fanhualta commented on a change in pull request #3191:
URL: https://github.com/apache/iotdb/pull/3191#discussion_r644860167
##########
File path:
cluster/src/main/java/org/apache/iotdb/cluster/server/member/DataGroupMember.java
##########
@@ -314,97 +321,61 @@ public synchronized boolean addNode(Node node,
NodeAdditionResult result) {
if (insertIndex > 0) {
allNodes.add(insertIndex, node);
peerMap.putIfAbsent(node, new Peer(logManager.getLastLogIndex()));
- // remove the last node because the group size is fixed to replication
number
- Node removedNode = allNodes.remove(allNodes.size() - 1);
- peerMap.remove(removedNode);
// if the local node is the last node and the insertion succeeds, this
node should leave
// the group
logger.debug("{}: Node {} is inserted into the data group {}", name,
node, allNodes);
- return removedNode.equals(thisNode);
}
- return false;
+ return insertIndex > 0;
}
}
/**
- * Process the election request from another node in the group. To win the
vote from the local
- * member, a node must have both meta and data logs no older than then local
member, or it will be
- * turned down.
+ * Try to add a Node into the group to which the member belongs.
*
- * @param electionRequest
- * @return Response.RESPONSE_META_LOG_STALE if the meta logs of the elector
fall behind
- * Response.RESPONSE_LOG_MISMATCH if the data logs of the elector fall
behind Response.SUCCESS
- * if the vote is given to the elector the term of local member if the
elector's term is no
- * bigger than the local member
+ * @return true if this node should leave the group because of the addition
of the node, false
+ * otherwise
*/
- @Override
- long checkElectorLogProgress(ElectionRequest electionRequest) {
- // to be a data group leader, a node should also be qualified to be the
meta group leader
- // which guarantees the data group leader has the newest partition table.
- long thatTerm = electionRequest.getTerm();
- long thatMetaLastLogIndex = electionRequest.getLastLogIndex();
- long thatMetaLastLogTerm = electionRequest.getLastLogTerm();
- long thatDataLastLogIndex = electionRequest.getDataLogLastIndex();
- long thatDataLastLogTerm = electionRequest.getDataLogLastTerm();
- logger.info(
- "{} received an dataGroup election request, term:{},
metaLastLogIndex:{}, metaLastLogTerm:{}, dataLastLogIndex:{},
dataLastLogTerm:{}",
- name,
- thatTerm,
- thatMetaLastLogIndex,
- thatMetaLastLogTerm,
- thatDataLastLogIndex,
- thatDataLastLogTerm);
-
- // check meta logs
- // term of the electors' MetaGroupMember is not verified, so 0 and 1 are
used to make sure
- // the verification does not fail
- long metaResponse = metaGroupMember.checkLogProgress(thatMetaLastLogIndex,
thatMetaLastLogTerm);
- if (metaResponse == Response.RESPONSE_LOG_MISMATCH) {
- return Response.RESPONSE_META_LOG_STALE;
- }
-
- long resp = checkLogProgress(thatDataLastLogIndex, thatDataLastLogTerm);
- if (resp == Response.RESPONSE_AGREE) {
- logger.info(
- "{} accepted an dataGroup election request, term:{}/{},
dataLogIndex:{}/{}, dataLogTerm:{}/{}, metaLogIndex:{}/{},metaLogTerm:{}/{}",
- name,
- thatTerm,
- term.get(),
- thatDataLastLogIndex,
- logManager.getLastLogIndex(),
- thatDataLastLogTerm,
- logManager.getLastLogTerm(),
- thatMetaLastLogIndex,
- metaGroupMember.getLogManager().getLastLogIndex(),
- thatMetaLastLogTerm,
- metaGroupMember.getLogManager().getLastLogTerm());
- setCharacter(NodeCharacter.FOLLOWER);
- lastHeartbeatReceivedTime = System.currentTimeMillis();
- setVoteFor(electionRequest.getElector());
- updateHardState(thatTerm, getVoteFor());
- } else {
- logger.info(
- "{} rejected an dataGroup election request, term:{}/{},
dataLogIndex:{}/{}, dataLogTerm:{}/{}, metaLogIndex:{}/{},metaLogTerm:{}/{}",
- name,
- thatTerm,
- term.get(),
- thatDataLastLogIndex,
- logManager.getLastLogIndex(),
- thatDataLastLogTerm,
- logManager.getLastLogTerm(),
- thatMetaLastLogIndex,
- metaGroupMember.getLogManager().getLastLogIndex(),
- thatMetaLastLogTerm,
- metaGroupMember.getLogManager().getLastLogTerm());
+ public boolean addNode(Node node, NodeAdditionResult result) {
+ if (logger.isDebugEnabled()) {
+ logger.debug("{}: start to add node {}", name, node);
+ }
+
+ // mark slots that do not belong to this group any more
+ Set<Integer> lostSlots =
+ ((SlotNodeAdditionResult) result)
+ .getLostSlots()
+ .getOrDefault(new RaftNode(getHeader(), getRaftGroupId()),
Collections.emptySet());
+ for (Integer lostSlot : lostSlots) {
+ slotManager.setToSending(lostSlot, false);
+ }
+ slotManager.save();
+
+ synchronized (allNodes) {
+ preAddNode(node);
+ if (allNodes.contains(node) && allNodes.size() >
config.getReplicationNum()) {
+ // remove the last node because the group size is fixed to replication
number
+ Node removedNode = allNodes.remove(allNodes.size() - 1);
+ peerMap.remove(removedNode);
+
+ if (removedNode.equals(leader.get()) && !removedNode.equals(thisNode))
{
+ // if the leader is removed, also start an election immediately
+ synchronized (term) {
+ setCharacter(NodeCharacter.ELECTOR);
Review comment:
When the leader of a data group is removed from the cluster, the data
group should elect a new leader. Each node in the data group will be an
elector and try to be the leader. If the log of this node is not new enough to
win an election, it just will not win the election and the node with the newest
log will win. Why data group will still have electionTimeout unavailable time?
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
