Skip to content

fix(inkless:consolidation): reject follower reads below the cross-tier earliest [KC-332] #720

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

Draft
viktorsomogyi wants to merge 1 commit into
base: svv/ts-unification-reclaim-floor
Choose a base branch
from
+310 −1
Draft

fix(inkless:consolidation): reject follower reads below the cross-tier earliest [KC-332] #720

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
31 changes: 30 additions & 1 deletion core/src/main/scala/kafka/server/ReplicaManager.scala
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2383,7 +2383,36 @@ class ReplicaManager(val config: KafkaConfig,
getPartitionOrError(tp.topicPartition) match {
case Right(partition) =>
val logEndOffset = partition.log.map(_.logEndOffset).getOrElse(0L)
if (fetchPartitionData.fetchOffset < logEndOffset) {
// A follower's local logStartOffset stays frozen at the switch. DeleteRecords and
// retention advance only the real leader's log start and the control-plane cross-tier
// earliest, never a follower's. With managed replicas the metadata transformer routes
// consumers to a hash-selected replica (usually a follower), which then serves the read
// from that stale local log, so the consumer could read records below the authoritative
// earliest. Reject those with OFFSET_OUT_OF_RANGE so the consumer resets via
// ListOffsets(EARLIEST). This covers every consumer fetch, not just pre-KIP-392 ones: a
// modern consumer is served from the follower too (allowReplica is true for it once the
// transformer points it at that replica). The leader is skipped since its own local
// logStartOffset already advanced. Best effort: the cache is only ever stale-low, so a
// few deleted records may survive until it refreshes.
val mayServeFromFollowerLocalLog =
params.isFromConsumer && !partition.isLeader
val crossTierEarliest =
if (mayServeFromFollowerLocalLog) crossTierEarliestOffset(tp.topicPartition())
else OptionalLong.empty()
if (crossTierEarliest.isPresent && fetchPartitionData.fetchOffset < crossTierEarliest.getAsLong) {
immediateFetchResponses += tp -> new FetchPartitionData(
Errors.OFFSET_OUT_OF_RANGE,
UnifiedLog.UNKNOWN_OFFSET,
UnifiedLog.UNKNOWN_OFFSET,
MemoryRecords.EMPTY,
Optional.empty(),
OptionalLong.empty(),
Optional.empty(),
OptionalInt.empty(),
false
)
partitionLookupFailed = true
} else if (fetchPartitionData.fetchOffset < logEndOffset) {
// Local log has data for this offset range — serve from local, track for diskless supplement
shouldReadFromUnifiedLog = true
// Skip supplement tracking when the fetch offset falls in the tiered-storage range
Expand Down
280 changes: 280 additions & 0 deletions core/src/test/scala/unit/kafka/server/ReplicaManagerInklessTest.scala
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -45,6 +45,8 @@ import org.apache.kafka.common.metadata.{PartitionChangeRecord, PartitionRecord,
import org.apache.kafka.common.metrics.Metrics
import org.apache.kafka.common.protocol.Errors
import org.apache.kafka.common.record._
import org.apache.kafka.common.replica.ClientMetadata.DefaultClientMetadata
import org.apache.kafka.common.security.auth.KafkaPrincipal
import org.apache.kafka.common.requests._
import org.apache.kafka.common.requests.FetchRequest.PartitionData
import org.apache.kafka.common.requests.ProduceResponse.PartitionResponse
Expand Down Expand Up @@ -72,6 +74,7 @@ import org.mockito.Mockito._
import org.mockito.{Answers, ArgumentMatchers, MockedConstruction, Mockito}

import java.io.File
import java.net.InetAddress
import java.util
import java.util.{Collections, Optional, OptionalInt, OptionalLong, Properties}
import java.util.concurrent.{CompletableFuture, CountDownLatch, TimeUnit}
Expand Down Expand Up @@ -1730,6 +1733,23 @@ class ReplicaManagerInklessTest {
ArgumentMatchers.eq(tp.topicPartition()))
}

