[
https://issues.apache.org/jira/browse/RATIS-2145?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17877941#comment-17877941
]
guangbao zhao commented on RATIS-2145:
--------------------------------------
[~szetszwo] I have submitted a pull request and used this code in the test. The
service is currently relatively stable.
> Follower hangs until the next trigger to take a snapshot
> ---------------------------------------------------------
>
> Key: RATIS-2145
> URL: https://issues.apache.org/jira/browse/RATIS-2145
> Project: Ratis
> Issue Type: Bug
> Components: gRPC
> Affects Versions: 3.0.1
> Reporter: guangbao zhao
> Assignee: guangbao zhao
> Priority: Major
> Time Spent: 10m
> Remaining Estimate: 0h
>
> We discovered a problem when writing tests with high concurrency. It often
> happens that a follower is running well and then triggers takeSnalshot.
> The following is the relevant log.
> follower: （as the follower log says, between 2024/08/22 20:18:14,044 and
> 2024/08/22 20:21:57,058, no other logs appeared in the follower, but the
> follower election was not triggered, indicating that the leader gave The
> heartbeat sent by the follower is successful）
> {code:java}
> 2024/08/22 20:18:13,987 [node1@group-4F53D3317400-StateMachineUpdater] INFO
> org.apache.ratis.server.raftlog.RaftLog:
> node1@group-4F53D3317400-SegmentedRaftLog: snapshotIndex: updateIncreasingly
> 22436696498 -> 22441096501
> 2024/08/22 20:18:13,999 [node1@group-4F53D3317400-SegmentedRaftLogWorker]
> INFO org.apache.ratis.server.raftlog.segmented.SegmentedRaftLogWorker:
> node1@group-4F53D3317400-SegmentedRaftLogWorker: created new log segment
> /home/work/ssd1/lavafs/aktst-private/metaserver/metadata/ratis/23d5405d-0e30-3d56-9a77-4f53d3317400/current/log_inprogress_22441098615
> 2024/08/22 20:18:14,044 [node1@group-4F53D3317400-SegmentedRaftLogWorker]
> INFO org.apache.ratis.server.raftlog.RaftLog:
> node1@group-4F53D3317400-SegmentedRaftLog: purgeIndex: updateToMax
> old=22432683959, new=22437078979, updated? true
> 2024/08/22 20:21:57,058 [grpc-default-executor-23] INFO
> com.xxx.RaftJournalManager: Received install snapshot notification from
> MetaStore leader: node3 with term index: (t:192, i:22441477801)
> 2024/08/22 20:21:57,059 [InstallSnapshotThread] INFO
> com.xxx.MetaStoreRatisSnapshotProvider: Downloading latest checkpoint from
> Leader MetaStore node3. Checkpoint address: leader:8170
> 2024/08/22 20:21:57,064 [grpc-default-executor-23] INFO
> org.apache.ratis.grpc.server.GrpcServerProtocolService: node1: Completed
> INSTALL_SNAPSHOT, lastRequest: node3->node1#0-t192,notify:(t:192,
> i:22441477801)
> 2024/08/22 20:21:57,065 [grpc-default-executor-23] INFO
> org.apache.ratis.grpc.server.GrpcServerProtocolService: node1: Completed
> INSTALL_SNAPSHOT, lastReply: null
> 2024/08/22 20:21:57,067 [node1-server-thread55] INFO
> org.apache.ratis.server.RaftServer$Division: node1@group-4F53D3317400: Failed
> appendEntries as snapshot (22441477801) installation is in progress
> 2024/08/22 20:21:57,068 [node1-server-thread55] INFO
> org.apache.ratis.server.RaftServer$Division: node1@group-4F53D3317400:
> inconsistency entries.
