I’m experimenting with Spring Kafka and I’m confused by the different ack modes.

By default, I believe it is set to BATCH which goes like:

• Call poll()
• poll() returns messages 1, 2, 3
• Process the messages sequentially (or concurrently)
• Call poll() again after all messages are processed
• Messages 1, 2, 3 are automatically committed, and messages 4, 5, 6 are returned
• Process repeats

One drawback I see to this though is that if you process messages 1 and 2, but 3 fails, you will have retry the entire batch.

I believe you can get around this by using AckMode RECORD or MANUAL_IMMEDIATE which “commits the offset after each record is processed by the listener” or by calling ack() on the acknowledgement object

So now it goes like:

• Call poll()
• poll() returns messages 1, 2, 3
• Process message 1, commit it
• Process message 2, commit it
• Process message 3, commit it
• Call poll() again after all messages are processed
• Messages 4, 5, 6 are returned
• Process repeats

What I’m confused about though is, what happens if I process the messages concurrently and something like this happens:

• Message 1 errors out
• Message 2 completes processing
• Message 3 completes processing

What is the offset set to in this case? 1? Will it return 2 and 3 again in that case? Or skip them and return 4 and 5?

I come from an SQS background where acking messages out of order is quite common and acceptable.