// Like stubConsolidatingPartitionWithLocalLeoFor but also pins leadership (the read-path
// cross-tier-earliest guard only fires for followers) and localLogStartOffset.
private def stubConsolidatingPartitionForFollowerReadGuard(replicaManager: ReplicaManager,
tp: TopicIdPartition,
localLeo: Long,
isLeader: Boolean,
localLogStartOffset: Long = 0L): Unit = {
val mockPartition = mock(classOf[Partition])
val mockLog = mock(classOf[UnifiedLog])
when(mockLog.logEndOffset).thenReturn(localLeo)
when(mockLog.localLogStartOffset).thenReturn(localLogStartOffset)
when(mockPartition.log).thenReturn(Some(mockLog))
when(mockPartition.isLeader).thenReturn(isLeader)
doReturn(Right(mockPartition)).when(replicaManager).getPartitionOrError(
ArgumentMatchers.eq(tp.topicPartition()))
}

private def stubConsolidatingPartitionWithoutLocalLog(replicaManager: ReplicaManager): Unit = {
val mockPartition = mock(classOf[Partition])
when(mockPartition.log).thenReturn(None)
Expand Down Expand Up @@ -1796,6 +1816,266 @@ class ReplicaManagerInklessTest {



// The metadata transformer can route an older consumer (no rackId -> empty clientMetadata) to a
// follower whose local logStartOffset is frozen at the classic-to-diskless seal. DeleteRecords /
// retention only advance the control-plane cross-tier earliest, so a fetch below it must be rejected
// as OFFSET_OUT_OF_RANGE (consumer then resets via ListOffsets(EARLIEST)) instead of serving
// logically-deleted records from the follower's stale local log.
@Test
def testFetchConsolidatingRejectsFollowerReadBelowCrossTierEarliest(): Unit = {
val fetchHandlerCtor = mockFetchHandler(Map.empty)
val cp = mock(classOf[ControlPlane])
val cache = mock(classOf[CrossTierLogStartCache])
when(cache.get(any())).thenReturn(java.lang.Long.valueOf(100L))
val replicaManager = spy(createReplicaManager(
List(disklessTopicPartition.topic()),
controlPlane = Some(cp),
disklessManagedReplicasEnabled = true,
disklessRemoteStorageConsolidationEnabled = true,
consolidatingDisklessTopics = Set(disklessTopicPartition.topic()),
topicIdMapping = Map(disklessTopicPartition.topic() -> disklessTopicPartition.topicId()),
crossTierLogStartCache = Some(cache),
))
try {
when(replicaManager.inklessMetadataView().getClassicToDisklessStartOffset(disklessTopicPartition.topicPartition()))
.thenReturn(100L)
stubConsolidatingPartitionForFollowerReadGuard(replicaManager, disklessTopicPartition, localLeo = 200L, isLeader = false)

// Older consumer: replicaId = -1 and empty clientMetadata.
val fetchParams = new FetchParams(
-1, -1L,
0L, 1, 1024, FetchIsolation.HIGH_WATERMARK, Optional.empty()
)
val fetchInfos = Seq(
disklessTopicPartition -> new PartitionData(disklessTopicPartition.topicId(), 50L, 0L, 1024, Optional.empty())
)

@volatile var responseData: Map[TopicIdPartition, FetchPartitionData] = null
val responseCallback = (response: Seq[(TopicIdPartition, FetchPartitionData)]) => responseData = response.toMap
replicaManager.fetchMessages(fetchParams, fetchInfos, QuotaFactory.UNBOUNDED_QUOTA, responseCallback)

waitForFetchResponse(responseData)
assertEquals(1, responseData.size)
assertEquals(Errors.OFFSET_OUT_OF_RANGE, responseData(disklessTopicPartition).error)
assertEquals(MemoryRecords.EMPTY, responseData(disklessTopicPartition).records)
verify(replicaManager, never()).readFromLog(any(), any(), any(), any())
verify(fetchHandlerCtor.constructed().get(0), never()).handle(any(), any())
} finally {
replicaManager.shutdown(checkpointHW = false)
fetchHandlerCtor.close()
}
}

