Hi, Mang
Thanks for your update, the FLIP looks good to me now.
Best,
Ron
Mang Zhang 于2023年6月9日周五 12:08写道:
> Hi Ron,
> Thanks for your reply!
> After our offline discussion, at present, there may be many of flink jobs
> using non-atomic CTAS functions, especially Stream jobs,
> If we only
Hi, Mang
In FLIP-214, we have discussed that atomicity is not needed in streaming
mode, so we have implemented the initial version that doesn't support
atomicity. In addition, we introduce the option
"table.ctas.atomicity-enabled" to enable the atomic ability. According to
your FLIP-315
Thank you for the great work, Mang! The updated proposal looks good to me.
Best,
Jark
> 2023年6月8日 11:49,Jingsong Li 写道:
>
> Thanks Mang for updating!
>
> Looks good to me!
>
> Best,
> Jingsong
>
> On Wed, Jun 7, 2023 at 2:31 PM Mang Zhang wrote:
>>
>> Hi Jingsong,
>>
>>> I have some
Thanks Mang for updating!
Looks good to me!
Best,
Jingsong
On Wed, Jun 7, 2023 at 2:31 PM Mang Zhang wrote:
>
> Hi Jingsong,
>
> >I have some doubts about the `TwoPhaseCatalogTable`. Generally, our
> >Flink design places execution in the TableFactory or directly in the
> >Catalog, so
Hi Jingsong,
>I have some doubts about the `TwoPhaseCatalogTable`. Generally, our
>Flink design places execution in the TableFactory or directly in the
>Catalog, so introducing an executable table makes me feel a bit
>strange. (Spark is this style, but Flink may not be)
On this issue, we
Hi Jingsong,
Thank you for your reply!
We introduced `TwoPhaseCatalogTable` for two reasons:
1. The `TwoPhaseCatalogTable` of different data sources can have more
operations, if through Catalog, there can only be simple create table and drop
table, not flexible enough; For example, deleting a
Hi Mang,
Thanks for starting this FLIP.
I have some doubts about the `TwoPhaseCatalogTable`. Generally, our
Flink design places execution in the TableFactory or directly in the
Catalog, so introducing an executable table makes me feel a bit
strange. (Spark is this style, but Flink may not be)
Hi Jing,
Currently, we cannot determine in the planner whether the source is bounded or
unbounded.
So when we design the API, we use the execution model to help determine if
atomicity can be supported.
Thank you very much for your reply!
--
Best regards,
Mang Zhang
At 2023-04-28
Hi Mang,
Boundedness and execution modes are two orthogonal concepts. Since atomic
CTAS will be only supported for bounded data, which means it does not
depend on the execution modes. I was wondering if it is possible to only
provide (or call) twoPhaseCreateTable for bounded data (in both
Hi Jing,
Yes, the atomic CTAS will be only supported for bounded data, but the execution
modes can be stream or batch.
I introduced the isStreamingMode parameter in the twoPhaseCreateTable API to
make it easier for users to provide different levels of atomicity
implementation depending on the
Hi Mang,
Thanks for clarifying it. I am trying to understand your thoughts. Do you
actually mean the boundedness[1] instead of the execution modes[2]? I.e.
the atomic CTAS will be only supported for bounded data.
Best regards,
Jing
[1]
Hi, Mang
I have a question about the implementation details. For the atomicity case,
since the target table is not created before the JobGraph is generated, but
then the target table is required to exist when optimizing plan to generate
the JobGraph. So how do you solve this problem?
Best,
Ron
Share some insights about the new TwoPhaseCatalogTable proposed after offline
discussion with Mang.
The main or important reason is that the TwoPhaseCatalogTable enables external
connectors to implement theirs own logic for commit / abort.
In FLIP-218, for atomic CTAS, the Catalog will then
Hi Mang,
This is the FLIP I was looking forward to after FLIP-218. Thanks for
driving it. I have two questions and would like to know your thoughts,
thanks:
1. It looks like you found another way to design the atomic CTAS with new
serializable TwoPhaseCatalogTable instead of making Catalog
Hi, Mang.
+1 for completing the support for atomicity of CTAS, this is very useful in
batch scenarios and integrate with the data lake which support transcation.
I just have one question, IIUC, the DynamiacTableSink will need to know it's
for normal case or the atomicity with CTAS as well as
Hi, Mang
+1 for completing the support for atomicity of CTAS, this is very useful in
batch scenarios.
I have two questions:
1. naming wise:
a) can we rename the `Catalog#getTwoPhaseCommitCreateTable` to
`Catalog#twoPhaseCreateTable` (and we may add
Hi, Mang
Atomicity is very important for CTAS, especially for batch jobs. This FLIP
is a continuation of FLIP-218, which is valuable for CTAS.
I just have one question, in the Motivation part of FLIP-218, we mentioned
three levels of atomicity semantics, can this current design do the same as
17 matches
Mail list logo