number=N_SHARDS, replication_factor=N_REPLICA)
wait_collection_exists_and_active_on_all_peers(
collection_name=COLLECTION_NAME,
peer_api_uris=peer_api_uris
)
# Insert some initial number of points
upsert_random_points(peer_api_uris[0], 100)
transfer_collection_cluster_info = get_collection_cluster_info(peer_api_uris[0], COLLECTION_NAME)
receiver_collection_cluster_info = get_collection_cluster_info(peer_api_uris[2], COLLECTION_NAME)
from_peer_id = transfer_collection_cluster_info['peer_id']
to_peer_id = receiver_collection_cluster_info['peer_id']
shard_id = transfer_collection_cluster_info['local_shards'][0]['shard_id']
# Transfer shard from one node to another
# Move shard `shard_id` to peer `target_peer_id`
r = requests.post(
f"{peer_api_uris[0]}/collections/{COLLECTION_NAME}/cluster", json={
"replicate_shard": {
"shard_id": shard_id,
"from_peer_id": from_peer_id,
"to_peer_id": to_peer_id,
"method": "stream_records",
}
})
assert_http_ok(r)
# Wait for end of shard transfer
wait_for_collection_shard_transfers_count(peer_api_uris[0], COLLECTION_NAME, 0)
receiver_collection_cluster_info = get_collection_cluster_info(peer_api_uris[2], COLLECTION_NAME)
number_local_shards = len(receiver_collection_cluster_info['local_shards'])
assert number_local_shards == 2
# Point counts must be consistent across nodes
counts = []
for uri in peer_api_uris:
r = requests.post(
f"{uri}/collections/{COLLECTION_NAME}/points/count", json={
"exact": True
}
)
assert_http_ok(r)
counts.append(r.json()["result"]['count'])
assert counts[0] == counts[1] == counts[2]
# Transfer shards from one node to another while applying throttled updates in parallel
#
# Updates are throttled to prevent sending updates faster than the queue proxy
# can handle. The transfer must therefore finish in 30 seconds without issues.
#
# Test that data on the both sides is consistent
def test_shard_stream_transfer_throttled_updates(tmp_path: pathlib.Path):
assert_project_root()
# seed port to reuse the same port for the restarted nodes
peer_api_uris, peer_dirs, bootstrap_uri = start_cluster(tmp_path, N_PEERS, 20000)
create_collection(peer_api_uris[0], shard_number=N_SHARDS, replication_factor=N_REPLICA)
wait_collection_exists_and_active_on_all_peers(
collection_name=COLLECTION_NAME,
peer_api_uris=peer_api_uris
)
# Insert some initial number of points
upsert_random_points(peer_api_uris[0], 10000)
# Start pushing points to the cluster
upload_process_1 = run_update_points_in_background(peer_api_uris[0], COLLECTION_NAME, init_offset=100, throttle=True)
upload_process_2 = run_update_points_in_background(peer_api_uris[1], COLLECTION_NAME, init_offset=10000, throttle=True)
upload_process_3 = run_update_points_in_background(peer_api_uris[2], COLLECTION_NAME, init_offset=20000, throttle=True)
transfer_collection_cluster_info = get_collection_cluster_info(peer_api_uris[0], COLLECTION_NAME)
receiver_collection_cluster_info = get_collection_cluster_info(peer_api_uris[2], COLLECTION_NAME)
from_peer_id = transfer_collection_cluster_info['peer_id']
to_peer_id = receiver_collection_cluster_info['peer_id']
shard_id = transfer_collection_cluster_info['local_shards'][0]['shard_id']
# Transfer shard from one node to another
# Move shard `shard_id` to peer `target_peer_id`
r = requests.post(
f"{peer_api_uris[0]}/collections/{COLLECTION_NAME}/cluster", json={
"replicate_shard": {
"shard_id": shard_id,
"from_peer_id": from_peer_id,
"to_peer_id": to_peer_id,
"method": "stream_records",
}
})
assert_http_ok(r)
# Wait for end of shard transfer
wait_for_collection_shard_transfers_count(peer_api_uris[0], COLLECTION_NAME, 0)
upload_process_1.kill()
upload_process_2.kill()
upload_process_3.kill()
sleep(1)
receiver_collection_cluster_info = get_collection_cluster_info(peer_api_uris[2], COLLECTION_NAME)
number_local_shards = len(receiver_collection_cluster_info['local_shards'])
assert number_local_shards == 2
# Point counts must be consistent across nodes
counts = []
for uri in peer_api_uris:
r = requests.post(
f"{uri}/collections/{COLLECTION_NAME}/points/count", json={
"exact": True
}
)
assert_http_ok(r)
counts.append(r.json()["result"]['count'])
assert counts[0] == counts[1] == counts[2]
