ISSUE-20094: [Bug] Topic ledger id will rollback in some cases

[sijie]

Original Issue: apache#20094

---

Search before asking

• I searched in the issues and found nothing similar.

Version

2.11.0

Minimal reproduce step

in my environment deployment: Ensemble: 3, write quorum: 3, ack quorum: 2
in Machine A : brokerA, bookieA
in Machine B: brokerB, bookieB
in Machine C: brokerC, bookieC

1、If Machine A goes offline for a minute, the ownership of the topic in broker A will be moved to broker B. Broker B will fence the old ledger X and use the new ledger Y to write data.
2、Ledger X on bookie B and bookie C is deleted by the GC thread.
3、After Machine A recovers from the network outage, and a period of time later, Machine B goes offline. The topic ownership will be transferred back to broker A, and broker A, for some reason, continues to use ledger X to write data. At this point, ledger X can be written successfully and returned to the client, causing message ID rollback.

What did you expect to see?

in the end, the topic owned by broker A should use ledger id larger than ledger Y

What did you see instead?

broker A use ledger X to write entry

Anything else?

some log in broker A:
2023-04-12T14:28:49,875+0800 [pulsar-io-4-3] DEBUG org.apache.pulsar.broker.service.ServerCnx - [/10.14.30.111:41134] Received send message request. producer: 0:38699 region1:default:38699 size: 65, partition key is: key38799, ordering key is null
2023-04-12T14:28:49,875+0800 [pulsar-io-4-3] DEBUG org.apache.bookkeeper.mledger.impl.ManagedLedgerImpl - [longhaul/test/persistent/testTopic] asyncAddEntry size=65 state=LedgerOpened
2023-04-12T14:28:49,875+0800 [BookKeeperClientWorker-OrderedExecutor-3-0] DEBUG org.apache.bookkeeper.mledger.impl.OpAddEntry - Created new OpAddEntry OpAddEntry{mlName=longhaul/test/persistent/testTopic, ledgerId=null, entryId=-1, startTime=67089676632364916, dataLength=65}
2023-04-12T14:28:49,875+0800 [BookKeeperClientWorker-OrderedExecutor-3-0] DEBUG org.apache.bookkeeper.mledger.impl.ManagedLedgerImpl - [longhaul/test/persistent/testTopic] Write into current ledger lh=13810 entries=16600
network down:
2023-04-12T14:29:06,538+0800 [ZKC-connect-executor-0-SendThread(10.14.30.107:2181)] WARN org.apache.zookeeper.ClientCnxn - Session 0x681d1fc5c69d0016 for server 10.14.30.107/10.14.30.107:2181, Closing socket connection. Attempting reconnect except it is a SessionExpiredException
java.net.SocketException: Network is unreachable
2023-04-12T14:29:21,303+0800 [BookKeeperClientWorker-OrderedExecutor-1-0] WARN org.apache.bookkeeper.client.PendingAddOp - Failed to write entry (13811, 185) to bookie (2, 10.14.30.111:11381): Bookie operation timeout
2023-04-12T14:29:53,302+0800 [ZKC-connect-executor-0-SendThread(10.14.30.107:2181)] ERROR org.apache.bookkeeper.client.MetadataUpdateLoop - UpdateLoop(ledgerId=13811,loopId=75e44191) Error writing metadata to store
org.apache.bookkeeper.client.BKException$ZKException: Error while using ZooKeeper
.....

2023-04-12T15:32:06,390+0800 [BookKeeperClientWorker-OrderedExecutor-1-0] DEBUG org.apache.bookkeeper.proto.PerChannelBookieClient - Got Read response from bookie:10.14.30.106:11381 rc:EOK, ledger:13810:entry:16593:entryLength:116
023-04-12T15:32:06,396+0800 [pulsar-io-4-5] DEBUG org.apache.pulsar.broker.service.ServerCnx - [/10.14.30.111:45287] Received send message request. producer: 0:377200 region1:default:377200 size: 67, partition key is: key377300, ordering key is null
2023-04-12T15:32:06,396+0800 [pulsar-io-4-5] DEBUG org.apache.bookkeeper.mledger.impl.ManagedLedgerImpl - [longhaul/test/persistent/testTopic] asyncAddEntry size=67 state=LedgerOpened
2023-04-12T15:32:06,396+0800 [BookKeeperClientWorker-OrderedExecutor-3-0] DEBUG org.apache.bookkeeper.mledger.impl.OpAddEntry - Created new OpAddEntry OpAddEntry{mlName=longhaul/test/persistent/testTopic, ledgerId=null, entryId=-1, startTime=67093473153018494, dataLength=67}
2023-04-12T15:32:06,396+0800 [BookKeeperClientWorker-OrderedExecutor-3-0] DEBUG org.apache.bookkeeper.mledger.impl.ManagedLedgerImpl - [longhaul/test/persistent/testTopic] Write into current ledger lh=13810 entries=16601
2023-04-12T15:32:06,402+0800 [BookKeeperClientWorker-OrderedExecutor-1-0] DEBUG org.apache.bookkeeper.proto.PerChannelBookieClient - Got Add response from bookie:11.167.133.184:11381 rc:EOK, ledger:13810:entry:16600
2023-04-12T15:32:06,408+0800 [BookKeeperClientWorker-OrderedExecutor-1-0] DEBUG org.apache.bookkeeper.proto.PerChannelBookieClient - Got Add response from bookie:10.14.30.106:11381 rc:EOK, ledger:13810:entry:16600

