Storage: Improve small partition table read performance by limit concurrency (#10489) by ti-chi-bot · Pull Request #10500 · pingcap/tiflash

ti-chi-bot · 2025-10-23T06:18:17Z

This is an automated cherry-pick of #10489

What problem does this PR solve?

Issue Number: close #10487

Problem Summary:

If a PartitionTableScan involve many partition but each partition only have 1 segment. Then DeltaMergeStore::read will generate num_partitions * num_streams * UnorderedSourceOp. Because there is only 1 segment, SegmentReadTaskPool will only have 1 concurrency for reading data from disk.

tiflash/dbms/src/Storages/DeltaMerge/DeltaMergeStore.cpp

Lines 1421 to 1454 in 80e1927

    
           size_t final_num_stream 
        
               = enable_read_thread ? std::max(1, num_streams) : std::max(1, std::min(num_streams, tasks.size())); 
        
           auto read_mode = getReadMode(db_context, is_fast_scan, keep_order, executor); 
        
           const auto & final_columns_to_read 
        
               = executor && executor->extra_cast ? *executor->columns_after_cast : columns_to_read; 
        
           auto read_task_pool = std::make_shared<SegmentReadTaskPool>( 
        
               extra_table_id_index, 
        
               final_columns_to_read, 
        
               executor, 
        
               start_ts, 
        
               expected_block_size, 
        
               read_mode, 
        
               std::move(tasks), 
        
               after_segment_read, 
        
               log_tracing_id, 
        
               enable_read_thread, 
        
               final_num_stream, 
        
               dm_context->scan_context->keyspace_id, 
        
               dm_context->scan_context->resource_group_name); 
        
           dm_context->scan_context->read_mode = read_mode; 
        
           if (enable_read_thread) 
        
           { 
        
               for (size_t i = 0; i < final_num_stream; ++i) 
        
               { 
        
                   group_builder.addConcurrency(std::make_unique<UnorderedSourceOp>( 
        
                       exec_context, 
        
                       read_task_pool, 
        
                       final_columns_to_read, 
        
                       extra_table_id_index, 
        
                       log_tracing_id, 
        
                       runtime_filter_list, 
        
                       rf_max_wait_time_ms)); 
        
               }

In order to avoid OOM issue when running queries on large PartitionTableScan (#8507), MultiplexInputStream and ConcatBuilderPool will process streams/source ops of partition tables one by one.

tiflash/dbms/src/Operators/ConcatSourceOp.h

Lines 190 to 199 in 80e1927

    
           void add(PipelineExecGroupBuilder & group_builder) 
        
           { 
        
               RUNTIME_CHECK(group_builder.groupCnt() == 1); 
        
               for (size_t i = 0; i < group_builder.concurrency(); ++i) 
        
               { 
        
                   pool[pre_index++].push_back(std::move(group_builder.getCurBuilder(i))); 
        
                   if (pre_index == pool.size()) 
        
                       pre_index = 0; 
        
               } 
        
           }

So there is only 1 concurrency for storage layer scanning data from 1 partition, and all compute thread wait for the blocks read from the partition. Only after the current partition finish reading, compute thread will call streams/source ops from the next partition. It result to the PartitionTableScan performance degrade along with the number of partitions, which is not expected.

What is changed and how it works?

* Limit the number of source ops by num of segment task * 4 in function `DeltaMergeStore::read`. In order to reduce concurrency overhead and let PartitionTableScan with small partitions that only contains 1~2 segments can schedule more segment read tasks in parallel.
  - For large partition, the storage layer still generate `num_streams` * `UnorderedSourceOp`, the behavior is the same as before.
  - For small partition, the storage layer only generate segment task * 4 * `UnorderedSourceOp`. And `ConcatBuilderPool` reorg the source ops and read the data from multiple partitions in parallel
* Introduce `DMReadOptions` and reduce the complexity of adding has_multiple_partitions from compute layer to storage layer.
* Add active_segment_limit, peak_active_segments, block_slot_limit, peak_blocks_in_queue when `SegmentReadTaskPool` finished

Main logic changes is this piece of code: https://github.com/pingcap/tiflash/pull/10489/files#diff-22b900e9e8020dc835612316f3bf151cae28b621bfac64d6a22b377d062f4b7eR1419-R1435

The source ops generated by each partition will be added to ConcatBuilderPool that wrap into ConcatSourceOp. Consider num_stream == 8.

