Closed fschueler closed 8 years ago
In 0fdb1e0f2211f65ad8c7b509386a8e8b624e1b34, the issue is addressed by using the estimators provided by Flink. The solution still is not perfect, as we rely on the default fraction of memory assigned to Flink, but provides save estimates even in the case of lazy allocation in Flink (which is the default).
If Flink provides access to the configuration values in the future (apart from the REST API), we should change the current implementation. For now, I think it is ok, as we should rely on the user to set reasonable values in their Flink setup.
For operator and broadcast selection we need to estimate/compute different memory characteristics for flink. Since there is now nice API way of doing this we are left with:
Right now we're using very conservative magic numbers that are based on Flink's default.