Thank you everybody for the questions and answers (especially Yanfei Lei), it
was very instructive to go over the discussion.
I am gonna add some questions on top of what happened and add some thoughts as
well below.
1) About locking VS reference counting:
I would like to clear out which mechanism prevents what:
The `KeyAccountingUnit` implements locking behavior on keys and ensures 2 state
requests on the same key happen in order. Double-locking the same key does not
result in deadlocks (thanks to the `previous == record` condition in your
pseudo-code), so, the same callback chain can update/read multiple times the
same piece of state.
On the other side we have the reference counting mechanism that is used to
understand whether a record has been fully processed, i.e., all state
invocations have been carried out.
Here is the question: am I correct if we say that key accounting is needed for
out-of-order while reference counting is needed for checkpointing and
watermarking?
2) Number of mails:
To expand on what Jing Ge already asked, in the example code in the FLIP:
```
state.value().then(
val -> {
return state.update(val + 1).then(
empty -> {
out.collect(val + 1);
};
}
}
```
Do you end up having two mails?:
• first wrapping `val -> {...}`
• second wrapping `empty -> {...}`
Did I get it correctly?
3) On the guarantees provided by the async execution framework:
Always referring to your example, say that when `val -> {...}` gets called the
state value is 0.
The callback will update to 1 and register another callback to collect while
still holding the lock on that piece of state, and preventing somebody else to
read the value during the entire process (implementing atomicity for a
transaction).
Now, say the code changes to:
```
state.value().then(
val -> {
state.update(val + 1);
out.collect(val + 1);
}
}
```
Would this change something on the consistency guarantees provided?
I guess not, as, the lock is held in any case until the value on the state
hasn't been updated.
This, instead (similar to your example):
```
int x = 0;
state.value().then( val -> { x = val + 1; } );
state.update(x);
out.collect(x);
```
Could lead to any inconsistency (most probably the state would be updated to 0).
4) On the local variables/attributes:
The last example above exemplifies one of my concerns: what about the values
enclosed in the callbacks?
That seems a bit counter-intuitive and brittle from a user perspective.
In the example we have a function-level variable, but what about fields:
```
int x = 0;
void processElement(...) {
state.value().then( val -> { x += val; } );
...
out.collect(x);
}
```
What could happen here?
Every callback would enclose the current value of the field?
Don't know exactly where I am heading, but it seems quite complex/convoluted :)
5) On watermarks:
It seems that, in order to achieve a good throughput, out-of-order mode should
be used.
In the FLIP I could not understand well how many things could go wrong if that
one is used.
Could you please clarify that?
Thank you for your availability and your great work!
On Mar 19, 2024 at 10:51 +0100, Yanfei Lei <fredia...@gmail.com>, wrote:
> Hi everyone,
>
> Thanks for your valuable discussion and feedback!
>
> Our discussions have been going on for a while and there have been no
> new comments for several days. So I would like to start a vote after
> 72 hours.
>
> Please let me know if you have any concerns, thanks!
>
> Yanfei Lei <fredia...@gmail.com> 于2024年3月13日周三 12:54写道：
>
> >
> > Hi Jing,
> > Thanks for the reply and follow up.
> >
> > > > What is the benefit for users to build a chain of mails instead of just
> > > > one mail(it is still async)?
> >
> > Just to make sure we're on the same page, I try to paraphrase your question:
> > A `then()` call will be encapsulated as a callback mail. Your question
> > is whether we can call then() as little as possible to reduce the
> > overhead of encapsulating it into a mail.
> >
> > In general, whether to call `then()` depends on the user's data
> > dependencies. The operations in a chain of `then()` are strictly
> > ordered.
> >
> >
> >
> > The following is an example without data dependencies, if written in
> > the form of a `then` chain:
> > stateA.update(1).then(stateB.update(2).then(stateC.update(3)));
> >
> > The execution order is:
> > ```
> > stateA update 1 -> stateB update 2-> stateC update 3
> > ```
> >
> > If written in the form without `then()` call, they will be placed in a
> > "mail/mailboxDefaultAction", and each state request will still be
> > executed asynchronously:
> > ```
> > stateA.update(1);
> > stateB.update(2);
> > stateC.update(3);
> > ```
> >
> > The order in which they are executed is undefined and may be:
> > ```
> > - stateA update 1 -> stateB update 2-> stateC update 3
> > - stateB update 2 -> stateC update 3-> stateA update 1
> > - stateC update 3 -> stateA update 1-> stateB update 2
> > ...