In the master, each partition will generate final_num_stream * UnorderedSourceOp. And ConcatBuilderPool will generate 8 ConcatSourceOp by

ConcatSourceOp on pool[0]: part-1-UnorderedSourceOp, part-2-UnorderedSourceOp, part-3-UnorderedSourceOp
ConcatSourceOp on pool[1]: part-1-UnorderedSourceOp, part-2-UnorderedSourceOp, part-3-UnorderedSourceOp
ConcatSourceOp on pool[2]: part-1-UnorderedSourceOp, part-2-UnorderedSourceOp, part-3-UnorderedSourceOp
ConcatSourceOp on pool[3]: part-1-UnorderedSourceOp, part-2-UnorderedSourceOp, part-3-UnorderedSourceOp
ConcatSourceOp on pool[4]: part-1-UnorderedSourceOp, part-2-UnorderedSourceOp, part-3-UnorderedSourceOp
ConcatSourceOp on pool[5]: part-1-UnorderedSourceOp, part-2-UnorderedSourceOp, part-3-UnorderedSourceOp
ConcatSourceOp on pool[6]: part-1-UnorderedSourceOp, part-2-UnorderedSourceOp, part-3-UnorderedSourceOp
ConcatSourceOp on pool[7]: part-1-UnorderedSourceOp, part-2-UnorderedSourceOp, part-3-UnorderedSourceOp

After this PR, ConcatBuilderPool will generate 8 ConcatSourceOp by

ConcatSourceOp on pool[0]: part-1-UnorderedSourceOp, part-3-UnorderedSourceOp
ConcatSourceOp on pool[1]: part-1-UnorderedSourceOp, part-3-UnorderedSourceOp
ConcatSourceOp on pool[2]: part-1-UnorderedSourceOp, part-3-UnorderedSourceOp
ConcatSourceOp on pool[3]: part-1-UnorderedSourceOp, part-3-UnorderedSourceOp
ConcatSourceOp on pool[4]: part-2-UnorderedSourceOp
ConcatSourceOp on pool[5]: part-2-UnorderedSourceOp
ConcatSourceOp on pool[6]: part-2-UnorderedSourceOp
ConcatSourceOp on pool[7]: part-2-UnorderedSourceOp

So after this PR, compute layer read part-1 and part-2 in parallel

Manual test

manual test of small partition table scan performance as described in #10487 (comment)

