* feat: Preserve number of streams after evaluation the window functions to allow parallel stream processing
* fix style
* fix style
* fix style
* setting query_plan_preserve_num_streams_after_window_functions default true
* fix tests by SETTINGS query_plan_preserve_num_streams_after_window_functions=0
* fix test references
* Resize the streams after the last window function, to keep the order between WindowTransforms (and WindowTransform works on single stream anyway).
* feat: Preserve number of streams after evaluation the window functions to allow parallel stream processing
* fix style
* fix style
* fix style
* setting query_plan_preserve_num_streams_after_window_functions default true
* fix tests by SETTINGS query_plan_preserve_num_streams_after_window_functions=0
* fix test references
* Resize the streams after the last window function, to keep the order between WindowTransforms (and WindowTransform works on single stream anyway).
* add perf test
* perf: change the dataset from 50M to 5M
* rename query_plan_preserve_num_streams_after_window_functions -> query_plan_enable_multithreading_after_window_functions
* update test reference
* fix clang-tidy
---------
Co-authored-by: Nikita Taranov <nikita.taranov@clickhouse.com>
I've noticed this in one of production setups, where lots of queries are
executed with distributed_group_by_no_merge=2 (automatically vai
optimize_skip_unused_shards optimization) and
optimize_aggregation_in_order=1, with this two settings initiator may
not read from some shards (i.e. X) for quite long period of time,
because it reads from other shards (i.e. Y) and it does not need
any data from X yet (due to it read everything in order), and this will
lead to query timeout, so timeouts had been increased.
Previously both timeouts had been tuned, but this leads to connection
hungs in case of abnormal machine reboots. So it is better to tune only
send_timeout (and this should be enough, since the only problem is the
sender) and this will allow to see that the connection is broken on the
initiator once it will read from this shard.
but after changing only send_timeout, the query still timedout, and the
reason is this place, which swaps this timeouts.
It had been introduced in 134efcd, with a comment:
NOTE: We use send_timeout for the receive timeout and vice versa (change arguments ordering in TimeoutSetter),
because send_timeout is client-side setting which has opposite meaning on the server side.
But it sounds odd to me, it may only make sense with the
clickhouse-client, since any other driver does not implement any server
settings.
Signed-off-by: Azat Khuzhin <a.khuzhin@semrush.com>
