Karol Sobczak

> So my take is that materialized views as they work now aren't really usable for anything with "layers". What do you think it will take to make MVs more...

Initial branch https://github.com/starburstdata/trino/tree/ks/rollup

Alternatively, we should just skip buffering there and just rely on `OutputStreamSliceOutput` (which is accounted for and already does buffering) to do buffering instead. cc @electrum @losipiuk @linzebing

See Slack discussion for context https://trinodb.slack.com/archives/CGB0QHWSW/p1662502274619329?thread_ts=1662298857.240529&cid=CGB0QHWSW

> We can work around that by adding to airlift the MemoryAware interface which exposes the getRetainedSize method. The we can make the TrinoS3StreamingOutputStream implement that interface. I think `OutputStreams`...

I don't think it's a duplicate. Improving memory efficiency, fixing leaks != accounting for memory

Yet another option is to pass query context in a static way. We already wrap `FileSystem` calls with `doAs`. Maybe we could pass memory context in same way.

cc @lukasz-stec

> I believe partial aggregation supports MarkDistinct too It does, but the problem is when you have multiple distinct aggregations, e.g: `aggr(distinct a), aggr(distinct b), ..`. Then for each you...

cc @lukasz-stec