log in brokerB:
2023-04-12T14:29:16,455+0800 [pulsar-io-4-7] INFO org.apache.bookkeeper.mledger.impl.ManagedLedgerImpl - Opening managed ledger longhaul/test/persistent/testTopic
2023-04-12T14:29:16,458+0800 [bookkeeper-ml-scheduler-OrderedScheduler-3-0] INFO org.apache.pulsar.broker.intercept.ManagedLedgerInterceptorImpl - onManagedLedgerP
ropertiesInitialize init timeID:1681280727036996
2023-04-12T14:29:16,458+0800 [bookkeeper-ml-scheduler-OrderedScheduler-3-0] DEBUG org.apache.bookkeeper.mledger.impl.ManagedLedgerImpl - [longhaul/test/persistent/t
estTopic] Opening ledger 13810
2023-04-12T14:29:16,469+0800 [ZKC-connect-executor-0-EventThread] DEBUG org.apache.bookkeeper.client.MetadataUpdateLoop - UpdateLoop(ledgerId=13810,loopId=6a93ef27)
success
2023-04-12T14:29:17,289+0800 [BookKeeperClientWorker-OrderedExecutor-1-0] DEBUG org.apache.bookkeeper.proto.PerChannelBookieClient - Got Read response from bookie:1
1.167.133.184:11381 rc:EOK, ledger:13810:entry:16599:entryLength:116
2023-04-12T14:29:17,293+0800 [BookKeeperClientWorker-OrderedExecutor-1-0] DEBUG org.apache.bookkeeper.proto.PerChannelBookieClient - Got Add response from bookie:11
.167.133.184:11381 rc:EOK, ledger:13810:entry:16599
2023-04-12T14:29:17,296+0800 [BookKeeperClientWorker-OrderedExecutor-1-0] DEBUG org.apache.bookkeeper.proto.PerChannelBookieClient - Got Add response from bookie:11
.164.30.111:11381 rc:EOK, ledger:13810:entry:16599
2023-04-12T14:29:17,491+0800 [BookKeeperClientWorker-OrderedExecutor-1-0] DEBUG org.apache.bookkeeper.proto.PerChannelBookieClient - Got Read response from bookie:10.14.30.111:11381 rc:ENOENTRY, ledger:13810:entry:16600:entryLength:0
2023-04-12T14:29:17,491+0800 [BookKeeperClientWorker-OrderedExecutor-1-0] DEBUG org.apache.bookkeeper.client.PendingReadOp - No such entry found on bookie. L13810E16600 bookie: 10.14.30.111:11381
2023-04-12T14:29:17,491+0800 [BookKeeperClientWorker-OrderedExecutor-1-0] INFO org.apache.bookkeeper.client.ReadOnlyLedgerHandle - Closing recovered ledger 13810 a
t entry 16599
2023-04-12T14:29:17,500+0800 [BookKeeperClientWorker-OrderedExecutor-3-0] DEBUG org.apache.bookkeeper.mledger.impl.ManagedLedgerImpl - [longhaul/test/persistent/testTopic] Opened ledger 13810: No problem
2023-04-12T14:29:17,770+0800 [pulsar-io-4-7] DEBUG org.apache.pulsar.broker.service.ServerCnx - [/10.14.30.111:42441] Received send message request. producer: 0:41100 region1:default:41100 size: 65, partition key is: key41200, ordering key is null
2023-04-12T14:29:17,770+0800 [pulsar-io-4-7] DEBUG org.apache.bookkeeper.mledger.impl.ManagedLedgerImpl - [longhaul/test/persistent/testTopic] asyncAddEntry size=65 state=LedgerOpened
2023-04-12T14:29:17,770+0800 [BookKeeperClientWorker-OrderedExecutor-3-0] DEBUG org.apache.bookkeeper.mledger.impl.OpAddEntry - Created new OpAddEntry OpAddEntry{mlName=longhaul/test/persistent/testTopic, ledgerId=null, entryId=-1, startTime=125088582810719907, dataLength=65}
2023-04-12T14:29:17,770+0800 [BookKeeperClientWorker-OrderedExecutor-3-0] DEBUG org.apache.bookkeeper.mledger.impl.ManagedLedgerImpl - [longhaul/test/persistent/testTopic] Write into current ledger lh=13814 entries=1
2023-04-12T14:29:26,470+0800 [BookKeeperClientWorker-OrderedExecutor-1-0] INFO org.apache.bookkeeper.client.PendingReadOp - Error: Bookie handle is not available w
hile reading L13810 E16592 from bookie: 10.14.30.106:11381
...
2023-04-12T14:29:26,471+0800 [BookKeeperClientWorker-OrderedExecutor-1-0] INFO org.apache.bookkeeper.client.PendingReadOp - Error: Bookie handle is not available w
hile reading L13810 E16599 from bookie: 10.14.30.106:11381

