Gabriel Commeau 2013-08-25, 14:24
Hari Shreedharan 2013-08-25, 16:07
Arvind Prabhakar 2013-08-27, 21:12
-Re: [jira] [Issue Comment Deleted] (FLUME-2173) Exactly once semantics for Flume
Hari Shreedharan 2013-08-28, 00:35
Thanks for your reply. You are right in the fact the global state check and update to the sink will require each sink to explicitly support it. We can, of course have this implementation be in an abstract class which is inherited, but yes, this would also mean that there needs to be code changes.
It makes sense to check state in the channels, pretty much the same way as in the sinks. What is a bit concerning is that we will need to do this check at every agent that the event passes through, and probably make some changes in the channel interface to get rid of race conditions (not sure if that is the case, but I think we will need to). Given that an event is likely to pass through 2-3 tiers, each event gets delayed by the time taken by that many ZK round-trips. I am open to this as well, especially considering that it is likely to be a better OOB experience for many users (the ones who have their own custom sinks). Would it suffice to check at the sinks at the terminal agent to make sure that an event gets written out only once?
Thinking about this, having a once-only delivery at the channel level also opens up some possibilities with regards to being able to do some sort of processing on events. Having a guarantee of seeing an event exactly once allows us to do some event processing like counters etc. That seems like a good side effect to have.
Either way, I am glad we agree on the aspect of checking a global state manager to verify that events are deduped.
On Tuesday, August 27, 2013 at 2:12 PM, Arvind Prabhakar wrote:
> Hi Hari,
> Thanks for bringing this up for discussion. I think it will be tremendously
> beneficial to Flume users if we can extend once-only guarantee. Your
> initial suggestion seems reasonable of having a Sink trap the events and
> reference a global state to drop duplicates. Rather than pushing this
> functionality to Sinks is there any other way by which we can make it more
> generally available? The reason I raise this concern is because otherwise
> this becomes a feature of a particular sink and not every sink will have
> the necessary implementation opportunity to get this.
> Alternatively what do you think about this being done at the channel level?
> Since we normally do not see custom implementations of channels, an
> implementation that works with the channel will likely be more useful for
> the broader community of Flume users.
> On Sun, Aug 25, 2013 at 9:07 AM, Hari Shreedharan <[EMAIL PROTECTED] (mailto:[EMAIL PROTECTED])
> > wrote:
> > Hi Gabriel,
> > Thanks for your input. The part where we use replicating channel selector
> > to purposefully replicate - we can easily make it configurable whether to
> > delete deplicate events or not. That should not be difficult to do.
> > The 2nd point where multiple agents/sinks could write the same event can
> > be solved by namespacing the events into different namespaces. So each sink
> > checks one namespace for the event, and multiple sinks can belong to the
> > same namespace - this way, if multiple events are going to write to the
> > same HDFS cluster, then if a duplicate occurs we can easily drop it.
> > Unfortunately, this also does not work around the who
> > HDFS-writing-but-throwing issue.
> > I agree updating ZK will hit latency, but that is the cost to build once
> > only semantics on a highly flexible system. If you look at the algorithm,
> > we actually go to ZK only once per event (to create, there are no updates)
> > - this can even happen per batch if needed to reduce ZK round trips (though
> > I am not sure if ZK provides a batch API).
> > The two phase commit approach sounds good, but it might require interface
> > changes which can now only be made in Flume 2.x. Alse, if we use a single
> > UUID combined with several flags we might be able to work duplicates caused
> > by this replication.
> > Thanks,
Hari Shreedharan 2013-08-28, 00:56