On 12/08/2016 10:13 AM, Gustavo Fernandes wrote:
I recently updated a proposal [1] based on several discussions we had
in the past that is essentially about introducing an event storage
mechanism (write ahead log) in order to improve reliability, failover
and "replayability" for the remote listeners, any feedback greatly
appreciated.
Hi Gustavo,
while I really like the pull-style architecture and reliable events, I
see some problematic parts here:
1) 'cache that would persist the events with a monotonically increasing id'
I assume that you mean globally (for all entries) monotonous. How will
you obtain such ID? Currently, commands have unique IDs that are
<Address, Long> where the number part is monotonous per node. That's
easy to achieve. But introducing globally monotonous counter means that
there will be a single contention point. (you can introduce another
contention points by adding backups, but this is probably unnecessary as
you can find out the last id from the indexed cache data). Per-segment
monotonous would be probably more scalabe, though that increases complexity.
It is complicated, but one way to do this is to have one “primary” node maintain the log and to have other replicate from it. The cluster does need to use consensus to agree which is the primary, and to know which secondary becomes the primary if the primary is failing. Consensus is not trivial, but JGroups Raft (
http://belaban.github.io/jgroups-raft/) may be an option. However, this approach ensures that the replica logs are identical to the primary since they are simply recording the primary’s log as-is. Of course, another challenge is what happens during a failure of the primary log node, and can any transactions be performed/completed while the primary is unavailable.
Another option is to have each node maintain their own log, and to have an aggregator log that merges/combines the various logs into one. Not sure how feasible it is to merge logs by getting rid of duplicates and determining a total order, but if it is then it may have better fault tolerance characteristics.
Of course, it is possible to have node-specific monotonic IDs. For example, MySQL Global Transaction IDs (GTIDs) use a unique UUID for each node, and then GTIDs consists of the node’s UUID plus a monotonically-increasing value (e.g., “31fc48cd-ecd4-46ad-b0a9-f515fc9497c4:1001”). The transaction log contains a mix of GTIDs, and MySQL replication uses a “GTID set” to describe the ranges of transactions known by a server (e.g., “u1:1-100,u2:1-10000,u3:3-5” where “u1”, “u2”, and “u3” are actually UUIDs). So, when a MySQL replica connects, it says “I know about this GTID set", and this tells the master where that client wants to start reading.
2) 'The write to the event log would be async in order to not affect
normal data writes'
Who should write to the cache?
a) originator - what if originator crashes (despite the change has been
added)? Besides, originator would have to do (async) RPC to primary
owner (which will be the primary owner of the event, too).
b) primary owner - with triangle, primary does not really know if the
change has been written on backup. Piggybacking that info won't be
trivial - we don't want to send another message explicitly. But even if
we get the confirmation, since the write to event cache is async, if the
primary owner crashes before replicating the event to backup, we lost
the event
c) all owners, but locally - that will require more complex
reconciliation if the event did really happen on all surviving nodes or
not. And backups could have some trouble to resolve order, too.
IIUC clustered listeners are called from primary owner before the change
is really confirmed on backups (@Pedro correct me if I am wrong,
please), but for this reliable event cache you need higher level of
consistency.
This could be handled by writing a confirmation or “commit” event to the log when the write is confirmed or the transaction is committed. Then, only those confirmed events/transactions would be exposed to client listeners. This requires some buffering, but this could be done in each HotRod client.
3) The log will also have to filter out retried operations (based on
command ID - though this can be indexed, too). Though, I would prefer to
see per-event command-id log to deal with retries properly.
IIUC, a “commit” event would work here, too.
4) Client should pull data, but I would keep push notifications that
'something happened' (throttled on server). There could be use case for
rarely updated caches, and polling the servers would be excessive there.
IMO the clients should poll, but if the server has nothing to return it blocks until there is something or until a timeout occurs. This makes it easy for clients and actually reduces network traffic compared to constantly polling.
BTW, a lot of this is replicating the functionality of Kafka, which is already quite mature and feature rich. It’s actually possible to *embed* Kafka to simplify operations, but I don’t think that’s recommended. And, it introduces a very complex codebase that would need to be supported.