streams-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Steve Blackmon <>
Subject Re: Proposing Changes to the StreamsProvider interface and StreamsProviderTask
Date Sat, 10 May 2014 23:07:17 GMT
On Tue, May 6, 2014 at 9:53 PM, Matthew Hager [W2O Digital]
<> wrote:
> Good Day!
> I would like to throw in my two pents in on this if it pleases the
> community.
> Here are my thoughts based on implementations that I have written with
> streams to ensure timely, high yield execution. Personally, I had to
> override much of the LocalStreamsBuilder to fit my use cases for many of
> the problems described below, except the opposite of which. I have a
> modality of a 'finite' stream which execution is hindered when being
> 'polled' in the manner that it is. This is further complicated by the
> excessive waiting caused by the current 'shutdown' the exists.
> There are essentially two major use-cases, that I can see, that are likely
> to take place. The first is a perpetual stream, that is technically never
> satisfied. The second, is the case of a finite stream (HDFS reader, S3
> reader, pulling a user's time-line, etc...) that has a definitive start
> and end. To solve these two models of execution here are my thoughts.

I think creating FiniteStream and PerpetualStream interfaces with
different methods would be a good start.  Interfaces for Stream and
sub-categories of Stream aren't there now but adding them in makes a
lot of sense.

The runtimes could then select an appropriate method of creating and
execute components depending on whether the entire pipeline was typed
as Stream, FiniteStream, PerpetualStream, or some more specific
sub-category that someone sees fit to invent.

> StreamsResultSet - I actually found this to be quite useful paradigm. A
> queue prevents a buffer overflow, an iterator makes it fun and easy to
> read (I love iterators), and it is simple and succinct. I do, however,
> feel it is best expressed as an interface instead of a class. Personally I
> had to override almost every function to fit the concept of a 'finite'
> stream. Without an expensive tear-down cost. The thing missing from this,
> as an interface, would be the notion of "isRunning" which could easily
> satisfy both of the aforementioned modalities. (As Ryan suggested) I
> actually have a reference implementation of this for finite streams if
> anyone would like to see it or use it.
> Event Driven - I concur with Matt 100% on this. As currently implemented,
> LocalStreamsBuilder is exceedingly inefficient from a memory perspective
> and time execution perspective.

In my opinion runtime-local is working decently well for simple
streams (such as those on the streams-examples github) - because I've
seen that it is capable of reading, processing, and writing hundreds
of GB through a single process to completion without crashing or
dropping any records.

There is absolutely room for improvement - and there are plenty of
simple improvements that could be made to the local runtime's memory
management and performance without altering the core interfaces.

For example, we could switch to disruptor for runtime-internal queuing
and make them the default for component queues.

I'm not opposed to a more extensive rewrite if someone really wants to do that.

> To me, it seems, that we could almost
> abstract out 2 common interfaces to make this happen.
>         * Listener { receive(StreamsDatum); }
>         * Producer { push(StreamsDatum); registerListener(Listener); }
> Where the following implementations would place:
>         * Reader implements Producer
>         * Processor implements Producer, Listener
>         * Writer implements Listener

This is pretty close to what is being proposed over at (same interfaces different names).  I think
adding another abstraction layer as you describe could allow for
runtimes that utilize an actor model and/or a pub-sub messaging
framework under the covers to be built more easily and think it's a
good idea.

> In the reference implementations, you can still have queues that are in
> place that could actually function as meaningful indicators of system
> performance and status. IE: the queue functions as, well, an actually
> queue, and processes are much more asynchronous than they currently are
> now. Then, LocalStreamsBuilder strings all the guys up together in their
> nice little workflows and the events just shoot the little Datums down
> their paths until they wind up wherever they are supposed to go as quickly
> as possible.
> Pardon the long response, I tend to be wordy, great discussion and thanks
> to everyone for indulging my thoughts!

Thanks for working to make the project better!

