mirror of
https://github.com/ClickHouse/ClickHouse.git
synced 2024-11-25 17:12:03 +00:00
fix order of ranges in MergeTreeDataSelectExecutor
This commit is contained in:
parent
cdcaef9349
commit
b9fc9b4569
@ -73,7 +73,10 @@ MergeTreeReadTask::MergeTreeReadTask(
|
||||
: data_part{data_part_}, mark_ranges{mark_ranges_}, part_index_in_query{part_index_in_query_},
|
||||
ordered_names{ordered_names_}, column_name_set{column_name_set_}, columns{columns_}, pre_columns{pre_columns_},
|
||||
remove_prewhere_column{remove_prewhere_column_}, should_reorder{should_reorder_}, size_predictor{std::move(size_predictor_)}
|
||||
{}
|
||||
{
|
||||
/// We need to save marks in reverse order.
|
||||
std::reverse(mark_ranges.begin(), mark_ranges.end());
|
||||
}
|
||||
|
||||
MergeTreeReadTask::~MergeTreeReadTask() = default;
|
||||
|
||||
|
@ -718,9 +718,6 @@ Pipes MergeTreeDataSelectExecutor::spreadMarkRangesAmongStreams(
|
||||
for (size_t i = 0; i < parts.size(); ++i)
|
||||
{
|
||||
total_rows += parts[i].getRowsCount();
|
||||
/// Let the ranges be listed from right to left so that the leftmost range can be dropped using `pop_back()`.
|
||||
std::reverse(parts[i].ranges.begin(), parts[i].ranges.end());
|
||||
|
||||
sum_marks_in_parts[i] = parts[i].getMarksCount();
|
||||
sum_marks += sum_marks_in_parts[i];
|
||||
|
||||
@ -897,6 +894,8 @@ Pipes MergeTreeDataSelectExecutor::spreadMarkRangesAmongStreamsWithOrder(
|
||||
}
|
||||
new_ranges.emplace_back(range.begin, range.end);
|
||||
}
|
||||
|
||||
/// Restore left-to-right order.
|
||||
std::reverse(new_ranges.begin(), new_ranges.end());
|
||||
}
|
||||
|
||||
|
@ -28,11 +28,6 @@ MergeTreeReadPool::MergeTreeReadPool(
|
||||
column_names{column_names_}, do_not_steal_tasks{do_not_steal_tasks_},
|
||||
predict_block_size_bytes{preferred_block_size_bytes_ > 0}, prewhere_info{prewhere_info_}, parts_ranges{parts_}
|
||||
{
|
||||
/// reverse from right-to-left to left-to-right
|
||||
/// because 'reverse' was done in MergeTreeDataSelectExecutor
|
||||
for (auto & part_ranges : parts_ranges)
|
||||
std::reverse(std::begin(part_ranges.ranges), std::end(part_ranges.ranges));
|
||||
|
||||
/// parts don't contain duplicate MergeTreeDataPart's.
|
||||
const auto per_part_sum_marks = fillPerPartInfo(parts_, check_columns_);
|
||||
fillPerThreadInfo(threads_, sum_marks_, per_part_sum_marks, parts_, min_marks_for_concurrent_read_);
|
||||
@ -79,10 +74,9 @@ MergeTreeReadTaskPtr MergeTreeReadPool::getTask(const size_t min_marks_to_read,
|
||||
{
|
||||
const auto marks_to_get_from_range = marks_in_part;
|
||||
|
||||
/** Ranges are in right-to-left order, because 'reverse' was done in MergeTreeDataSelectExecutor
|
||||
* and that order is supported in 'fillPerThreadInfo'.
|
||||
*/
|
||||
/// Ranges are in right-to-left order, because 'reverse' was done in 'fillPerThreadInfo'.
|
||||
ranges_to_get_from_part = thread_task.ranges;
|
||||
std::reverse(ranges_to_get_from_part.begin(), ranges_to_get_from_part.end());
|
||||
|
||||
marks_in_part -= marks_to_get_from_range;
|
||||
|
||||
@ -113,11 +107,7 @@ MergeTreeReadTaskPtr MergeTreeReadPool::getTask(const size_t min_marks_to_read,
|
||||
marks_in_part -= marks_to_get_from_range;
|
||||
need_marks -= marks_to_get_from_range;
|
||||
}
|
||||
|
||||
/** Change order to right-to-left, for MergeTreeThreadSelectBlockInputStream to get ranges with .pop_back()
|
||||
* (order was changed to left-to-right due to .pop_back() above).
