ratis
ratis copied to clipboard
RATIS-1291. Send heartbeat when there is no reply
What changes were proposed in this pull request?
Currently, when there always exist log, leader only send log, and will not send heartbeat. The problem is if follower need a long time to process log, leader maybe receives response from follower after a long time.
default long getHeartbeatRemainingTimeMs() {
return getServer().properties().minRpcTimeoutMs()/2 - getFollower().getLastRpcTime().elapsedTimeMs();
}
In this pr, I record lastRpcSendTimeWithResponse, i.e. when leader receive response of request1, leader record the send time of request1. Besides, I record lastHeartBeatSendTime, i.e. the send time of heartbeat.
Use this two timestamp, leader send heartbeat when can not receive response after minRpcTimeoutMs()/2. If log cost a short time from request to reply, leader also does not need to send heartbeat.
default long getHeartbeatRemainingTimeMs() {
return getServer().properties().minRpcTimeoutMs()/2 -
Math.min(getFollower().getLastRpcSendTimeWithResponse().elapsedTimeMs(),
getFollower().getLastHeartBeatSendTime().elapsedTimeMs());
}
What is the link to the Apache JIRA
https://issues.apache.org/jira/browse/RATIS-1291
How was this patch tested?
no need new ut
will fix the failed ut.
cc @amaliujia I can help review when the UT is fixed.