apex-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Pramod Immaneni <pra...@datatorrent.com>
Subject Re: Schema Discovery Support in Apex Applications
Date Fri, 03 Feb 2017 16:47:23 GMT
On Fri, Feb 3, 2017 at 7:59 AM, Thomas Weise <thw@apache.org> wrote:

> Agreed. As noted the main concern was the ability to support idempotency.
> It isn't really "re-ordering" because when you have multiple input ports,
> there isn't any ordering guarantee within a streaming window.
>

The reordering I was referring to is the reordering that would happen
either within the container that receives the tuples or the one that sends
the tuples, by holding on to the control tuple(s) till the window boundary
and not the order in which data is received across the different paths.
Also, the idempotency concern is that the operator developer might make a
mistake by making the incorrect ordering assumption, that you mentioned,
and do the wrong thing isn't it? It is not that this approach will break
idempotency or make it not possible to achieve it.

It looks like we have an agreement on this approach so far. Do folks see
the need to summarize this in the control tuple discussion thread as well.
I can do that.

Thanks


> The end window boundary is good when the control tuple needs to be
> processed after all associated data tuples (which is the case for
> watermarks).
>
> For schema it is the opposite, the schema needs to be seen before all data
> tuples. The scenario of multiple input ports needs to be considered here as
> well.
>
> Thomas
>
>
> On Thu, Feb 2, 2017 at 9:59 AM, Vlad Rozov <v.rozov@datatorrent.com>
> wrote:
>
> > I second the proposal to revisit custom control tuple delivery and
> > re-ordering. Schema support brings a use case that was missing when we
> > discussed custom control tuples.
> >
> > Thank you,
> >
> > Vlad
> >
> >
> > On 2/1/17 21:56, Pramod Immaneni wrote:
> >
> >> This can be done neatly and possibly completely outside the engine if we
> >> are able to deliver schema information via the control tuple mechanism.
> >> Current control tuple proposal reorders the control tuple to be
> delivered
> >> at the end of the window to the operator. This would not be feasible for
> >> schemas as the schema would need to be delivered before the data. If we
> >> can
> >> reconsider this behavior and consider not reordering the control tuple
> it
> >> would work in this use case. We can have further discussions on the
> >> scenarios this raises like what to do when there are multiple paths for
> >> data, how control tuples get delivered to unifiers and look into
> >> suggestions like synchronizing on control tuple boundaries and other
> ways
> >> to solve these. What do you guys think?
> >>
> >> On Wed, Feb 1, 2017 at 8:27 PM, Thomas Weise <thw@apache.org> wrote:
> >>
> >> I think dynamic schema would be good to consider (schema known and
> >>> possibly
> >>> changing at runtime). Some applications cannot be written under the
> >>> assumption that the schema is known upfront.
> >>>
> >>> Also, does this really need to leak into the engine? I think it would
> be
> >>> good to consider alternatives and tradeoffs.
> >>>
> >>> Thomas
> >>>
> >>>
> >>> On Mon, Jan 30, 2017 at 10:44 PM, Chinmay Kolhatkar <
> >>> chinmay@datatorrent.com
> >>>
> >>>> wrote:
> >>>> Consumer of output port operator schema is going next downstream
> >>>>
> >>> operator.
> >>>
> >>>>
> >>>> On Tue, Jan 31, 2017 at 4:01 AM, Sergey Golovko <
> sergey@datatorrent.com
> >>>> >
> >>>> wrote:
> >>>>
> >>>> Sorry, I’m a new person in the APEX team. And I don't understand
> >>>>>
> >>>> clearly
> >>>
> >>>> who are consumers of the output port operator schema(s).
> >>>>>
> >>>>> 1. If the consumers are non-run-time callers like the application
> >>>>>
> >>>> manager
> >>>
> >>>> or UI designer, maybe it makes sense to use Java static method(s) to
> >>>>> retrieve the output port operator schema(s). I guess the performance
> >>>>>
> >>>> of a
> >>>
> >>>> single call of a static method via reflection can be ignored.
> >>>>>
> >>>>> 2. If the consumer is next downstream operator, maybe it makes sense
> to
> >>>>> send an output port operator schema from upstream operator to next
> >>>>> downstream operator via the stream. The corresponded methods that
> would
> >>>>> send and receive the schema should be declared in the
> >>>>> interface/abstract-class of the upstream and downstream operators.
> The
> >>>>> sending/receiving of an output schema should be processed right
> before
> >>>>>
> >>>> the
> >>>>
> >>>>> sending of the first data record via the stream.
> >>>>>
> >>>>> One of examples of a typical implementation for sending of metadata
> >>>>>
> >>>> with
> >>>
> >>>> a
> >>>>
> >>>>> regular result set is the sending of JDBC metadata as a part of
