I’m working on a topology that will be similar to this application so I was thinking about this yesterday.
I’m thinking that if there is any significant work to do on messages in making them into tuples, shouldn’t the message be emitted and the work be in a bolt? I don’t think that bolt execute functions have the same limitations as spout nextTuple functions. Now with that said, bolt executes should not be long running computations either, but can be longer than the spouts nextTuple function.
So the Spout documentation (assuming its correct...) here (http://storm.apache.org/releases/current/Concepts.html#spouts) mentions this:
"The main method on spouts is
nextTupleeither emits a new tuple into the topology or simply returns if there are no new tuples to emit. It is imperative that
nextTupledoes not block for any spout implementation, because Storm calls all the spout methods on the same thread."
When developing a custom spout we interpreted it to mean that any "real work" done by a spout should be done in a separate thread, and decided on the following pattern which seems some what relevant to what you are trying to do in your bolts.
On Spout prepare, we create a concurrent/thread safe queue. We then create a new Thread passing it a reference to our thread safe queue. This thread handles finding new data that needs to be emitted. When that thread finds data, it adds it to the shared queue. When the spout's nextTuple() method is called, it looks for data on the shared queue and emits it.
I imagine doing async processing in a bolt using one or more threads could work with a similar pattern. On prepare you setup your thread(s) with references to a shared queue. The bolt passes work to be completed to the thread(s), the thread(s) communicate back to the bolt the result via a shared queue. Add in the concept of tick tuples to ensure your bolt checks for completed work on a regular basis?
Is there a better way to do this?
On Thu, Apr 28, 2016 at 11:22 AM, Julien Nioche <email@example.com> wrote:
Thanks for the clarification
On 28 April 2016 at 15:12, P. Taylor Goetz <firstname.lastname@example.org> wrote:
The documentation is wrong. See:
At some point it looks like the change made there got reverted. I will reopen it to make sure the documentation is corrected.
OutputCollector is NOT thread-safe.
On Apr 28, 2016, at 9:06 AM, Stephen Powis <email@example.com> wrote:
"Its perfectly fine to launch new threads in bolts that do processing asynchronously. OutputCollector is thread-safe and can be called at any time."
From the docs for 0.9.6: http://storm.apache.org/releases/0.9.6/Concepts.html#bolts
On Thu, Apr 28, 2016 at 9:03 AM, P. Taylor Goetz <firstname.lastname@example.org> wrote:
IIRC there was discussion about making it thread safe, but I don't believe it was implemented.
On Apr 28, 2016, at 3:52 AM, Julien Nioche <email@example.com> wrote:
I asked the same question in February but did not get a reply
Anyone who could confirm this?
On 27 April 2016 at 14:05, Steven Lewis <Steven.Lewis@walmart.com> wrote:
I have conflicting information, and have not checked personally but has the output collector finally been made thread safe for emitting in version 1.0 or 0.10? I know it was a huge problem in 0.9.5 when trying to do threading in a bolt for async future calls and emitting once it returns.
This email and any files transmitted with it are confidential and intended solely for the individual or entity to whom they are addressed. If you have received this email in error destroy it immediately. *** Walmart Confidential ***