I like TP's dual-subclass approach. I see it as the most well-scoped and the cleanest solution here and I think it resolves most of the open questions. Raise a specific exception that inherits from both AirflowException and the target type, the existing catch based code will keep working, new code gets specificity, no deprecation cycles or windows to manage.
On Jarek's suggestions database idea, I think that's worth a separate thread. The exception hierarchy should not be designed around a storage/retrieval system that does not exist yet. And the database idea does not actually depend on a new exception hierarchy either, it could work off existing error messages. Conflating the two makes both harder to evaluate. Jarek, a couple of questions if you want to push the suggestions database idea forward: - Where does the database live and who owns it? Is this a new repo, part of the docs, something in CLAUDE.md? - What does the write path look like? Agents write during PR review, but what would be the trigger? any new exception type, any failing test, or something else? Happy to ask further details and see it in a different proposal if we have appetite at the moment. Thanks & Regards, Amogh Desai On Sun, May 31, 2026 at 10:07 PM Jarek Potiuk <[email protected]> wrote: > > Okay. That sounds great. It would be interesting to see what the system > > Well.. If others agree that is :). That is just one of the options :) > > On Sun, May 31, 2026 at 6:23 PM Sameer Mesiah <[email protected]> > wrote: > > > Okay. That sounds great. It would be interesting to see what the system > > looks like when it is live. > > > > Thanks, > > Sameer Mesiah. > > > > On Sun, 31 May 2026 at 17:15, Jarek Potiuk <[email protected]> wrote: > > > > > Hi Sameer, > > > > > > Thanks for your thoughtful questions! Here are my thoughts on those > > points: > > > > > > 1) I see the value here mainly in terms of efficiency and > > sustainability. > > > By computing a solution once and storing it, we can save the time and > > > tokens an LLM would otherwise use to solve the same problem over and > > over. > > > It's a nice way to be more environmentally friendly, too! > > > > > > 2) Since many contributions are already agent-assisted, we can address > > > hallucinations by adding some guidance to AGENTS.md. Rather than > relying > > on > > > periodic checks, we can ask agents to perform their own consistency and > > > hallucination checks during the PR process. This keeps a human > naturally > > in > > > the loop during review, turning token usage into a great opportunity > for > > > automated quality control—which we ask our contributors' agents to > > perform. > > > > > > This is actually a very interesting "twist" - because this also nicely > > > approaches the "assymetry" we have by having a lot more AI assisted > > > contributions. We can ask - in our AGENTS.md those agents to do **way** > > mor > > > things to make our projects better - all of them done on the local > > machines > > > of the contributors, with their tokens being spent. > > > > > > Best, > > > Jarek Potiuk > > > > > > > > > On Sun, May 31, 2026 at 4:55 PM Sameer Mesiah <[email protected]> > > > wrote: > > > > > > > Hi Jarek, > > > > > > > > I do have a few questions about the "suggestions database" point: > > > > > > > > 1) Previously, it was agreed that since LLMs work well with > > unstructured > > > > error messages, there was minimal benefit to an additional taxonomy > > layer > > > > on top of existing exceptions. I was just hoping you could elaborate > on > > > the > > > > marginal value of a structured data store over better exception > > messages? > > > > > > > > 2) If we were to implement this "suggestions database", do you > believe > > > the > > > > hallucination rate of the current agentic systems we have available > is > > > low > > > > enough that it requires minimal manual intervention to the extent > that > > it > > > > is worthwhile to introduce? If a maintainer/committer/contributor has > > to > > > > skim through the database periodically to remove hallucinations, > would > > we > > > > not be expending tokens in favour of negligible time savings? > > > > > > > > Granted, I could be overestimating the hallucination rate of these > > > agentic > > > > systems (I have never used them personally) so it might be the ideal > > > > solution. It's an interesting idea anyway so I would love to hear > more > > > > about it. > > > > > > > > On Sun, 31 May 2026 at 15:37, Jarek Potiuk <[email protected]> wrote: > > > > > > > > > Hi all, > > > > > > > > > > TP's suggestion aligns with Omkar’s earlier proposal ( > > > > > https://lists.apache.org/thread/f275xo3tjd4olsmrc8nggncs62fjnl0x). > I > > > see > > > > > two main approaches: > > > > > > > > > > 1. Generic Python exceptions: Simple and concise, but difficult > for > > > > > providing nuanced