Maybe it is better if we walk through an example in which you think we can have a better observer-specific implementation. Feel free to start a jira so that we can discuss it.
-Flavio > On 27 Jun 2016, at 23:02, Ryan Zhang <[email protected]> wrote: > > Hi, We (twitter) zookeeper cluster ( based on 3.4.x) recently encountered a > prolonged leader election downtime because of the lead machine was > accidentally wiped. After looking at the logs, we noticed that the observer > was trying to connect to the wrong leader for a prolonged time. The specific > bug of taking too long to connect to the lead was fixed in the trunk. > However, I wonder why should the observer accept the lead sid when the > notifications that it get are all from the “LOOKING” quorum machine? The > election algorithm (lookForLeader) seems to be same for the observer or > participants. Is that on purpose? Would it be good to have a specific > observer logic to only act on notifications from “LEADING” and “FOLLOWING” > machine? Thanks.
