helix-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From kishore g <g.kish...@gmail.com>
Subject Re: Add performant IPC (Helix actors)
Date Sat, 12 Jul 2014 06:40:28 GMT
Good point Vlad, I was thinking of defining the right abstraction and
possibly provide one implementation based. I think the synchronous and
asynchronous messaging should be covered as part of this.

Also I think Finagle and likes of it cater more towards the client server
communication but what we lack today is a good solution for peer to peer
transport. For example, if some one has to build a consensus layer they
have build everything from ground up. Providing a base layer that exchange
messages between peers on a per partition basis can be a great building
block to build different replication schemes.

Overall messaging can be used for two cases
-- data and control.


This is needed for control messages that occur rarely, for these type of
messages latency/throughput is not important but its important for it to be


This is mainly exchanging data between different roles,
controller-participant, participant-participant, spectator-participant.
These types of message exchanges occur quite frequently and having high
throughput and low latency is a requirement.

For example having the following api and guarantee would allow one to build
synchronous replication, asynchronous, quorum etc.

send(partition, message, ACK_MODE, callback) ACK_MODE can be like ACK from
ALL, QUORUM, 1, NONE etc. callback is fired when one gets the response back
from the receiver.

This simple api can allow both synchronous and asynchronous mode of
communication. We dont have to do the serialization/deserialization. The
hard part here would be to guarantee ordering between the messages. One
should be able to specify the message ordering requirement, FIFO on a per
partition level or dont care about ordering. Having this makes it easy for
one to implement replication schemes.

Kishore G

On Fri, Jul 11, 2014 at 12:49 PM, Greg Brandt <brandt.greg@gmail.com> wrote:

> Vlad,
> I'm not sure that the goal here is to cover all possible use cases. This is
> intended as basically a replacement for the current Helix cluster messaging
> service, which doesn't rely on ZooKeeper. I would argue that Helix is
> already that general framework for discovering the service endpoint, and
> that this is just one implementation of an underlying messaging stack (for
> which there is currently only the ZooKeeper-based implementation).
> Moreover, keeping it too general (i.e. providing the APIs outlined in the
> JIRA, but no implementation) puts much greater onus on the user. It would
> be nice for someone just starting out with Helix to have a pretty good
> messaging service readily available, as well as the APIs to implement more
> specific solutions if he or she so chooses.
> -Greg
> On Fri, Jul 11, 2014 at 11:07 AM, vlad.gm@gmail.com <vlad.gm@gmail.com>
> wrote:
> > This sounds like service discovery for a messaging solution. At this
> point
> > there would be some overlap with message buses such as Finagle. Perhaps
> > this should be a general framework for discovering the service endpoint
> > that could leave the actual underlaying messaging stack open (be it
> > Finagle, Netty-based or ZeroMQ, for example). My only fear is that if the
> > messaging framework is encapsulated completely into Helix, it would be
> hard
> > to compete with tuned messaging bus solutions and cover all possible use
> > cases (for example, in my company, we use a large number of sub cases on
> > the synchronous call to asynchronous call spectrum).
> >
> > Regards,
> > Vlad
> >
> >
> > On Fri, Jul 11, 2014 at 10:56 AM, Greg Brandt <brandt.greg@gmail.com>
> > wrote:
> >
> >> (copied from HELIX-470)
> >>
> >> Helix is missing a high-performance way to exchange messages among
> >> resource partitions, with a user-friendly API.
> >>
> >> Currently, the Helix messaging service relies on creating many nodes in
> >> ZooKeeper, which can lead to ZooKeeper outages if messages are sent too
> >> frequently.
> >>
> >> In order to avoid this, high-performance NIO-based HelixActors should be
> >> implemented (in rough accordance with the actor model). HelixActors
> exchange
> >> messages asynchronously without waiting for a response, and are
> >> partition/state-addressable.
> >>
> >> The API would look something like this:
> >>
> >> public interface HelixActor<T> {
> >>     void send(Partition partition, String state, T message);
> >>     void register(String resource, HelixActorCallback<T> callback);
> >> }
> >> public interface HelixActorCallback<T> {
> >>     void onMessage(Partition partition, State state, T message);
> >> }
> >>
> >> #send should likely support wildcards for partition number and state, or
> >> its method signature might need to be massaged a little bit for more
> >> flexibility. But that's the basic idea.
> >>
> >> Nothing is inferred about the format of the messages - the only metadata
> >> we need to be able to interpret is (1) partition name and (2) state. The
> >> user provides a codec to encode / decode messages, so it's nicer to
> >> implementHelixActor#send and HelixActorCallback#onMessage.
> >>
> >> public interface HelixActorMessageCodec<T> {
> >>     byte[] encode(T message);
> >>     T decode(byte[] message);
> >> }
> >>
> >> Actors should support somewhere around 100k to 1M messages per second.
> >> The Netty framework is a potential implementation candidate, but should
> be
> >> thoroughly evaluated w.r.t. performance.
> >>
> >
> >

View raw message