Ah cool Jungtaek, from the looks of it this might help alleviate the problem but I think it won't eliminate it... Keep up the good work :).

Warm regards.


On Wed, Feb 1, 2017 at 12:48 AM, Jungtaek Lim <kabhwan@gmail.com> wrote:

I ported back STORM-1742 to 1.0.x branch since I got many questions regarding complete latency from user@, dev@, and Stack Overflow.

Release vote for Storm 1.0.3 RC1 is starting today (link), so you can wait for 1.0.3 or even use 1.0.3 RC1 before releasing officially.
I'd encourage you to help testing 1.0.3 RC1 if you have time.

STORM-1742 changes the way to calculate complete latency: 
- start time: Acker receiving ACK_INIT from Spout
- end time: Acker receiving ACK message which make the tuple tree completed

yes I was also aware of clock difference on nodes so I picked this way though it doesn't account the latency from Spout to Acker. If you're interested, please refer the issue and relevant discussion link in description. Feedbacks are always appreciated!

Jungtaek Lim (HeartSaVioR)

2017년 2월 1일 (수) 오전 8:12, Andrew Xor <andreas.grammenos@gmail.com>님이 작성:

 Unfortunately this is a bit more complicated than it might initially seem; first of all where acks are processed, if at all, depends on the level of the flow consistency you have implemented for your Spout; which applies to Trident. In regular Storm, it might be perfectly OK to discard missed messages, thereby ignoring the "acks" for each of the tuples. In the other extreme Trident might even stall all of the incoming batches if a tuple fails to be ack'ed and will have to be replayed, ack'ed and then resume the computation -- this is done to enforce the in-order exactly once processing. Additional factor might also impact your latency as network speed used, packet size and so much more... I think Nick from Upenn did a similar study and had a more thorough investigation regarding latency tracking so if he's still subscribed in this list maybe can chip in to give you more details.

From my experience the timing is vastly impacted by each node "clock", thus it cannot really be "trusted" because this can cause big deviations in the latency measurements -- and to be fair to Storm devs do say its an approximation!

Hope this helped...

On Tue, Jan 31, 2017 at 3:55 PM, Thomas Cooper (PGR) <t.cooper@newcastle.ac.uk> wrote:


First a bit of background:

I am a PhD student working on modelling the performance of Storm topologies. I am having reasonable success in modelling the complete latency, however depending on the load (throughput) I can be up to 50% out. 

After exhausting all other sources of possible latency (mostly remote transfer delays between workers on separate machines) it seems the final path of ack tuples from the end component to the Acker and through to the Spout is the last source of unknown latency. Under heavy load and a relatively low number of spout tasks, each task will be busy calling next_tuple and acking, so ack complete messages may back up at the spout. This will artificially extend the complete latency. As the spout does not report metrics for the delay/processing of acks, I cannot account for this effect in my models. 

I thought I might have to resort to implementing my own spout (a custom Storm fork is something I would prefer to avoid). However, after seeing issue 1742 (https://issues.apache.org/jira/browse/STORM-1742) it seems Jungtaek Lim and the Storm devs have already spotted this problem and implemented a solution in the master and 1.x branches. Having the Ackers stop the complete latency clock makes more sense (particularly under heavy load) and makes the complete latency match more closely that of the sojourn time (spout to final component) through the whole topology.

However, I was hoping to get these models working with the latest storm release (1.0.2). It doesn't appear that these changes have been backported to the 1.0.x branch yet?

My Question (TL;DR):

Where in the 1.0.x codebase does the ack_ack message to the spout tasks get processed? I know that implementations of ISpout have an ack() method that gets called. However, in my test topologies when I leave this method unimplemented the system still reports a complete latency for that spout? The timestamp in the ack_ack message must be getting processes somewhere, but I am struggling to identify where.

Any help locating this would be most appreciated.


Thomas Cooper
PhD Student
Newcastle University, School of Computer Science
W: http://www.tomcooper.org.uk | A: 4th Floor, The Core, Science Central, Bath Lane, Newcastle upon Tyne, NE4 5TF