> > ```
> > And the final results are "stateA = 1, stateB = 2, stateC = 3". In
> > this case, the two ways of writing are equivalent.
> >
> >
> >
> > If there are data dependencies, for example:
> > ```
> > stateA.update(1).then(stateA.update(2))
> > ```
> >
> > Then the execution order is:
> > ```
> > stateA update 1 -> stateA update 2
> > ```
> >
> > If written in the form without `then()` call:
> > ```
> > stateA.update(1);
> > stateA.update(2);
> > ```
> >
> > The order in which they are executed is undefined and may be:
> > ```
> > - stateA update 1 -> stateA update 2
> > - stateA update 2-> stateA update 1
> > ```
> > The final result may be "stateA = 1" *OR* "stateA = 2". In this case,
> > the way without `then()` chain to limit the execution order, and the
> > results may be wrong.
> >
> > In summary, how many mails are encapsulated depends on how the user
> > writes the code, and how the user writes the code depends on their
> > data dependencies. [1][2] may be helpful for asynchronous programming
> > practice.
> >
> >
> > > > I was wondering if exceptions in the mail chain would have an impact on
> > > > the reference counting?
> >
> > We will catch exceptions that can be handled, they don't have impacts
> > on the reference counting.
> > For exceptions that cannot be handled, we will directly fail the job.
> >
> > > > Maybe a UT to cover all kinds of cases, i.e. happy paths and unhappy
> > > > paths, would make sense.
> >
> > Nice suggestions, we will add a UT to cover those cases.
> >
> >
> > [1]
> > https://docs.oracle.com/javase/8/docs/api/java/util/concurrent/CompletableFuture.html
> > [2] https://www.codingjunkie.net/completable-futures-part1/
> >
> > Jing Ge <j...@ververica.com.invalid> 于2024年3月13日周三 07:05写道：
> > > >
> > > > Hi Yanfei,
> > > >
> > > > Thanks for your clarification! Now I got a much clear picture and I am
> > > > still trying to understand your thoughts for some of those questions:
> > > >
> > > >
> > > > > > How many mails are encapsulated depends on how the user writes the
> > > > > > code. The statements in a `then()` will be wrapped into a mail.
> > > > > > StateFuture is a restricted version of CompletableFuture, their
> > > > > > basic
> > > > > > semantics are consistent.
> > > > > >
> > > >
> > > > Conceptually, users can write a chain of many async calls, i.e. many
> > > > then()
> > > > calls. And all these calls for Record A must be executed in order, while
> > > > Record B should stay at the Blocking buffer. What is the benefit for
> > > > users
> > > > to build a chain of mails instead of just one mail(it is still async)?
> > > > Is
> > > > there any best practices or guidelines to teach/tell users when and how
> > > > many async calls in a chain could/should be built?
> > > >
> > > > > > The challenge arises in determining when all the processing logic
> > > > associated with Record A is fully executed. To address this, we have
> > > > adopted a reference counting mechanism that tracks ongoing operations
> > > > (either processing input or executing a callback) related to a single
> > > > record.
> > > >
> > > > > > We describe this in the "Error handling"[2] section. This FLIP also
> > > > > > adopts the design from FLIP-368, ensuring that all state interfaces
> > > > > > throw unchecked exceptions and, consequently, do not declare any
> > > > > > exceptions in their signatures. In cases where an exception occurs
> > > > > > while accessing the state, the job should fail.
> > > > > >
> > > >
> > > > My question was not about how exceptions will be defined. I am not sure
> > > > how
> > > > unchecked exceptions handling will be implemented. I was wondering if
> > > > exceptions in the mail chain would have an impact on the reference
> > > > counting? E.g. in Fig 5, if an exception happened in the value(),
> > > > update(int), or function within then(), any -1 counting might be missed?
> > > > Maybe a UT to cover all kinds of cases, i.e. happy paths and unhappy
> > > > paths,
> > > > would make sense.
> > > >
> > > > Best regards,
> > > > Jing
> > > >
> > > > On Mon, Mar 11, 2024 at 3:58 AM Yanfei Lei <fredia...@gmail.com> wrote:
> > > >
> > > > > > Hi Jing,
> > > > > >
> > > > > > Thanks for your thoughtful feedback!
> > > > > >
> > > > > > > > does it mean that there will be three mails for Read, Update,
> > > > > > > > and Output
> > > > > > ?
