def test_changing_topic_retention(self, property, acks):
"""
Test changing topic retention duration for topics with data produced
with ACKS=1 and ACKS=-1. This test produces data until 10 segments
appear, then it changes retention topic property and waits for
segments to be removed
"""
kafka_tools = KafkaCliTools(self.redpanda)
# produce until segments have been compacted
produce_until_segments(
self.redpanda,
topic=self.topic,
partition_idx=0,
count=10,
acks=acks,
)
# change retention time
kafka_tools.alter_topic_config(self.topic, {
property: 10000,
})
wait_for_segments_removal(self.redpanda,
self.topic,
partition_idx=0,
count=5)
def test_incremental_config(self):
topic = self.topics[0].name
kafka_tools = KafkaCliTools(self.redpanda)
kafka_tools.alter_topic_config(
topic, {
TopicSpec.PROPERTY_DATA_POLICY_FUNCTION_NAME: "1",
TopicSpec.PROPERTY_DATA_POLICY_SCRIPT_NAME: "2"
})
spec = kafka_tools.describe_topic(topic)
assert spec.redpanda_datapolicy == self._get_data_policy(1, 2)
# Expect that trying to set with a function name but no script fails.
try:
r = kafka_tools.alter_topic_config(
topic, {TopicSpec.PROPERTY_DATA_POLICY_FUNCTION_NAME: "3"})
except subprocess.CalledProcessError as e:
# Expected: request fails to update topic
self.logger.info(f"Kafka CLI alter failed as expected: {e.stdout}")
assert "unable to parse property" in e.stdout
else:
raise RuntimeError(f"Expected API error, got {r}")
# Expect that the failed alter operation has not modified the topic
spec = kafka_tools.describe_topic(topic)
assert spec.redpanda_datapolicy == self._get_data_policy(1, 2)
def test_fetch_after_committed_offset_was_removed(self,
transactions_enabled):
"""
Test fetching when consumer offset was deleted by retention
"""
self.redpanda._extra_rp_conf[
"enable_transactions"] = transactions_enabled
self.redpanda._extra_rp_conf[
"enable_idempotence"] = transactions_enabled
self.redpanda.start()
topic = TopicSpec(partition_count=1,
replication_factor=3,
cleanup_policy=TopicSpec.CLEANUP_DELETE)
self.client().create_topic(topic)
kafka_tools = KafkaCliTools(self.redpanda)
# produce until segments have been compacted
produce_until_segments(
self.redpanda,
topic=topic.name,
partition_idx=0,
count=10,
)
consumer_group = 'test'
rpk = RpkTool(self.redpanda)
def consume(n=1):
out = rpk.consume(topic.name, group=consumer_group, n=n)
split = out.split('}')
split = filter(lambda s: "{" in s, split)
return map(lambda s: json.loads(s + "}"), split)
#consume from the beggining
msgs = consume(10)
last = list(msgs).pop()
offset = last['offset']
# change retention time
kafka_tools.alter_topic_config(
topic.name, {
TopicSpec.PROPERTY_RETENTION_BYTES: 2 * self.segment_size,
})
wait_for_segments_removal(self.redpanda,
topic.name,
partition_idx=0,
count=5)
partitions = list(rpk.describe_topic(topic.name))
p = partitions[0]
assert p.start_offset > offset
# consume from the offset that doesn't exists,
# the one that was committed previously was already removed
out = list(consume(1))
assert out[0]['offset'] == p.start_offset
def alter_topic_configs(self, topic: str,
props: dict[str, typing.Union[str, int]]):
"""
Alter multiple topic configuration properties.
"""
kafka_tools = KafkaCliTools(self._redpanda)
kafka_tools.alter_topic_config(topic, props)
def test_altering_topic_configuration(self, property, value):
topic = self.topics[0].name
kafka_tools = KafkaCliTools(self.redpanda)
kafka_tools.alter_topic_config(topic, {property: value})
spec = kafka_tools.describe_topic(topic)
# e.g. retention.ms is TopicSpec.retention_ms
attr_name = property.replace(".", "_")
assert getattr(spec, attr_name, None) == value
def test_altering_multiple_topic_configurations(self):
topic = self.topics[0].name
kafka_tools = KafkaCliTools(self.redpanda)
kafka_tools.alter_topic_config(
topic, {
TopicSpec.PROPERTY_SEGMENT_SIZE: 1024,
TopicSpec.PROPERTY_RETENTION_TIME: 360000,
TopicSpec.PROPERTY_TIMESTAMP_TYPE: "LogAppendTime"
})
spec = kafka_tools.describe_topic(topic)
assert spec.segment_bytes == 1024
assert spec.retention_ms == 360000
assert spec.message_timestamp_type == "LogAppendTime"
def test_timequery_after_segments_eviction(self):
"""
Test checking if the offset returned by time based index is
valid during applying log cleanup policy
"""
segment_size = 1048576
# produce until segments have been compacted
produce_until_segments(
self.redpanda,
topic=self.topic,
partition_idx=0,
count=10,
acks=-1,
)
# restart all nodes to force replicating raft configuration
self.redpanda.restart_nodes(self.redpanda.nodes)
kafka_tools = KafkaCliTools(self.redpanda)
# Wait for controller, alter configs doesn't have a retry loop
kafka_tools.describe_topic(self.topic)
# change retention bytes to preserve 15 segments
kafka_tools.alter_topic_config(
self.topic, {
TopicSpec.PROPERTY_RETENTION_BYTES: 2 * segment_size,
})
def validate_time_query_until_deleted():
def done():
kcat = KafkaCat(self.redpanda)
ts = 1638748800 # 12.6.2021 - old timestamp, query first offset
offset = kcat.query_offset(self.topic, 0, ts)
# assert that offset is valid
assert offset >= 0
topic_partitions = segments_count(self.redpanda, self.topic, 0)
partitions = []
for p in topic_partitions:
partitions.append(p <= 5)
return all([p <= 5 for p in topic_partitions])
wait_until(done,
timeout_sec=30,
backoff_sec=5,
err_msg="Segments were not removed")
validate_time_query_until_deleted()
def test_configuration_properties_name_validation(self):
topic = self.topics[0].name
kafka_tools = KafkaCliTools(self.redpanda)
spec = kafka_tools.describe_topic(topic)
for i in range(0, 5):
key = self.random_string(5)
try:
kafka_tools.alter_topic_config(topic, {key: "123"})
except Exception as inst:
self.logger.info(
"alter failed as expected: expected exception %s", inst)
else:
raise RuntimeError("Alter should have failed but succeeded!")
