Skip to content

When decommission is attempted again, clear any existing transferred_ranges to force re-streaming and avoid consistency violation #5000

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
Jollyplum wants to merge 1 commit into
base: cassandra-5.0
Choose a base branch
from
Open

When decommission is attempted again, clear any existing transferred_ranges to force re-streaming and avoid consistency violation #5000

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
24 changes: 22 additions & 2 deletions src/java/org/apache/cassandra/db/SystemKeyspace.java
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1726,6 +1726,18 @@ public static void resetAvailableStreamedRangesForKeyspace(String keyspace)
executeInternal(format(cql, SchemaConstants.SYSTEM_KEYSPACE_NAME, AVAILABLE_RANGES_V2), keyspace);
}

/**
* Wipes the whole table rather than deleting per (operation, keyspace): keyspace_name is part of the
* partition key, so an operation-scoped delete would have to iterate keyspaces. Safe because the
* decommission path is the only reader - see getTransferredRanges. Revisit if another topology
* change starts consulting this table.
*/
public static synchronized void resetTransferredRanges()
{
for (String table : Arrays.asList(LEGACY_TRANSFERRED_RANGES, TRANSFERRED_RANGES_V2))
Keyspace.open(SchemaConstants.SYSTEM_KEYSPACE_NAME).getColumnFamilyStore(table).truncateBlockingWithoutSnapshot();
}

public static synchronized void updateTransferredRanges(StreamOperation streamOperation,
InetAddressAndPort peer,
String keyspace,
Expand All @@ -1742,11 +1754,19 @@ public static synchronized void updateTransferredRanges(StreamOperation streamOp
executeInternal(String.format(cql, TRANSFERRED_RANGES_V2), rangesToUpdate, streamOperation.getDescription(), peer.getAddress(), peer.getPort(), keyspace);
}

public static synchronized Map<InetAddressAndPort, Set<Range<Token>>> getTransferredRanges(String description, String keyspace, IPartitioner partitioner)
// Only consulted on the decommission path, where the leaving node is the only streamer and its
// local transferred_ranges_v2 is therefore a complete record of what has already moved.
//
// Being node-local rules out the other topology changes. Under removenode the surviving replicas
// stream, and the node running it need not be one of them, so its local table may record nothing;
// even when it is a replica it can only account for its own streams. Move could use it for the
// ranges the moving node gives up, but the move path registers no listener so nothing is recorded,
// and it would still miss the ranges moving the other way.
public static synchronized Map<InetAddressAndPort, Set<Range<Token>>> getTransferredRanges(StreamOperation streamOperation, String keyspace, IPartitioner partitioner)
{
Map<InetAddressAndPort, Set<Range<Token>>> result = new HashMap<>();
String query = "SELECT * FROM system.%s WHERE operation = ? AND keyspace_name = ?";
UntypedResultSet rs = executeInternal(String.format(query, TRANSFERRED_RANGES_V2), description, keyspace);
UntypedResultSet rs = executeInternal(String.format(query, TRANSFERRED_RANGES_V2), streamOperation.getDescription(), keyspace);
for (UntypedResultSet.Row row : rs)
{
InetAddress peerAddress = row.getInetAddress("peer");
Expand Down
19 changes: 16 additions & 3 deletions src/java/org/apache/cassandra/service/StorageService.java
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -278,6 +278,7 @@
import static org.apache.cassandra.service.StorageService.Mode.DECOMMISSIONED;
import static org.apache.cassandra.service.StorageService.Mode.DECOMMISSION_FAILED;
import static org.apache.cassandra.service.StorageService.Mode.JOINING_FAILED;
import static org.apache.cassandra.service.StorageService.Mode.NORMAL;
import static org.apache.cassandra.utils.Clock.Global.currentTimeMillis;
import static org.apache.cassandra.utils.Clock.Global.nanoTime;
import static org.apache.cassandra.utils.FBUtilities.getBroadcastAddressAndPort;
Expand Down Expand Up @@ -5331,6 +5332,19 @@ public void decommission(boolean force) throws InterruptedException
{
PendingRangeCalculatorService.instance.blockUntilFinished();

