Bug Report

Describe the bug
I have tail input with multiline filter. Suppose we have two subsequent records R1 and R2 that must be concatenated into one by multiline parser. If Buffer_Chunk_Size is such that R1 and R2 are in different chunks, multiline parser breaks log into two records instead of one.

To Reproduce

• Download and extract gist: https://gist.github.com/starteleport/aeadc1647ac1530fe8de0c0c562cfe29
• Run fluent-bit -c fluent-bit-repro.conf
• Notice the following lines in stdout:

[3] kube.repro.log: [1636558784.789656241, {"log"=>"info: Microsoft.AspNetCore.Hosting.Diagnostics[1]
", "stream"=>"stdout", "time"=>"2021-11-10T15:39:44.789656241Z"}]
[4] kube.repro.log: [1636558784.789656241, {"log"=>"      => SpanId:5018968b4ab3f342, TraceId:94b3b5b666fba84ca420f0a336559e7b, ParentId:0000000000000000 => ConnectionId:0HMD3I6TBKKDD => RequestPath:/ RequestId:0HMD3I6TBKKDD:00000002
      Request starting HTTP/1.1 GET http://10.0.2.125:80/ - -
"}]

Expected behavior
Lines [3] and [4] are concatenated and output as single log message.

Screenshots
Not applicable

Your Environment

• Version used: reproduced on 1.8.3+
• Configuration: please see gist above
• Environment name and version (e.g. Kubernetes? What version?): local
• Server type and version:
• Operating System and version: macOS Monterey
• Filters and plugins: multiline, multiline_parser

Additional context
In production this leads to multiline log records being randomly split when they shouldn't. As I understand, I could just set large Buffer_Chunk_Size but this is clearly a duct tape fix for me as it does not eliminate problem completely, just reducing the frequency.