[ https://issues.apache.org/jira/browse/KAFKA-7045?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]
Dhruvil Shah updated KAFKA-7045: -------------------------------- Description: When down-conversion is required, the consumer might fail consuming messages under certain conditions. Couple such cases are outlined below: (1) When consuming from a compacted topic, it is possible that the consumer wants to fetch messages that fall in the middle of a batch but the messages have been compacted by the cleaner. For example, let's say we have the following two segments. The brackets indicate a single batch of messages and the numbers within are the message offsets. Segment #1: [0, 1, 2], [3, 4, 5], [6, 7, 8] Segment #2: [9, 10, 11], [12, 13, 14] If the cleaner were to come in now and clean up messages with offsets 7 and 8, the segments would look like the following: Segment #1: [0, 1, 2], [3, 4, 5], [6] Segment #2: [9, 10, 11], [12, 13, 14] A consumer attempting to fetch messages at offset 7 will start reading the batch starting at offset 6. During down-conversion, we will drop the record starting at 6 it is less than the current fetch start offset. However, there are no messages in the log following offset 6. In such cases, we return the `FileRecords` itself which would cause the consumer to throw an exception because it does not understand the stored message format. (2) When consuming from a topic with transactional messages, down-conversion usually drops control batches because these did not exist in V0 and V1 message formats. If there are no message batches following the control batch in the particular segment (or if we are at the end of the log), we would again get no records after down-conversion and will return the `FileRecords`. Because the consumer is not able to interpret control batches, it will again throw an exception. was: When down-conversion is required, the consumer might fail consuming messages under certain conditions. Couple such cases are outlined below: # When consuming from a compacted topic, it is possible that the consumer wants to fetch messages that fall in the middle of a batch but the messages have been compacted by the cleaner. For example, let's say we have the following two segments. The brackets indicate a single batch of messages and the numbers within are the message offsets. Segment #1: [0, 1, 2], [3, 4, 5], [6, 7, 8] Segment #2: [9, 10, 11], [12, 13, 14] If the cleaner were to come in now and clean up messages with offsets 7 and 8, the segments would look like the following: Segment #1: [0, 1, 2], [3, 4, 5], [6] Segment #2: [9, 10, 11], [12, 13, 14] A consumer attempting to fetch messages at offset 7 will start reading the batch starting at offset 6. During down-conversion, we will drop the record starting at 6 it is less than the current fetch start offset. However, there are no messages in the log following offset 6. In such cases, we return the `FileRecords` itself which would cause the consumer to throw an exception because it does not understand the stored message format. (1) When consuming from a topic with transactional messages, down-conversion usually drops control batches because these did not exist in V0 and V1 message formats. If there are no message batches following the control batch in the particular segment (or if we are at the end of the log), we would again get no records after down-conversion and will return the `FileRecords`. Because the consumer is not able to interpret control batches, it will again throw an exception. (2) When consuming from a topic with transactional messages, down-conversion usually drops control batches because these did not exist in V0 and V1 message formats. If there are no message batches following the control batch in the particular segment (or if we are at the end of the log), we would again get no records after down-conversion and will return the `FileRecords`. Because the consumer is not able to interpret control batches, it will again throw an exception. > Consumer may not be able to consume all messages when down-conversion is > required > --------------------------------------------------------------------------------- > > Key: KAFKA-7045 > URL: https://issues.apache.org/jira/browse/KAFKA-7045 > Project: Kafka > Issue Type: Bug > Components: consumer, core > Affects Versions: 0.11.0.0, 0.11.0.1, 1.0.0, 0.11.0.2, 1.1.0, 2.0.0, 1.0.1 > Reporter: Dhruvil Shah > Priority: Major > Fix For: 2.1.0 > > > When down-conversion is required, the consumer might fail consuming messages > under certain conditions. Couple such cases are outlined below: > > (1) When consuming from a compacted topic, it is possible that the consumer > wants to fetch messages that fall in the middle of a batch but the messages > have been compacted by the cleaner. For example, let's say we have the > following two segments. The brackets indicate a single batch of messages and > the numbers within are the message offsets. > Segment #1: [0, 1, 2], [3, 4, 5], [6, 7, 8] > Segment #2: [9, 10, 11], [12, 13, 14] > If the cleaner were to come in now and clean up messages with offsets 7 and > 8, the segments would look like the following: > Segment #1: [0, 1, 2], [3, 4, 5], [6] > Segment #2: [9, 10, 11], [12, 13, 14] > A consumer attempting to fetch messages at offset 7 will start reading the > batch starting at offset 6. During down-conversion, we will drop the record > starting at 6 it is less than the current fetch start offset. However, there > are no messages in the log following offset 6. In such cases, we return the > `FileRecords` itself which would cause the consumer to throw an exception > because it does not understand the stored message format. > > (2) When consuming from a topic with transactional messages, down-conversion > usually drops control batches because these did not exist in V0 and V1 > message formats. If there are no message batches following the control batch > in the particular segment (or if we are at the end of the log), we would > again get no records after down-conversion and will return the `FileRecords`. > Because the consumer is not able to interpret control batches, it will again > throw an exception. -- This message was sent by Atlassian JIRA (v7.6.3#76005)