// This check needs to happen above startLeaving
boolean resumingInFlightDecommission = tokenMetadata.isLeaving(FBUtilities.getBroadcastAddressAndPort())
&& operationMode != NORMAL;

// We reset transferred ranges upon starting a new decommission so that we fully stream
// anything written since a previous attempt, which may not have been persisted to a pending endpoint.
// See CASSANDRA-16290.
if (!resumingInFlightDecommission)
{
logger.info("resetting transferred ranges to force re-streaming");
SystemKeyspace.resetTransferredRanges();
}

String dc = DatabaseDescriptor.getEndpointSnitch().getLocalDatacenter();

// If we're already decommissioning there is no point checking RF/pending ranges
Expand Down Expand Up @@ -6385,8 +6399,7 @@ private Future<StreamState> streamRanges(Map<String, EndpointsByReplica> rangesT
if (rangesWithEndpoints.isEmpty())
continue;

//Description is always Unbootstrap? Is that right?
Map<InetAddressAndPort, Set<Range<Token>>> transferredRangePerKeyspace = SystemKeyspace.getTransferredRanges("Unbootstrap",
Map<InetAddressAndPort, Set<Range<Token>>> transferredRangePerKeyspace = SystemKeyspace.getTransferredRanges(StreamOperation.DECOMMISSION,
keyspace,
StorageService.instance.getTokenMetadata().partitioner);
RangesByEndpoint.Builder replicasPerEndpoint = new RangesByEndpoint.Builder();
Expand All @@ -6397,7 +6410,7 @@ private Future<StreamState> streamRanges(Map<String, EndpointsByReplica> rangesT
Set<Range<Token>> transferredRanges = transferredRangePerKeyspace.get(remote.endpoint());
if (transferredRanges != null && transferredRanges.contains(local.range()))
{
logger.debug("Skipping transferred range {} of keyspace {}, endpoint {}", local, keyspace, remote);
logger.info("Skipping transferred range {} of keyspace {}, endpoint {}", local, keyspace, remote);
continue;
}

Expand Down
101 changes: 100 additions & 1 deletion test/distributed/org/apache/cassandra/distributed/test/DecommissionTest.java
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -18,9 +18,18 @@

package org.apache.cassandra.distributed.test;

import java.io.IOException;
import java.nio.ByteBuffer;
import java.util.Arrays;
import java.util.List;
import java.util.concurrent.Callable;
import java.util.function.Supplier;
import java.util.stream.Collectors;

import org.apache.cassandra.dht.Murmur3Partitioner;
import org.apache.cassandra.distributed.api.ConsistencyLevel;
import org.apache.cassandra.distributed.shared.ClusterUtils;
import org.apache.cassandra.utils.concurrent.ImmediateFuture;
import org.junit.Test;

import net.bytebuddy.ByteBuddy;
Expand All @@ -36,9 +45,11 @@
import static net.bytebuddy.matcher.ElementMatchers.named;
import static org.apache.cassandra.db.SystemKeyspace.BootstrapState.COMPLETED;
import static org.apache.cassandra.db.SystemKeyspace.BootstrapState.DECOMMISSIONED;
import static org.apache.cassandra.db.SystemKeyspace.TRANSFERRED_RANGES_V2;
import static org.apache.cassandra.distributed.api.Feature.GOSSIP;
import static org.apache.cassandra.distributed.api.Feature.NETWORK;
import static org.apache.cassandra.distributed.shared.ClusterUtils.stopUnchecked;
import static org.apache.cassandra.distributed.test.ring.BootstrapTest.populate;
import static org.apache.cassandra.service.StorageService.Mode.DECOMMISSION_FAILED;
import static org.apache.cassandra.service.StorageService.Mode.NORMAL;
import static org.junit.Assert.assertEquals;
Expand Down Expand Up @@ -186,8 +197,80 @@ public void testDecommissionAfterNodeRestart() throws Throwable
}