# Transfer shards from one node to another while applying updates in parallel
#
# A fast burst of updates is sent in the first 5 seconds, the queue proxy will
# not be able to keep up with this. After that, updates are throttled. The
# transfer must still finish in 30 seconds without issues.
#
# Test that data on the both sides is consistent
def test_shard_stream_transfer_fast_burst(tmp_path: pathlib.Path):
assert_project_root()
# seed port to reuse the same port for the restarted nodes
peer_api_uris, peer_dirs, bootstrap_uri = start_cluster(tmp_path, N_PEERS, 20000)
create_collection(peer_api_uris[0], shard_number=N_SHARDS, replication_factor=N_REPLICA)
wait_collection_exists_and_active_on_all_peers(
collection_name=COLLECTION_NAME,
peer_api_uris=peer_api_uris
)
# Insert some initial number of points
upsert_random_points(peer_api_uris[0], 10000)
# Start pushing points to the cluster
upload_process_1 = run_update_points_in_background(peer_api_uris[0], COLLECTION_NAME, init_offset=100, duration=5)
upload_process_2 = run_update_points_in_background(peer_api_uris[1], COLLECTION_NAME, init_offset=10000, duration=5)
upload_process_3 = run_update_points_in_background(peer_api_uris[2], COLLECTION_NAME, init_offset=20000, throttle=True)
transfer_collection_cluster_info = get_collection_cluster_info(peer_api_uris[0], COLLECTION_NAME)
receiver_collection_cluster_info = get_collection_cluster_info(peer_api_uris[2], COLLECTION_NAME)
from_peer_id = transfer_collection_cluster_info['peer_id']
to_peer_id = receiver_collection_cluster_info['peer_id']
shard_id = transfer_collection_cluster_info['local_shards'][0]['shard_id']
# Transfer shard from one node to another
# Move shard `shard_id` to peer `target_peer_id`
r = requests.post(
f"{peer_api_uris[0]}/collections/{COLLECTION_NAME}/cluster", json={
"replicate_shard": {
"shard_id": shard_id,
"from_peer_id": from_peer_id,
"to_peer_id": to_peer_id,
"method": "stream_records",
}
})
assert_http_ok(r)
# Wait for end of shard transfer
wait_for_collection_shard_transfers_count(peer_api_uris[0], COLLECTION_NAME, 0)
upload_process_1.kill()
upload_process_2.kill()
upload_process_3.kill()
sleep(1)
receiver_collection_cluster_info = get_collection_cluster_info(peer_api_uris[2], COLLECTION_NAME)
number_local_shards = len(receiver_collection_cluster_info['local_shards'])
assert number_local_shards == 2
# Point counts must be consistent across nodes
counts = []
for uri in peer_api_uris:
r = requests.post(
f"{uri}/collections/{COLLECTION_NAME}/points/count", json={
"exact": True
}
)
assert_http_ok(r)
counts.append(r.json()["result"]['count'])
if counts[0] != counts[1] or counts[1] != counts[2]:
data = []
for uri in peer_api_uris:
r = requests.post(
f"{uri}/collections/{COLLECTION_NAME}/points/scroll", json={
"limit": 999999999,
"with_vectors": False,
"with_payload": False,
}
)
assert_http_ok(r)
data.append(r.json()["result"])
check_data_consistency(data)
