exactly once and storing offsets in db (transactionally with computation results)

["Marasoiu, Nicu" <ni...@metrosystems.net>]

Hi,
We would consider one of 2 or 3 flows to ensure an "exactly once" process from an input kafka topic to a database storing results (using kafka consumer, but also evaluated kafka streams and details at the end) and wanted to gather your input on them:
(for simplicity let's assume that any exception exits the process except if the exception comes out of step 5)
The outlined flows are executed in a loop.

First flow/solution:
1. read from kafka
2. start transaction in db
3. update target tables
4. commit transaction
5. commit offset to Kafka
6. if commit offset failed, attempt another transaction to revert the previous one in db. (compensate)

Solution 2 - offsets persisted in db in the same transaction, consumer reads from explicit offsets at init
If it is possible for the consumer to configure its offsets before starting to consume, then this flow would be possible:
0. at consumer process boot, read the latest offsets for partitions from db and configs consumer to start from those.
1. read from kafka (first read, from explicit offsets, the next polls just continue)
2. start transaction in db
3. update target tables
3'. update an "offsets" table, for consumer group and partition id
4. commit transaction (which includes offsets)

Solution 3 - If it would be possible to commit an explicit value of the offset to kafka for a (partition, consumer group), not just the current offset, but a previously saved one (at step 0), than another flow would be possible, with 4 and 5 reversed:
4. commit offset to Kafka
5. commit transaction
6. if commit transaction failed, attempt to commit the old offset back to kafka. (compensate). Exit or rewind the consumer.

Solution 4 - use Kafka Streams configured with exactly once. This seems to imply that the aggregates (the results of the processing), currently stored in the db, would also need to be duplicated in kafka as output topics & local Rocksdb instances. Since the data volume even on the aggregates is significant, we are exploring solutions close to exactly once which would not imply the cost of doubly storing the result "tables".

Do you see any other possibility? What do you suggest for improving the options above, or what is your advice?

Indeed, solution 2 seems feasible using db transaction (e.g. Cassandra batch) to include an offset update atomically.
A sophisticated implementation is for instance under the hood of Akka Streams Kafka Source: https://doc.akka.io/docs/akka-stream-kafka/current/consumer.htmlhttps://doc.akka.io/docs/akka-stream-kafka/current/consumer.html

But a "manual" implementation just on top of consumer seems feasible:
- on new partition during subscribe or rebalance, get the latest offset for partition from db and do consumer.seek on the partition to that offset
(using only onPartitionsAssigned rebalancing callback passed to subscribe - does it include the initial partitions allocation to a consumer, at subscribe time?)
- repeat
  - poll a batch of messages from kafka
  - compute the results
  - update the db with results and offset in same transaction

I would have a few questions:
- does the plan sound ok to you?
- is there a limit on the number of records Kafka consumer pulls from the brokers in advance?
- what about the number of ConsumerRecords returned by consumer.poll ?
- is there a risk that messages coming from the same partition reach multiple consumers doing poll if a rebalancing moves a partition?
- is it indeed sufficient to use onPartitionsAssigned and not onPartitionsRevoked (given we update offset in transaction with a batch of results only)
- does onPartitionsAssigned cover the startup/subscribe phase - the initial partitions with which the consumer starts?
- if we would have a batch of messages from a single partition, we could have smaller transactions - a way I think about is doing consumer.pause in onPartitionsAssigned, and in the main loop, iterate through partitions and resume one partition, do poll, process, then next partition, in rotation? but this would need a blocking in poll right? to have time from resuming that partition until data can be downloaded by the consumer; is there a memory-efficient way to get partition-by-partition results, and not in a large memory structure with multiple partitions there?

Please advise,
Thank you,
Nicu

Geschäftsanschrift/Business address: METRO SYSTEMS GmbH, Metro-Straße 12, 40235 Düsseldorf, Germany
Aufsichtsrat/Supervisory Board: Heiko Hutmacher (Vorsitzender/ Chairman)
Geschäftsführung/Management Board: Dr. Dirk Toepfer (Vorsitzender/CEO), Wim van Herwijnen
Sitz Düsseldorf, Amtsgericht Düsseldorf, HRB 18232/Registered Office Düsseldorf, Commercial Register of the Düsseldorf Local Court, HRB 18232

Betreffend Mails von *@metrosystems.net
Die in dieser E-Mail enthaltenen Nachrichten und Anhänge sind ausschließlich für den bezeichneten Adressaten bestimmt. Sie können rechtlich geschützte, vertrauliche Informationen enthalten. Falls Sie nicht der bezeichnete Empfänger oder zum Empfang dieser E-Mail nicht berechtigt sind, ist die Verwendung, Vervielfältigung oder Weitergabe der Nachrichten und Anhänge untersagt. Falls Sie diese E-Mail irrtümlich erhalten haben, informieren Sie bitte unverzüglich den Absender und vernichten Sie die E-Mail.

Regarding mails from *@metrosystems.net
This e-mail message and any attachment are intended exclusively for the named addressee. They may contain confidential information which may also be protected by professional secrecy. Unless you are the named addressee (or authorised to receive for the addressee) you may not copy or use this message or any attachment or disclose the contents to anyone else. If this e-mail was sent to you by mistake please notify the sender immediately and delete this e-mail.