diagnostic or resolution guidance in logs. > > > > > 2. Custom exception hierarchy: Allows us to accumulate specific > > > > > troubleshooting documentation for users. This also enables the > > > > non-breaking > > > > > compatibility TP mentioned. > > > > > > > > > > While I previously favored simplicity, I now lean toward a > structured > > > > > hierarchy. In an agentic context, we could automate the maintenance > > of > > > a > > > > > "suggestions database" by scanning past issues. This would provide > > > users > > > > > (and agents) with clear, documented resolution steps, ultimately > > saving > > > > > time and tokens. > > > > > > > > > > Best, > > > > > Jarek Potiuk > > > > > > > > > > > > > > > On Sun, May 31, 2026 at 4:19 PM Tzu-ping Chung via dev < > > > > > [email protected]> wrote: > > > > > > > > > > > I think we can make this non-breaking entirely? We can simply > > raise a > > > > > > subclass of both AirflowException and whatever we want to migrate > > to. > > > > > > > > > > > > If we want to, we can emit a deprecation warning if > > AirflowException > > > is > > > > > > imported directly by the user, but I feel even that is > particularly > > > > > needed. > > > > > > > > > > > > Raising a custom exception that’s as specific as practical is > > always > > > > the > > > > > > best practice. Using a custom exception that subclasses both > > built-in > > > > and > > > > > > AirflowException (but not AirflowException itself) is a good idea > > on > > > > its > > > > > > own. > > > > > > > > > > > > TP > > > > > > > > > > > > > > > > > > > On 31 May 2026, at 21:25, Jens Scheffler <[email protected]> > > > > wrote: > > > > > > > > > > > > > > Hi, > > > > > > > > > > > > > > I do not have a stron opinion on this, with the past decision > we > > > > mainly > > > > > > wanted to ensure code is clean. > > > > > > > > > > > > > > For me also Option (3) would be viable: We leave existing code > as > > > it > > > > is > > > > > > and just ensure that new code is clean. But I am also okay to > join > > > the > > > > > > other majority opinion. > > > > > > > > > > > > > > Jens > > > > > > > > > > > > > > On 31.05.26 14:08, Sameer Mesiah wrote: > > > > > > >> Hi, > > > > > > >> > > > > > > >> Thanks for opening this thread. This was long overdue > > considering > > > > that > > > > > > >> current migration efforts for AirflowException are more > > scattered > > > > > rather > > > > > > >> than a consolidated pre-planned wave. > > > > > > >> > > > > > > >> In general, I agree with your overall approach. But there are > a > > > few > > > > > > things > > > > > > >> I want to raise in response to your proposal: > > > > > > >> > > > > > > >> On 1) > > > > > > >> > > > > > > >> I think Amogh already expanded on thiss somewhat. But I > believe > > it > > > > is > > > > > > >> important to understand that AirflowException is not > constrained > > > to > > > > > > >> the provider layer but often bubble up from Airflow core.For > > > example > > > > > > >> task_runner.py explicitly catches AirflowException as part of > > task > > > > > > >> execution and retry handling, and there are also usages in the > > > hook > > > > > and > > > > > > >> connection layers. > > > > > > >> > > > > > > >> But depending on whether we scope this to core + providers vs > > > > > providers > > > > > > >> only, the migration strategy looks quite different. If the > > > > intention > > > > > is > > > > > > >> provider-only migration, users will continue to encounter > > > > > > AirflowException > > > > > > >> from core Airflow components, which may make the overall > > exception > > > > > model > > > > > > >> feel somewhat inconsistent. If the intention is project-wide > > > > > > deprecation, I > > > > > > >> think we need a broader plan than provider-level migrations > > alone. > > > > > > >> > > > > > > >> Also, as Amogh already implied (he is free to correct me in > > case I > > > > > > >> misinterpreted his reservations), deprecation warnings may not > > be > > > > > > >> worthwhile here. Unlike a deprecated API or method, this is > not > > > > > > something > > > > > > >> users are actively choosing to call. In most cases, Airflow > > itself > > > > is > > > > > > >> deciding which exception type to raise. I am leaning against > > > > warnings > > > > > > due > > > > > > >> to the potential for log pollution and the relatively limited > > > value > > > > > they > > > > > > >> provide. > > > > > > >> > > > > > > >> On 2) > > > > > > >> > > > > > > >> I agree with your per-package migration proposal. But as Amogh > > > said, > > > > > we > > > > > > >> should ensure new exceptions inherit from AirflowException to > > > > prevent > > > > > > >> breaking changes across the provider ecosystem. I must add > that > > > > there > > > > > > will > > > > > > >> inevitably be a period where users run multiple providers, > some > > > > > migrated > > > > > > >> and some not. This may result in a mixture of exception types > > > > > appearing > > > > > > in > > > > > > >> logs and stack traces. While not ideal, this is probably > > > unavoidable > > > > > > given > > > > > > >> independent provider versioning. As long as migrations are > > clearly > > > > > > >> documented, I do not see this as a major issue. > > > > > > >> > > > > > > >> My answers to your questions: > > > > > > >> > > > > > > >> *Should we add a deprecation warning to AirflowException?