|
||||
*/
|
||||
std::reverse(std::begin(ranges_to_get_from_part), std::end(ranges_to_get_from_part));
|
||||
/// Order of ranges was changed to left-to-right due to .pop_back() above.
|
||||
}
|
||||
|
||||
auto curr_task_size_predictor = !per_part_size_predictor[part_idx] ? nullptr
|
||||
@ -211,7 +201,6 @@ std::vector<size_t> MergeTreeReadPool::fillPerPartInfo(
|
||||
|
||||
/// Read marks for every data part.
|
||||
size_t sum_marks = 0;
|
||||
/// Ranges are in right-to-left order, due to 'reverse' in MergeTreeDataSelectExecutor.
|
||||
for (const auto & range : part.ranges)
|
||||
sum_marks += range.end - range.begin;
|
||||
|
||||
@ -251,6 +240,10 @@ void MergeTreeReadPool::fillPerThreadInfo(
|
||||
{
|
||||
threads_tasks.resize(threads);
|
||||
|
||||
/// Let the ranges be listed from right to left so that the leftmost range can be dropped using `pop_back()`.
|
||||
for (auto & part : parts)
|
||||
std::reverse(part.ranges.begin(), part.ranges.end());
|
||||
|
||||
const size_t min_marks_per_thread = (sum_marks - 1) / threads + 1;
|
||||
|
||||
for (size_t i = 0; i < threads && !parts.empty(); ++i)
|
||||
|
@ -95,11 +95,6 @@ try
|
||||
|
||||
task_columns = getReadTaskColumns(storage, data_part, required_columns, prewhere_info, check_columns);
|
||||
|
||||
/** @note you could simply swap `reverse` in if and else branches of MergeTreeDataSelectExecutor,
|
||||
* and remove this reverse. */
|
||||
MarkRanges remaining_mark_ranges = all_mark_ranges;
|
||||
std::reverse(remaining_mark_ranges.begin(), remaining_mark_ranges.end());
|
||||
|
||||
auto size_predictor = (preferred_block_size_bytes == 0)
|
||||
? nullptr
|
||||
: std::make_unique<MergeTreeBlockSizePredictor>(data_part, ordered_names, data_part->storage.getSampleBlock());
|
||||
@ -109,7 +104,7 @@ try
|
||||
column_name_set = NameSet{column_names.begin(), column_names.end()};
|
||||
|
||||
task = std::make_unique<MergeTreeReadTask>(
|
||||
data_part, remaining_mark_ranges, part_index_in_query, ordered_names, column_name_set, task_columns.columns,
|
||||
data_part, all_mark_ranges, part_index_in_query, ordered_names, column_name_set, task_columns.columns,
|
||||
task_columns.pre_columns, prewhere_info && prewhere_info->remove_prewhere_column,
|
||||
task_columns.should_reorder, std::move(size_predictor));
|
||||
|
||||
|
@ -0,0 +1 @@
|
||||
300
|
@ -0,0 +1,16 @@
|
||||
drop table if exists t;
|
||||
|
||||
create table t (a Int, b Int) engine = MergeTree order by (a, b) settings index_granularity = 400;
|
||||
|
||||
insert into t select 0, 0 from numbers(50);
|
||||
insert into t select 0, 1 from numbers(350);
|
||||
insert into t select 1, 2 from numbers(400);
|
||||
insert into t select 2, 2 from numbers(400);
|
||||
insert into t select 3, 0 from numbers(100);
|
||||
|
||||
select sleep(1) format Null; -- sleep a bit to wait possible merges after insert
|
||||
|
||||
set max_threads = 1;
|
||||
optimize table t final;
|
||||
|
||||
select sum(a) from t where a in (0, 3) and b = 0;
|
Loading…
Reference in New Issue
Block a user