new_spec = kafka_tools.describe_topic(topic)
# topic spec shouldn't change
assert new_spec == spec
def test_changing_topic_retention(self, property, acks):
"""
Test changing topic retention duration for topics with data produced
with ACKS=1 and ACKS=-1. This test produces data until 10 segments
appear, then it changes retention topic property and waits for
segments to be removed
"""
# operate on a partition. doesn't matter which one
partition = self.redpanda.partitions(self.topic)[0]
# produce until segments have been compacted
self._produce_until_segments(self.topic, 0, 10, acks)
kafka_tools = KafkaCliTools(self.redpanda)
# change retention time
kafka_tools.alter_topic_config(self.topic, {
property: 10000,
})
self._wait_for_segments_removal(self.topic, 0, 5)
def test_set_data_policy(self):
topic = self.topics[0].name
kafka_tools = KafkaCliTools(self.redpanda)
res = kafka_tools.alter_topic_config(
topic, {
TopicSpec.PROPERTY_DATA_POLICY_FUNCTION_NAME: "1",
TopicSpec.PROPERTY_DATA_POLICY_SCRIPT_NAME: "2"
})
spec = kafka_tools.describe_topic(topic)
assert spec.redpanda_datapolicy == self._get_data_policy(1, 2)
def test_configuration_properties_name_validation(self):
topic = self.topics[0].name
kafka_tools = KafkaCliTools(self.redpanda)
spec = kafka_tools.describe_topic(topic)
for i in range(0, 5):
key = self.random_string(5)
try:
res = kafka_tools.alter_topic_config(topic, {key: "123"})
except Exception as inst:
test_logger.info("exception %s", inst)
new_spec = kafka_tools.describe_topic(topic)
# topic spec shouldn't change
assert new_spec == spec
def test_changing_topic_retention_with_restart(self):
"""
Test changing topic retention duration for topics with data produced
with ACKS=1 and ACKS=-1. This test produces data until 10 segments
appear, then it changes retention topic property and waits for some
segmetnts to be removed
"""
segment_size = 1048576
# produce until segments have been compacted
produce_until_segments(
self.redpanda,
topic=self.topic,
partition_idx=0,
count=20,
acks=-1,
)
# restart all nodes to force replicating raft configuration
self.redpanda.restart_nodes(self.redpanda.nodes)
kafka_tools = KafkaCliTools(self.redpanda)
# Wait for controller, alter configs doesn't have a retry loop
kafka_tools.describe_topic(self.topic)
# change retention bytes to preserve 15 segments
kafka_tools.alter_topic_config(
self.topic, {
TopicSpec.PROPERTY_RETENTION_BYTES: 15 * segment_size,
})
wait_for_segments_removal(redpanda=self.redpanda,
topic=self.topic,
partition_idx=0,
count=16)
# change retention bytes again to preserve 10 segments
kafka_tools.alter_topic_config(
self.topic, {
TopicSpec.PROPERTY_RETENTION_BYTES: 10 * segment_size,
})
wait_for_segments_removal(redpanda=self.redpanda,
topic=self.topic,
partition_idx=0,
count=11)
# change retention bytes again to preserve 5 segments
kafka_tools.alter_topic_config(
self.topic, {
TopicSpec.PROPERTY_RETENTION_BYTES: 4 * segment_size,
})
wait_for_segments_removal(redpanda=self.redpanda,
topic=self.topic,
partition_idx=0,
count=5)
def test_altering_topic_configuration(self, property, value):
topic = self.topics[0].name
kafka_tools = KafkaCliTools(self.redpanda)
res = kafka_tools.alter_topic_config(topic, {property: value})
spec = kafka_tools.describe_topic(topic)
class ArchivalTest(RedpandaTest):
log_segment_size = 1048576 # 1MB
log_compaction_interval_ms = 10000
s3_host_name = "minio-s3"
s3_access_key = "panda-user"
s3_secret_key = "panda-secret"
s3_region = "panda-region"
s3_topic_name = "panda-topic"
topics = (TopicSpec(name='panda-topic',
partition_count=1,
replication_factor=3), )
def __init__(self, test_context):
self.s3_bucket_name = f"panda-bucket-{uuid.uuid1()}"
self._extra_rp_conf = dict(
cloud_storage_enabled=True,
cloud_storage_access_key=ArchivalTest.s3_access_key,
cloud_storage_secret_key=ArchivalTest.s3_secret_key,
cloud_storage_region=ArchivalTest.s3_region,
cloud_storage_bucket=self.s3_bucket_name,
cloud_storage_disable_tls=True,
cloud_storage_api_endpoint=ArchivalTest.s3_host_name,
cloud_storage_api_endpoint_port=9000,
cloud_storage_reconciliation_interval_ms=500,
cloud_storage_max_connections=5,
log_compaction_interval_ms=self.log_compaction_interval_ms,
log_segment_size=self.log_segment_size,
)