// At or above the cross-tier earliest the follower read is legitimate and served from the local log.
@Test
def testFetchConsolidatingServesFollowerReadAtOrAboveCrossTierEarliest(): Unit = {
val fetchHandlerCtor = mockFetchHandler(Map.empty)
val cp = mock(classOf[ControlPlane])
val cache = mock(classOf[CrossTierLogStartCache])
when(cache.get(any())).thenReturn(java.lang.Long.valueOf(50L))
val replicaManager = spy(createReplicaManager(
List(disklessTopicPartition.topic()),
controlPlane = Some(cp),
disklessManagedReplicasEnabled = true,
disklessRemoteStorageConsolidationEnabled = true,
consolidatingDisklessTopics = Set(disklessTopicPartition.topic()),
topicIdMapping = Map(disklessTopicPartition.topic() -> disklessTopicPartition.topicId()),
crossTierLogStartCache = Some(cache),
))
try {
when(replicaManager.inklessMetadataView().getClassicToDisklessStartOffset(disklessTopicPartition.topicPartition()))
.thenReturn(100L)
stubConsolidatingPartitionForFollowerReadGuard(replicaManager, disklessTopicPartition, localLeo = 200L, isLeader = false)

doReturn(Seq(disklessTopicPartition ->
new LogReadResult(
new FetchDataInfo(new LogOffsetMetadata(1L, 0L, 0), RECORDS),
Optional.empty(), 10L, 0L, 10L, 0L, 0L, OptionalLong.empty(), Errors.NONE
))
).when(replicaManager).readFromLog(any(), any(), any(), any())

// fetchOffset (50) == cross-tier earliest (50): not below, so the read proceeds.
val fetchParams = new FetchParams(
-1, -1L,
0L, 1, 1024, FetchIsolation.HIGH_WATERMARK, Optional.empty()
)
val fetchInfos = Seq(
disklessTopicPartition -> new PartitionData(disklessTopicPartition.topicId(), 50L, 0L, 1024, Optional.empty())
)

@volatile var responseData: Map[TopicIdPartition, FetchPartitionData] = null
val responseCallback = (response: Seq[(TopicIdPartition, FetchPartitionData)]) => responseData = response.toMap
replicaManager.fetchMessages(fetchParams, fetchInfos, QuotaFactory.UNBOUNDED_QUOTA, responseCallback)

waitForFetchResponse(responseData)
assertEquals(1, responseData.size)
assertEquals(RECORDS, responseData(disklessTopicPartition).records)
verify(replicaManager, times(1)).readFromLog(any(), any(), any(), any())
verify(fetchHandlerCtor.constructed().get(0), never()).handle(any(), any())
} finally {
replicaManager.shutdown(checkpointHW = false)
fetchHandlerCtor.close()
}
}

// The leader's own local logStartOffset already advanced with DeleteRecords/retention, so the guard
// is scoped to followers: a leader read below the cross-tier earliest is served from local without
// even consulting the cross-tier earliest.
@Test
def testFetchConsolidatingLeaderReadNotGatedByCrossTierEarliest(): Unit = {
val fetchHandlerCtor = mockFetchHandler(Map.empty)
val cp = mock(classOf[ControlPlane])
val cache = mock(classOf[CrossTierLogStartCache])
// Would reject offset 50 if consulted -- it must not be.
when(cache.get(any())).thenReturn(java.lang.Long.valueOf(100L))
val replicaManager = spy(createReplicaManager(
List(disklessTopicPartition.topic()),
controlPlane = Some(cp),
disklessManagedReplicasEnabled = true,
disklessRemoteStorageConsolidationEnabled = true,
consolidatingDisklessTopics = Set(disklessTopicPartition.topic()),
topicIdMapping = Map(disklessTopicPartition.topic() -> disklessTopicPartition.topicId()),
crossTierLogStartCache = Some(cache),
))
try {
when(replicaManager.inklessMetadataView().getClassicToDisklessStartOffset(disklessTopicPartition.topicPartition()))
.thenReturn(100L)
stubConsolidatingPartitionForFollowerReadGuard(replicaManager, disklessTopicPartition, localLeo = 200L, isLeader = true)

