Tail input is paused even with filesystem buffer

[LukoJy3D]

Bug Report

Describe the bug
Tail input is paused when input bytes exceed or gets close to ~1MB, no matter what value it is set to (I tried 5MB and 10MB). Then I thought it had to do with the chunks limit in memory, yet it's far away from reaching limits (check ss).
From metrics, it is visible that input storage goes over the limit, but it seems like that limit should be Mem_Buf_Limit according to docs, yet it can not be configured, or I am missing something. Also, there is no backpressure from outputs.

Configuring the filesystem is not a problem, but I would like to understand why Mem_Buf_Limit is not taken into the account 🤔

    [INPUT]
        Name             tail
        Tag              kube.*
        Path             /var/log/containers/*.log
        Exclude_Path     /var/log/containers/*kubeshark*.log
        DB               /var/log/flb_kube_containers.db
        Parser           CRI
        Mem_Buf_Limit    10MB
        Buffer_Max_Size  5MB
        Rotate_Wait      30
        Refresh_Interval 10
        Skip_Long_Lines  On

To Reproduce

• Have input burst over 1MB/s
• Input is paused with the following message

[ warn] [input] tail.0 paused (mem buf overlimit)

Expected behavior
Input is not paused unless incoming bytes exceed Mem_Buf_Limit

Screenshots

Input bytes

Ingestion pauses

Storage over the limit

chunks in memory

Your Environment

• Version used: 3.2.2
• Environment name and version (e.g. Kubernetes? What version?): Kubernetes, helm chart 0.48.3

[LukoJy3D]

Switched tail input storage type to filesystem, yet I still see input pauses. Based on the docs mem_buf_Limit applies only when storage.type is set to the default value of memory., thus my original issue no longer applies to mem_buf_Limit, but rather something else, which I can not explain yet.

as you can see pipeline is paused, but input overflow does not exist , chunk count remain stable, none of the chunks are marked as "down"

[LukoJy3D]

I checked Loki with the bytes_over_time function, which showed way higher log volume than our metrics capture (scrape every 15 seconds) thus increasing Emitter_Mem_Buf_Limit on all filters solved the issue. It seems like the filesystem buffer works as expected