Hi,
Thanks Wei for taking the lead in starting to implement! Hope I can
review the next days.
as I was writing the AIP together with Vikram I was and still am for
(=+1) to keep it "human" centric. Also adding an API such that somebody
else is able to roll their whatever UI and not being locked into Airflow
UI but still with the aim to loop-in humans.
For the provider question I am for a separate provider because (1) it
was as such in the AIP, (2) I see it is optional and we should not force
every install to have this (as it has not been there the last 10 years I
assume there are many installs not needing it actually and some
objections were raised in the discussion that it is accepted if it is an
optional feature which it would not be if in standard provider) as well
as (3) we need to adjust the DB and slightly extend this for the human
response data storage - and I would feel uncomfortable to force this DB
extension with every install... then we could also directly package this
into core - but as (2) it should be an optional feature.
As well as (4) other common things like http, ftp, git, sftp, smtp are
also pretty like stdnard but are also separate providers. From point of
security (5) every additional thing adds a bit of complexity and if you
want to make your setup secure you want to slim it down to the functions
you need. Even though minimal if no human interaction is needed then I
think we should not force every install to have this.
TLDR I'd therefore favor (+1) a separate "human" provider; not favoring
(but +0) adding this to standard.
Jens
On 23.06.25 08:02, Amogh Desai wrote:
I was not strongly against using "human" -- it just felt a little odd and
confusing to me at first.
Jarek's email has convinced me that having HITH is contextual in the AI
space and it is kinda what
we are doing with this AIP - 90. In fact, using "interactive" now seems odd
that it is not descriptive enough
or doesn't highlight the intention of the operator enough.
I do not have concerns with whatever we decide to name it :D
Thanks & Regards,
Amogh Desai
On Mon, Jun 23, 2025 at 9:42 AM Jarek Potiuk <ja...@potiuk.com> wrote:
In standard Provider, yes
Re: name: I changed my opinion. Previously I raised concerns about it, but
they are gone. The name is IMHO perfect.
Why do I think "Human-In-The-Loop" is the **right** name. It's a very
popular term in AI workflows, and used to interact with the "real" human,
and it has a very concrete meaning. Also I think it's really, really worth
looking at the talk by Andrey Karpathy published a few days ago
https://www.youtube.com/watch?v=LCEmiRjPEtQ - I think it is very
insightful. Andrey coined the term "Vibe Coding", and I think he is one of
the smartest people in the AI space who is not hype-driven - i.e. he seems
to genuinely think that AI is another technology change that is reinventing
how we do software. Unlike many of others he is not "selling" their product
in AI, he seems to be focused on one thing that I believe also is important
i.e. "Keeping Human in the Loop".
One of the very interesting things I've learned from that talk is how
important it is to provide a good User Interface to AI. I.e. that the chat
interface is cool, and everything but the crucial part of the AI
interaction is to wrap the AI results into actionable, quick and "nice" way
of interacting with various aspects of AI by Human(s), when the input is
not only important, but crucial to get the real value of AI.
In this context I think we should focus to make sure that our "Human In
the Loop" is indeed designed for the Human - not for LLM imitating Human,
not for Agents. It should have a nice, pleasant and efficient
UI, that should allow surfacing all the information that is necessary for
the Human to make the decision. That information should be
nicely formattable, and you should be able to use the typical way
that people interact with it - with controls and everything they are
used to. A good interface example of a UI is when you use Copilot in your
IDE for the translation. For example, the information you get (as human) is
targeted for humans and is very actionable. It is put in context, you can
interact with it individually by accepting individual suggestions (or
rejecting them) or accept/reject things in bulk.
Here is an example: (for those who do not see embedded picture - link here
https://ibb.co/3Y03xN06)
[image: Screenshot 2025-06-23 at 06.05.38.png]
We should design "Human In the Loop" of ours in a very similar way - i.e.
give the author of the "HIL" interaction capability of adding UI
components, surfacing the right information - and having rich interaction.
Maybe not all the bells and whistles initially (for example it's ok for now
to just have bulk decisions on the whole interaction, but I think this
should be our long-term design goal to allow for richer interactions. And -
in this context - "Human in the loop" is a very appropriate name.
BTW. Slightly related - there is a blog post coming about it from a few of
us about AI with internationalisation and how we made it to follow that
(pretty naturally) with Open-Source spirit - by making sure that we keep
Human In the Loop and that it is designed to follow the Open Source Spirit,
Foster collaboration.
On Mon, Jun 23, 2025 at 4:42 AM Wei Lee <weilee...@gmail.com> wrote:
Got it! Yes, it makes sense to keep the phrase widely used. Thanks a lot!
As a compromise, I will try something like `HITHOperator`, which may
address some of the concerns. We can always rename it to whatever we decide
before the release. I will also send a follow-up email to this thread once
it's ready for review, so that anyone interested can take a look.
Best,
Wei
On Jun 23, 2025, at 10:27 AM, Vikram Koka <vik...@astronomer.io.INVALID>
wrote:
I agree with the standard provider approach.
On Sun, Jun 22, 2025 at 7:26 PM Vikram Koka <vik...@astronomer.io>
wrote:
Thanks Wei, I really appreciate the work, and will review it as soon as
possible.
With respect to the naming, I believe the Human-in-the-loop is the
right
phrase, because that is recognized as such both in older "legacy"
systems,
as well as the new AI solutions. I agree that it may be less than ideal
from a technical perspective, but from a user perspective, I believe
it is
better to stick with a known term, rather than to invent our own.
On Sun, Jun 22, 2025 at 7:07 PM Wei Lee <weilee...@gmail.com> wrote:
Hi fellow Airflower,
I am currently working on a PoC for AIP-90. I've incorporated some
suggestions based on comments in the voting thread and Jira page.
Since
they have not yet been included in the AIP, I want to confirm with
everyone
to ensure I'm on the right track.
1. Many have expressed concerns about the term “Human,” so I'm now
using
the term “Interactive” as suggested by Among. For example, I change
"HumanOperator" to “InteractiveOperator".
2. This functionality is now part of the standard provider rather than
being a separate provider as suggested by Bas and Ash.
Here is the PoC PR. https://github.com/apache/airflow/pull/52053
It's not ready to be reviewed yet, but I'll try to wrap it up over the
next few days. Several features are still missing and will be
implemented
in the following pull requests. Thanks!
Best regards,
Wei Lee
---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscr...@airflow.apache.org
For additional commands, e-mail: dev-h...@airflow.apache.org
---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscr...@airflow.apache.org
For additional commands, e-mail: dev-h...@airflow.apache.org
