Hi Ash,

Nice, thanks for the explanation. If it does not need to go back to a
worker, it will be perfect.

Thanks,

Ping

Best wishes

Ping Zhang


On Fri, Dec 17, 2021 at 2:16 AM Ash Berlin-Taylor <[email protected]> wrote:

> > A few lightweight long running smart sensors can help to reduce the
> burden of the scheduler and workers.
>
> That is exactly what deferrable tasks are. When a task is deferred it gets
> picked up by a triggerer process, and it runs an in an async IO loop until
> the task is ready to continue.
>
> We could add an optimization for deferred task that don't have any further
> steps and this don't need to go back to a worker (i.e. they also don't need
> to run any further callbacks)
>
> Would that address your concern?
>
> -ash
>
> On Thu, Dec 16 2021 at 22:06:21 -0800, Ping Zhang <[email protected]>
> wrote:
>
> I think the Smart Sensor should not be deprecated.
>
> Based on the doc
> https://airflow.apache.org/docs/apache-airflow/stable/concepts/deferring.html#deferrable-operators-triggers
>  ,
>
> The trigger is run until it fires, at which point its source task is
>> re-scheduled
>> The scheduler queues the task to resume on a worker node
>
>
> deferrable operators/tasks still need to be scheduled by the schedulers.
> Will this incur additional load to the scheduler?
>
> When rescheduling those tasks, it will create bursty traffic to the
> worker, which might cause issues. Also now, workers need to parse those dag
> files again, which could be expensive.
>
> A few lightweight long running smart sensors can help to reduce the burden
> of the scheduler and workers.
>
> Best wishes
>
> Ping Zhang
>
>
> On Thu, Dec 9, 2021 at 12:47 PM Jarek Potiuk <[email protected]> wrote:
>
>> Love it ! Thanks Jed :).
>>
>> On Thu, Dec 9, 2021 at 9:35 PM Jed Cunningham <[email protected]>
>> wrote:
>> >
>> > Thanks everyone for the feedback. Being more user-centric is certainly
>> not a bad thing. (@Jarek, I didn't take your comments personally.)
>> >
>> > I think I've addressed everything discussed here and in the PR, so
>> please take another look. High level, I've added a docs section on
>> migrating and linked to it across the board.
>> >
>> >
>> > On Thu, Dec 9, 2021 at 10:08 AM Kaxil Naik <[email protected]> wrote:
>> >>
>> >> In 100% agreement with the sentiments, Jarek and you have been doing
>> great work at improving that recently.
>> >>
>> >> Looking forward to a single Updating guide from Jed too.
>> >>
>> >>
>> >>
>> >> On Thu, Dec 9, 2021 at 5:04 PM Jarek Potiuk <[email protected]> wrote:
>> >>>
>> >>> This is cool, Jed. Looking forward to it!
>> >>>
>> >>> One more comment. Apologies if my words were understood as "critique"
>> >>> of anyone's job. Jed - if it came to you as such - it was never
>> >>> intended. So - sincere apologies.
>> >>>
>> >>> I am as guilty as anyone of the "poor communication with our users".
>> >>> So the "we have done quite a bad job in the past" applies to me in my
>> >>> mind mostly (And I am personally embarrassed with some of those). I've
>> >>> done a very poor job on that multiple times.
>> >>>
>> >>> My point is really - if there are - really small - things that we can
>> >>> do now to improve what we have without introducing new ways (which
>> >>> will take time I am sure) - why not?
>> >>> Creating a small chapter on "Why no instructions?" and "What should I
>> >>> do now?" - thinking from our poor users' perspective - who have no
>> >>> time to read the docs - takes almost no time. Linking to it from the
>> >>> error message - also simple.
>> >>> This might prevent some questions from our users. So why not do it?
>> >>>
>> >>> J.
>> >>>
>> >>> On Thu, Dec 9, 2021 at 5:56 PM Jed Cunningham
>> <[email protected]> wrote:
>> >>> >
>> >>> > As for UPDATING only being on github, I have a separate proposal in
>> that area coming soon. It likely won't be an issue come time to release
>> 2.3.0 👍.
>> >>> >
>> >>> > On Thu, Dec 9, 2021 at 9:34 AM Jarek Potiuk <[email protected]>
>> wrote:
>> >>> >>
>> >>> >> And I agree with you :) (but with a twist).
>> >>> >>
>> >>> >> I do not say we should remove "UPDATING.md" information.  Not at
>> all.
>> >>> >>
>> >>> >> Providing that:
>> >>> >>
>> >>> >> * UPDATING.md contains both
>> >>> >> * It is available in our User-facing docs
>> >>> >> * it has an anchor to this particular "piece of upgrading"
>> >>> >> * the deprecated error message has a direct link to it to help to
>> find it
>> >>> >>
>> >>> >> I (and our users I hope) would be perfectly happy.
>> >>> >>
>> >>> >> As far as I know. UPDATING.md is only in Github (I just checked
>> and I
>> >>> >> could not find in in airflow.apache.org. So by definition it's
>> not a
>> >>> >> User documentation. It's developer documentation only.
>> >>> >>
>> >>> >> J
>> >>> >>
>> >>> >>
>> >>> >>
>> >>> >>
>> >>> >> On Thu, Dec 9, 2021 at 5:20 PM Kaxil Naik <[email protected]>
>> wrote:
>> >>> >> >
>> >>> >> > Partially agree -- not completely.
>> >>> >> >
>> >>> >> > Firstly what I agree - (1) and (2) points from your email.
>> >>> >> >
>> >>> >> > Disagree the (3) point and the para after that.
>> >>> >> >