if test_context.function_name == "test_timeboxed_uploads":
self._extra_rp_conf.update(
log_segment_size=1024 * 1024 * 1024,
cloud_storage_segment_max_upload_interval_sec=1)
super(ArchivalTest, self).__init__(test_context=test_context,
extra_rp_conf=self._extra_rp_conf)
self.kafka_tools = KafkaCliTools(self.redpanda)
self.rpk = RpkTool(self.redpanda)
self.s3_client = S3Client(
region='panda-region',
access_key=u"panda-user",
secret_key=u"panda-secret",
endpoint=f'http://{ArchivalTest.s3_host_name}:9000',
logger=self.logger)
def setUp(self):
self.s3_client.empty_bucket(self.s3_bucket_name)
self.s3_client.create_bucket(self.s3_bucket_name)
# Deletes in S3 are eventually consistent so we might still
# see previously removed objects for a while.
validate(self._check_bucket_is_emtpy, self.logger, 300)
super().setUp() # topic is created here
# enable archival for topic
for topic in self.topics:
self.rpk.alter_topic_config(topic.name, 'redpanda.remote.write',
'true')
def tearDown(self):
self.s3_client.empty_bucket(self.s3_bucket_name)
super().tearDown()
@cluster(num_nodes=3)
def test_write(self):
"""Simpe smoke test, write data to redpanda and check if the
data hit the S3 storage bucket"""
self.kafka_tools.produce(self.topic, 10000, 1024)
validate(self._quick_verify, self.logger, 90)
@cluster(num_nodes=3)
def test_isolate(self):
"""Verify that our isolate/rejoin facilities actually work"""
with firewall_blocked(self.redpanda.nodes, self._get_s3_endpoint_ip()):
self.kafka_tools.produce(self.topic, 10000, 1024)
time.sleep(10) # can't busy wait here
# Topic manifest can be present in the bucket because topic is created before
# firewall is blocked. No segments or partition manifest should be present.
topic_manifest_id = "d0000000/meta/kafka/panda-topic/topic_manifest.json"
objects = self.s3_client.list_objects(self.s3_bucket_name)
keys = [x.Key for x in objects]
assert len(keys) < 2, \
f"Bucket should be empty or contain only {topic_manifest_id}, but contains {keys}"
if len(keys) == 1:
assert topic_manifest_id == keys[0], \
f"Bucket should be empty or contain only {topic_manifest_id}, but contains {keys[0]}"
@cluster(num_nodes=3)
def test_reconnect(self):
"""Disconnect redpanda from S3, write data, connect redpanda to S3
and check that the data is uploaded"""
with firewall_blocked(self.redpanda.nodes, self._get_s3_endpoint_ip()):
self.kafka_tools.produce(self.topic, 10000, 1024)
time.sleep(10) # sleep is needed because we need to make sure that
# reconciliation loop kicked in and started uploading
# data, otherwse we can rejoin before archival storage
# will even try to upload new segments
validate(self._quick_verify, self.logger, 90)
@cluster(num_nodes=3)
def test_one_node_reconnect(self):
"""Disconnect one redpanda node from S3, write data, connect redpanda to S3
and check that the data is uploaded"""
self.kafka_tools.produce(self.topic, 1000, 1024)
leaders = list(self._get_partition_leaders().values())
with firewall_blocked(leaders[0:1], self._get_s3_endpoint_ip()):
self.kafka_tools.produce(self.topic, 9000, 1024)
time.sleep(10) # sleep is needed because we need to make sure that
# reconciliation loop kicked in and started uploading
# data, otherwse we can rejoin before archival storage
# will even try to upload new segments
validate(self._quick_verify, self.logger, 90)
@cluster(num_nodes=3)
def test_connection_drop(self):
"""Disconnect redpanda from S3 during the active upload, restore connection
and check that everything is uploaded"""
self.kafka_tools.produce(self.topic, 10000, 1024)
with firewall_blocked(self.redpanda.nodes, self._get_s3_endpoint_ip()):
time.sleep(10) # sleep is needed because we need to make sure that
# reconciliation loop kicked in and started uploading
# data, otherwse we can rejoin before archival storage
# will even try to upload new segments
validate(self._quick_verify, self.logger, 90)
@cluster(num_nodes=3)
def test_connection_flicker(self):
"""Disconnect redpanda from S3 during the active upload for short period of time
during upload and check that everything is uploaded"""
con_enabled = True
for _ in range(0, 20):
# upload data in batches
if con_enabled:
with firewall_blocked(self.redpanda.nodes,
self._get_s3_endpoint_ip()):
self.kafka_tools.produce(self.topic, 500, 1024)
else:
