Status

Current state: Under Discussion

Discussion thread: here

JIRA: KAFKA-2511

Please keep the discussion on the mailing list rather than commenting on the wiki (wiki discussions get unwieldy fast).

Motivation

Today the broker needs to decompress compressed messages, assign offsets to each message and recompress the messages again. This causes additional CPU cost. This KIP is trying to avoid server side recompression.

This KIP is a distilled/improved version of an earlier discussion that we started.

Public Interfaces

We propose the following change to the message format

MessageAndOffset => MessageSize Offset Message
  MessageSize => int32
  Offset => int64 <------ CHANGE: Base offset for wrapper message of compressed message, relative offset for inner compressed message.
  
  Message => Crc MagicByte Attributes Timestamp KeyLength Key ValueLength Value
    Crc => int32
    MagicByte => int8
    Attributes => int8
    KeyLength => int32
    Key => bytes
    ValueLength => int32
    Value => bytes

Proposed Changes

Wire protocol change

Change the usage of offset field

• When the producer compresses a message, write the relative offset value in the raw message's offset field. Leave the wrapped message's offset blank.
• When broker receives a compressed message, it only needs to
  
  1. Decompress the message to verify the CRC and relative offset.
     NOTE: If the relative offset is not contiguous (e.g., if this is a mirrored compacted topic), the broker will reassign the relative offsets. There are two ways to handle this - (i) reject the ProducerRequest or (ii) just assign the relative offsets. We chose to reassign offsets rather than reject the request because there is a useful use case where mirror maker can do a direct copy from source cluster to destination cluster without even decompressing the message. In this case, the compressed message can have noncontinuous relative offsets (for compacted topics).
  2. Set outer message's base offset. The outer message's base offset will be the offset of the last inner message.  (Since the broker only needs to update the message-set header, there is no need to re-compress message sets.)
• When the log cleaner compacts log segments, if multiple message sets are integrated into one message set, the broker needs to update the inner message's relative offset values. (This will leave "holes" inside the new wrapped message).
• When the consumer receives a message, it converts the relative offset back to actual offset.

Compatibility, Deprecation, and Migration Plan

NOTE: This part is drafted based on the assumption that KIP-31 and KIP-32 will be implemented in one patch.

The proposed protocol is not backward compatible. The migration plan are as below:

1. Bump up ProducerRequest and FetchRequest version to V2, which uses MessageAndOffset V1.
2. Create ApiVersion 0.9.0-V1 which uses ProducerRequest V2 and FetchRequest V2.
3. Configure the broker to use ApiVersion 0.9.0 (ProduceRequest V1 and FetchRequest V1).
4. Do a rolling upgrade of the brokers to let the broker pick up the new code supporting ApiVersion 0.9.0-V1.
5. Bump up ApiVersion of broker to 0.9.0-V1
6. Do a rolling bounce of the brokers to let the broker use FetchRequest V2 for replication.
7. Upgraded brokers support both ProducerRequest V2 and FetchRequest V2 which uses magic byte 1 for MessageAndOffset.
   
   1. When broker sees a producer request V1 (MessageAndOffset = V0), it will decompress the message, assign offsets using relative offsets and re-compress the message. i.e. upconvert the message format to MessageAndOffset V1.
      
   2. When broker sees a producer request V2 (MessageAndOffset = V1), it will decompress the message for verification, assign the offset to outer message and NOT do recompression. 
   3. When broker sees a fetch request V1 (MessageAndOffset = V0), because the data format on disk is MessageAndOffset V1, it will not use the zero-copy transfer, but read the message to memory, do down-conversion, then send fetch response V1.
   4. When broker sees a fetch request V2 (MessageAndOffset = V1), it will use zero-copy transfer to reply with fetch response V2.
      
8. Upgrade consumer to send FetchRequest V2.
9. Upgrade producer to send ProducerRequest V2.

For producer, there will be no impact.

For consumers using MessageAndOffset V0, there will be some performance penalty because there is no zero-copy transfer.

During step 7 and step 8, the majority of the consumers may be still using consumers using MessageAndOffset V0, broker could consume more memory.

Rejected Alternatives

None