* > > > > > > >> > > > > > > >> I am +0. Not strongly opposed, but I am leaning no due to the > > risk > > > > of > > > > > > log > > > > > > >> pollution and the fact that the warning would be emitted by > > > Airflow > > > > > > itself > > > > > > >> rather than as a result of an explicit user action. > > > > > > >> > > > > > > >> *Is the per-package migration approach reasonable?* > > > > > > >> > > > > > > >> Yes. No blocking concerns from my side. > > > > > > >> > > > > > > >> *Should we define a project-wide exception taxonomy, or leave > > > > > exception > > > > > > >> choices to individual provider teams?* > > > > > > >> > > > > > > >> I would leave this to individual provider teams. The providers > > > have > > > > > very > > > > > > >> different APIs, error models, and opinions around exception > > > > handling. > > > > > > >> Attempting to enforce a central taxonomy across all providers > > may > > > be > > > > > > >> impractical and may not provide much practical value over what > > we > > > > > > currently > > > > > > >> have in place. > > > > > > >> > > > > > > >> Thanks, > > > > > > >> Sameer Mesiah. > > > > > > >> > > > > > > >> On Sat, 30 May 2026 at 10:59, Shahar Epstein < > [email protected] > > > > > > > > wrote: > > > > > > >> > > > > > > >>> Hi all, > > > > > > >>> > > > > > > >>> This follows up on the earlier lazy-consensus > > > > > > >>> < > > > https://lists.apache.org/thread/m86gs4mqt8hq1n26g0pp0fq5h5g0x2q9> > > > > > and > > > > > > >>> related discussions around deprecating AirflowException. > > > > > > >>> > > > > > > >>> While working on a cleanup PR for the Google Cloud Run > operator > > > in > > > > > the > > > > > > >>> Google provider (#67769), Amogh pointed out that replacing > > > > > > AirflowException > > > > > > >>> with built-in or custom exceptions is a breaking change. > Users > > > may > > > > > > rely on > > > > > > >>> AirflowException in callbacks (on_failure_callback, > > > > > on_retry_callback) > > > > > > or > > > > > > >>> in custom wrappers that catch it explicitly. > > > > > > >>> > > > > > > >>> Rather than making these changes incrementally across > > individual > > > > PRs, > > > > > > I'd > > > > > > >>> like to discuss a coordinated migration strategy. > > > > > > >>> Proposal > > > > > > >>> > > > > > > >>> *1. Emit a deprecation warning from > > > **AirflowException.__init__**.* > > > > > > >>> > > > > > > >>> Although emitting warnings when exceptions are raised is > > > uncommon, > > > > it > > > > > > is > > > > > > >>> justified here because users may depend on AirflowException > in > > > > > > callbacks > > > > > > >>> and exception handlers. This provides advance notice before > > > > > > >>> package-specific migrations occur. > > > > > > >>> > > > > > > >>> *2. Migrate per package in the next breaking release of that > > > > > package.* > > > > > > >>> > > > > > > >>> Require a single migration PR per package (airflow-core and > > each > > > > > > provider), > > > > > > >>> rather than spreading exception changes across multiple PRs > and > > > > > > releases. > > > > > > >>> The migration would be included in that package's next > breaking > > > > > > release and > > > > > > >>> documented accordingly. > > > > > > >>> Questions > > > > > > >>> > > > > > > >>> - Should we add a deprecation warning to AirflowException? > > > > > > >>> - Is the per-package migration approach reasonable? > > > > > > >>> - Should we define a project-wide exception taxonomy, or > > leave > > > > > > exception > > > > > > >>> choices to individual provider teams? > > > > > > >>> > > > > > > >>> If there's agreement, I'll follow up with a lazy-consensus > > > > proposal. > > > > > > >>> > > > > > > >>> Thanks, > > > > > > >>> > > > > > > >>> > > > > > > >>> Shahar > > > > > > >>> > > > > > > > > > > > > > > > > > --------------------------------------------------------------------- > > > > > > > To unsubscribe, e-mail: [email protected] > > > > > > > For additional commands, e-mail: [email protected] > > > > > > > > > > > > > > > > > > > > > > > > > > > --------------------------------------------------------------------- > > > > > > To unsubscribe, e-mail: [email protected] > > > > > > For additional commands, e-mail: [email protected] > > > > > > > > > > > > > > > > > > > > > > > > > > >