doReturn(Seq(disklessTopicPartition ->
new LogReadResult(
new FetchDataInfo(new LogOffsetMetadata(1L, 0L, 0), RECORDS),
Optional.empty(), 10L, 0L, 10L, 0L, 0L, OptionalLong.empty(), Errors.NONE
))
).when(replicaManager).readFromLog(any(), any(), any(), any())

val fetchParams = new FetchParams(
-1, -1L,
0L, 1, 1024, FetchIsolation.HIGH_WATERMARK, Optional.empty()
)
val fetchInfos = Seq(
disklessTopicPartition -> new PartitionData(disklessTopicPartition.topicId(), 50L, 0L, 1024, Optional.empty())
)

@volatile var responseData: Map[TopicIdPartition, FetchPartitionData] = null
val responseCallback = (response: Seq[(TopicIdPartition, FetchPartitionData)]) => responseData = response.toMap
replicaManager.fetchMessages(fetchParams, fetchInfos, QuotaFactory.UNBOUNDED_QUOTA, responseCallback)

waitForFetchResponse(responseData)
assertEquals(1, responseData.size)
assertEquals(RECORDS, responseData(disklessTopicPartition).records)
verify(replicaManager, times(1)).readFromLog(any(), any(), any(), any())
verify(cache, never()).get(any())
} finally {
replicaManager.shutdown(checkpointHW = false)
fetchHandlerCtor.close()
}
}

// A modern consumer (non-empty clientMetadata) is served from the follower's local log too: once the
// transformer points it at a non-leader replica, allowReplica is true for it. So the guard must reject
// its below-earliest reads exactly like an older consumer's, not just pre-KIP-392 ones.
@Test
def testFetchConsolidatingRejectsModernFollowerReadBelowCrossTierEarliest(): Unit = {
val fetchHandlerCtor = mockFetchHandler(Map.empty)
val cp = mock(classOf[ControlPlane])
val cache = mock(classOf[CrossTierLogStartCache])
when(cache.get(any())).thenReturn(java.lang.Long.valueOf(100L))
val replicaManager = spy(createReplicaManager(
List(disklessTopicPartition.topic()),
controlPlane = Some(cp),
disklessManagedReplicasEnabled = true,
disklessRemoteStorageConsolidationEnabled = true,
consolidatingDisklessTopics = Set(disklessTopicPartition.topic()),
topicIdMapping = Map(disklessTopicPartition.topic() -> disklessTopicPartition.topicId()),
crossTierLogStartCache = Some(cache),
))
try {
when(replicaManager.inklessMetadataView().getClassicToDisklessStartOffset(disklessTopicPartition.topicPartition()))
.thenReturn(100L)
stubConsolidatingPartitionForFollowerReadGuard(replicaManager, disklessTopicPartition, localLeo = 200L, isLeader = false)

// Modern consumer: non-empty clientMetadata (rackId present).
val clientMetadata = new DefaultClientMetadata(
"rack-a", "client-id", InetAddress.getByName("localhost"), KafkaPrincipal.ANONYMOUS, "PLAINTEXT")
val fetchParams = new FetchParams(
-1, -1L,
0L, 1, 1024, FetchIsolation.HIGH_WATERMARK, Optional.of(clientMetadata)
)
val fetchInfos = Seq(
disklessTopicPartition -> new PartitionData(disklessTopicPartition.topicId(), 50L, 0L, 1024, Optional.empty())
)

