markRead: negative count still occurs when reading compacted topic from start

[dovmashleh]

Describe the bug

panic: markRead: negative count

goroutine 68 [running]:
github.com/segmentio/kafka-go.(*messageSetReader).markRead(0xc0004587a0)
    external/com_github_segmentio_kafka_go/message_reader.go:345 +0x11a
github.com/segmentio/kafka-go.(*messageSetReader).readMessageV2(0xc0004587a0, 0x2c709, 0xc000507ac8, 0xc000507ab8, 0x2c708, 0x2c708, 0xc000059800, 0xc0005078f8, 0x416b3b, 0xc0005078f8, ...)
    external/com_github_segmentio_kafka_go/message_reader.go:329 +0x49d
github.com/segmentio/kafka-go.(*messageSetReader).readMessage(0xc0004587a0, 0x2c709, 0xc000507ac8, 0xc000507ab8, 0x2c708, 0xc000507a5c, 0x17fd1b29700, 0x0, 0xc0004922a0, 0x0, ...)
    external/com_github_segmentio_kafka_go/message_reader.go:136 +0xc5
github.com/segmentio/kafka-go.(*Batch).readMessage(0xc000195880, 0xc000507ac8, 0xc000507ab8, 0x0, 0x17fd1b29700, 0x100000001, 0xc0004922a0, 0xc0004922a0, 0xc000507ab8, 0x2)
    external/com_github_segmentio_kafka_go/batch.go:240 +0x79
github.com/segmentio/kafka-go.(*Batch).ReadMessage(0xc000195880, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, ...)
    external/com_github_segmentio_kafka_go/batch.go:192 +0x11a
github.com/segmentio/kafka-go.(*reader).read(0xc000507ed8, 0xe30160, 0xc0004882c0, 0x2c709, 0xc0000e21e0, 0x0, 0x0, 0x0)
    external/com_github_segmentio_kafka_go/reader.go:1492 +0x3ec
github.com/segmentio/kafka-go.(*reader).run(0xc000507ed8, 0xe30160, 0xc0004882c0, 0x0)
    external/com_github_segmentio_kafka_go/reader.go:1310 +0x2d9
github.com/segmentio/kafka-go.(*Reader).start.func1(0xc0004d8000, 0xe30160, 0xc0004882c0, 0xc00004402c, 0x10, 0x0, 0xfffffffffffffffe, 0xc0004d8138)
    external/com_github_segmentio_kafka_go/reader.go:1211 +0x1d8
created by github.com/segmentio/kafka-go.(*Reader).start
    external/com_github_segmentio_kafka_go/reader.go:1191 +0x1a5

Kafka Version

3.0

To Reproduce

Start reding compact topic from start when several compactifications already occured

Expected Behavior

No panic

Additional Context

there is a solution #1292 - forks with implementing this work well

[Abdulsametileri]

could you help with that @collinvandyck

[collinvandyck]

@Abdulsametileri if you're asking for help merging #1292 sadly I can't be of assistance as I don't work at Segment/Twilio. I will say that the PR should have tests for the edge case it addresses.

[Dmitriy-Kulagin]

Probably the #1411 will fix this issue also.

Actually I've received the same panic but the reason of that is endless reading the same compacted batch as incorrect handling of lengthRemain into readMessageV2 brakes skipping to the next batch by batch.lastOffset into batch.readMessage:283