Closed raunaqmorarka closed 9 months ago
We can try this, but I'm not sure it'll have the same effect as it did for Hive. In Hive we have a queue of up to 1000 splits in memory for some amount of time, and there might be a large gap between when they are generated and when they are served to a worker. In Iceberg we don't have that queue, splits are generated on demand, so the gap between when they are generated and when they are served should be small.
@alexjo2144 splits get queued on worker node too
@raunaqmorarka that's interesting, I am volunteering to work on this issue. wondering is it worth to identify the slowness can be there when this feature is not there.
@osscm are you working on this?
@osscm are you working on this?
Hi @sopel39, yes I started on it.
Implemented by https://github.com/trinodb/trino/pull/20212
Port hive optimization from https://github.com/trinodb/trino/pull/9869 It should help in improving utilisation of dynamic filters which arrive late
cc: @sopel39 @radek-starburst @alexjo2144