self.kafka_tools.produce(self.topic, 500, 1024)
con_enabled = not con_enabled
time.sleep(1)
time.sleep(10)
validate(self._quick_verify, self.logger, 90)
@cluster(num_nodes=3)
def test_single_partition_leadership_transfer(self):
"""Start uploading data, restart leader node of the partition 0 to trigger the
leadership transfer, continue upload, verify S3 bucket content"""
self.kafka_tools.produce(self.topic, 5000, 1024)
time.sleep(5)
leaders = self._get_partition_leaders()
node = leaders[0]
self.redpanda.stop_node(node)
time.sleep(1)
self.redpanda.start_node(node)
time.sleep(5)
self.kafka_tools.produce(self.topic, 5000, 1024)
validate(self._cross_node_verify, self.logger, 90)
@cluster(num_nodes=3)
def test_all_partitions_leadership_transfer(self):
"""Start uploading data, restart leader nodes of all partitions to trigger the
leadership transfer, continue upload, verify S3 bucket content"""
self.kafka_tools.produce(self.topic, 5000, 1024)
time.sleep(5)
leaders = self._get_partition_leaders()
for ip, node in leaders.items():
self.logger.debug(f"going to restart node {ip}")
self.redpanda.stop_node(node)
time.sleep(1)
self.redpanda.start_node(node)
time.sleep(5)
self.kafka_tools.produce(self.topic, 5000, 1024)
validate(self._cross_node_verify, self.logger, 90)
@cluster(num_nodes=3)
def test_timeboxed_uploads(self):
"""This test checks segment upload time limit. The feature is enabled in the
configuration. The configuration defines maximum time interval between uploads.
If the option is set then redpanda will start uploading a segment partially if
configured amount of time passed since previous upload and the segment has some
new data.
The test sets the timeout value to 1s. Then it uploads data in batches with delays
between the batches. The segment size is set to 1GiB. We upload 10MiB total. So
normally, there won't be any data uploaded to Minio. But since the time limit for
a segment is set to 1s we will see a bunch of segments in the bucket. The offsets
of the segments won't align with the segment in the redpanda data directory. But
their respective offset ranges should align and the sizes should make sense.
"""
# The offsets of the segments in the Minio bucket won't necessary
# correlate with the write bursts here. The upload depends on the
# timeout but also on raft and current high_watermark. So we can
# expect that the bucket won't have 9 segments with 1000 offsets.
# The actual segments will be larger.
for i in range(0, 10):
self.kafka_tools.produce(self.topic, 1000, 1024)
time.sleep(1)
time.sleep(5)
def check_upload():
# check that the upload happened
ntps = set()
sizes = {}
for node in self.redpanda.nodes:
checksums = self._get_redpanda_log_segment_checksums(node)
self.logger.info(
f"Node: {node.account.hostname} checksums: {checksums}")
lst = [
_parse_normalized_segment_path(path, md5, size)
for path, (md5, size) in checksums.items()
]
lst = sorted(lst, key=lambda x: x.base_offset)
segments = defaultdict(int)
sz = defaultdict(int)
for it in lst:
ntps.add(it.ntp)
sz[it.ntp] += it.size
segments[it.ntp] += 1
for ntp, s in segments.items():
assert s != 0, f"expected to have at least one segment per partition, got {s}"
for ntp, s in sz.items():
if ntp not in sizes:
sizes[ntp] = s
# Download manifest for partitions
for ntp in ntps:
manifest = self._download_partition_manifest(ntp)
self.logger.info(f"downloaded manifest {manifest}")
segments = []
for _, segment in manifest['segments'].items():
segments.append(segment)
segments = sorted(segments, key=lambda s: s['base_offset'])
self.logger.info(f"sorted segments {segments}")
prev_committed_offset = -1
size = 0
for segment in segments:
self.logger.info(
f"checking {segment} prev: {prev_committed_offset}")
base_offset = segment['base_offset']
assert prev_committed_offset + 1 == base_offset, f"inconsistent segments, " +\
"expected base_offset: {prev_committed_offset + 1}, actual: {base_offset}"
prev_committed_offset = segment['committed_offset']
size += segment['size_bytes']
assert sizes[ntp] >= size
assert size > 0
validate(check_upload, self.logger, 90)
@cluster(num_nodes=3)
def test_retention_archival_coordination(self):
"""
Test that only archived segments can be evicted and that eviction
restarts once the segments have been archived.
