Message exceeds the maximum size when using Compression about sarama

[eapache]

Which revision of Sarama are you using? Specifically, does the revision you run include 396afc5 ? If it does, you may want to set producerConfig.BackPressureThresholdBytes to match the broker's configured message.max.bytes (which it looks like you've decreased significantly from the default - is there a reason for that?).

[wkuranowski]

I am using latest Sarama.
message.max.bytes is set to 1000000 by default, it's not decreased.

[wkuranowski]

Setting producerConfig.BackPressureThresholdBytes to 1000000 (default value for message.max.bytes) solves the problem. I can even increase this value to about 15000000 and it works (strange, needs more testing).

I wonder about two things:

1. What is the difference between producerConfig.MaxBufferedBytes and producerConfig.BackPressureThresholdBytes? Especially when using asynchronous producer (QueueMessage).
2. How does compression works? Does it compress single message or compress all messages in a buffer at once? It's strange that I see this problem only with Snappy or GZip enabled.

[eapache]

message.max.bytes is set to 1000000 by default, it's not decreased.

Ah, sorry, I was confusing socket.request.max.bytes with message.max.bytes.

1. MaxBufferedBytes is how many bytes have to accumulate in order for the producer to send the current batch to the broker. BackPressureThresholdBytes is how many bytes have to accumulate in order for the producer to stop accepting new messages entirely (until it can clear its queue). Both are only used for async (QueueMessage).
2. It compresses all messages at once, then has to send that as a single "message" because Kafka's protocol is kind of weird. @wvanbergen should we reduce the default back-pressure threshold to 1MB to fix this case?

[eapache]

CC also @graemej would this affect us at all?

[wkuranowski]

@eapache thanks for clarification. If i understand correctly, MaxBufferedBytes should be a maximum size of a batch to send and BackPressureThresholdBytes is a buffer for all messages to send?

If yes, then I am not sure if reducing BackPressureThresholdBytes is a good solution. I think that size of a back-pressure threshold should be much bigger than size of a single batch.

[eapache]

MaxBufferedBytes is just a trigger saying "when I have this many bytes, send another message". BackPressureThresholdBytes is, effectively, the maximum size of the buffer.

Perhaps adding a MaxMessageBytes which defaults to 1MB to match kafka would be a more appropriate, though also more complex fix.

[wkuranowski]

MaxMessageBytes sounds good. I think it is the only way to not exceed limits defined on Kafka broker.

[wkuranowski]

Any progress on that issue?

[eapache]

Unfortunately not, we've all been quite busy with other things and this isn't a terribly high priority for us since setting BackPressureThresholdBytes is a functional work-around. We're happy to accept pull requests if you want to do it yourself.

[eapache]

Just merged the new producer which should fix this issue among many others.