Skip to content
1 change: 1 addition & 0 deletions CHANGES.txt
Original file line number Diff line number Diff line change
@@ -1,5 +1,6 @@
0.5.0
-----
* Support Sidecar behind a load balancer for time-skew validation in single-cluster and coordinated writes (CASSANALYTICS-181)
* Opening one keyspace instance rebuilds every keyspace instance in the Cassandra 6.0 bridge (CASSANALYTICS-202)
* CDC logs NPE for deleted column values (CASSANALYTICS-178)
* Add Cassandra 6.0 support (CASSANALYTICS-37)
Expand Down
Original file line number Diff line number Diff line change
Expand Up @@ -100,6 +100,6 @@ public String clusterId()
@Override
public ClusterInfo reconstruct()
{
return new CassandraClusterInfo(this);
return CassandraClusterInfo.create(this);
}
}
Original file line number Diff line number Diff line change
Expand Up @@ -146,6 +146,7 @@ public class BulkSparkConf implements Serializable
public final double importCoordinatorTimeoutMultiplier;
public boolean quoteIdentifiers;
public final boolean skipSecondaryIndexCheck;
public final boolean sidecarBehindLoadBalancer;
protected final String keystorePassword;
protected final String keystorePath;
protected final String keystoreBase64Encoded;
Expand Down Expand Up @@ -229,6 +230,7 @@ public BulkSparkConf(SparkConf conf, Map<String, String> options, @Nullable Logg
this.timestamp = MapUtils.getOrDefault(options, WriterOptions.TIMESTAMP.name(), null);
this.quoteIdentifiers = MapUtils.getBoolean(options, WriterOptions.QUOTE_IDENTIFIERS.name(), false, "quote identifiers");
this.skipSecondaryIndexCheck = MapUtils.getBoolean(options, WriterOptions.SKIP_SECONDARY_INDEX_CHECK.name(), false, "skip secondary index check");
this.sidecarBehindLoadBalancer = MapUtils.getBoolean(options, WriterOptions.SIDECAR_BEHIND_LOAD_BALANCER.name(), false, "sidecar behind load balancer");
int storageClientConcurrency = MapUtils.getInt(options, WriterOptions.STORAGE_CLIENT_CONCURRENCY.name(),
DEFAULT_STORAGE_CLIENT_CONCURRENCY, "storage client concurrency");
long storageClientKeepAliveSeconds = MapUtils.getLong(options, WriterOptions.STORAGE_CLIENT_THREAD_KEEP_ALIVE_SECONDS.name(),
Expand Down
Original file line number Diff line number Diff line change
Expand Up @@ -59,7 +59,7 @@ protected CassandraBulkWriterContext(@NotNull BulkWriterConfig config)
@Override
protected ClusterInfo buildClusterInfo()
{
return new CassandraClusterInfo(bulkSparkConf());
return CassandraClusterInfo.create(bulkSparkConf());
}

@Override
Expand Down
Original file line number Diff line number Diff line change
Expand Up @@ -138,6 +138,57 @@ public CassandraClusterInfo(BroadcastableClusterInfo broadcastable)
this.allNodeSettingFutures = null;
}

/**
* Creates a {@link CassandraClusterInfo} for a single cluster, selecting the concrete type based on the
* {@link BulkSparkConf#sidecarBehindLoadBalancer} flag.
*
* @param conf bulk write conf
* @return {@link LoadBalancedCassandraClusterInfo} when Sidecar is behind a load balancer, otherwise a plain
* {@link CassandraClusterInfo}
*/
public static CassandraClusterInfo create(BulkSparkConf conf)

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Maybe a nit, but at least I'd like for us to consider.

We are adding a new class LoadBalancedCassandraClusterInfo that extends CassandraClusterInfo, which is fine, but:
CassandraClusterInfo may return an instance of itself or an instance of the new class, forcing it to "know" (depend) on LoadBalancedCassandraClusterInfo. This feels like a factory pattern with a potential abstract CassandraClusterInfo class and two classes extending it: LoadBalancedCassandraClusterInfo and maybe StandaloneCassandraClusterInfo, and the factory static create method returning one or the other? I think that approach would be cleaner. Opinions?