-- master
-- we can observe performance regression as the number of partition increased
-- and scanning the same number of rows on partition table is slower than non-partition table
TiDB root@10.2.12.81:test> select "p0-0",count(*) from reports_part partition(p0);
                        -> select "p0-1",count(*) from reports_part partition(p0,p1);
                        -> select "p0-2",count(*) from reports_part partition(p0,p1,p2);
                        -> select "p0-3",count(*) from reports_part partition(p0,p1,p2,p3);
                        -> select "p0-4",count(*) from reports_part partition(p0,p1,p2,p3,p4);
                        -> select "p0-5",count(*) from reports_part partition(p0,p1,p2,p3,p4,p5);
                        -> select "p0-6",count(*) from reports_part partition(p0,p1,p2,p3,p4,p5,p6);
                        -> select "p0-7",count(*) from reports_part partition(p0,p1,p2,p3,p4,p5,p6,p7);
                        -> select "p0-8",count(*) from reports_part partition(p0,p1,p2,p3,p4,p5,p6,p7,p8);
                        -> select "p0-9",count(*) from reports_part partition(p0,p1,p2,p3,p4,p5,p6,p7,p8,p9);
                        -> select "p0-10",count(*) from reports_part partition(p0,p1,p2,p3,p4,p5,p6,p7,p8,p9,p10);
                        -> select "p0-11",count(*) from reports_part partition(p0,p1,p2,p3,p4,p5,p6,p7,p8,p9,p10,p11);
                        -> select "p0-12",count(*) from reports_part partition(p0,p1,p2,p3,p4,p5,p6,p7,p8,p9,p10,p11,p12);
                        -> select "non-part",count(*) from reports;
+------+----------+
| p0-0 | count(*) |
+------+----------+
| p0-0 | 0        |
+------+----------+
1 row in set
Time: 0.025s
+------+----------+
| p0-1 | count(*) |
+------+----------+
| p0-1 | 524288   |
+------+----------+
1 row in set
Time: 0.014s
+------+----------+
| p0-2 | count(*) |
+------+----------+
| p0-2 | 1048576  |
+------+----------+
1 row in set
Time: 0.017s
+------+----------+
| p0-3 | count(*) |
+------+----------+
| p0-3 | 1572864  |
+------+----------+
1 row in set
Time: 0.019s
+------+----------+
| p0-4 | count(*) |
+------+----------+
| p0-4 | 2097152  |
+------+----------+
1 row in set
Time: 0.023s
+------+----------+
| p0-5 | count(*) |
+------+----------+
| p0-5 | 2621440  |
+------+----------+
1 row in set
Time: 0.023s
+------+----------+
| p0-6 | count(*) |
+------+----------+
| p0-6 | 3145728  |
+------+----------+
1 row in set
Time: 0.024s
+------+----------+
| p0-7 | count(*) |
+------+----------+
| p0-7 | 3670016  |
+------+----------+
1 row in set
Time: 0.027s
+------+----------+
| p0-8 | count(*) |
+------+----------+
| p0-8 | 4194304  |
+------+----------+
1 row in set
Time: 0.029s
+------+----------+
| p0-9 | count(*) |
+------+----------+
| p0-9 | 4718592  |
+------+----------+
1 row in set
Time: 0.031s
+-------+----------+
| p0-10 | count(*) |
+-------+----------+
| p0-10 | 5242880  |
+-------+----------+
1 row in set
Time: 0.036s
+-------+----------+
| p0-11 | count(*) |
+-------+----------+
| p0-11 | 5767168  |
+-------+----------+
1 row in set
Time: 0.036s
+-------+----------+
| p0-12 | count(*) |
+-------+----------+
| p0-12 | 6291456  |
+-------+----------+
1 row in set
Time: 0.040s
+----------+----------+
| non-part | count(*) |
+----------+----------+
| non-part | 6291456  |
+----------+----------+
1 row in set
Time: 0.017s

-- after the fix
-- there is no performance regression as the number of partition increased
TiDB root@10.2.12.81:test> select "p0-0",count(*) from reports_part partition(p0);
                        -> select "p0-1",count(*) from reports_part partition(p0,p1);
                        -> select "p0-2",count(*) from reports_part partition(p0,p1,p2);
                        -> select "p0-3",count(*) from reports_part partition(p0,p1,p2,p3);
                        -> select "p0-4",count(*) from reports_part partition(p0,p1,p2,p3,p4);
                        -> select "p0-5",count(*) from reports_part partition(p0,p1,p2,p3,p4,p5);
                        -> select "p0-6",count(*) from reports_part partition(p0,p1,p2,p3,p4,p5,p6);
                        -> select "p0-7",count(*) from reports_part partition(p0,p1,p2,p3,p4,p5,p6,p7);
                        -> select "p0-8",count(*) from reports_part partition(p0,p1,p2,p3,p4,p5,p6,p7,p8);
                        -> select "p0-9",count(*) from reports_part partition(p0,p1,p2,p3,p4,p5,p6,p7,p8,p9);
                        -> select "p0-10",count(*) from reports_part partition(p0,p1,p2,p3,p4,p5,p6,p7,p8,p9,p10);
                        -> select "p0-11",count(*) from reports_part partition(p0,p1,p2,p3,p4,p5,p6,p7,p8,p9,p10,p11);
                        -> select "p0-12",count(*) from reports_part partition(p0,p1,p2,p3,p4,p5,p6,p7,p8,p9,p10,p11,p12);
                        -> select "non-part",count(*) from reports;
+------+----------+
| p0-0 | count(*) |
+------+----------+
| p0-0 | 0        |
+------+----------+
1 row in set
Time: 0.018s
+------+----------+
| p0-1 | count(*) |
+------+----------+
| p0-1 | 524288   |
+------+----------+
1 row in set
Time: 0.011s
+------+----------+
| p0-2 | count(*) |
+------+----------+
| p0-2 | 1048576  |
+------+----------+
1 row in set
Time: 0.013s
+------+----------+
| p0-3 | count(*) |
+------+----------+
| p0-3 | 1572864  |
+------+----------+
1 row in set
Time: 0.012s
+------+----------+
| p0-4 | count(*) |
+------+----------+
| p0-4 | 2097152  |
+------+----------+
1 row in set
Time: 0.012s
+------+----------+
| p0-5 | count(*) |
+------+----------+
| p0-5 | 2621440  |
+------+----------+
1 row in set
Time: 0.013s
+------+----------+
| p0-6 | count(*) |
+------+----------+
| p0-6 | 3145728  |
+------+----------+
1 row in set
Time: 0.013s
+------+----------+
| p0-7 | count(*) |
+------+----------+
| p0-7 | 3670016  |
+------+----------+
1 row in set
Time: 0.014s
+------+----------+
| p0-8 | count(*) |
+------+----------+
| p0-8 | 4194304  |
+------+----------+
1 row in set
Time: 0.015s
+------+----------+
| p0-9 | count(*) |
+------+----------+
| p0-9 | 4718592  |
+------+----------+
1 row in set
Time: 0.014s
+-------+----------+
| p0-10 | count(*) |
+-------+----------+
| p0-10 | 5242880  |
+-------+----------+
1 row in set
Time: 0.017s
+-------+----------+
| p0-11 | count(*) |
+-------+----------+
| p0-11 | 5767168  |
+-------+----------+
1 row in set
Time: 0.014s
+-------+----------+
| p0-12 | count(*) |
+-------+----------+
| p0-12 | 6291456  |
+-------+----------+
1 row in set
Time: 0.017s
+----------+----------+
| non-part | count(*) |
+----------+----------+
| non-part | 6291456  |
+----------+----------+
1 row in set
Time: 0.016s

