Hi Guozhang,
Sorry, I should have been more clear. By "an unreadable end of the log", I
mean the `endOffsets` method returns an offset for a record that is never
surfaced to the caller of `poll`.
I've done some more digging and I think I understand why that is now. The
API `endOffsets` calls tell
Hi Chris,
Yes, the transaction marker should be hidden to clients.
There is similar issues reported:
https://issues.apache.org/jira/browse/KAFKA-10683
Welcome to submit KIP/PR to improve it.
Thank you.
Luke
On Tue, Mar 22, 2022 at 5:16 PM Chris Jansen
wrote:
> Hi Guozhang,
>
> Sorry, I should
Thanks Luke,
If the transaction marker should be hidden, does it follow that aborted
transaction at the end of the log should also be hidden for clients that
are in read committed mode?
Happy to do a KIP/PR.
Thanks again,
Chris
On Tue, Mar 22, 2022 at 10:21 AM Luke Chen wrote:
> Hi Chris,
>
Hi Chris,
Since you are using read_committed mode, the txn marker from the
`endOffsets()` should indeed be skipped, no matter for committed of aborted
txns. For example if the log looks like this:
offsets: 0, 1, 2, 3, t_c(4) // or t_a(4) which means "abort marker"
then the endOffset should ret
