Re: [DISCUSS] KIP-1050: Consistent error handling for Transactions

[Kaushik Raina]

Thanks Matthias for feedback
- We have updates KIP and grouped exceptions
https://cwiki.apache.org/confluence/display/KAFKA/KIP-1050%3A+Consistent+error+handling+for+Transactions#KIP1050:ConsistenterrorhandlingforTransactions-ExceptionTable

- Regarding compatibility,  all changes in KIP are expected to be *backword
compatible*.  We have updated KIP to make it clear.


On Tue, Jun 11, 2024 at 4:50 AM Matthias J. Sax  wrote:

> Thanks for this KIP. Great to see it. I would assume it will make
> KIP-691 unnecessary?
>
> I don't think I fully understand the proposal yet. It's clear, that you
> propose to add new sub-classed to group existing exceptions. But it's
> not clear to me, which of the existing exceptions (which implement
> ApiException directly right now) will get a new parent class and go into
> the same group. You only list `InvalidProducerEpochException` which gets
> `AbortableTransactionException` as new parent. It would help a lot, if
> you could list out explicitly, which existing exceptions are grouped
> together via which sub-class.
>
> It should be sufficient to just add a list for each group. For the newly
> added exception classes, I would also omit all constructors etc and just
> add a comment about it -- having constructors listed out does not add
> much value to the KIP itself but makes it harder to read (it's
> effectively noise we can avoid IMHO).
>
>
>
> I am also wondering about compatibility? If I read the section
> correctly, you actually propose to introduce a non-backward-compatible
> change?
>
> > Based on type of exception thrown, user needs to change their exception
> catching logic to take actions against their exception handling.
>
> Ie, an application cannot be upgrade w/o code changes? I am not sure if
> this is acceptable?
>
> I think it would be much better (not sure if feasible) to keep the old
> behavior and let users opt-in / enable the new semantics via a config.
> If the new behavior is disabled, we could log a WARN that the app should
> upgrade to work with the new semantics, and we would only enforce the
> new behavior in a later major release.
>
> Thoughts?
>
>
>
> -Matthias
>
>
>
>
>
>
> On 6/7/24 4:06 AM, Kaushik Raina wrote:
> > Thank you Andrew for feedback
> >
> > 1. We are suggesting to only update subclasses of
> > o.a.k.common.errors.ApiException, which are used in transactions. All
> such
> > subclasses are mentioned in Exception table
> > <
> https://cwiki.apache.org/confluence/display/KAFKA/KIP-1050%3A+Consistent+error+handling+for+Transactions#KIP1050:ConsistenterrorhandlingforTransactions-ExceptionTable
> >
> >
> > 2. "Producer-Recoverable" corresponds to the AbortableException. I have
> > updated comments on each exception type.
> >
> > 3. Yes, it's correct that by adding a "Retriable" exception, it
> simplifies
> > the determination of which errors can be retried internally. In the
> Exception
> > table
> > <
> https://cwiki.apache.org/confluence/display/KAFKA/KIP-1050%3A+Consistent+error+handling+for+Transactions#KIP1050:ConsistenterrorhandlingforTransactions-ExceptionTable
> >
> > mentioned
> > in the "Proposed Changes" section, the "Expected Handling" column
> signifies
> > the handling for each error type. Please let me know if any further
> > clarification is needed.
> >
> > 4a. Yes, that is correct. For clarity, only one constructor has been
> > mentioned in the KIP. An ellipsis has been added as a placeholder,
> > indicating that there are additional functions in the class but they are
> > not explicitly specified.
> > 4b. Updated in the KIP.
> >
> > 5. TopicAuthorizationException extends "Invalid Configuration". "Invalid
> > Configuration" type can be resolved either by dynamically updating the
> > configuration, which does not require a restart, or by statically
> updating
> > it by restarting the application. It is at the application's discretion
> how
> > they want to handle each "Invalid Configuration" type.
> >
> > I have added Client side handling example
> > <
> https://cwiki.apache.org/confluence/display/KAFKA/KIP-1050%3A+Consistent+error+handling+for+Transactions#KIP1050:ConsistenterrorhandlingforTransactions-Clientsidecodeexample
> >
> > in
> > KIP. Hope that helps.
> >
>

Re: [DISCUSS] KIP-1050: Consistent error handling for Transactions

[Matthias J. Sax]

Thanks for this KIP. Great to see it. I would assume it will make 
KIP-691 unnecessary?


I don't think I fully understand the proposal yet. It's clear, that you 
propose to add new sub-classed to group existing exceptions. But it's 
not clear to me, which of the existing exceptions (which implement 
ApiException directly right now) will get a new parent class and go into 
the same group. You only list `InvalidProducerEpochException` which gets 
`AbortableTransactionException` as new parent. It would help a lot, if 
you could list out explicitly, which existing exceptions are grouped 
together via which sub-class.


It should be sufficient to just add a list for each group. For the newly 
added exception classes, I would also omit all constructors etc and just 
add a comment about it -- having constructors listed out does not add 
much value to the KIP itself but makes it harder to read (it's 
effectively noise we can avoid IMHO).




I am also wondering about compatibility? If I read the section 
correctly, you actually propose to introduce a non-backward-compatible 
change?



Based on type of exception thrown, user needs to change their exception 
catching logic to take actions against their exception handling.


