Http sink retries #181
Http sink retries #181
Conversation
… compatible constants, deprecate old constants, consolidate HTTP Code evaluation logic
| "failed requests: {}, throwing BatchHttpStatusCodeValidationFailedException from sink", | ||
| failedRequests | ||
| ); | ||
| getFatalExceptionCons().accept(new BatchHttpStatusCodeValidationFailedException( |
There was a problem hiding this comment.
For context, the README reads like an error should trigger a job failure, which is why I threw a fatal exception here
There was a problem hiding this comment.
I would suggest that you raise an issue for each of the 3 things you would like to put in. It will be far easier to review smaller pieces.
- the retries for sink should be very self contained piece of code.
- for the ratelimitter -I see we have apache/flink#27134 which is not yet merged. Is there a generic way we can do Sink in the Flink code base ?
- fyi - we are porting over changes to https://github.com/apache/flink-connector-http and are hoping to release this in the near future
| You can configure HTTP status code handling for HTTP sink table and enable automatic retries with delivery guarantees. | ||
|
|
||
| #### Retries and delivery guarantee | ||
| HTTP Sink supports automatic retries when `sink.delivery-guarantee` is set to `at-least-once`. Failed requests will be automatically retried based on the configured status codes. |
There was a problem hiding this comment.
sink.delivery-guarantee is interesting and looks like it is a bringing this connector in line with others. When do we think we will get more that one request issued? I assume we would need more processing for exactly once.
There was a problem hiding this comment.
I am lifting a big part of the code from an earlier PR, which was the source of using a delivery guarantee. However, for my personal use case, I do in fact need to guarantee at least once behavior 🙂. The current implementation has an indefinite number of retries. If we don't have the delivery guarantee, we'll need to have something like number of attempts imo. I'm fine with whatever though! I just really need retries!
There was a problem hiding this comment.
Exactly once is pretty challenging given that we are using HTTP requests. Imo, it would require coordination with the receiving service to response with some kind of ack in order to enforce it. I don't believe exactly once can be achieved on just the client side with HTTP requests if that makes sense. However, duplicated success message emissions should be really infrequent
That's great to hear this could make it in the Apache foundation! You probably meant the config changes, the rate limiting, and the retry behavior 🤦 . I can do that |
2 participants
Description
Attempting to re-vitalize the http sink updates. Branch ended up getting bigger than I hoped, but please provide any feedback you have!
Resolves
<issue nr here>PR Checklist