Queries on taxi dataset are slow with ps=1

[githubmanticore]

Queries on partial taxi dataset (6 indexes) are almost 2x slower with pseudo_sharding enabled:

mysql> set global pseudo_sharding=0; SELECT avg(total_amount) FROM taxi WHERE trip_distance = 5; show meta; 
Query OK, 0 rows affected (0.00 sec) 
 
+-------------------+ 
| avg(total_amount) | 
+-------------------+ 
|       17.94111949 | 
+-------------------+ 
1 row in set (0.07 sec) 
 
+----------------+-------------------------------------+ 
| Variable_name  | Value                               | 
+----------------+-------------------------------------+ 
| total          | 1                                   | 
| total_found    | 1                                   | 
| total_relation | eq                                  | 
| time           | 0.065                               | 
| index          | trip_distance:SecondaryIndex (100%) | 
+----------------+-------------------------------------+ 
5 rows in set (0.00 sec) 
 
mysql> set global pseudo_sharding=1; SELECT avg(total_amount) FROM taxi WHERE trip_distance = 5; show meta; 
Query OK, 0 rows affected (0.00 sec) 
 
+-------------------+ 
| avg(total_amount) | 
+-------------------+ 
|       17.94111949 | 
+-------------------+ 
1 row in set (0.12 sec) 
 
+----------------+----------------------------------------------------------------------+ 
| Variable_name  | Value                                                                | 
+----------------+----------------------------------------------------------------------+ 
| total          | 1                                                                    | 
| total_found    | 1                                                                    | 
| total_relation | eq                                                                   | 
| time           | 0.122                                                                | 
| index          | trip_distance:SecondaryIndex (18%), trip_distance:ColumnarScan (81%) | 
+----------------+----------------------------------------------------------------------+ 
5 rows in set (0.00 sec) 

This needs to be fixed.

[githubmanticore]

➤ Ilya Kuznetsov commented:

Incorrect CBO path selection was due to 2 things:

1. CBO didn't account for the fact that implicit group sorter is faster than plain group sorter
2. Histograms sometimes gave VERY misleading estimates when value distribution is uneven. Changed max histogram size from 1k to 8k which helped a lot (requires reindexing to work).

Fixed those issues in 92d722a