Re: Empty state restore seems to be broken for Kafka source (1.3.2)

Still not nice, though, and it took a while to finalise discovery for 1.4. ;-) If you need that now you might be able to back port the 1.4 consumer to 1.3. > On 12. Oct 2017, at 17:05, Gyula Fóra wrote: > > Ok, thanks for the clarification. :) > > Gyula > > > On Thu, Oct

Re: Empty state restore seems to be broken for Kafka source (1.3.2)

Ok, thanks for the clarification. :) Gyula On Thu, Oct 12, 2017, 17:04 Aljoscha Krettek wrote: > It might be old but it's not forgotten, the issue I created is actually > marked as a blocker so we won't forget it when releasing 1.3.3 and 1.4.0. > > The issue in Kafka is

Re: Empty state restore seems to be broken for Kafka source (1.3.2)

It might be old but it's not forgotten, the issue I created is actually marked as a blocker so we won't forget it when releasing 1.3.3 and 1.4.0. The issue in Kafka is about new topics/partitions not being discovered or something else? That would be the expected behaviour in Flink < 1.4.0.

Re: Empty state restore seems to be broken for Kafka source (1.3.2)

Hey, I know it's old discussion but there also seems to be a problem with the logic in the kafka source alone regarding new topics added after a checkpoint. Maybe there is a ticket for this already and I just missed it. Cheers, Gyula Gyula Fóra ezt írta (időpont: 2017.

Re: Empty state restore seems to be broken for Kafka source (1.3.2)

Good job for figuring this out! This certainly seems to explain our problems. Thanks! Gyula Aljoscha Krettek ezt írta (időpont: 2017. szept. 14., Cs, 14:46): > After a bit more digging I found that the "isRestored" flag doesn't work > correctly if there are operators

Re: Empty state restore seems to be broken for Kafka source (1.3.2)

After a bit more digging I found that the "isRestored" flag doesn't work correctly if there are operators chained to the sink that have state: https://issues.apache.org/jira/browse/FLINK-7623 Blocker issue for 1.3.3 and 1.4.0. Best, Aljoscha

Re: Empty state restore seems to be broken for Kafka source (1.3.2)

After discussing this between Stefan and me we think that this should actually work. Do you have the log output from restoring the Kafka Consumer? It would be interesting to see whether any of those print: -

Re: Empty state restore seems to be broken for Kafka source (1.3.2)

Yes, and that's essentially what's happening in the 1.4-SNAPSHOT consumer which also has discovery of new partitions. Starting from 1.4-SNAPSHOT we store state in a union state, i.e. all sources get all partition on restore and if they didn't get any they know that they are new. There is no

Re: Empty state restore seems to be broken for Kafka source (1.3.2)

Wouldnt it be enough that Kafka sources store some empty container for there state if it is empty, compared to null when it should be bootstrapped again? Gyula Aljoscha Krettek ezt írta (időpont: 2017. szept. 6., Sze, 14:31): > The problem here is that context.isRestored()

Re: Empty state restore seems to be broken for Kafka source (1.3.2)

The problem here is that context.isRestored() is a global flag and not local to each operator. It says "yes this job was restored" but the source would need to know that it is actually brand new and never had any state. This is quite tricky to do, since there is currently no way (if I'm

Re: Empty state restore seems to be broken for Kafka source (1.3.2)

Thanks for the report, I will take a look. > Am 06.09.2017 um 11:48 schrieb Gyula Fóra : > > Hi all, > > We are running into some problems with the kafka source after changing the > uid and restoring from the savepoint. > What we are expecting is to clear the partition

Empty state restore seems to be broken for Kafka source (1.3.2)

Hi all, We are running into some problems with the kafka source after changing the uid and restoring from the savepoint. What we are expecting is to clear the partition state, and set it up all over again, but what seems to happen is that the consumer thinks that it doesnt have any partitions