> > > > > >
> > > > > > With this example, there are two mails. The Read is processed by
> > > > > > `mailboxDefaultAction`[1], and the Update and Output are
> > > > > > encapsulated
> > > > > > as mail.
> > > > > >
> > > > > > > > does it make sense to encapsulate one mail instead of 3 mails
> > > > > > > > with more
> > > > > > overhead?
> > > > > >
> > > > > >
> > > >
> > > > > > How many mails are encapsulated depends on how the user writes the
> > > > > > code. The statements in a `then()` will be wrapped into a mail.
> > > > > > StateFuture is a restricted version of CompletableFuture, their
> > > > > > basic
> > > > > > semantics are consistent.
> > > > > >
> > > > > >
> > > >
> > > > > > > > Would you like to add more description for cases when exceptions
> > > > > > happened? E.g. when reading or/and updating State throws
> > > > > > IOExceptions.
> > > > > >
> > > > > >
> > > >
> > > > > > We describe this in the "Error handling"[2] section. This FLIP also
> > > > > > adopts the design from FLIP-368, ensuring that all state interfaces
> > > > > > throw unchecked exceptions and, consequently, do not declare any
> > > > > > exceptions in their signatures. In cases where an exception occurs
> > > > > > while accessing the state, the job should fail.
> > > > > >
> > > >
> > > >
> > > >
> > > > > > > > Is it correct to understand that AEC is stateless?
> > > > > >
> > > > > > Great perspective, yes, it can be understood that way.
> > > > > > AEC is a task-level component. When the job fails or is restarted,
> > > > > > the
> > > > > > runtime status in AEC will be reset.
> > > > > > In fact, we have considered taking a snapshot of the status in AEC
> > > > > > and
> > > > > > storing it in a checkpoint like "unaligned checkpoint", but since
> > > > > > callback cannot be serialized, this idea is not feasible for the
> > > > > > time
> > > > > > being.
> > > > > >
> > > > > > > > would you like to add Pseudo-code for the inFilghtReocordNum
> > > > > > > > decrement
> > > > > > to help us understand the logic better?
> > > > > >
> > > > > > This part of the code is a bit scattered, we will try to abstract a
> > > > > > pseudo-code. You can first refer to the RecordContext-related code
> > > > > > [3]
> > > > > > in the PoC to understand it.
> > > > > >
> > > > > > [1]
> > > > > > https://github.com/apache/flink/blob/master/flink-streaming-java/src/main/java/org/apache/flink/streaming/runtime/tasks/mailbox/MailboxProcessor.java#L81
> > > > > > [2]
> > > > > > https://cwiki.apache.org/confluence/display/FLINK/FLIP-425%3A+Asynchronous+Execution+Model#FLIP425:AsynchronousExecutionModel-ErrorHandling
> > > > > > [3]
> > > > > > https://github.com/ververica/flink-poc/blob/disagg-poc-2/flink-runtime/src/main/java/org/apache/flink/runtime/state/async/RecordContext.java#L77
> > > > > >
> > > > > >
> > > > > > Best,
> > > > > > Yanfei
> > > > > >
> > > > > > Jing Ge <j...@ververica.com.invalid> 于2024年3月10日周日 23:47写道：
> > > > > > > >
> > > > > > > > Hi Yanfei,
> > > > > > > >
> > > > > > > > Thanks for your proposal! The FLIP contains a lot of great new
> > > > > > > > ideas. I'd
> > > > > > > > like to ask some questions to make sure we are on the same page.
> > > > > > > >
> > > > > > > > > > For the asynchronous interface, Record A should run with
> > > > > > > > > > Read, Update
> > > > > > and
> > > > > > > > Output, while Record B should stay at the Blocking buffer.
> > > > > > > >
> > > > > > > > 1. With this example, does it mean that there will be three
> > > > > > > > mails for
> > > > > > Read,
> > > > > > > > Update, and Output ?
> > > > > > > > 2. If yes, since the Read, Update, and Output have to be
> > > > > > > > executed before
> > > > > > > > Record B, does it make sense to encapsulate one mail instead of
> > > > > > > > 3 mails
> > > > > > > > with more overhead? There must be some thoughts behind the
> > > > > > > > design. Look
> > > > > > > > forward to it.
