Hi all, Several people mentioned Spark event listeners. After looking more closely into this feature, I think it actually looks very similar to what I'm proposing:
- The primary intended use cases that I could find are : monitoring job progress, tracking stages and task completion, gathering metrics [1]. - Events are distributed asynchronously via an internal component called SparkListenerBus [2], which manages an event bus and a single-threaded event queue. - The API consists solely of methods that return void without checked exceptions: IOW, the API wasn't designed to allow for listeners to interact with the server (other than throwing an unchecked exception, of course). - Afaict if a listener throws, the bus catches the exception and moves on. I'm not an expert in this topic so I might be wrong here, but the suggestion that Spark event listeners were designed to allow listeners to modify Spark's behavior doesn't look accurate to me. Thanks, Alex [1]: https://www.ibm.com/think/insights/apache-spark-monitoring-using-listeners-and-data-quality-libraries [2]: https://spark.apache.org/docs/latest/api/java/org/apache/spark/scheduler/SparkListenerBus.html On Tue, Nov 11, 2025 at 1:30 PM Dmitri Bourlatchkov <[email protected]> wrote: > > Hi Eric, > > I would very much prefer not to use the event listener SPI as a means to > control the operation of the Polaris Server. > > More specifically, I believe that any error / exception in an event > listener should not affect the processing of the request. > > If we need custom callbacks to control some aspects of the server > behaviour, let's define a dedicated SPI for that, but, IMHO, it should be > outside the scope of events. WDYT? > > Thanks, > Dmitri. > > On Mon, Nov 10, 2025 at 8:55 PM Eric Maynard <[email protected]> > wrote: > > > In fact, shouldn’t it be exclusively a listener’s decision on whether an > > event is handled in a blocking way or not? As was noted in a past thread on > > events, much of the utility of the event framework comes from the ability > > to introduce custom logic and hooks into the normal operation of Polaris. > > > > If you wish, for example, to prevent the creation of more than 1k tables > > with some given prefix, you can do so using a listener. If the event which > > might trigger that logic becomes non-blocking, you would no longer be able > > to block/fail the create table request. > > > > I think maybe it’s the name “event”, but we seem to keep conflating these > > hooks with the iceberg events or auditing events when they are not exactly > > the same thing. > > > > —EM > > > > On Mon, Nov 10, 2025 at 8:47 PM Adnan Hemani > > <[email protected]> wrote: > > > > > Hi Alex, > > > > > > Thanks for writing down the proposal for this! As I had previously > > > suggested this when implementing the Persistence of Polaris Events > > > <https://github.com/apache/polaris/pull/1844>, I am obviously very much > > in > > > favor of doing this :) > > > > > > A few questions I have regarding your vision of how we should implement > > > this: > > > * Are you envisioning anything for being able to make dependencies > > between > > > event listeners? Or are we taking a set direction that Event Listeners > > > should be independent of each other? > > > * In some listeners we have the ability to make events emission > > synchronous > > > [example > > > < > > > > > https://github.com/apache/polaris/blob/main/runtime/service/src/main/java/org/apache/polaris/service/events/jsonEventListener/aws/cloudwatch/AwsCloudWatchEventListener.java#L186 > > > >]. > > > How do we plan to support/advise (or not...) that with the introduction > > > of @Blocking annotations. > > > > > > Best, > > > Adnan Hemani > > > > > > On Mon, Nov 10, 2025 at 11:29 AM Yufei Gu <[email protected]> wrote: > > > > > > > Thanks for the reply. It's overall a good idea to have async event > > > > listeners so that they are not blocking each other. > > > > > > > > One downside of the async ones is that event order isn't deterministic. > > > > For example, event listeners of Spark need the order to understand the > > > > execution semantics. I think Polaris is fine with that, given the ts of > > > > each event is generated by Polaris. The downstream can still figure out > > > the > > > > order. > > > > > > > > Thanks Pierre for sharing, I think any I/O-bound or potentially slow > > > > listener should be annotated with @Blocking. That ensures we keep the > > > event > > > > loop responsive and avoid impacting REST latency. > > > > > > > > Yufei > > > > > > > > > > > > On Mon, Nov 10, 2025 at 9:43 AM Alexandre Dutra <[email protected]> > > > wrote: > > > > > > > > > Hi all, > > > > > > > > > > Answering the questions above: > > > > > > > > > > > However, we can easily make sure that we use Quarkus's SmallRye > > Fault > > > > > Tolerance > > > > > > > > > > Yes, that was my idea. It's not so much the bus itself that needs to > > > > > be fault tolerant, but the receiving end, that is, the listeners. A > > > > > listener can fail for a variety of reasons (e.g. remote broker > > > > > unavailable), it would be nice to be able to backoff and retry > > > > > automatically. > > > > > > > > > > > Since the Vert.x event bus runs on event-loop threads [...] could > > > > > blocking or slow event listeners potentially stall REST requests and > > > > impact > > > > > latency? > > > > > > > > > > What Pierre said: this could indeed happen, but it's possible to > > > > > annotate the receiving end with @Blocking, in which case, the > > listener > > > > > will be invoked in a separate