# Move one replica between nodes a few times, update pending point and ensure
# completion
#
# During the shard move we keep sending an operation that changes the very last
# point in the collection. More specifically, we invoke the set payload
# operation on the very last point.
#
# While the transfer is ongoing, the target replica might not have received the
# last point yet. The set payload operation must not mark the target replica as
# dead because it cannot find the point, as point may be transferred later as
# part of the transfer.
#
# If the test runs successfully, we see that the replica is properly moved. If the
# test does not run successfully, the point not found errors cascade into
# cancelling the shard transfers, and the replica will stay on the source node.
#
# Tests bug: <https://github.com/qdrant/qdrant/pull/5991>
#
# Updates are throttled to prevent overloading the cluster during the test.
def test_transfer_change_pending_point(tmp_path: pathlib.Path):
assert_project_root()
N_TRANSFERS = 5
N_POINTS = 5000
def set_payload_in_loop(peer_url, point_id):
while True:
r = requests.post(
f"{peer_url}/collections/{COLLECTION_NAME}/points/payload?wait=true",
json={
"points": [point_id],
"payload": {"n": random.random()},
},
)
assert_http_ok(r)
sleep(0.1)
def run_set_payload_in_background(peer_url, point_id):
p = multiprocessing.Process(target=set_payload_in_loop, args=(peer_url, point_id))
p.start()
return p
# seed port to reuse the same port for the restarted nodes
peer_api_uris, peer_dirs, bootstrap_uri = start_cluster(tmp_path, N_PEERS, 20000)
create_collection(peer_api_uris[0], shard_number=1, replication_factor=1)
wait_collection_exists_and_active_on_all_peers(
collection_name=COLLECTION_NAME,
peer_api_uris=peer_api_uris
)
# Insert some initial number of points
upsert_random_points(peer_api_uris[0], N_POINTS)
# Start pushing points to the cluster
update_process_1 = run_update_points_in_background(peer_api_uris[0], COLLECTION_NAME, init_offset=100, throttle=True)
update_process_2 = run_set_payload_in_background(peer_api_uris[1], point_id=N_POINTS - 1)
update_process_3 = run_set_payload_in_background(peer_api_uris[2], point_id=N_POINTS - 1)
# Move the shard a few times as it may not always trigger
for _ in range(0, N_TRANSFERS):
# Find what peer shard is currently on, select next peer as target
from_index = None
for index, uri in enumerate(peer_api_uris):
collection_cluster_info = get_collection_cluster_info(uri, COLLECTION_NAME)
if len(collection_cluster_info['local_shards']) == 1:
from_index = index
break
assert from_index is not None, "No shard found"
to_index = (from_index + 1) % len(peer_api_uris)
# Grab peer IDs
from_peer_uri = peer_api_uris[from_index]
to_peer_uri = peer_api_uris[to_index]
from_info = get_collection_cluster_info(from_peer_uri, COLLECTION_NAME)
to_info = get_collection_cluster_info(to_peer_uri, COLLECTION_NAME)
from_peer_id = from_info['peer_id']
to_peer_id = to_info['peer_id']
# Move shard from source to target peer
r = requests.post(
f"{from_peer_uri}/collections/{COLLECTION_NAME}/cluster", json={
"move_shard": {
"shard_id": 0,
"from_peer_id": from_peer_id,
"to_peer_id": to_peer_id,
"method": "stream_records",
}
})
assert_http_ok(r)
# Wait both peers to have recognized end of transfer
wait_for_collection_shard_transfers_count(from_peer_uri, COLLECTION_NAME, 0)
wait_for_collection_shard_transfers_count(to_peer_uri, COLLECTION_NAME, 0)
# Assert shard is actually moved
# If the transfer succeeded, the replica is moved to the other node
# If the transfer failed due to a missing point error, the replica is not moved
from_info = get_collection_cluster_info(from_peer_uri, COLLECTION_NAME)
to_info = get_collection_cluster_info(to_peer_uri, COLLECTION_NAME)
assert len(from_info['local_shards']) == 0, "Shard must be moved off of source peer"
assert len(to_info['local_shards']) == 1, "Shard must be moved onto target peer"
# Cleanup
update_process_1.kill()
update_process_2.kill()
update_process_3.kill()
# Transfer a shard, and assert pending operations are also transferred
#
# When we start the transfer, there may still be a (large) number of operations
# pending. Such pending operations may be in the update queue (channel) until
# the update worker applies them. In Qdrant <1.17.0 these operations were
# missed. This test proves all operations in the update queue are properly
# transferred to the target node, and applied there.
#
# See: <https://github.com/qdrant/qdrant/pull/8103>
def test_shard_stream_transfer_pending_queue_data_race(tmp_path: pathlib.Path):
assert_project_root()
# seed port to reuse the same port for the restarted nodes
peer_api_uris, peer_dirs, bootstrap_uri = start_cluster(tmp_path, N_PEERS, 20000)
create_collection(peer_api_uris[0], shard_number=N_SHARDS, replication_factor=N_REPLICA)
wait_collection_exists_and_active_on_all_peers(
collection_name=COLLECTION_NAME,
peer_api_uris=peer_api_uris
)
# Insert some initial number of points
upsert_random_points(peer_api_uris[0], 100, offset=100)
transfer_collection_cluster_info = get_collection_cluster_info(peer_api_uris[0], COLLECTION_NAME)
receiver_collection_cluster_info = get_collection_cluster_info(peer_api_uris[2], COLLECTION_NAME)
from_peer_id = transfer_collection_cluster_info['peer_id']
to_peer_id = receiver_collection_cluster_info['peer_id']
shard_id = transfer_collection_cluster_info['local_shards'][0]['shard_id']
# Send delay update operation which will lag all following operations
if check_feature_enabled(peer_api_uris[0], "staging"):
r = requests.post(
f"{peer_api_uris[0]}/collections/{COLLECTION_NAME}/debug", json={
"delay": {
"duration_sec": 2.0
}
})
assert_http_ok(r)
# Insert points with wait=false
# These updates will be pending for some time as they hang in the queue
# Insert with low point IDs here to make it more likely to conflict with the transfer
upsert_random_points(peer_api_uris[0], 100, offset=0, wait="false", batch_size=1)
# NOTE: the core issue happened here!
# We now have pending updates in the queue which are not applied yet. Below
# we start a stream records transfer. The transfer will only transfer
# records that have been applied. Only new incoming operations are forwarded
# to the remote over a separate channel. It means that all updates in this
# gap (the operations pending in the queue) do not appear on the receiver.
# Move shard `shard_id` to peer `target_peer_id`
r = requests.post(
f"{peer_api_uris[0]}/collections/{COLLECTION_NAME}/cluster", json={
"replicate_shard": {
"shard_id": shard_id,
"from_peer_id": from_peer_id,
"to_peer_id": to_peer_id,
"method": "stream_records",
}
})
assert_http_ok(r)
# Wait for end of shard transfer
wait_for_collection_shard_transfers_count(peer_api_uris[0], COLLECTION_NAME, 0)
# Assert target peer actually has two shards now
receiver_collection_cluster_info = get_collection_cluster_info(peer_api_uris[2], COLLECTION_NAME)
number_local_shards = len(receiver_collection_cluster_info['local_shards'])
assert number_local_shards == 2
# Synchronize peers
# Send one more wait=true operation to ensure all peers have processed all operations
upsert_random_points(peer_api_uris[0], 1, offset=200, wait="true")
# Point counts must be consistent across nodes
counts = []
for uri in peer_api_uris:
r = requests.post(
f"{uri}/collections/{COLLECTION_NAME}/points/count", json={
"exact": True
}
)
assert_http_ok(r)
counts.append(r.json()["result"]['count'])
assert counts[0] == counts[1] == counts[2]