What version of Flume are you using? If the channel does not accept the events, 
the transaction does get rolled back (so that the channel drops the references 
to the events), but the source would not retry the events again - since we do 
not throw a ChannelException to the source. You will see the rolled back log 
message, but the events are dropped and not tried again - the next set would 
get tried. 


Thanks,
Hari


On Monday, December 2, 2013 at 9:21 AM, James Estes wrote:

> Hoping someone can point me in the right direction. We're indexing our logs 
> into elastic search just for added real time convenience and want to make 
> that step optional. Essentially, if we fall behind writing to ES, we would 
> prefer to just skip ES (since we have a more durable channel for higher 
> latency querying of the same data). Optional Channels seemed to fit, but we 
> haven't had much success. 
> 
> First, we set our config to have a Memory Channel and made it optional. If 
> the ES sink fell behind, the channel would fill and reject new events. 
> However, the channel throws an exception and the Channel Processor rolls back 
> the transaction, causing the events to be put back on the queue to be 
> attempted again. The doc for getOptionalChannels says "A failure in writing 
> the event to these channels must be ignored." Should the transaction just 
> always commit when optional channels fail (basically a best-effort 
> commit-what-you-could since it was optional anyway)?
> 
> Second, we tried the PseudoTxMemoryChannel, but found it to also continue to 
> bottleneck on ES. Turns out that it uses queue.put instead of queue.offer, 
> which means it will block until there is room in the queue to add the event. 
> MemoryChannel uses offer. Should PseudoTxMemoryChannel switch to using offer 
> always, or at least have an optional 'failFast' to enable that behavior?
> 
> Is there another way I can accomplish truly optional channels? I do find it 
> encouraging it takes this much effort to make Flume drop events :)
> 
> Thanks,
> 
> James 

Reply via email to