"""
self.kafka_tools.alter_topic_config(
self.topic,
{
TopicSpec.PROPERTY_RETENTION_BYTES: 5 * self.log_segment_size,
},
)
with firewall_blocked(self.redpanda.nodes, self._get_s3_endpoint_ip()):
produce_until_segments(redpanda=self.redpanda,
topic=self.topic,
partition_idx=0,
count=10)
# Sleep some time sufficient for log eviction under normal conditions
# and check that no segment has been evicted (because we can't upload
# segments to the cloud storage).
time.sleep(3 * self.log_compaction_interval_ms / 1000.0)
counts = list(
segments_count(self.redpanda, self.topic, partition_idx=0))
self.logger.info(f"node segment counts: {counts}")
assert len(counts) == len(self.redpanda.nodes)
assert all(c >= 10 for c in counts)
# Check that eviction restarts after we restored the connection to cloud
# storage.
wait_for_segments_removal(redpanda=self.redpanda,
topic=self.topic,
partition_idx=0,
count=6)
def _check_bucket_is_emtpy(self):
allobj = self._list_objects()
for obj in allobj:
self.logger.debug(
f"found object {obj} in bucket {self.s3_bucket_name}")
assert len(allobj) == 0
def _get_partition_leaders(self):
kcat = KafkaCat(self.redpanda)
m = kcat.metadata()
self.logger.info(f"kcat.metadata() == {m}")
brokers = {}
for b in m['brokers']:
id = b['id']
ip = b['name']
ip = ip[:ip.index(':')]
for n in self.redpanda.nodes:
n_ip = n.account.hostname
self.logger.debug(f"matching {n_ip} over {ip}")
if n_ip == ip:
brokers[id] = n
break
self.logger.debug(f"found brokers {brokers}")
assert len(brokers) == 3
leaders = {}
for topic in m['topics']:
if topic['topic'] == ArchivalTest.s3_topic_name:
for part in topic['partitions']:
leader_id = part['leader']
partition_id = part['partition']
leader = brokers[leader_id]
leaders[partition_id] = leader
return leaders
def _download_partition_manifest(self, ntp):
"""Find and download individual partition manifest"""
expected = f"{ntp.ns}/{ntp.topic}/{ntp.partition}_{ntp.revision}/manifest.json"
id = None
objects = []
for loc in self._list_objects():
objects.append(loc)
if expected in loc:
id = loc
break
if id is None:
objlist = "\n".join(objects)
self.logger.debug(
f"expected path {expected} is not found in the bucket, bucket content: \n{objlist}"
)
assert not id is None
manifest = self.s3_client.get_object_data(self.s3_bucket_name, id)
self.logger.info(f"manifest found: {manifest}")
return json.loads(manifest)
def _verify_manifest(self, ntp, manifest, remote):
"""Check that all segments that present in manifest are available
in remote storage"""
for sname, _ in manifest['segments'].items():
spath = f"{ntp.ns}/{ntp.topic}/{ntp.partition}_{ntp.revision}/{sname}"
self.logger.info(f"validating manifest path {spath}")
assert spath in remote
ranges = [(int(m['base_offset']), int(m['committed_offset']))
for _, m in manifest['segments'].items()]
ranges = sorted(ranges, key=lambda x: x[0])
last_offset = -1
num_gaps = 0
for base, committed in ranges:
if last_offset + 1 != base:
self.logger.debug(
f"gap between {last_offset} and {base} detected")
num_gaps += 1
last_offset = committed
assert num_gaps == 0
def _cross_node_verify(self):
"""Verify data on all nodes taking into account possible alignment issues
caused by leadership transitions.
The verification algorithm is following:
- Download and verify partition manifest;
- Partition manifest has all segments and metadata like committed offset
and base offset. We can also retrieve MD5 hash of every segment;
- Load segment metadata for every redpanda node.
- Scan every node's metadata and match segments with manifest, on success
remove matched segment from the partition manifest.
The goal #1 is to remove all segments from the manifest. The goal #2 is to
find the last segment that's supposed to be uploaded from the leader node,
it's base offset should be equal to manifest's last offset + 1.
The segments match if:
- The base offset and md5 hashes are the same;
- The committed offset of both segments are the same, md5 hashes are different,
and base offset of the segment from manifest is larger than base offset of the
segment from redpanda node. In this case we should also compare the data
directly by scanning both segments.