@volatile var responseData: Map[TopicIdPartition, FetchPartitionData] = null
val responseCallback = (response: Seq[(TopicIdPartition, FetchPartitionData)]) => responseData = response.toMap
replicaManager.fetchMessages(fetchParams, fetchInfos, QuotaFactory.UNBOUNDED_QUOTA, responseCallback)

waitForFetchResponse(responseData)
assertEquals(1, responseData.size)
assertEquals(Errors.OFFSET_OUT_OF_RANGE, responseData(disklessTopicPartition).error)
assertEquals(MemoryRecords.EMPTY, responseData(disklessTopicPartition).records)
verify(replicaManager, never()).readFromLog(any(), any(), any(), any())
verify(fetchHandlerCtor.constructed().get(0), never()).handle(any(), any())
} finally {
replicaManager.shutdown(checkpointHW = false)
fetchHandlerCtor.close()
}
}

// Fail-open: when the cross-tier earliest is unavailable (cache miss + control plane returns nothing,
// e.g. a control-plane outage), the guard does not reject. The cache is only ever stale-low, so the
// worst case is a few logically-deleted records survive until the earliest resolves -- preferable to
// spuriously failing legitimate reads.
@Test
def testFetchConsolidatingServesFollowerReadWhenCrossTierEarliestUnavailable(): Unit = {
val fetchHandlerCtor = mockFetchHandler(Map.empty)
val cp = mock(classOf[ControlPlane])
val cache = mock(classOf[CrossTierLogStartCache])
// Cache miss; control-plane listOffsets left unstubbed -> returns null -> earliest resolves empty.
when(cache.get(any())).thenReturn(null.asInstanceOf[java.lang.Long])
val replicaManager = spy(createReplicaManager(
List(disklessTopicPartition.topic()),
controlPlane = Some(cp),
disklessManagedReplicasEnabled = true,
disklessRemoteStorageConsolidationEnabled = true,
consolidatingDisklessTopics = Set(disklessTopicPartition.topic()),
topicIdMapping = Map(disklessTopicPartition.topic() -> disklessTopicPartition.topicId()),
crossTierLogStartCache = Some(cache),
))
try {
when(replicaManager.inklessMetadataView().getClassicToDisklessStartOffset(disklessTopicPartition.topicPartition()))
.thenReturn(100L)
stubConsolidatingPartitionForFollowerReadGuard(replicaManager, disklessTopicPartition, localLeo = 200L, isLeader = false)

doReturn(Seq(disklessTopicPartition ->
new LogReadResult(
new FetchDataInfo(new LogOffsetMetadata(1L, 0L, 0), RECORDS),
Optional.empty(), 10L, 0L, 10L, 0L, 0L, OptionalLong.empty(), Errors.NONE
))
).when(replicaManager).readFromLog(any(), any(), any(), any())

val fetchParams = new FetchParams(
-1, -1L,
0L, 1, 1024, FetchIsolation.HIGH_WATERMARK, Optional.empty()
)
val fetchInfos = Seq(
disklessTopicPartition -> new PartitionData(disklessTopicPartition.topicId(), 50L, 0L, 1024, Optional.empty())
)

@volatile var responseData: Map[TopicIdPartition, FetchPartitionData] = null
val responseCallback = (response: Seq[(TopicIdPartition, FetchPartitionData)]) => responseData = response.toMap
replicaManager.fetchMessages(fetchParams, fetchInfos, QuotaFactory.UNBOUNDED_QUOTA, responseCallback)

waitForFetchResponse(responseData)
assertEquals(1, responseData.size)
assertEquals(RECORDS, responseData(disklessTopicPartition).records)
verify(replicaManager, times(1)).readFromLog(any(), any(), any(), any())
} finally {
replicaManager.shutdown(checkpointHW = false)
fetchHandlerCtor.close()
}
}

// When local log has data but below minBytes and diskless data is available, the fetch must
// respond immediately (merged local+diskless) without parking in the delayed-fetch purgatory.
@Test
Expand Down
Loading