log in bookieB and bookieC, we can find the same message:
2023-04-12T14:29:17,291+0800 [BookieJournal-11381] DEBUG org.apache.bookkeeper.bookie.Journal - Written and queuing for flush Ledger: 13810 Entry: 16599
2023-04-12T14:29:17,291+0800 [ForceWriteThread] DEBUG org.apache.bookkeeper.bookie.JournalChannel - Journal ForceWrite
2023-04-12T14:29:17,296+0800 [bookie-journal-callback-4] DEBUG org.apache.bookkeeper.bookie.Journal - Acknowledge Ledger: 13810, Entry: 16599
2023-04-12T14:29:17,490+0800 [BookieHighPriorityThread-OrderedExecutor-1-0] DEBUG org.apache.bookkeeper.proto.ReadEntryProcessor - Received new read request: Op(2)[
Ledger:13810,Entry:16600]
2023-04-12T14:29:17,490+0800 [BookieHighPriorityThread-OrderedExecutor-1-0] WARN org.apache.bookkeeper.proto.ReadEntryProcessor - Ledger: 13810 fenced by: /10.14
.30.104:36794
2023-04-12T14:29:17,490+0800 [BookieHighPriorityThread-OrderedExecutor-1-0] DEBUG org.apache.bookkeeper.bookie.storage.ldb.SingleDirectoryDbLedgerStorage - Set fenc
ed. ledger: 13810
2023-04-12T14:29:17,490+0800 [BookieHighPriorityThread-OrderedExecutor-1-0] DEBUG org.apache.bookkeeper.bookie.storage.ldb.SingleDirectoryDbLedgerStorage - Get Entr
y: 13810@16600
2023-04-12T14:29:17,490+0800 [BookieHighPriorityThread-OrderedExecutor-1-0] DEBUG org.apache.bookkeeper.bookie.storage.ldb.EntryLocationIndex - Entry not found 1381
0@16600 in db index
...
2023-04-12T14:52:19,980+0800 [GarbageCollectorThread-9-1] DEBUG org.apache.bookkeeper.bookie.GarbageCollectorThread - delete ledger : 13810
2023-04-12T14:52:19,980+0800 [GarbageCollectorThread-9-1] DEBUG org.apache.bookkeeper.bookie.storage.ldb.SingleDirectoryDbLedgerStorage - Deleting ledger 13810
2023-04-12T14:52:19,980+0800 [GarbageCollectorThread-9-1] DEBUG org.apache.bookkeeper.bookie.storage.ldb.LedgerMetadataIndex - Removed ledger 13810
...
2023-04-12T15:32:06,397+0800 [bookie-io-11-2] DEBUG org.apache.bookkeeper.bookie.storage.ldb.SingleDirectoryDbLedgerStorage - Set master key. ledger: 13810
2023-04-12T15:32:06,397+0800 [bookie-io-11-2] DEBUG org.apache.bookkeeper.bookie.storage.ldb.LedgerMetadataIndex - Inserting new ledger 13810
2023-04-12T15:32:06,397+0800 [bookie-io-11-2] DEBUG org.apache.bookkeeper.bookie.storage.ldb.SingleDirectoryDbLedgerStorage - isFenced. ledger: 13810
2023-04-12T15:32:06,397+0800 [bookie-io-11-2] DEBUG org.apache.bookkeeper.bookie.storage.ldb.SingleDirectoryDbLedgerStorage - hasLimboState. ledger: 13810
2023-04-12T15:32:06,398+0800 [bookie-io-11-2] DEBUG org.apache.bookkeeper.bookie.storage.ldb.SingleDirectoryDbLedgerStorage - Add entry. 13810@16600, lac = 16599
2023-04-12T15:32:06,399+0800 [BookieJournal-11381] DEBUG org.apache.bookkeeper.bookie.Journal - Written and queuing for flush Ledger: 13810 Entry: 16600
2023-04-12T15:32:06,399+0800 [ForceWriteThread] DEBUG org.apache.bookkeeper.bookie.JournalChannel - Journal ForceWrite
2023-04-12T15:32:06,403+0800 [bookie-journal-callback-5] DEBUG org.apache.bookkeeper.bookie.Journal - Acknowledge Ledger: 13810, Entry: 16600

in the log, we can know that:
broker A in machine A
14:28:49: key38799 was written into ledger=13810 entryID=16599
14:29:00~14:30:00 : network was down
15:32:06: key377300 was written into ledger=13810 entryID=16600

brokerB in machine B:
14:29:16: Opening managed ledger longhaul/test/persistent/testTopic
14:29:17: Closing recovered ledger 13810 at entry 16599
14:29:17: key41200 was written into ledger=13814 entryID=0
15:32:00~15:33:00: netdown

the ledger id rollback from 13814 to 13810.
the question is, why broker A hold the topic ledger after zk session is timeout, how to solve this problem.

Are you willing to submit a PR?

• I'm willing to submit a PR!