"""
nodes = {}
ntps = set()
for node in self.redpanda.nodes:
checksums = self._get_redpanda_log_segment_checksums(node)
self.logger.info(
f"Node: {node.account.hostname} checksums: {checksums}")
lst = [
_parse_normalized_segment_path(path, md5, size)
for path, (md5, size) in checksums.items()
]
lst = sorted(lst, key=lambda x: x.base_offset)
nodes[node.account.hostname] = lst
for it in lst:
ntps.add(it.ntp)
# Download metadata from S3
remote = self._get_redpanda_s3_checksums()
# Download manifest for partitions
manifests = {}
for ntp in ntps:
manifest = self._download_partition_manifest(ntp)
manifests[ntp] = manifest
self._verify_manifest(ntp, manifest, remote)
for ntp in ntps:
self.logger.debug(f"verifying {ntp}")
manifest = manifests[ntp]
segments = manifest['segments']
manifest_segments = [
_parse_manifest_segment(manifest, sname, meta, remote,
self.logger)
for sname, meta in segments.items()
]
manifest_segments = sorted(manifest_segments,
key=lambda x: x.base_offset)
for node_key, node_segments in nodes.items():
self.logger.debug(f"checking {ntp} on {node_key}")
for mix, msegm in enumerate(manifest_segments):
if not msegm is None:
segments = sorted([
segment
for segment in node_segments if segment.ntp == ntp
],
key=lambda x: x.base_offset)
self.logger.debug(
f"checking manifest segment {msegm} over {node_key} segments {segments}"
)
found = False
for ix in range(0, len(segments)):
nsegm = segments[ix]
if nsegm.ntp != ntp:
continue
nsegm_co = -1 if (ix + 1) == len(segments) else (
segments[ix + 1].base_offset - 1)
self.logger.debug(
f"comparing {msegm.base_offset}:{msegm.committed_offset}:{msegm.md5} to {nsegm.base_offset}:{nsegm_co}:{nsegm.md5}"
)
if msegm.base_offset == nsegm.base_offset and msegm.md5 == nsegm.md5:
# Success
self.logger.info(
f"found match for segment {msegm.ntp} {msegm.base_offset} on {node_key}"
)
manifest_segments[mix] = None
found = True
break
if msegm.committed_offset == nsegm_co and msegm.base_offset > nsegm.base_offset:
# Found segment with truncated head (due to leadership transition)
actual_hash = self._get_partial_checksum(
node_key, nsegm.normalized_path,
msegm.size)
self.logger.info(
f"partial hash {actual_hash} retreived, s3 hash {msegm.md5}"
)
if actual_hash == msegm.md5:
manifest_segments[mix] = None
self.logger.info(
f"partial match for segment {msegm.ntp} {msegm.base_offset}-"
+
f"{msegm.committed_offset} on {node_key}"
)
found = True
break
if not found:
self.logger.debug(
f"failed to match {msegm.base_offset}:{msegm.committed_offset}"
)
else:
self.logger.debug(
f"matched {msegm.base_offset}:{msegm.committed_offset} successfully"
)
# All segments should be matched and set to None
if any(manifest_segments):
self.logger.debug(
f"manifest segments that fail to validate: {manifest_segments}"
)
assert not any(manifest_segments)
# Verify goal #2, the last segment on a leader node is manifest.last_offset + 1
ntp_offsets = []
for node_key, node_segments in nodes.items():
offsets = [
segm.base_offset for segm in node_segments
if segm.ntp == ntp
]
if offsets:
max_offset = max([
segm.base_offset for segm in node_segments
if segm.ntp == ntp
])
ntp_offsets.append(max_offset)
self.logger.debug(
f"NTP {ntp} has the largest offset {max_offset} on node {node_key}"
)
else:
self.logger.debug(
f"NTP {ntp} has no offsets on node {node_key}")
last_offset = int(manifest['last_offset'])
self.logger.debug(
f"last offset: {last_offset}, ntp offsets: {ntp_offsets}")
assert (last_offset + 1) in ntp_offsets
def _list_objects(self):
"""Emulate ListObjects call by fetching the topic manifests and
iterating through its content"""
try:
topic_manifest_id = "d0000000/meta/kafka/panda-topic/topic_manifest.json"
partition_manifest_id = "d0000000/meta/kafka/panda-topic/0_9/manifest.json"
manifest = self.s3_client.get_object_data(self.s3_bucket_name,
partition_manifest_id)
results = [topic_manifest_id, partition_manifest_id]
for id in manifest['segments'].keys():
results.append(id)
self.logger.debug(f"ListObjects(source: manifest): {results}")
except:
results = [
loc.Key
for loc in self.s3_client.list_objects(self.s3_bucket_name)
]
self.logger.debug(f"ListObjects: {results}")
return results
def _quick_verify(self):
"""Verification algorithm that works only if no leadership
transfer happend during the run. It works by looking up all
segments from the remote storage in local redpanda storages.
It's done by using md5 hashes of the nodes.
"""
local = {}
for node in self.redpanda.nodes:
checksums = self._get_redpanda_log_segment_checksums(node)
self.logger.info(
f"Node: {node.account.hostname} checksums: {checksums}")
for k, v in checksums.items():
local.setdefault(k, set()).add(v)
remote = self._get_redpanda_s3_checksums()
self.logger.info(f"S3 checksums: {remote}")
self.logger.info(f"Local checksums: {local}")
assert len(local) != 0
assert len(remote) != 0
md5fails = 0
lookup_fails = 0
for path, csum in remote.items():
self.logger.info(f"checking remote path: {path} csum: {csum}")
if path not in local:
self.logger.debug(
f"remote path {path} can't be found in any of the local storages"
)
lookup_fails += 1
else:
if len(local[path]) != 1:
self.logger.info(
f"remote segment {path} have more than one variant {local[path]}"
)
if not csum in local[path]:
self.logger.debug(
f"remote md5 {csum} doesn't match any local {local[path]}"
)
md5fails += 1
if md5fails != 0:
self.logger.debug(
f"Validation failed, {md5fails} remote segments doesn't match")
if lookup_fails != 0:
self.logger.debug(
f"Validation failed, remote {lookup_fails} remote locations doesn't match local"
)
assert md5fails == 0 and lookup_fails == 0
# Validate partitions
# for every partition the segment with largest base offset shouldn't be
# available in remote storage
local_partitions = {}
remote_partitions = {}
for path, items in local.items():
meta = _parse_normalized_segment_path(path, '', 0)
local_partitions.setdefault(meta.ntp, []).append((meta, items))
for path, items in remote.items():
meta = _parse_normalized_segment_path(path, '', 0)
remote_partitions.setdefault(meta.ntp, []).append((meta, items))
self.logger.info(
f"generated local partitions {local_partitions.keys()}")
self.logger.info(
f"generated remote partitions {remote_partitions.keys()}")
# Download manifest for partitions
manifests = {}
for ntp in local_partitions.keys():
manifest = self._download_partition_manifest(ntp)
manifests[ntp] = manifest
self._verify_manifest(ntp, manifest, remote)
# Check that all local partition are archived
assert len(local_partitions) == 1
assert len(remote_partitions) == 1
missing_partitions = 0
for key in local_partitions.keys():
if key not in remote_partitions:
self.logger.debug(f"partition {key} not found in remote set")
missing_partitions += 1
assert missing_partitions == 0
def _get_redpanda_log_segment_checksums(self, node):
"""Get MD5 checksums of log segments that match the topic. The paths are
normalized (<namespace>/<topic>/<partition>_<rev>/...)."""