pool. > > > > > > > > > > > With asynchronous event listeners, is there a guarantee of delivery > > > to > > > > > all listeners for a given event? > > > > > > > > > > If I understand the question correctly: with asynchronous delivery, a > > > > > slow or failing listener wouldn't impact the delivery of the same > > > > > event to other listeners. > > > > > > > > > > Thanks, > > > > > Alex > > > > > > > > > > On Mon, Nov 10, 2025 at 10:12 AM Pierre Laporte < > > [email protected] > > > > > > > > > wrote: > > > > > > > > > > > > Thanks for the proposal, Alex. This sounds like a great > > improvement. > > > > > > > > > > > > @Yufei As per Quarkus documentation, slow event listeners should be > > > > > marked > > > > > > with @Blocking so that they are not run on the event loop threads. > > > > > > -- > > > > > > > > > > > > Pierre > > > > > > > > > > > > > > > > > > On Sat, Nov 8, 2025 at 2:14 AM Michael Collado < > > > [email protected] > > > > > > > > > > > wrote: > > > > > > > > > > > > > With asynchronous event listeners, is there a guarantee of > > delivery > > > > to > > > > > all > > > > > > > listeners for a given event? The downside of synchronous > > listeners > > > is > > > > > that > > > > > > > everything is serial, but also if something fails, processing > > > stops. > > > > > This > > > > > > > feels important for auditing purposes, though less important for > > > > other > > > > > > > cases. > > > > > > > > > > > > > > Mike > > > > > > > > > > > > > > On Fri, Nov 7, 2025 at 2:28 PM Yufei Gu <[email protected]> > > > > wrote: > > > > > > > > > > > > > > > Thanks, Alex and Adam. One concern I have is about the shared > > > > runtime > > > > > > > > thread pool. > > > > > > > > Since the Vert.x event bus runs on event-loop threads that are > > > also > > > > > used > > > > > > > by > > > > > > > > Quarkus’ reactive REST endpoints, could blocking or slow event > > > > > listeners > > > > > > > > potentially stall REST requests and impact latency? > > > > > > > > > > > > > > > > Yufei > > > > > > > > > > > > > > > > > > > > > > > > On Fri, Nov 7, 2025 at 11:25 AM Adam Christian < > > > > > > > > [email protected]> wrote: > > > > > > > > > > > > > > > > > I think that this would be a great enhancement. Thanks for > > > > > proposing > > > > > > > it! > > > > > > > > > > > > > > > > > > The only concern I would have is around fault-tolerance. From > > > > what > > > > > I > > > > > > > can > > > > > > > > > tell, from the Quarkus documentation, the Quarkus event bus > > > uses > > > > > Vert.x > > > > > > > > > EventBus which does not guarantee message delivery if failure > > > of > > > > > part > > > > > > > of > > > > > > > > > the event bus occurs [1]. However, we can easily make sure > > that > > > > we > > > > > use > > > > > > > > > Quarkus's SmallRye Fault Tolerance [2]. Is my rough > > > understanding > > > > > > > inline > > > > > > > > > with your proposal? > > > > > > > > > > > > > > > > > > Go community, > > > > > > > > > > > > > > > > > > Adam > > > > > > > > > > > > > > > > > > [1]: > > > > > > > > > https://vertx.io/docs/apidocs/io/vertx/core/eventbus/EventBus.html > > > > > > > > > [2]: https://quarkus.io/guides/smallrye-fault-tolerance > > > > > > > > > > > > > > > > > > On Fri, Nov 7, 2025 at 11:49 AM Alexandre Dutra < > > > > [email protected] > > > > > > > > > > > > > > wrote: > > > > > > > > > > > > > > > > > > > Hi all, > > > > > > > > > > > > > > > > > > > > I'd like to propose an enhancement to our existing events > > > > > feature: > > > > > > > the > > > > > > > > > > ability to support multiple listeners. > > > > > > > > > > > > > > > > > > > > Currently, only a single listener can be active at a time, > > > > which > > > > > is > > > > > > > > > > quite limiting. For example, we might need to persist > > events > > > > for > > > > > > > audit > > > > > > > > > > purposes and simultaneously send them to a message queue > > for > > > > > > > > > > optimization. With the current setup, this isn't easily > > > > > achievable. > > > > > > > > > > > > > > > > > > > > While a composite listener could be created, it feels like > > a > > > > less > > > > > > > > > > elegant solution, and the delivery would be strictly > > serial, > > > > > > > > > > processing one listener after another. > > > > > > > > > > > > > > > > > > > > My suggestion is to leverage Quarkus internal event bus > > [1]: > > > > > > > > > > > > > > > > > > > > 1) There will be one central event emitter responsible for > > > > > publishing > > > > > > > > > > events to the bus. > > > > > > > > > > > > > > > > > > > > 2) We will have zero to N listeners, each independently > > > > watching > > > > > the > > > > > > > > > > event bus for relevant events. They will be discovered by > > > CDI. > > > > > > > > > > > > > > > > > > > > 3) We could apply filters to each listener, e.g. listener A > > > > > listens > > > > > > > > > > for event types X and Y, listener B only listens to event > > > type > > > > Y. > > > > > > > > > > > > > > > > > > > > 4) This approach would ensure fully asynchronous delivery > > of > > > > > events > > > > > > > to > > > > > > > > > > all interested listeners. > > > > > > > > > > > > > > > > > > > > 5) Fault-tolerance could also be easily implemented (event > > > > > delivery > > > > > > > > > > retries, timeouts, etc.). > > > > > > > > > > > > > > > > > > > > What do you all think? > > > > > > > > > > > > > > > > > > > > Thanks, > > > > > > > > > > Alex > > > > > > > > > > > > > > > > > > > > [1]: https://quarkus.io/guides/reactive-event-bus > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > >
