You are viewing a plain text version of this content. The canonical link for it is here.
Posted to dev@pulsar.apache.org by Michael Marshall <mm...@apache.org> on 2022/08/11 05:36:15 UTC
[DISCUSS] Message redeliveryCount semantics and DLQ
Hi Pulsar Community,
In Pulsar 2.5.0, the semantics for a message's redeliveryCount changed
at the request of this issue [0]. Please see the issue for relevant
context.
In summary, the issue is whether a message that is neither positively
nor negatively acknowledged should get counted as "redelivered" when a
consumer disconnects from the broker.
The issue asserts that because a message could lead to a fatal error
in the consumer application that could prevent negative
acknowledgement getting sent to the broker, every delivery of a
message should count towards the redeliveryCount.
In my view, this edge case should not justify changing the
redeliveryCount semantics. It is the responsibility of the application
to handle all messages, even those that are malformed.
My primary motivation for revisiting this design is to discuss one of
its consequences. After the change, messages that are only ever sent
to an application's receiverQueue are counted against the delivery
count used to determine whether a message ought to go to the DLQ
topic. Because the DLQ implementation uses the redeliveryCount to
determine when a message has been attempted "enough" times, there is a
chance that messages can get sent to the DLQ without ever having been
"received" by the consuming application. While this scenario is
unlikely, I think it introduces doubt into this feature's design
because it is possible for messages to get classified as dead letters
without delivery at least the maxRedeliverCount.
Here is my PR to adopt the original behavior [1]. Please take a look.
Thanks,
Michael
[0] https://github.com/apache/pulsar/issues/5881
[1] https://github.com/apache/pulsar/pull/17060
Re: [DISCUSS] Message redeliveryCount semantics and DLQ
Posted by PengHui Li <pe...@apache.org>.
Hi Michael,
I agree with cherry-picking to release branches.
And should be highlighted in the release note.
Thanks,
Penghui
On Thu, Aug 18, 2022 at 12:14 PM Michael Marshall <mm...@apache.org>
wrote:
> The delivery semantics are a very succinct way to describe my concerns,
> thanks for your comment Penghui.
>
> My only other question is whether we classify the behavior change as a
> regression that should be fixed in existing releases or if it should only
> be reverted for 2.12 and beyond. I propose it should get cherry picked to
> existing release branches because of the impacts to delivery semantics.
>
> Thanks,
> Michael
>
> On Sun, Aug 14, 2022 at 7:21 PM PengHui Li <pe...@apache.org> wrote:
>
> > I support this motivation. We should avoid any cases which
> > might break the at-least-one delivery semantic from the user's
> perspective.
> >
> > Thanks,
> > Penghui
> >
> > On Thu, Aug 11, 2022 at 1:36 PM Michael Marshall <mm...@apache.org>
> > wrote:
> >
> > > Hi Pulsar Community,
> > >
> > > In Pulsar 2.5.0, the semantics for a message's redeliveryCount changed
> > > at the request of this issue [0]. Please see the issue for relevant
> > > context.
> > >
> > > In summary, the issue is whether a message that is neither positively
> > > nor negatively acknowledged should get counted as "redelivered" when a
> > > consumer disconnects from the broker.
> > >
> > > The issue asserts that because a message could lead to a fatal error
> > > in the consumer application that could prevent negative
> > > acknowledgement getting sent to the broker, every delivery of a
> > > message should count towards the redeliveryCount.
> > >
> > > In my view, this edge case should not justify changing the
> > > redeliveryCount semantics. It is the responsibility of the application
> > > to handle all messages, even those that are malformed.
> > >
> > > My primary motivation for revisiting this design is to discuss one of
> > > its consequences. After the change, messages that are only ever sent
> > > to an application's receiverQueue are counted against the delivery
> > > count used to determine whether a message ought to go to the DLQ
> > > topic. Because the DLQ implementation uses the redeliveryCount to
> > > determine when a message has been attempted "enough" times, there is a
> > > chance that messages can get sent to the DLQ without ever having been
> > > "received" by the consuming application. While this scenario is
> > > unlikely, I think it introduces doubt into this feature's design
> > > because it is possible for messages to get classified as dead letters
> > > without delivery at least the maxRedeliverCount.
> > >
> > > Here is my PR to adopt the original behavior [1]. Please take a look.