checksums = self.redpanda.data_checksum(node)
# Filter out all unwanted paths
def included(path):
controller_log_prefix = os.path.join(RedpandaService.DATA_DIR,
"redpanda")
log_segment_extension = ".log"
return not path.startswith(
controller_log_prefix) and path.endswith(log_segment_extension)
# Remove data dir from path
def normalize_path(path):
return os.path.relpath(path, RedpandaService.DATA_DIR)
return {
normalize_path(path): value
for path, value in checksums.items() if included(path)
}
def _get_redpanda_s3_checksums(self):
"""Get MD5 checksums of log segments stored in S3 (minio). The paths are
normalized (<namespace>/<topic>/<partition>_<rev>/...)."""
def normalize(path):
return path[9:] # 8-character hash + /
def included(path):
manifest_extension = ".json"
return not path.endswith(manifest_extension)
return {
normalize(it.Key): (it.ETag, it.ContentLength)
for it in self.s3_client.list_objects(self.s3_bucket_name)
if included(it.Key)
}
def _get_partial_checksum(self, hostname, normalized_path, tail_bytes):
"""Compute md5 checksum of the last 'tail_bytes' of the file located
on a node."""
node = None
for n in self.redpanda.nodes:
if n.account.hostname == hostname:
node = n
full_path = os.path.join(RedpandaService.DATA_DIR, normalized_path)
cmd = f"tail -c {tail_bytes} {full_path} | md5sum"
line = node.account.ssh_output(cmd)
tokens = line.split()
return tokens[0].decode()
def _isolate(self, nodes, ips):
"""Isolate certain ips from the nodes using firewall rules"""
cmd = []
for ip in ips:
cmd.append(f"iptables -A INPUT -s {ip} -j DROP")
cmd.append(f"iptables -A OUTPUT -d {ip} -j DROP")
cmd = " && ".join(cmd)
for node in nodes:
node.account.ssh_output(cmd, allow_fail=False)
def _rejoin(self, nodes, ips):
"""Remove firewall rules that isolate ips from the nodes"""
cmd = []
for ip in ips:
cmd.append(f"iptables -D INPUT -s {ip} -j DROP")
cmd.append(f"iptables -D OUTPUT -d {ip} -j DROP")
cmd = " && ".join(cmd)
for node in nodes:
node.account.ssh_output(cmd, allow_fail=False)
def _host_name_to_ip_address(self, hostname):
ip_host = self.redpanda.nodes[0].account.ssh_output(
f'getent hosts {hostname}')
return ip_host.split()[0].decode()
def _get_s3_endpoint_ip(self):
return self._host_name_to_ip_address(ArchivalTest.s3_host_name)
def _get_rp_cluster_ips(self, nhosts=4):
lst = []
for ix in range(1, nhosts + 1):
h = f"rp_n{ix}_1"
lst.append(self._host_name_to_ip_address(h))
return lst
class EndToEndShadowIndexingTest(EndToEndTest):
segment_size = 1048576 # 1 Mb
s3_host_name = "minio-s3"
s3_access_key = "panda-user"
s3_secret_key = "panda-secret"
s3_region = "panda-region"
s3_topic_name = "panda-topic"
topics = (TopicSpec(
name=s3_topic_name,
partition_count=1,
replication_factor=3,
), )
def __init__(self, test_context):
super(EndToEndShadowIndexingTest,
self).__init__(test_context=test_context)
self.s3_bucket_name = f"panda-bucket-{uuid.uuid1()}"
self.topic = EndToEndShadowIndexingTest.s3_topic_name
self._extra_rp_conf = dict(
cloud_storage_enabled=True,
cloud_storage_enable_remote_read=True,
cloud_storage_enable_remote_write=True,
cloud_storage_access_key=EndToEndShadowIndexingTest.s3_access_key,
cloud_storage_secret_key=EndToEndShadowIndexingTest.s3_secret_key,
cloud_storage_region=EndToEndShadowIndexingTest.s3_region,
cloud_storage_bucket=self.s3_bucket_name,
cloud_storage_disable_tls=True,
cloud_storage_api_endpoint=EndToEndShadowIndexingTest.s3_host_name,
cloud_storage_api_endpoint_port=9000,
cloud_storage_reconciliation_interval_ms=500,
cloud_storage_max_connections=5,
log_segment_size=EndToEndShadowIndexingTest.segment_size, # 1MB
)
self.scale = Scale(test_context)
self.redpanda = RedpandaService(
context=test_context,
num_brokers=3,
extra_rp_conf=self._extra_rp_conf,
)
self.kafka_tools = KafkaCliTools(self.redpanda)
self.s3_client = S3Client(
region=EndToEndShadowIndexingTest.s3_region,
access_key=EndToEndShadowIndexingTest.s3_access_key,
secret_key=EndToEndShadowIndexingTest.s3_secret_key,
endpoint=f"http://{EndToEndShadowIndexingTest.s3_host_name}:9000",
logger=self.logger,
)
def setUp(self):
self.s3_client.empty_bucket(self.s3_bucket_name)
self.s3_client.create_bucket(self.s3_bucket_name)
self.redpanda.start()
for topic in EndToEndShadowIndexingTest.topics:
self.kafka_tools.create_topic(topic)
def tearDown(self):
self.s3_client.empty_bucket(self.s3_bucket_name)
@cluster(num_nodes=5)
def test_write(self):
"""Write at least 10 segments, set retention policy to leave only 5
segments, wait for segments removal, consume data and run validation,
that everything that is acked is consumed."""
