Hey everyone,

We had some discussions recently in Slack and GitHub Issues, about a
new approach that apparently we should take when it comes to hooks.
Seems that the "new" approach for current and future Hooks should be
to make them stateless. and not store connection information (or any
other non-serializable (and possibly state) information, because
keeping them "stateful" makes it not possible to fit into the
"Deferrable Operators"  approach.

I "feel" where it is coming from and there are some examples of "true"
deferrable (with remote connections) Operators being developed -
notably: Databricks operators :
https://github.com/apache/airflow/pull/19736. But I have no full
understanding/don't realize the full consequences yet. This might have
an impact on the reviews of committers when it comes to new
operators/hooks.

So maybe we should discuss it here and let everyone know about this
change of approach (we have many hooks that have state and keep
connections once created).

Is my understanding correct? Is it about "non-serializable" fields of
Hooks? Or any state fields ? Can someone who already has a deeper
understanding of this issue explain the context to others as well
(maybe an example too where it could go wrong :) ?

J.

Reply via email to