As I understand, this is more about the documented semantics of checkpoint feature that Flink community agreed to guarantee for users so far in the last releases. Although technically the rescaling is currently possible for full and incremental checkpoints, it can change in future if checkpoints would need breaking optimisations or on contrary it might be fixed in the docs as supported from some release and on. In general, checkpoint API is currently designed for failure recovery of running job, not as savepoints for manual restart including rescaling.
> On 27 Aug 2018, at 11:23, Chesnay Schepler <ches...@apache.org> wrote: > > Well that's a single StateBackend. In order to say that "Checkpoints support > rescaling" this must apply to all StateBackends, which AFAIK is not the case. > > On 27.08.2018 10:52, Yun Tang wrote: >> Hi, all >> >> >From the latest 1.6 (or all the previous verions) documentation >> >https://ci.apache.org/projects/flink/flink-docs-release-1.6/ops/state/checkpoints.html#difference-to-savepoints, >> > it said "Checkpoints have a few differences from savepoints, they do not >> >support Flink specific features like rescaling." >> >> However, if you just take a look at RocksDB-StateBackend's code, you could >> see clearly that RocksDBStateBackend support rescaling from previous >> checkpoint. Could anyone give an explanation? >> >> Best >> Yun >> >
