I would suggest that using an epoch number stored in ZK might be helpful. Every operation that the master takes could be made conditional on the epoch number using a multi-transaction.
Unfortunately, as you say, you have to have the update of the epoch be atomic with becoming leader. The natural way to do this is to have an update of an epoch file be part of the leader election, but that probably isn't possible using Curator. The way I would tend to do it would be have a persistent file that is updated atomically as part of leader election. The version of that persistent file could then be used as the epoch number. All updates to files that are gated on the epoch number would only proceed if no other master has been elected, at least if you use the sync option. On Fri, Sep 20, 2019 at 1:31 AM Zili Chen <wander4...@gmail.com> wrote: > Hi ZooKeepers, > > Recently there is an ongoing refactor[1] in Flink community aimed at > overcoming several inconsistent state issues on ZK we have met. I come > here to share our design of leader election and leader operation. For > leader operation, it is operation that should be committed only if the > contender is the leader. Also CC Curator mailing list because it also > contains the reason why we cannot JUST use Curator. > > The rule we want to keep is > > **Writes on ZK must be committed only if the contender is the leader** > > We represent contender by an individual ZK client. At the moment we use > Curator for leader election so the algorithm is the same as the > optimized version in this page[2]. > > The problem is that this algorithm only take care of leader election but > is indifferent to subsequent operations. Consider the scenario below: > > 1. contender-1 becomes the leader > 2. contender-1 proposes a create txn-1 > 3. sender thread suspended for full gc > 4. contender-1 lost leadership and contender-2 becomes the leader > 5. contender-1 recovers from full gc, before it reacts to revoke > leadership event, txn-1 retried and sent to ZK. > > Without other guard txn will success on ZK and thus contender-1 commit > a write operation even if it is no longer the leader. This issue is > also documented in this note[3]. > > To overcome this issue instead of just saying that we're unfortunate, > we draft two possible solution. > > The first is document here[4]. Briefly, when the contender becomes the > leader, we memorize the latch path at that moment. And for > subsequent operations, we do in a transaction first checking the > existence of the latch path. Leadership is only switched if the latch > gone, and all operations will fail if the latch gone. > > The second is still rough. Basically it relies on session expire > mechanism in ZK. We will adopt the unoptimized version in the > recipe[2] given that in our scenario there are only few contenders > at the same time. Thus we create /leader node as ephemeral znode with > leader information and when session expired we think leadership is > revoked and terminate the contender. Asynchronous write operations > should not succeed because they will all fail on session expire. > > We cannot adopt 1 using Curator because it doesn't expose the latch > path(which is added recently, but not in the version we use); we > cannot adopt 2 using Curator because although we have to retry on > connection loss but we don't want to retry on session expire. Curator > always creates a new client on session expire and retry the operation. > > I'd like to learn from ZooKeeper community that 1. is there any > potential risk if we eventually adopt option 1 or option 2? 2. is > there any other solution we can adopt? > > Best, > tison. > > [1] https://issues.apache.org/jira/browse/FLINK-10333 > [2] > https://zookeeper.apache.org/doc/current/recipes.html#sc_leaderElection > [3] https://cwiki.apache.org/confluence/display/CURATOR/TN10 > [4] > https://docs.google.com/document/d/1cBY1t0k5g1xNqzyfZby3LcPu4t-wpx57G1xf-nmWrCo/edit?usp=sharing > > >
