The new setting should allow to control query complexity on leaf nodes
excluding the final merging stage on the root-node. For example, distributed
query that reads 1k rows from 5 shards will breach the `max_rows_to_read=5000`,
while effectively every shard reads only 1k rows. With setting `max_rows_to_read_leaf=1500`
this limit won't be reached and query will succeed since every shard reads
not more that ~1k rows.
The motivation behind this change is to skip ranges scan for all selected parts
if it is clear that `max_rows_to_read` is already exceeded. The change is quite
noticeable for queries over big number of parts.
This change makes skipping index efficiency more obvious, changing this:
```
Selected 30 parts by date, 30 parts by key, 592 marks to read from 541 ranges
```
Into this:
```
Selected 30 parts by date, 30 parts by key, 48324 marks by primary key, 592 marks to read from 541 ranges
```
This change makes skipping index efficiency more obvious, changing this:
```
Index `idx_duration` has dropped 59 granules.
```
Into this:
```
Index `idx_duration` has dropped 59 / 61 granules.
```
This runs PK lookup and skipping index stages on parts
in parallel, as described in #11564.
While #12277 sped up PK lookups, skipping index stage
may still be a bottleneck in a select query. Here we
parallelize both stages between parts.
On a query that uses a bloom filter skipping index to pick
2,688 rows out of 8,273,114,994 on a two day time span,
this change reduces latency from 10.5s to 1.5s.
Existing code that looks up marks that match the query has a pathological
case, when most of the part does in fact match the query.
The code works by recursively splitting a part into ranges and then discarding
the ranges that definitely do not match the query, based on primary key.
The problem is that it requires visiting every mark that matches the query,
making the complexity of this sort of look up O(n).
For queries that match exact range on the primary key, we can find
both left and right parts of the range with O(log 2) complexity.
This change implements exactly that.
To engage this optimization, the query must:
* Have a prefix list of the primary key.
* Have only range or single set element constraints for columns.
* Have only AND as a boolean operator.
Consider a table with `(service, timestamp)` as the primary key.
The following conditions will be optimized:
* `service = 'foo'`
* `service = 'foo' and timestamp >= now() - 3600`
* `service in ('foo')`
* `service in ('foo') and timestamp >= now() - 3600 and timestamp <= now`
The following will fall back to previous lookup algorithm:
* `timestamp >= now() - 3600`
* `service in ('foo', 'bar') and timestamp >= now() - 3600`
* `service = 'foo'`
Note that the optimization won't engage when PK has a range expression
followed by a point expression, since in that case the range is not continuous.
Trace query logging provides the following messages types of messages,
each representing a different kind of PK usage for a part:
```
Used optimized inclusion search over index for part 20200711_5710108_5710108_0 with 9 steps
Used generic exclusion search over index for part 20200711_5710118_5710228_5 with 1495 steps
Not using index on part 20200710_5710473_5710473_0
```
Number of steps translates to computational complexity.
Here's a comparison for before and after for a query over 24h of data:
```
Read 4562944 rows, 148.05 MiB in 45.19249672 sec., 100966 rows/sec., 3.28 MiB/sec.
Read 4183040 rows, 135.78 MiB in 0.196279627 sec., 21311636 rows/sec., 691.75 MiB/sec.
```
This is especially useful for queries that read data in order
and terminate early to return "last X things" matching a query.
See #11564 for more thoughts on this.