self.start_producer()
produce_until_segments(
redpanda=self.redpanda,
topic=self.topic,
partition_idx=0,
count=10,
)
self.kafka_tools.alter_topic_config(
self.topic,
{
TopicSpec.PROPERTY_RETENTION_BYTES:
5 * EndToEndShadowIndexingTest.segment_size,
},
)
wait_for_segments_removal(redpanda=self.redpanda,
topic=self.topic,
partition_idx=0,
count=6)
self.start_consumer()
self.run_validation()
def test_shadow_indexing_aborted_txs(self):
"""Check that messages belonging to aborted transaction are not seen by clients
when fetching from remote segments."""
topic = self.topics[0]
class Producer:
def __init__(self, brokers, logger):
self.keys = []
self.cur_offset = 0
self.brokers = brokers
self.logger = logger
self.num_aborted = 0
self.reconnect()
def reconnect(self):
self.producer = ck.Producer({
'bootstrap.servers':
self.brokers,
'transactional.id':
'shadow-indexing-tx-test',
})
self.producer.init_transactions()
def produce(self, topic):
"""produce some messages inside a transaction with increasing keys
and random values. Then randomly commit/abort the transaction."""
n_msgs = random.randint(50, 100)
keys = []
self.producer.begin_transaction()
for _ in range(n_msgs):
val = ''.join(
map(chr, (random.randint(0, 256)
for _ in range(random.randint(100, 1000)))))
self.producer.produce(topic.name, val,
str(self.cur_offset))
keys.append(str(self.cur_offset).encode('utf8'))
self.cur_offset += 1
self.logger.info(
f"writing {len(keys)} msgs: {keys[0]}-{keys[-1]}...")
self.producer.flush()
if random.random() < 0.1:
self.producer.abort_transaction()
self.num_aborted += 1
self.logger.info("aborted txn")
else:
self.producer.commit_transaction()
self.keys.extend(keys)
producer = Producer(self.redpanda.brokers(), self.logger)
def done():
for _ in range(100):
try:
producer.produce(topic)
except ck.KafkaException as err:
self.logger.warn(f"producer error: {err}")
producer.reconnect()
self.logger.info("producer iteration complete")
topic_partitions = segments_count(self.redpanda,
topic.name,
partition_idx=0)
partitions = []
for p in topic_partitions:
partitions.append(p >= 10)
return all(partitions)
wait_until(done,
timeout_sec=120,
backoff_sec=1,
err_msg="producing failed")
assert producer.num_aborted > 0
kafka_tools = KafkaCliTools(self.redpanda)
kafka_tools.alter_topic_config(
self.topic,
{
TopicSpec.PROPERTY_RETENTION_BYTES: 3 * self.segment_size,
},
)
wait_for_segments_removal(redpanda=self.redpanda,
topic=self.topic,
partition_idx=0,
count=6)
consumer = ck.Consumer(
{
'bootstrap.servers': self.redpanda.brokers(),
'group.id': 'shadow-indexing-tx-test',
'auto.offset.reset': 'earliest',
},
logger=self.logger)
consumer.subscribe([topic.name])
consumed = []
while True:
msgs = consumer.consume(timeout=5.0)
if len(msgs) == 0:
break
consumed.extend([(m.key(), m.offset()) for m in msgs])
first_mismatch = ''
for p_key, (c_key, c_offset) in zip_longest(producer.keys, consumed):
if p_key != c_key:
first_mismatch = f"produced: {p_key}, consumed: {c_key} (offset: {c_offset})"
break
assert (not first_mismatch), (
f"produced and consumed messages differ, "
f"produced length: {len(producer.keys)}, consumed length: {len(consumed)}, "
f"first mismatch: {first_mismatch}")
