Building a Distributed Log from Scratch: Data Replication
bravenewgeek.com
bravenewgeek.com
1. Leader replicates to followers
2. Followers get the messages and send ACK to leader
3. Leader gets the confirmation from followers, increments high watermark (HW) and replies client that messages is commited.
4. NOW: Leader fails before it could piggyback to followers that HW has been incremented.
5. The question is: Since potential leader is not aware of the fact that HW has been incremented, it becomes the new leader and truncates the log to old HW, which means we have lost an entry in the log that has been confirmed by previous leader.
As a result, client has been confirmed that the message has been successfully written, but it has been lost.
In step 5 of your example, the new leader does not truncate its log; only the new followers do.
Basically, every message that was committed is guaranteed to be in every in-sync replica's log. (Some of these may be after the replica's HW.) But the converse is not true: some replica's logs will contain messages that may not have been committed.
So the new leader has to keep its entire log -- which includes all acknowledged messages, plus some unacknowledged ones that become committed as a result of the leadership transition. But the new followers have to discard messages above their HW, because those messages might not be present on the new leader.
> The replica that registers first becomes the new leader. The new leader chooses its LEO as the new HW.
Isn't it a better idea to elect a process who has the largest offset? Basically during leader election, a process should get confirmation of majority of correct processes. If another process observes that it has larger offset in the log, rejects the election and starts another round of election. Eventualy a correct process who has largest offset will become leader. Regarding correctness, this sounds to me as a safer election.
None of the replicas are guaranteed to have all of the messages that reached the old leader (some of those might never have been replicated anywhere). So clients that didn't receive acknowledgements might need to retry on the new leader, no matter what. Since the clients have to have retry logic anyway, there's no safety issue with discarding some of the messages that weren't yet acknowledged.
Your suggestion would try to minimize the number of necessary retries, but at the cost of adding delay and complexity to the leader election process.
Kafka is a at-least-once system, not transactional.
It's been a while since I've read the code though, so it might have changed.
Imagine Kafka but I am able to run my processors within the same JVM. You might even call them servlets for Kafka.
If you want to work on streaming data (streaming logs), you can use Spark and others...
Raft/multi-paxos are already the cheapest communication bound consensus. Theoretically you can't achieve any better with current network guarantees.
Even if you assume that all nodes are benevolent, and try their best to converge on the single longest chain without needing mining fees as an incentive, you still can't know whether the chain you're querying is the "winner" unless you have an omniscient view of the entire system. So you only get eventual consistency, as opposed to sequential consistency.
However, this part is wrong:
> Because of finite propagation delays and network packet loss, there are normally multiple valid chains at any given moment in time.
How often forks occur is a property of a specific system, not of "blockchain" in general. Bitcoin, for instance, usually does not have multiple chains. They do sometimes happen but generally the orphan rate is quite low. Ethereum has a much higher block rate so propagation delays are more important and forks happen there much more often.
There's an improvement called Bobtail [1] which makes the network much better at estimating how much work has been done. A blockchain which adopted Bobtail would still have probabilistic consensus but would almost never have forks.
Casper, Ethereum's proposed Proof of Stake system, is designed to provide "finality", what the rest of us call "consensus". Once adopted, and assuming it works, we'll have a blockchain where you can be sure that every block you see which you determine is valid is a block which the rest of the network will also decide is valid (assuming less than 1/3 of the network is byzantine).