|Hi mates, we got some Flink jobs, that are writing data from kafka into hdfs, using Bucketing-Sink.|
For some reasons, those jobs are running without checkpointing. For now, it not a big problem for us, if some files are remained opened in case of job reloading.
Periodically, those jobs fail with OutOfMemory exception, and seems, that I found a strange thing in the implementation of BucketingSink.
During the sink lifecycle, we have a state object, implemented as a map, where key is a bucket path, and value is a state, that contains information about opened files and list of pending files.
After researching of the heap dump, I found, that those state stores information about ~ 1_000 buckets and their state, all this stuff weights ~ 120 Mb.
I’ve looked through the code, and found, that we removing the buckets from the state, in notifyCheckpointComplete method.
Iterator<Map.Entry<String, BucketState<T>>> bucketStatesIt = state.bucketStates.entrySet().iterator();
if (!bucketState.isWriterOpen &&
So, this looks like an issue, when you are using this sink in checkpointless environment, because the data always added to the state, but never removed.
Of course, we could enabled checkpointing, and use one of available backends, but as for me, it seems like a non expected behaviour, like I have an opportunity to run the job without checkpointing, but really, if I do so,
I got an exception in sink component.
What do you think about this ? Do anyone got the same problem, and how’ve you solved it ?