Copy link
Copy Markdown
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Good point, thanks.
I agree that having the base class reference its subclass in create() isn’t ideal.

I think the abstract CassandraClusterInfo + StandaloneCassandraClusterInfo / LoadBalancedCassandraClusterInfo approach makes sense. My only hesitation is the scope. CassandraClusterInfo is quite central, and making it abstract would mean introducing a separate StandaloneCassandraClusterInfo and updating the places that currently instantiate or mock CassandraClusterInfo directly. That would likely touch a number of tests and mocks as well.

Since the PR is otherwise ready, I’d prefer to keep this change smaller for now and revisit the full split separately. But happy to go with the cleaner approach if you think it’s worth doing as part of this PR.

{
return create(conf, null);
}

/**
* Creates a {@link CassandraClusterInfo}, selecting the concrete type based on the
* {@link BulkSparkConf#sidecarBehindLoadBalancer} flag. Kept centralized so the driver-side factory and the
* executor-side broadcast reconstruction stay in lockstep, for both single-cluster and coordinated writes.
*
* @param conf bulk write conf
* @param clusterId cluster identifier, or {@code null} for a single (non-coordinated) cluster
* @return {@link LoadBalancedCassandraClusterInfo} when Sidecar is behind a load balancer, otherwise a plain
* {@link CassandraClusterInfo}
*/
public static CassandraClusterInfo create(BulkSparkConf conf, String clusterId)
{
if (conf.sidecarBehindLoadBalancer)
{
LOGGER.info("Using LoadBalancedCassandraClusterInfo for load-balanced Sidecar. clusterId={}", clusterId);
return new LoadBalancedCassandraClusterInfo(conf, clusterId);
}
return new CassandraClusterInfo(conf, clusterId);
}

/**
* Reconstructs a {@link CassandraClusterInfo} on an executor from broadcast, selecting the concrete type based on
* the {@link BulkSparkConf#sidecarBehindLoadBalancer} flag so it matches the driver-side selection in
* {@link #create(BulkSparkConf, String)}.
*
* @param broadcastable the broadcastable cluster info from broadcast
* @return {@link LoadBalancedCassandraClusterInfo} when Sidecar is behind a load balancer, otherwise a plain
* {@link CassandraClusterInfo}
*/
public static CassandraClusterInfo create(BroadcastableClusterInfo broadcastable)
{
if (broadcastable.getConf().sidecarBehindLoadBalancer)
{
return new LoadBalancedCassandraClusterInfo(broadcastable);
}
return new CassandraClusterInfo(broadcastable);
}

@Override
public void checkBulkWriterIsEnabledOrThrow()
{
Expand Down Expand Up @@ -246,16 +297,7 @@ void validateTimeSkewWithLocalNow(Range<BigInteger> range, Instant localNow) thr
TimeSkewResponse timeSkew;
try
{
TokenRangeMapping<RingInstance> topology = getTokenRangeMapping(true);
List<SidecarInstance> instances = topology.getSubRanges(range)
.asMapOfRanges()
.values()
.stream()
.flatMap(Collection::stream)
.distinct() // remove duplications
.map(replica -> new SidecarInstanceImpl(replica.nodeName(), getCassandraContext().sidecarPort()))
.collect(Collectors.toList());
timeSkew = getCassandraContext().getSidecarClient().timeSkew(instances).get();
timeSkew = fetchTimeSkew(range).get();
}
catch (InterruptedException | ExecutionException exception)
{
Expand All @@ -270,6 +312,26 @@ void validateTimeSkewWithLocalNow(Range<BigInteger> range, Instant localNow) thr
}
}

/**
* Fetches time-skew information from Sidecar. The default implementation queries the replicas
* that own {@code range}. Subclasses may override to target a different endpoint set — for
* example, coordinated writes route through shared contact points when replica FQDNs are not
* reachable from Spark executors.
*/
protected CompletableFuture<TimeSkewResponse> fetchTimeSkew(Range<BigInteger> range)
{
TokenRangeMapping<RingInstance> topology = getTokenRangeMapping(true);
List<SidecarInstance> instances = topology.getSubRanges(range)
.asMapOfRanges()
.values()
.stream()
.flatMap(Collection::stream)
.distinct() // remove duplications
.map(replica -> new SidecarInstanceImpl(replica.nodeName(), getCassandraContext().sidecarPort()))
.collect(Collectors.toList());
return getCassandraContext().getSidecarClient().timeSkew(instances);
}

