[KAFKA-4293] ByteBufferMessageSet.deepIterator burns CPU catching EOFExceptions

• Type: Bug
• Status: Open
• Priority: Major
• Resolution: Unresolved
• Affects Version/s: 0.10.0.1
• Component/s: core
• Assignee: Radai Rosenblatt
• Reporter: Radai Rosenblatt
• Created: 11/Oct/16 22:08
• Updated: 06/May/17 23:18

Description

around line 110:

try {
    while (true)
        innerMessageAndOffsets.add(readMessageFromStream(compressed))
} catch {
    case eofe: EOFException =>
    // we don't do anything at all here, because the finally
    // will close the compressed input stream, and we simply
    // want to return the innerMessageAndOffsets

the only indication the code has that the end of the oteration was reached is by catching EOFException (which will be thrown inside readMessageFromStream()).

profiling runs performed at linkedIn show 10% of the total broker CPU time taken up by Throwable.fillInStack() because of this behaviour.

unfortunately InputStream.available() cannot be relied upon (concrete example - GZipInputStream will not correctly return 0) so the fix would probably be a wire format change to also encode the number of messages.

Issue Links

• is related to: KAFKA-5150 LZ4 decompression is 4-5x slower than Snappy on small batches / messages
• links to: GitHub Pull Request #2025