JDBC
> >>>>>
> >>>> result
> >>>>
> >>>>> set. And I hope the output schema (metadata of the streamed data)
in
> >>>>>
> >>>> the
> >>>
> >>>> implementation should contain not only a signature of the streamed
> >>>>>
> >>>> objects
> >>>>
> >>>>> (like field names and data types), but also any other properties
of
> the
> >>>>> data that can be useful by the schema receiver to process the data
> (for
> >>>>> instance, a delimiter for CSV record stream).
> >>>>>
> >>>>> Thanks,
> >>>>> Sergey
> >>>>>
> >>>>> On 2017-01-25 01:47 (-0800), Chinmay Kolhatkar <
> >>>>>
> >>>> chinmay@datatorrent.com>
> >>>
> >>>> wrote:
> >>>>>
> >>>>>> Thank you all for the feedback.
> >>>>>>
> >>>>>> I've created a Jira for this: APEXCORE-623 and I'll attach the
same
> >>>>>> document and link to this mailchain there.
> >>>>>>
> >>>>>> As a first part of this Jira, there are 2 steps I would like
to
> >>>>>>
> >>>>> propose:
> >>>>
> >>>>> 1. Add following interface at com.datatorrent.common.util.
> >>>>>>
> >>>>> SchemaAware.
> >>>
> >>>> interface SchemaAware {
> >>>>>>
> >>>>>> Map<OutputPort, Schema> registerSchema(Map<InputPort,
Schema>
> >>>>>>
> >>>>> inputSchema);
> >>>>>
> >>>>>> }
> >>>>>>
> >>>>>> This interface can be implemented by Operators to communicate
its
> >>>>>>
> >>>>> output
> >>>>
> >>>>> schema(s) to engine.
> >>>>>> Input to this schema will be schema at its input port.
> >>>>>>
> >>>>>> 2. After LogicalPlan is created call SchemaAware method from
> upstream
> >>>>>>
> >>>>> to
> >>>>
> >>>>> downstream operator in the DAG to propagate the Schema.
> >>>>>>
> >>>>>> Once this is done, changes can be done in Malhar for the operators
> in
> >>>>>> question.
> >>>>>>
> >>>>>> Please share your opinion on this approach.
> >>>>>>
> >>>>>> Thanks,
> >>>>>> Chinmay.
> >>>>>>
> >>>>>>
> >>>>>>
> >>>>>>
> >>>>>> On Wed, Jan 18, 2017 at 2:31 PM, Priyanka Gugale <priyag@apache.org
> >
> >>>>>>
> >>>>> wrote:
> >>>>>
> >>>>>> +1 to have this feature.
> >>>>>>>
> >>>>>>> -Priyanka
> >>>>>>>
> >>>>>>> On Tue, Jan 17, 2017 at 9:18 PM, Pramod Immaneni <
> >>>>>>>
> >>>>>> pramod@datatorrent.com>
> >>>>>
> >>>>>> wrote:
> >>>>>>>
> >>>>>>> +1
> >>>>>>>>
> >>>>>>>> On Mon, Jan 16, 2017 at 1:23 AM, Chinmay Kolhatkar <
> >>>>>>>>
> >>>>>>> chinmay@apache.org>
> >>>>>
> >>>>>> wrote:
> >>>>>>>>
> >>>>>>>> Hi All,
> >>>>>>>>>
> >>>>>>>>> Currently a DAG that is generated by user, if contains
any
> >>>>>>>>>
> >>>>>>>> POJOfied
> >>>>
> >>>>> operators, TUPLE_CLASS attribute needs to be set on each and
> >>>>>>>>>
> >>>>>>>> every
> >>>>
> >>>>> port
> >>>>>
> >>>>>> which receives or sends a POJO.
> >>>>>>>>>
> >>>>>>>>> For e.g., if a DAG is like File -> Parser ->
Transform -> Dedup
> >>>>>>>>>
> >>>>>>>> ->
> >>>>
> >>>>> Formatter -> Kafka, then TUPLE_CLASS attribute needs to be set
> >>>>>>>>>
> >>>>>>>> by
> >>>
> >>>> user
> >>>>>
> >>>>>> on
> >>>>>>>
> >>>>>>>> both input and output ports of transform, dedup operators
and
> >>>>>>>>>
> >>>>>>>> also
> >>>>
> >>>>> on
> >>>>>
> >>>>>> parser output and formatter input.
> >>>>>>>>>
> >>>>>>>>> The proposal here is to reduce work that is required
by user to
> >>>>>>>>>
> >>>>>>>> configure
> >>>>>>>
> >>>>>>>> the DAG. Technically speaking if an operators knows
input
> >>>>>>>>>
> >>>>>>>> schema
> >>>
> >>>> and
> >>>>>
> >>>>>> processing properties, it can determine output schema and
> >>>>>>>>>
> >>>>>>>> convey
> >>>
> >>>> it to
> >>>>>
> >>>>>> downstream operators. This way the complete pipeline can be
> >>>>>>>>>
> >>>>>>>> configured
> >>>>>
> >>>>>> without user setting TUPLE_CLASS or even creating POJOs and
> >>>>>>>>>
> >>>>>>>> adding
> >>>>
> >>>>> them
> >>>>>
> >>>>>> to
> >>>>>>>>
> >>>>>>>>> classpath.
> >>>>>>>>>
> >>>>>>>>> On the same idea, I want to propose an approach
where the
> >>>>>>>>>
> >>>>>>>> pipeline
> >>>>
> >>>>> can
> >>>>>
> >>>>>> be
> >>>>>>>
> >>>>>>>> configured without user setting TUPLE_CLASS or even
creating
> >>>>>>>>>
> >>>>>>>> POJOs
> >>>>
> >>>>> and
> >>>>>
> >>>>>> adding them to classpath.
> >>>>>>>>> Here is the document which at a high level explains
the idea
> >>>>>>>>>
> >>>>>>>> and
> >>>
> >>>> a
> >>>>
> >>>>> high
> >>>>>
> >>>>>> level design:
> >>>>>>>>> https://docs.google.com/document/d/1ibLQ1KYCLTeufG7dLoHyN_
> >>>>>>>>> tRQXEM3LR-7o_S0z_porQ/edit?usp=sharing
> >>>>>>>>>
> >>>>>>>>> I would like to get opinion from community about
feasibility
> >>>>>>>>>
> >>>>>>>> and
> >>>
> >>>> applications of this proposal.
> >>>>>>>>> Once we get some consensus we can discuss the design
in
> >>>>>>>>>
> >>>>>>>> details.
> >>>
> >>>> Thanks,
> >>>>>>>>> Chinmay.
> >>>>>>>>>
> >>>>>>>>>
> >
>

Mime
  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message