> Reply:node3<-node1#19406445:FAIL-t192,INCONSISTENCY,nextIndex=22441098642,followerCommit=22441098595,matchIndex=-1{code}
> leader:
> {code:java}
> 2024/08/22 20:18:16,958 [timer5] WARN
> org.apache.ratis.grpc.server.GrpcLogAppender:
> node3@group-4F53D3317400->node1-GrpcLogAppender: Timed out appendEntries,
> errorCount=1,
> request=AppendEntriesRequest:cid=16867241,entriesCount=25,entries=(t:192,
> i:22441098598)...(t:192, i:22441098622)
> 2024/08/22 20:18:16,964 [timer3] WARN
> org.apache.ratis.grpc.server.GrpcLogAppender:
> node3@group-4F53D3317400->node1-GrpcLogAppender: Timed out appendEntries,
> errorCount=1,
> request=AppendEntriesRequest:cid=16867246,entriesCount=1,entry=(t:192,
> i:22441098624)
> 2024/08/22 20:18:16,964 [timer6] WARN
> org.apache.ratis.grpc.server.GrpcLogAppender:
> node3@group-4F53D3317400->node1-GrpcLogAppender: Timed out appendEntries,
> errorCount=1,
> request=AppendEntriesRequest:cid=16867247,entriesCount=1,entry=(t:192,
> i:22441098625)
> 2024/08/22 20:18:16,964 [timer7] WARN
> org.apache.ratis.grpc.server.GrpcLogAppender:
> node3@group-4F53D3317400->node1-GrpcLogAppender: Timed out appendEntries,
> errorCount=1,
> request=AppendEntriesRequest:cid=16867245,entriesCount=1,entry=(t:192,
> i:22441098623)
> 2024/08/22 20:18:16,965 [timer3] WARN
> org.apache.ratis.grpc.server.GrpcLogAppender:
> node3@group-4F53D3317400->node1-GrpcLogAppender: Timed out appendEntries,
> errorCount=1,
> request=AppendEntriesRequest:cid=16867255,entriesCount=1,entry=(t:192,
> i:22441098627)
> 2024/08/22 20:18:16,965 [timer7] WARN
> org.apache.ratis.grpc.server.GrpcLogAppender:
> node3@group-4F53D3317400->node1-GrpcLogAppender: Timed out appendEntries,
> errorCount=1,
> request=AppendEntriesRequest:cid=16867253,entriesCount=1,entry=(t:192,
> i:22441098626)
> 2024/08/22 20:18:16,979 [timer0] WARN
> org.apache.ratis.grpc.server.GrpcLogAppender:
> node3@group-4F53D3317400->node1-GrpcLogAppender: Timed out appendEntries,
> errorCount=1,
> request=AppendEntriesRequest:cid=16867406,entriesCount=13,entries=(t:192,
> i:22441098628)...(t:192, i:22441098640)
> 2024/08/22 20:18:16,981 [timer5] WARN
> org.apache.ratis.grpc.server.GrpcLogAppender:
> node3@group-4F53D3317400->node1-GrpcLogAppender: Timed out appendEntries,
> errorCount=1,
> request=AppendEntriesRequest:cid=16867411,entriesCount=1,entry=(t:192,
> i:22441098641)
> 2024/08/22 20:21:52,849 [node3@group-4F53D3317400-StateMachineUpdater] INFO
> com.xxx.RaftJournalManager: Current LastApplied Index (t:192, i:22445496444)
> 2024/08/22 20:21:57,021 [node3@group-4F53D3317400-StateMachineUpdater] INFO
> org.apache.ratis.server.impl.StateMachineUpdater:
> node3@group-4F53D3317400-StateMachineUpdater: Took a snapshot at index
> 22445496444
> 2024/08/22 20:21:57,021 [node3@group-4F53D3317400-StateMachineUpdater] INFO
> org.apache.ratis.server.impl.StateMachineUpdater:
> node3@group-4F53D3317400-StateMachineUpdater: snapshotIndex:
> updateIncreasingly 22441096438 -> 22445496444
> 2024/08/22 20:21:57,021 [node3@group-4F53D3317400-StateMachineUpdater] INFO
> org.apache.ratis.server.raftlog.RaftLog:
> node3@group-4F53D3317400-SegmentedRaftLog: purge 22441496502
> 2024/08/22 20:21:57,021 [node3@group-4F53D3317400-StateMachineUpdater] INFO
> org.apache.ratis.server.raftlog.RaftLog:
> node3@group-4F53D3317400-SegmentedRaftLog: snapshotIndex: updateIncreasingly
> 22441096438 -> 22445496444
> 2024/08/22 20:21:57,021
> [node3@group-4F53D3317400->node1-GrpcLogAppender-LogAppenderDaemon] INFO
> org.apache.ratis.grpc.server.GrpcLogAppender:
> node3@group-4F53D3317400->node1-GrpcLogAppender: notifyInstallSnapshot with
> firstAvailable=(t:192, i:22441477801), followerNextIndex=22441098642 {code}
> Because the leader did not receive the onNext callback within the
> requestTimeoutDuration（3s） time, the above warn log was
> generated.（raft.grpc.server.leader.outstanding.appends.max is set to 8）