> > >
> > > Thanks,
> > > Michael
> > >
> > > [0] https://github.com/apache/pulsar/issues/5881
> > > [1] https://github.com/apache/pulsar/pull/17060
> > >
> >
>
Re: [DISCUSS] Message redeliveryCount semantics and DLQ
Posted by Michael Marshall <mm...@apache.org>.
The delivery semantics are a very succinct way to describe my concerns,
thanks for your comment Penghui.
My only other question is whether we classify the behavior change as a
regression that should be fixed in existing releases or if it should only
be reverted for 2.12 and beyond. I propose it should get cherry picked to
existing release branches because of the impacts to delivery semantics.
Thanks,
Michael
On Sun, Aug 14, 2022 at 7:21 PM PengHui Li <pe...@apache.org> wrote:
> I support this motivation. We should avoid any cases which
> might break the at-least-one delivery semantic from the user's perspective.
>
> Thanks,
> Penghui
>
> On Thu, Aug 11, 2022 at 1:36 PM Michael Marshall <mm...@apache.org>
> wrote:
>
> > Hi Pulsar Community,
> >
> > In Pulsar 2.5.0, the semantics for a message's redeliveryCount changed
> > at the request of this issue [0]. Please see the issue for relevant
> > context.
> >
> > In summary, the issue is whether a message that is neither positively
> > nor negatively acknowledged should get counted as "redelivered" when a
> > consumer disconnects from the broker.
> >
> > The issue asserts that because a message could lead to a fatal error
> > in the consumer application that could prevent negative
> > acknowledgement getting sent to the broker, every delivery of a
> > message should count towards the redeliveryCount.
> >
> > In my view, this edge case should not justify changing the
> > redeliveryCount semantics. It is the responsibility of the application
> > to handle all messages, even those that are malformed.
> >
> > My primary motivation for revisiting this design is to discuss one of
> > its consequences. After the change, messages that are only ever sent
> > to an application's receiverQueue are counted against the delivery
> > count used to determine whether a message ought to go to the DLQ
> > topic. Because the DLQ implementation uses the redeliveryCount to
> > determine when a message has been attempted "enough" times, there is a
> > chance that messages can get sent to the DLQ without ever having been
> > "received" by the consuming application. While this scenario is
> > unlikely, I think it introduces doubt into this feature's design
> > because it is possible for messages to get classified as dead letters
> > without delivery at least the maxRedeliverCount.
> >
> > Here is my PR to adopt the original behavior [1]. Please take a look.
> >
> > Thanks,
> > Michael
> >
> > [0] https://github.com/apache/pulsar/issues/5881
> > [1] https://github.com/apache/pulsar/pull/17060
> >
>
Re: [DISCUSS] Message redeliveryCount semantics and DLQ
Posted by PengHui Li <pe...@apache.org>.
I support this motivation. We should avoid any cases which
might break the at-least-one delivery semantic from the user's perspective.
Thanks,
Penghui
On Thu, Aug 11, 2022 at 1:36 PM Michael Marshall <mm...@apache.org>
wrote:
> Hi Pulsar Community,
>
> In Pulsar 2.5.0, the semantics for a message's redeliveryCount changed
> at the request of this issue [0]. Please see the issue for relevant
> context.
>
> In summary, the issue is whether a message that is neither positively
> nor negatively acknowledged should get counted as "redelivered" when a
> consumer disconnects from the broker.
>
> The issue asserts that because a message could lead to a fatal error
> in the consumer application that could prevent negative
> acknowledgement getting sent to the broker, every delivery of a
> message should count towards the redeliveryCount.
>
> In my view, this edge case should not justify changing the
> redeliveryCount semantics. It is the responsibility of the application
> to handle all messages, even those that are malformed.
>
> My primary motivation for revisiting this design is to discuss one of
> its consequences. After the change, messages that are only ever sent
> to an application's receiverQueue are counted against the delivery
> count used to determine whether a message ought to go to the DLQ
> topic. Because the DLQ implementation uses the redeliveryCount to
> determine when a message has been attempted "enough" times, there is a
> chance that messages can get sent to the DLQ without ever having been
> "received" by the consuming application. While this scenario is
> unlikely, I think it introduces doubt into this feature's design
> because it is possible for messages to get classified as dead letters
> without delivery at least the maxRedeliverCount.
>
> Here is my PR to adopt the original behavior [1]. Please take a look.
>
> Thanks,
> Michael
>
> [0] https://github.com/apache/pulsar/issues/5881
> [1] https://github.com/apache/pulsar/pull/17060
>