@Test
public void testAbortingDecommissionRestreams() throws Exception
{
// https://issues.apache.org/jira/browse/CASSANDRA-16290
// We demonstrate here that decommissioning and then aborting decommission is unsafe
// if we've persisted transferred ranges and then skip them for something which was delivered after we aborted the decommission but before we resumed
try (Cluster cluster = builder().withNodes(4)
.withConfig(config -> config.with(NETWORK, GOSSIP)
// disable hints to simplify test
.set("hinted_handoff_enabled", false)
)
// only install on the first generation of node2: a restarted instance gets a
// fresh classloader, so any static "already failed once" state would be reset
// and the resumed decommission would fail again
.withInstanceInitializer((cl, threadGroup, num, generation) -> {
if (num == 2 && generation == 0)
BB.streamHintsInstall(cl);
})
.start())
{
// We need blob columns here so later we can do Murmur3Partitioner.LongToken.keyForToken(token);
populate(cluster, 0, 100, 1, 2, ConsistencyLevel.QUORUM, "pk blob, ck blob, v blob");

IInvokableInstance leavingNode = cluster.get(2);

leavingNode.nodetoolResult("decommission").asserts().failure();
leavingNode.runOnInstance(() -> assertEquals(DECOMMISSION_FAILED, StorageService.Mode.valueOf(StorageService.instance.getOperationMode())));

// abort the decommission
ClusterUtils.stopUnchecked(leavingNode);
ClusterUtils.start(leavingNode, props -> {});
ClusterUtils.awaitRingHealthy(leavingNode);

// Stop the non leaving nodes so we can write at ONE and fail to stream that datum
ClusterUtils.stopUnchecked(cluster.get(1));
ClusterUtils.stopUnchecked(cluster.get(3));
ClusterUtils.stopUnchecked(cluster.get(4));

List<Murmur3Partitioner.LongToken> tokens = ClusterUtils.getLocalTokens(leavingNode).stream().map(t -> new Murmur3Partitioner.LongToken(Long.parseLong(t))).collect(Collectors.toList());
for (Murmur3Partitioner.LongToken token : tokens)
{
ByteBuffer key = Murmur3Partitioner.LongToken.keyForToken(token);
leavingNode.coordinator().execute("INSERT INTO " + KEYSPACE + ".tbl (pk, ck, v) VALUES (?, ?, ?)", ConsistencyLevel.ONE, key, key, key);
}

ClusterUtils.start(cluster.get(1), props -> {});
ClusterUtils.start(cluster.get(3), props -> {});
ClusterUtils.start(cluster.get(4), props -> {});

ClusterUtils.awaitRingHealthy(leavingNode);

Object[][] ranges = leavingNode.executeInternal("SELECT keyspace_name from system." + TRANSFERRED_RANGES_V2);

assertTrue("transferred ranges missing entirely", ranges.length > 0);
assertTrue("transferred ranges present for keyspace", Arrays.stream(ranges).anyMatch(x -> x[0].equals(KEYSPACE)));

// Resume decomm
leavingNode.nodetoolResult("decommission").asserts().success();