> {code:java}
> scheduler.onTimeout(requestTimeoutDuration,
> () -> timeoutAppendRequest(request.getCallId(), request.isHeartbeat()),
> LOG, () -> "Timeout check failed for append entry request: " + request);
> {code}
> and see through code：
> {code:java}
> private boolean haveTooManyPendingRequests() {
> final int size = pendingRequests.logRequestsSize();
> if (size == 0) {
> return false;
> } else if (size >= maxPendingRequestsNum) {
> return true;
> } else {
> // queue is non-empty and non-full
> return !replyState.isFirstReplyReceived();
> }
> } {code}
> When the size of pendingRequests.logRequestsSize() is >= 8, log entries will
> not be sent, only heartbeats will be sent.（This also verifies that there is a
> log gap in the follower log and no election is taking place）
> {quote}leader:
> 2024/08/22 20:18:16,981 [timer5] WARN
> org.apache.ratis.grpc.server.GrpcLogAppender:
> node3@group-4F53D3317400->node1-GrpcLogAppender: Timed out appendEntries,
> errorCount=1,
> request=AppendEntriesRequest:cid=16867411,entriesCount=1,entry=(t:192,
> i:22441098641)
> ...
> 2024/08/22 20:21:57,021
> [node3@group-4F53D3317400->node1-GrpcLogAppender-LogAppenderDaemon] INFO
> org.apache.ratis.grpc.server.GrpcLogAppender:
> node3@group-4F53D3317400->node1-GrpcLogAppender: notifyInstallSnapshot with
> firstAvailable=(t:192, i:22441477801),
> followerNextIndex={color:#ff0000}22441098642{color}
> follower:
> 2024/08/22 20:21:57,068 [node1-server-thread55] INFO
> org.apache.ratis.server.RaftServer$Division: node1@group-4F53D3317400:
> inconsistency entries.
> Reply:node3<-node1#19406445:FAIL-t192,INCONSISTENCY,nextIndex={color:#ff0000}22441098642{color},followerCommit=22441098595,matchIndex=-1
> {quote}
> And when the leader meets the purge log conditions, the leader will trigger
> notifyInstallSnapshot to the follower. At this time, the follower's nextIndex
> is 22441098642, and the last warn log is 22441098641 (exactly 22441098642 -
> 1), indicating that the follower has received the above warning log and
> updated the follower's nextIndex, but the logRequests of these 8 warns are
> still in logRequests , if it is the next 8 logRequests, then the
> corresponding warn log will appear in the following log, but I did not find
> it in the log.
> As mentioned above, pendingRequests.logRequestsSize() >=
> maxPendingRequestsNum is always satisfied during the period of 2024/08/22
> 20:18:14,044 and 2024/08/22 20:21:57,058, until the leader's firstAvailable >
> followerNextIndex
> {quote}
> private void timeoutAppendRequest(long cid, boolean heartbeat) {
> final AppendEntriesRequest pending = pendingRequests.handleTimeout(cid,
> heartbeat);
> if (pending != null) {
> final int errorCount = replyState.process(Event.TIMEOUT);
> LOG.warn("{}: Timed out {}appendEntries, errorCount={}, request={}",
> this, heartbeat ? "HEARTBEAT " : "", errorCount, pending);
> grpcServerMetrics.onRequestTimeout(getFollowerId().toString(), heartbeat);
> pending.stopRequestTimer();
> }
> }
>
> {quote}
> [~szetszwo] Is it possible to replace handleTimeout with the remove method
> here, so that it will not cause pendingRequests.logRequestsSize() >=
> maxPendingRequestsNum, causing the leader to keep sending heartbeats to the
> follower until notifyInstallSnapshot is triggered? leave the rest to the
> follower's checkInconsistentAppendEntries to ensure, do you think this is ok?
> When this phenomenon occurs, the last log of the follower always stays at:
> {quote}2024/08/22 20:18:14,044
> [node1@group-4F53D3317400-SegmentedRaftLogWorker] INFO
> org.apache.ratis.server.raftlog.RaftLog:
> node1@group-4F53D3317400-SegmentedRaftLog: purgeIndex: updateToMax
> old=22432683959, new=22437078979, updated? true
> {quote}
> Or sometimes there are hidden bugs in the purge call logic that have not been
> discovered.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