>> >>> >> > UPDATING.md is our source of breaking changes. Instead of users
>> just having to rely and checking "deprecation" for 100s of commands, we
>> should be helpful to users by also having a single page where we list all
>> the deprecations.
>> >>> >> >
>> >>> >> > That is another way of being helpful in finding the "right"
>> information and context quickly too. And "Guiding the users" in a different
>> way.
>> >>> >> >
>> >>> >> > On Thu, Dec 9, 2021 at 4:13 PM Jarek Potiuk <[email protected]>
>> wrote:
>> >>> >> >>
>> >>> >> >> I really think about a chapter (Which was missing):
>> >>> >> >>
>> >>> >> >> "How should I approach this migration?"
>> >>> >> >>
>> >>> >> >> 1) explain why there is no 1-1 migration instruction
>> >>> >> >> 2) explain that for every smart sensor they need to use or write
>> >>> >> >> deferrable operator
>> >>> >> >> 3) link to this information from "deprecation message" they
>> will see
>> >>> >> >> in the logs when they use smart sensor (rather than relying on
>> the
>> >>> >> >> fact that they will look at UPDATING.md and find the right part
>> >>> >> >>
>> >>> >> >> That's it, Guiding the users. Being helpful in finding the right
>> >>> >> >> information and context quickly (at the place where they hit
>> the error
>> >>> >> >> and not in one of the 100 pages of documentation that they will
>> only
>> >>> >> >> find by googling.
>> >>> >> >>
>> >>> >> >> J.
>> >>> >> >>
>> >>> >> >> On Thu, Dec 9, 2021 at 5:09 PM Kaxil Naik <[email protected]>
>> wrote:
>> >>> >> >> >
>> >>> >> >> > Just as an FYI - the commit 18 hours ago on that PR already
>> had added "deprecation" in the docs too.
>> >>> >> >> >
>> >>> >> >> > Not only docs, but UPDATING.md, even in the Scheduler logs,
>> so kudos to Jed for taking care of it.
>> >>> >> >> >
>> >>> >> >> > So I don't agree with your comment or suggestion Jarek at
>> least in the context of this discussion as it makes me (at least) read that
>> the PR does not do those things.
>> >>> >> >> >
>> >>> >> >> > re: Tomek's question - it is a very valid question.
>> Unfortunately, I don't see a like-by-like replacement for DAG Authors as
>> different work needs to be done to write an Async operator and make a
>> sensor "smart sensor compatible".
>> >>> >> >> > However, agree that we try to be as clear as possible on what
>> a user might need to do - I just don't know what that would be other than
>> what I suggested in last email and would love the feedback on the PR of
>> what else can be included.
>> >>> >> >> >
>> >>> >> >> > Thanks.
>> >>> >> >> >
>> >>> >> >> > Regards,
>> >>> >> >> > Kaxil
>> >>> >> >> >
>> >>> >> >> > On Thu, Dec 9, 2021 at 4:00 PM Kaxil Naik <
>> [email protected]> wrote:
>> >>> >> >> >>
>> >>> >> >> >> I don't think there is a 1-1 migration path. Async operators
>> supersede what Smart sensors were written to achieve - Cost Savings.
>> >>> >> >> >>
>> >>> >> >> >> Smart Sensors were marked experimental feature for the same
>> reason and there are currently just two Sensors that are Smart
>> >>> >> >> >> sensors compatible.
>> >>> >> >> >>
>> >>> >> >> >> The only thing I can currently think of is writing an async
>> version of the Smart Sensor Hook and Operator differs based on the
>> underlying library that is used and
>> https://airflow.apache.org/docs/apache-airflow/stable/concepts/deferring.html
>> explains how you can write one. Also -
>> https://airflow.apache.org/docs/apache-airflow/stable/concepts/deferring.html#smart-sensors
>> >>> >> >> >>
>> >>> >> >> >>
>> >>> >> >> >>> I believe we have done quite a bad job in the past assuming
>> that our
>> >>> >> >> >>> users read all the discussions and AIPs we write. They
>> don't. They
>> >>> >> >> >>> need some guidance.
>> >>> >> >> >>
>> >>> >> >> >>
>> >>> >> >> >> Which instances? I am just curious to know what are those
>> bad instances where we "assumed" that our users read mailing list and not
>> covered it in UPDATING.md or docs.
>> >>> >> >> >>
>> >>> >> >> >> Regards,
>> >>> >> >> >> Kaxil
>> >>> >> >> >>
>> >>> >> >> >> On Thu, Dec 9, 2021 at 3:46 PM Jarek Potiuk <
>> [email protected]> wrote:
>> >>> >> >> >>>
>> >>> >> >> >>> Extremely good point Tomek.
>> >>> >> >> >>>
>> >>> >> >> >>> Also as Ephraim pointed out in the PR - IMHO any time when
>> we do
>> >>> >> >> >>> deprecation we should have a note in our docs, explaining
>> at the very
>> >>> >> >> >>> least how the users should approach the migration as
>> correctly pointed
>> >>> >> >> >>> out by @turbaszek in the devlist.
>> >>> >> >> >>> I think this should be a standard of any deprecation we do.
>> >>> >> >> >>>
>> >>> >> >> >>> I believe we have done quite a bad job in the past assuming
>> that our
>> >>> >> >> >>> users read all the discussions and AIPs we write. They
>> don't. They
>> >>> >> >> >>> need some guidance.
>> >>> >> >> >>>
>> >>> >> >> >>> J.
>> >>> >> >> >>>
>> >>> >> >> >>> On Wed, Dec 8, 2021 at 11:44 PM Tomasz Urbaszek <
>> [email protected]> wrote:
>> >>> >> >> >>> >
>> >>> >> >> >>> > Do we have documentation about how to migrate from smart
>> sensors to deferrable operators?
>>
>

Reply via email to