@Override
public synchronized void refreshClusterInfo()
{
Expand Down
Original file line number Diff line number Diff line change
@@ -0,0 +1,60 @@
/*
* Licensed to the Apache Software Foundation (ASF) under one
* or more contributor license agreements. See the NOTICE file
* distributed with this work for additional information
* regarding copyright ownership. The ASF licenses this file
* to you under the Apache License, Version 2.0 (the
* "License"); you may not use this file except in compliance
* with the License. You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing,
* software distributed under the License is distributed on an
* "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
* KIND, either express or implied. See the License for the
* specific language governing permissions and limitations
* under the License.
*/

package org.apache.cassandra.spark.bulkwriter;

import java.math.BigInteger;
import java.util.concurrent.CompletableFuture;

import com.google.common.collect.Range;

import o.a.c.sidecar.client.shaded.common.response.TimeSkewResponse;

/**
* Variant of {@link CassandraClusterInfo} used when Sidecars are fronted by a load balancer.
* Replica FQDNs from the token map are not routable from Spark executors in that topology,
* so requests that would otherwise fan out to per-replica Sidecar addresses are routed through
* the configured contact points (load balancer endpoints) instead.
* <p>
* This applies to both single-cluster and coordinated writes. It is selected by
* {@link CassandraClusterInfo#create(BulkSparkConf, String)} and
* {@link CassandraClusterInfo#create(BroadcastableClusterInfo)} when
* {@link org.apache.cassandra.spark.bulkwriter.WriterOptions#SIDECAR_BEHIND_LOAD_BALANCER}
* is set.
*/
public class LoadBalancedCassandraClusterInfo extends CassandraClusterInfo
{
public LoadBalancedCassandraClusterInfo(BulkSparkConf conf, String clusterId)
{
super(conf, clusterId);
}

public LoadBalancedCassandraClusterInfo(BroadcastableClusterInfo broadcastable)
{
super(broadcastable);
}

@Override
protected CompletableFuture<TimeSkewResponse> fetchTimeSkew(Range<BigInteger> range)
{
// range is irrelevant; the load balancer contact points are queried directly rather
// than the per-range replicas, whose FQDNs are not routable from Spark executors.
return getCassandraContext().getSidecarClient().timeSkew();

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The same problem exists for Single cluster behind a load balancer. Hence instead of using 'Coordinated' word, good to develop this as a generic framework for handling cluster/clusters behind a load balancer, then we can invoke it from anywhere needed

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Also, range param is unused, hence mention a comment something like 'range is irrelevant; contact points are queried directly'

Copy link
Copy Markdown
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Made it generic: the load-balanced variant is now selected by CassandraClusterInfo.create(...), which is invoked from both the single-cluster path (CassandraBulkWriterContext / BroadcastableClusterInfo.reconstruct) and the coordinated path (CassandraClusterInfoGroup), so it can be reused wherever needed.

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Nice work

}
}
Original file line number Diff line number Diff line change
Expand Up @@ -154,4 +154,13 @@ public enum WriterOptions implements WriterOption
* </ul>
*/
STORAGE_CREDENTIAL_TYPE,
/**
* Option declaring that the Sidecar cluster is fronted by a load balancer, so replica FQDNs
* from the token map are not directly routable from Spark executors. When {@code true},
* requests that would otherwise fan out to per-replica Sidecar addresses are routed through
* the configured contact points instead. Defaults to {@code false}.
* <p>
* Today this affects time-skew validation, and is honored for both single-cluster and coordinated writes.
*/
SIDECAR_BEHIND_LOAD_BALANCER,
}
Original file line number Diff line number Diff line change
Expand Up @@ -95,7 +95,7 @@ public class CassandraClusterInfoGroup implements ClusterInfo, MultiClusterSuppo
*/
public static CassandraClusterInfoGroup fromBulkSparkConf(BulkSparkConf conf)
{
return fromBulkSparkConf(conf, clusterId -> new CassandraClusterInfo(conf, clusterId));
return fromBulkSparkConf(conf, clusterId -> CassandraClusterInfo.create(conf, clusterId));
}

/**
Expand Down Expand Up @@ -170,7 +170,7 @@ private CassandraClusterInfoGroup(BroadcastableClusterInfoGroup broadcastable)
// Build list of ClusterInfo from broadcastable data
List<ClusterInfo> clusterInfosList = new ArrayList<>();
broadcastable.forEach((clusterId, broadcastableInfo) -> {
clusterInfosList.add(new CassandraClusterInfo((BroadcastableClusterInfo) broadcastableInfo));
clusterInfosList.add(CassandraClusterInfo.create((BroadcastableClusterInfo) broadcastableInfo));
});
this.clusterInfos = Collections.unmodifiableList(clusterInfosList);

Expand Down
Original file line number Diff line number Diff line change
Expand Up @@ -24,29 +24,39 @@
import java.util.ArrayList;
import java.util.Collections;
import java.util.List;
import java.util.Map;
import java.util.concurrent.CompletableFuture;
import java.util.concurrent.TimeUnit;
import java.util.stream.Stream;

import com.google.common.collect.ImmutableMap;
import com.google.common.collect.Maps;
import com.google.common.util.concurrent.Uninterruptibles;
import org.junit.jupiter.api.Test;
import org.junit.jupiter.api.Timeout;
import org.junit.jupiter.params.ParameterizedTest;
import org.junit.jupiter.params.provider.Arguments;
import org.junit.jupiter.params.provider.MethodSource;

import o.a.c.sidecar.client.shaded.client.SidecarClient;
import o.a.c.sidecar.client.shaded.common.response.NodeSettings;
import o.a.c.sidecar.client.shaded.common.response.TimeSkewResponse;
import org.apache.cassandra.bridge.CassandraVersion;
import org.apache.cassandra.spark.bulkwriter.token.TokenRangeMapping;
import org.apache.cassandra.spark.data.partitioner.Partitioner;
import org.apache.cassandra.spark.exception.TimeSkewTooLargeException;
import org.apache.spark.SparkConf;

import static org.apache.cassandra.spark.TestUtils.range;
import static org.assertj.core.api.Assertions.assertThat;
import static org.assertj.core.api.Assertions.assertThatNoException;
import static org.assertj.core.api.Assertions.assertThatThrownBy;
import static org.mockito.ArgumentMatchers.any;
import static org.mockito.ArgumentMatchers.anyList;
import static org.mockito.Mockito.RETURNS_DEEP_STUBS;
import static org.mockito.Mockito.mock;
import static org.mockito.Mockito.never;
import static org.mockito.Mockito.verify;
import static org.mockito.Mockito.when;

public class CassandraClusterInfoTest
Expand All @@ -63,6 +73,107 @@ void testTimeSkewAcceptable()
.isThrownBy(() -> ci.validateTimeSkewWithLocalNow(range(10, 20), localNow));
}

@Test
void testLoadBalancedClusterInfoUsesSidecarClientContactPointsForTimeSkew()
{
Instant localNow = Instant.now();
int allowanceMinutes = 10;

SidecarClient sidecarClient = mock(SidecarClient.class);
CassandraContext ctx = mock(CassandraContext.class);
when(ctx.getSidecarClient()).thenReturn(sidecarClient);
TimeSkewResponse tsr = new TimeSkewResponse(localNow.toEpochMilli(), allowanceMinutes);
when(sidecarClient.timeSkew()).thenReturn(CompletableFuture.completedFuture(tsr));

CassandraClusterInfo ci = new LoadBalancedCassandraClusterInfo((BulkSparkConf) null, null)
{
@Override
protected CassandraContext buildCassandraContext()
{
return ctx;
}

@Override
public TokenRangeMapping<RingInstance> getTokenRangeMapping(boolean cached)
{
return TokenRangeMappingUtils.buildTokenRangeMapping(0, ImmutableMap.of("dc1", 3), 5);
}
};

assertThatNoException()
.describedAs("Load-balanced cluster info must use timeSkew() so load balancer contact points stay reachable")
.isThrownBy(() -> ci.validateTimeSkewWithLocalNow(range(10, 20), localNow));
verify(sidecarClient).timeSkew();
verify(sidecarClient, never()).timeSkew(anyList());
}