> > > > > > > >
> > > > > > > > > > The challenge arises in determining when all the processing
> > > > > > > > > > logic
> > > > > > > > associated with Record A is fully executed. To address this, we
> > > > > > > > have
> > > > > > > > adopted a reference counting mechanism that tracks ongoing
> > > > > > > > operations
> > > > > > > > (either processing input or executing a callback) related to a
> > > > > > > > single
> > > > > > > > record.
> > > > > > > >
> > > > > > > > The design reminds me of the JVM reference counting for GC.
> > > > > > > > Would you
> > > > > > like
> > > > > > > > to add more description for cases when exceptions happened?
> > > > > > > > E.g. when
> > > > > > > > reading or/and updating State throws IOExceptions.
> > > > > > > >
> > > > > > > > > > In more detail, AEC uses a inFilghtReocordNum variable to
> > > > > > > > > > trace the
> > > > > > > > current number of records in progress. Every time the AEC
> > > > > > > > receives a new
> > > > > > > > record, the inFilghtReocordNum increases by 1; when all
> > > > > > > > processing and
> > > > > > > > callback for this record have completed, the inFilghtReocordNum
> > > > > > decreases
> > > > > > > > by 1. When processing one checkpoint mail, the current task
> > > > > > > > thread will
> > > > > > > > give up the time slice through the yield() method of the mailbox
> > > > > > executor,
> > > > > > > > so that the ongoing state request’s callback and the blocking
> > > > > > > > state
> > > > > > > > requests will be drained first until inFlightRecordNum reduces
> > > > > > > > to 0.
> > > > > > > >
> > > > > > > > 1. Speaking of draining, is it correct to understand that AEC is
> > > > > > stateless?
> > > > > > > > E.g. AEC could be easily scaled out if it became a bottleneck.
> > > > > > > > 2. There are Pseudo-code for the inFilghtReocordNum increment,
> > > > > > > > would you
> > > > > > > > like to add Pseudo-code for the inFilghtReocordNum decrement to
> > > > > > > > help us
> > > > > > > > understand the logic better?
> > > > > > > >
> > > > > > > > The FLIP shows overall a great design! +1 for it! Looking
> > > > > > > > forward to your
> > > > > > > > thoughts, thanks!
> > > > > > > >
> > > > > > > > Best regards,
> > > > > > > > Jing
> > > > > > > >
> > > > > > > > On Thu, Mar 7, 2024 at 10:05 AM Yanfei Lei
> > > > > > > > <fredia...@gmail.com> wrote:
> > > > > > > >
> > > > > > > > > > Hi devs,
> > > > > > > > > >
> > > > > > > > > > I'd like to start a discussion on FLIP-425: Asynchronous
> > > > > > > > > > Execution
> > > > > > > > > > Model[1], which is a sub-FLIP of FLIP-423: Disaggregated
> > > > > > > > > > State Storage
> > > > > > > > > > and Management[2].
> > > > > > > > > >
> > > > > > > > > > FLIP-425 introduces a non-blocking execution model
> > > > > > > > > > leveraging the
> > > > > > > > > > asynchronous APIs introduced in FLIP-424[3].
> > > > > > > > > > For the whole story please read the FLIP-423[2], and this
> > > > > > > > > > thread is
> > > > > > > > > > aimed to discuss the details of "FLIP-425: Asynchronous
> > > > > > > > > > Execution
> > > > > > > > > > Model".
> > > > > > > > > >
> > > > > > > > > > Regarding the details of this FLIP, there have been some
> > > > > > > > > > discussions
> > > > > > > > > > here[4], mainly focusing on framework overhead profiling,
> > > > > > > > > > watermark
> > > > > > > > > > processing, etc. Please see link[4] for the context.
> > > > > > > > > >
> > > > > > > > > > Looking forward to hearing from you!
> > > > > > > > > >
> > > > > > > > > >
> > > > > > > > > > [1] https://cwiki.apache.org/confluence/x/S4p3EQ
> > > > > > > > > > [2] https://cwiki.apache.org/confluence/x/R4p3EQ
> > > > > > > > > > [3] https://cwiki.apache.org/confluence/x/SYp3EQ
> > > > > > > > > > [4]
> > > > > > > > > > https://lists.apache.org/thread/ct8smn6g9y0b8730z7rp9zfpnwmj8vf0
> > > > > > > > > >
> > > > > > > > > > Best,
> > > > > > > > > > Yanfei
> > > > > > > > > >
> > > > > >
> >
> >
> >
> > --
> > Best,
> > Yanfei
>
>
>
> --
> Best,
> Yanfei
