Hello, i highly recommend Apache Kafka, to me it's the best. You can deploy it in cluster mode inside K8S, thus you can have a Highly available system (also auto scalable).

Good luck

ZeroMQ is fast but you need to build build reliability yourself. There are a number of patterns described in the zeromq guide. I have used RabbitMQ before which gives lot of functionality out of the box, you can probably use the worker queues example from the tutorial, it can also persists messages in the queue.

I haven't used Amazon SQS before. Another tool you could use is Kafka.

Both would do the trick, but there are some nuances. We work with both.

From the sound of it, your main focus is "not losing messages". In that case, I would go with RabbitMQ with a high availability policy (ha-mode=all) and a main/retry/error queue pattern.

Push messages to an exchange, which sends them to the main queue. If an error occurs, push the errored out message to the retry exchange, which forwards it to the retry queue. Give the retry queue a x-message-ttl and set the main exchange as a dead-letter-exchange. If your message has been retried several times, push it to the error exchange, where the message can remain until someone has time to look at it.

This is a very useful and resilient pattern that allows you to never lose messages. With the high availability policy, you make sure that if one of your rabbitmq nodes dies, another can take over and messages are already mirrored to it.

This is not really possible with SQS, because SQS is a lot more focused on throughput and scaling. Combined with SNS it can do interesting things like deduplication of messages and such. That said, one thing core to its design is that messages have a maximum retention time. The idea is that a message that has stayed in an SQS queue for a while serves no more purpose after a while, so it gets removed - so as to not block up any listener resources for a long time. You can also set up a DLQ here, but these similarly do not hold onto messages forever. Since you seem to depend on messages surviving at all cost, I would suggest that the scaling/throughput benefit of SQS does not outweigh the difference in approach to messages there.

Mithiridi, I believe you are talking about two different things. 1. If you need to process messages with delays of more 15m or at specific times, it's not a good idea to use queues, independently of tool SQM, Rabbit or Amazon MQ. you should considerer another approach using a scheduled job. 2. For dead queues and policy retries RabbitMQ, for example, doesn't support your use case. https://medium.com/@kiennguyen88/rabbitmq-delay-retry-schedule-with-dead-letter-exchange-31fb25a440fc I'm not sure if that is possible SNS/SQS support, they have a maximum delay for delivery (maxDelayTarget) in seconds but it's not clear the number. You can check this out: https://docs.aws.amazon.com/sns/latest/dg/sns-message-delivery-retries.html

For a non-streaming approach:

You could consider using more checkpoints throughout your spark jobs. Furthermore, you could consider separating your workload into multiple jobs with an intermittent data store (suggesting cassandra or you may choose based on your choice and availability) to store results , perform aggregations and store results of those.

Spark Job 1 - Fetch Data From 10 URLs and store data and metadata in a data store (cassandra) Spark Job 2..n - Check data store for unprocessed items and continue the aggregation

Alternatively for a streaming approach: Treating your data as stream might be useful also. Spark Streaming allows you to utilize a checkpoint interval - https://spark.apache.org/docs/latest/streaming-programming-guide.html#checkpointing