@Test
void testCreateSelectsLoadBalancedClusterInfoWhenSidecarBehindLoadBalancer()

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Can you please write test cases to ensure

  • Coordinated case creates LoadBalancedCassandraClusterInfo for each cluster when LB is enabled
  • BroadcastableClusterInfo creates LoadBalancedCassandraClusterInfo on executors

Copy link
Copy Markdown
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Done, added the tests. The coordinated one checks fromBulkSparkConf(...) builds a LoadBalancedCassandraClusterInfo per cluster when the flag is on, and the executor one checks BroadcastableClusterInfo.reconstruct() returns one too. Also added the flag-off cases on both paths to catch the inverse regression.

{
// Single-cluster path: the SIDECAR_BEHIND_LOAD_BALANCER flag must be honored, not just for coordinated writes
try (CassandraClusterInfo ci = CassandraClusterInfo.create(bulkSparkConf(true)))
{
assertThat(ci)
.describedAs("Sidecar behind a load balancer must select the load-balanced variant on the single-cluster path")
.isExactlyInstanceOf(LoadBalancedCassandraClusterInfo.class);
}
}

@Test
void testCreateSelectsPlainClusterInfoByDefault()
{
try (CassandraClusterInfo ci = CassandraClusterInfo.create(bulkSparkConf(false)))
{
assertThat(ci)
.describedAs("Without the load balancer flag, the plain per-replica variant must be selected")
.isExactlyInstanceOf(CassandraClusterInfo.class);
}
}

@Test
void testBroadcastableClusterInfoReconstructsLoadBalancedClusterInfoOnExecutorWhenSidecarBehindLoadBalancer()
{
// Executor path: reconstructing from a broadcastable whose conf has the flag set must yield the load-balanced
// variant, matching the driver-side selection so executors also route through the load balancer contact points.
ClusterInfo source = mock(ClusterInfo.class);
when(source.getPartitioner()).thenReturn(Partitioner.Murmur3Partitioner);
when(source.getBridgeVersion()).thenReturn(CassandraVersion.FIVEZERO);
BroadcastableClusterInfo broadcastable = BroadcastableClusterInfo.from(source, bulkSparkConf(true));
try (CassandraClusterInfo ci = (CassandraClusterInfo) broadcastable.reconstruct())
{
assertThat(ci)
.describedAs("Reconstructing on an executor with Sidecar behind a load balancer must select the load-balanced variant")
.isExactlyInstanceOf(LoadBalancedCassandraClusterInfo.class);
}
}

@Test
void testBroadcastableClusterInfoReconstructsPlainClusterInfoOnExecutorByDefault()
{
ClusterInfo source = mock(ClusterInfo.class);
when(source.getPartitioner()).thenReturn(Partitioner.Murmur3Partitioner);
when(source.getBridgeVersion()).thenReturn(CassandraVersion.FIVEZERO);
BroadcastableClusterInfo broadcastable = BroadcastableClusterInfo.from(source, bulkSparkConf(false));
try (CassandraClusterInfo ci = (CassandraClusterInfo) broadcastable.reconstruct())
{
assertThat(ci)
.describedAs("Without the load balancer flag, reconstructing on an executor must use the plain per-replica variant")
.isExactlyInstanceOf(CassandraClusterInfo.class);
}
}

private static BulkSparkConf bulkSparkConf(boolean sidecarBehindLoadBalancer)
{
// No keystore options: this keeps SSL disabled so create() can build a real (non-TLS) Sidecar client
// offline. We only assert on the concrete type create() selects, never issuing a request.
Map<String, String> options = Maps.newTreeMap(String.CASE_INSENSITIVE_ORDER);
options.put(WriterOptions.SIDECAR_CONTACT_POINTS.name(), "127.0.0.1");
options.put(WriterOptions.KEYSPACE.name(), "ks");
options.put(WriterOptions.TABLE.name(), "table");
options.put(WriterOptions.SIDECAR_BEHIND_LOAD_BALANCER.name(), String.valueOf(sidecarBehindLoadBalancer));
return new BulkSparkConf(new SparkConf(), options);
}

@Test
void testTimeSkewTooLarge()
{
Expand Down
Loading