-1 (non-binding)
Apologies for late comments. I do like the idea of adding cost sensitive
alignment mode and fixing the silent fallback logic of existing
EVENLY_SPREAD mode.
Sharing my concerns with the current design:
1. EVENLY_SPREAD mode logic is changed silently. This forces the end
users to follow a migration path mentioned in FLIP that may go unnoticed as
part of release upgrade. It is preferable to create new config such as
STRICT_EVENLY_SPREAD.
2. The design introduces more complexity than simplifying due to
addition of fallback mode creating an M * N combinations. It forces user to
figure out which combinations are relevant or not. An example of this
confusion is visible in the Flip itself where the below mode is mentioned
as valid but doesn't make sense.
1. "EVENLY_SPREAD + ADAPTIVE_UPWARD_SPREAD" - "Exact divisor first,
relaxed upward search if none." - Mentioned in FLIP.
1. The "ADAPTIVE_UPWARD_SPREAD" mode, by default, performs exact
division first.
2. User can also specify
"ADAPTIVE_UPWARD_SPREAD+ADAPTIVE_UPWARD_SPREAD" which makes no logical
sense.
3. It becomes very difficult to validate and reason about the best
config given all these options
3. Given that these alignment configs apply only to a subset of
use-cases (especially the source partitions or keyBy operator), we should
try to simplify the config by using 3-5 valid modes or perhaps leverage
existing global config and simplify the existing logic.
On a side note, the current autoscaling logic in the JobVertexScaler class
needs refactoring as it has become bloated over time by taking on too many
responsibilities.
Thanks,
Vivek Jhaver
On Wed, Jun 3, 2026 at 11:09 PM Ryan van Huuksloot via dev <
[email protected]> wrote:
> +1 (non-binding)
>
> Thanks
> Ryan van Huuksloot
> Staff Engineer, Infrastructure | Streaming Platform
> [image: Shopify]
> <https://www.shopify.com/?utm_medium=salessignatures&utm_source=hs_email>
>
>
> On Wed, Jun 3, 2026 at 7:54 AM Gyula Fóra <[email protected]> wrote:
>
> > +1
> >
> > Gyula
> >
> > On Wed, Jun 3, 2026 at 1:10 PM Dennis-Mircea Ciupitu <
> > [email protected]> wrote:
> >
> > > Hi all,
> > >
> > > I'd like to start the vote to approve FLIP-586: Composable Parallelism
> > > Alignment Modes for Flink Autoscaler [1].
> > >
> > > The discussion thread [2] has been active for a while, received
> positive
> > > feedback, and the proposal is ready for a formal vote.
> > >
> > > Thanks in advance,
> > > Dennis
> > >
> > > [1]
> > >
> > >
> >
> https://cwiki.apache.org/confluence/display/FLINK/FLIP-586%3A+Composable+Parallelism+Alignment+Modes+for+Flink+Autoscaler
> > > [2] https://lists.apache.org/thread/blq5jfthmhl06l23778js4ox98dqw2v8
> > >
> >
>