Ie, an application cannot be upgrade w/o code changes? I am not sure if 
this is acceptable?


I think it would be much better (not sure if feasible) to keep the old 
behavior and let users opt-in / enable the new semantics via a config. 
If the new behavior is disabled, we could log a WARN that the app should 
upgrade to work with the new semantics, and we would only enforce the 
new behavior in a later major release.


Thoughts?



-Matthias






On 6/7/24 4:06 AM, Kaushik Raina wrote:

Thank you Andrew for feedback

1. We are suggesting to only update subclasses of
o.a.k.common.errors.ApiException, which are used in transactions. All such
subclasses are mentioned in Exception table


2. "Producer-Recoverable" corresponds to the AbortableException. I have
updated comments on each exception type.

3. Yes, it's correct that by adding a "Retriable" exception, it simplifies
the determination of which errors can be retried internally. In the Exception
table

mentioned
in the "Proposed Changes" section, the "Expected Handling" column signifies
the handling for each error type. Please let me know if any further
clarification is needed.

4a. Yes, that is correct. For clarity, only one constructor has been
mentioned in the KIP. An ellipsis has been added as a placeholder,
indicating that there are additional functions in the class but they are
not explicitly specified.
4b. Updated in the KIP.

5. TopicAuthorizationException extends "Invalid Configuration". "Invalid
Configuration" type can be resolved either by dynamically updating the
configuration, which does not require a restart, or by statically updating
it by restarting the application. It is at the application's discretion how
they want to handle each "Invalid Configuration" type.

I have added Client side handling example

in
KIP. Hope that helps.

Re: [DISCUSS] KIP-1050: Consistent error handling for Transactions

[Kaushik Raina]

Thank you Andrew for feedback

1. We are suggesting to only update subclasses of
o.a.k.common.errors.ApiException, which are used in transactions. All such
subclasses are mentioned in Exception table


2. "Producer-Recoverable" corresponds to the AbortableException. I have
updated comments on each exception type.

3. Yes, it's correct that by adding a "Retriable" exception, it simplifies
the determination of which errors can be retried internally. In the Exception
table

mentioned
in the "Proposed Changes" section, the "Expected Handling" column signifies
the handling for each error type. Please let me know if any further
clarification is needed.

4a. Yes, that is correct. For clarity, only one constructor has been
mentioned in the KIP. An ellipsis has been added as a placeholder,
indicating that there are additional functions in the class but they are
not explicitly specified.
4b. Updated in the KIP.

5. TopicAuthorizationException extends "Invalid Configuration". "Invalid
Configuration" type can be resolved either by dynamically updating the
configuration, which does not require a restart, or by statically updating
it by restarting the application. It is at the application's discretion how
they want to handle each "Invalid Configuration" type.

I have added Client side handling example

in
KIP. Hope that helps.

Re: [DISCUSS] KIP-1050: Consistent error handling for Transactions

[Andrew Schofield]

Hi Kaushik,
Thanks for the KIP. This is definitely an area that needs clearing up so it’s 
good to see it.

A few initial questions.

1. If I understand correctly, you are proposing to change the superclass of all
subclasses of o.a.k.common.errors.ApiException which can be thrown by the 
producer
or transaction APIs. Is this accurate?

2. You have 4 exception types (and 2 subtypes) in the list in Proposed Changes. 
So, I
would expect 5 new exception classes with names with a direct correspondence to
the list of types, or at least an explicit statement of which of the new 
exceptions
maps to each of the types. For example, “Producer-Recoverable” doesn’t seem to
have a new exception. I’m a bit confused.

3. Some of the error types, the “producer retriable” ones, can be handled
directly within the producer code and do not need to be surfaced to the 
applications.
I suppose this means there is no need to make any change at all for these 
exceptions.
By changing these exceptions too, I suppose it makes it simple in the producer 
to
figure out which errors can be retried internally.

Could you be explicit about which exceptions are going to be changed and which
class is the new superclass? For one thing, having a table would make it obvious
to the code reviewers whether the intended change was being made.

4. A few nits about the Java code snippet for the new exception types.

a. ApiException has 4 constructors: (), (String), (Throwable) and (String, 
Throwable).
I think you’ll need 4 constructors for each of your new exceptions.

b. In some cases, the class names and the constructor names are inconsistent.

5. I believe that a producer application which received 
TopicAuthorizationException
could continue to use the same Producer attempting to use the topic while the
administrator fixed the ACLs to grant access, and then the operations would 
start
working as soon as the access was in place. Is this true for the authorization
exceptions in this KIP? Does it vary by resource type, such as transactional ID?

Thanks,
Andrew

> On 6 Jun 2024, at 12:57, Kaushik Raina  wrote:
>
> Hi everyone, I would like to start a discussion thread for KIP-1050:
> Consistent error handling for Transactions
> https://cwiki.apache.org/confluence/display/KAFKA/KIP-1050%3A+Consistent+error+handling+for+Transactions
>
>
> Thanks
> Kaushik Raina

[DISCUSS] KIP-1050: Consistent error handling for Transactions

[Kaushik Raina]

Hi everyone, I would like to start a discussion thread for KIP-1050:
Consistent error handling for Transactions
https://cwiki.apache.org/confluence/display/KAFKA/KIP-1050%3A+Consistent+error+handling+for+Transactions


Thanks
Kaushik Raina