> Cheers!
> Smashew (Matthew Hager)
> On 5/6/14, 8:24 AM, "Matt Franklin" <> wrote:
>>On Mon, May 5, 2014 at 1:15 PM, Steve Blackmon <>
>>> What I meant to say re #1 below is that batch-level metadata could be
>>> useful for modules downstream of the StreamsProvider /
>>> StreamsPersistReader, and the StreamsResultSet gives us a class to
>>> which we can add new metadata in core as the project evolves, or
>>> supplement on a per-module or per-implementation basis via
>>> subclassing.  Within a provider there's no need to modify or extend
>>> StreamsResultSet to maintain and utilize state from a third-party API.
>>I agree that in batch mode, metadata might be important.  In conversations
>>with other people, I think what might be missing is a completely reactive,
>>event-driven mode where a provider pushes to the rest of the stream rather
>>than gets polled.
>>> I think I would support making StreamsResultSet an interface rather
>>> than a class.
>>+1 on interface
>>> Steve Blackmon
>>> On Mon, May 5, 2014 at 12:07 PM, Steve Blackmon <>
>>> wrote:
>>> > Comments on this in-line below.
>>> >
>>> > On Thu, May 1, 2014 at 4:38 PM, Ryan Ebanks <>
>>> wrote:
>>> >> The use and implementations of the StreamsProviders seems to have
>>> drifted
>>> >> away from what it was originally designed for.  I recommend that we
>>> change
>>> >> the StreamsProvider interface and StreamsProvider task to reflect the
>>> >> current usage patterns and to be more efficient.
>>> >>
>>> >> Current Problems:
>>> >>
>>> >> 1.) newPerpetualStream in LocalStream builder is not perpetual.  The
>>> >> StreamProvider task will shut down after a certain amount of empty
>>> returns
>>> >> from the provider.  A perpetual stream implies that it will run in
>>> >> perpetuity.  If I open a Twitter Gardenhose that is returning tweets
>>> with
>>> >> obscure key words, I don't want my stream shutting down if it is just
>>> quiet
>>> >> for a few time periods.
>>> >>
>>> >> 2.) StreamsProviderTasks assumes that a single read*, will return all
>>> the
>>> >> data for that request.  This means that if I do a readRange for a
>>> the
>>> >> provider has to hold all of that data in memory and return it as one
>>> >> StreamsResultSet.  I believe the readPerpetual was designed to get
>>> around
>>> >> this problem.
>>> >>
>>> >> Proposed Fixes/Changes:
>>> >>
>>> >> Fix 1.) Remove the StreamsResultSet.  No implementations in the
>>> >> currently use it for anything other than a wrapper around a Queue
>>> is
>>> >> then iterated over.  StreamsProvider will now return a
>>> Queue<StreamsDatum>
>>> >> instead of a StreamsResultSet.  This will allow providers to queue
>>> as
>>> >> they receive it, and the StreamsProviderTask can pop them off as
>>>soon as
>>> >> they are available.  It will help fix problem #2, as well as help to
>>> lower
>>> >> memory usage.
>>> >>
>>> >
>>> > I'm not convinced this is a good idea.  StreamsResultSet is a useful
>>> > abstraction even if no modules are using it as more than a wrapper for
>>> > Queue at the moment.  For example read* in a provider or persistReader
>>> > could return batch-level (as opposed to datum-level) metadata from the
>>> > underlying API which would be useful state for the provider.
>>> > Switching to Queue would eliminate our ability to add those
>>> > capabilities at the core level or at the module level.
>>> >
>>> >> Fix 2.) Add a method, public boolean isRunning(), to the
>>> >> interface.  The StreamsProviderTask can call this function to see if
>>> >> provider is still operating. This will help fix problems #1 and #2.
>>> >> will allow the provider to run mulitthreaded, queue data as it's
>>> available,
>>> >> and notify the task when it's done so that it can be closed down
>>> properly.
>>> >>  It will also allow the stream to be run in perpetuity as the
>>> >> won't shut down providers that have not been producing data for a
>>> >>
>>> >
>>> > I think this is a good idea.  +1
>>> >
>>> >> Right now the StreamsProvider and StreamsProviderTask seem to be
>>>full of
>>> >> short term fixes that need to be redesigned into long term solutions.
>>>  With
>>> >> enough positive feedback, I will create Jira tasks, a feature branch,
>>> and
>>> >> begin work.
>>> >>
>>> >> Sincerely,
>>> >> Ryan Ebanks

View raw message