// Try and read data we wrote at ONE at ALL
for (Murmur3Partitioner.LongToken token : tokens)
{
ByteBuffer key = Murmur3Partitioner.LongToken.keyForToken(token);
Object[][] resp = cluster.get(1).coordinator().execute("SELECT pk from " + KEYSPACE + ".tbl where pk=?", ConsistencyLevel.ALL, key);
assertTrue("We should get a response for this key we wrote it at ONE", resp.length > 0);
assertEquals(key, resp[0][0]);
}
}
}

public static class BB
{
private static int invocations = 0;

public static void install(ClassLoader classLoader, Integer num)
{
new ByteBuddy().rebase(StorageService.class)
Expand All @@ -197,7 +280,15 @@ public static void install(ClassLoader classLoader, Integer num)
.load(classLoader, ClassLoadingStrategy.Default.INJECTION);
}

private static int invocations = 0;
static void streamHintsInstall(ClassLoader cl)
{
new ByteBuddy().rebase(StorageService.class)
.method(named("streamHints"))
.intercept(MethodDelegation.to(BB.class))
.make()
.load(cl, ClassLoadingStrategy.Default.INJECTION);
}


@SuppressWarnings("unused")
public static Supplier<Future<StreamState>> prepareUnbootstrapStreaming(@SuperCall Callable<Supplier<Future<StreamState>>> zuper)
Expand All @@ -216,5 +307,13 @@ public static Supplier<Future<StreamState>> prepareUnbootstrapStreaming(@SuperCa
throw new RuntimeException(e);
}
}

@SuppressWarnings("unused")
public static Future<?> streamHints(@SuperCall Callable<Future<?>> zuper)
{
// this is only installed on the first startup of the leaving node, so every invocation
// here belongs to the decommission attempt we want to fail at the last possible moment
return ImmediateFuture.failure(new IOException("failing hints so that decomm fails at last moment possible"));
}
}
}
14 changes: 11 additions & 3 deletions test/distributed/org/apache/cassandra/distributed/test/ring/BootstrapTest.java
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -67,7 +67,6 @@
import org.apache.cassandra.distributed.shared.JMXUtil;
import org.apache.cassandra.distributed.shared.NetworkTopology;
import org.apache.cassandra.distributed.shared.WithProperties;
import org.apache.cassandra.distributed.test.DecommissionTest;
import org.apache.cassandra.distributed.test.TestBaseImpl;
import org.apache.cassandra.schema.SchemaConstants;
import org.apache.cassandra.service.StorageService;
Expand Down Expand Up @@ -329,9 +328,14 @@ public static void populate(ICluster cluster, int from, int to)
}

public static void populate(ICluster cluster, int from, int to, int coord, int rf, ConsistencyLevel cl)
{
populate(cluster, from, to, coord, rf, cl, "pk int, ck int, v int");
}

public static void populate(ICluster cluster, int from, int to, int coord, int rf, ConsistencyLevel cl, String columnDefinitions)
{
cluster.schemaChange("CREATE KEYSPACE IF NOT EXISTS " + KEYSPACE + " WITH replication = {'class': 'SimpleStrategy', 'replication_factor': " + rf + "};");
cluster.schemaChange("CREATE TABLE IF NOT EXISTS " + KEYSPACE + ".tbl (pk int, ck int, v int, PRIMARY KEY (pk, ck))");
cluster.schemaChange("CREATE TABLE IF NOT EXISTS " + KEYSPACE + ".tbl (" + columnDefinitions + ", PRIMARY KEY (pk, ck))");
for (int i = from; i < to; i++)
{
cluster.coordinator(coord).execute("INSERT INTO " + KEYSPACE + ".tbl (pk, ck, v) VALUES (?, ?, ?)",
Expand All @@ -356,9 +360,13 @@ public static void install(ClassLoader classLoader, Integer num)
{
return;
}
// bootstrapFinished is implemented by this class, so delegate here. The previous
// target (DecommissionTest.BB) was wrong but bound anyway while that class had
// exactly one @SuperCall interceptor method; CASSANDRA-16290 adds a second one,
// which makes the binding ambiguous.
new ByteBuddy().rebase(StorageService.class)
.method(named("bootstrapFinished"))
.intercept(MethodDelegation.to(DecommissionTest.BB.class))
.intercept(MethodDelegation.to(BB.class))
.make()
.load(classLoader, ClassLoadingStrategy.Default.INJECTION);
}
Expand Down