Check List

Tests

Unit test
Integration test
Manual test (add detailed scripts or steps below)
See the manual test describe above
No code

Side effects

Performance regression: Consumes more CPU
Performance regression: Consumes more Memory
Breaking backward compatibility

Documentation

Release note

Fix the bug that table scan performance on small partition table is not optimal

…urrency (pingcap#10489) close pingcap#10487 * Limit the number of source ops by num of segment task * 4 in function `DeltaMergeStore::read`. In order to reduce concurrency overhead and let PartitionTableScan with small partitions that only contains 1~2 segments can schedule more segment read tasks in parallel. - For large partition, the storage layer still generate `num_streams` * `UnorderedSourceOp`, the behavior is the same as before. - For small partition, the storage layer only generate segment task * 4 * `UnorderedSourceOp`. And `ConcatBuilderPool` reorg the source ops and read the data from multiple partitions in parallel * Introduce `DMReadOptions` and reduce the complexity of adding has_multiple_partitions from compute layer to storage layer. * Add active_segment_limit, peak_active_segments, block_slot_limit, peak_blocks_in_queue when `SegmentReadTaskPool` finished Signed-off-by: JaySon-Huang <tshent@qq.com>

ti-chi-bot · 2025-10-23T06:40:09Z

[APPROVALNOTIFIER] This PR is APPROVED

This pull-request has been approved by: JaySon-Huang, Lloyd-Pottiger

The full list of commands accepted by this bot can be found here.

The pull request process is described here

Details

Needs approval from an approver in each of these files:

~~OWNERS~~ [JaySon-Huang,Lloyd-Pottiger]

Approvers can indicate their approval by writing /approve in a comment
Approvers can cancel approval by writing /approve cancel in a comment

ti-chi-bot · 2025-10-23T06:40:13Z

[LGTM Timeline notifier]

Timeline:

2025-10-23 06:25:42.529969931 +0000 UTC m=+939448.607222491: ☑️ agreed by JaySon-Huang.
2025-10-23 06:40:12.2277413 +0000 UTC m=+940318.304993860: ☑️ agreed by Lloyd-Pottiger.

JaySon-Huang · 2025-10-23T07:49:56Z

/test pull-unit-next-gen

ti-chi-bot added release-note Denotes a PR that will be considered when it comes time to generate release notes. size/XXL Denotes a PR that changes 1000+ lines, ignoring generated files. type/cherry-pick-for-release-nextgen-20251011 labels Oct 23, 2025

ti-chi-bot mentioned this pull request Oct 23, 2025

Storage: Improve small partition table read performance by limit concurrency #10489

Merged

12 tasks

ti-chi-bot assigned JaySon-Huang Oct 23, 2025

JaySon-Huang requested review from CalvinNeo, JinheLin and Lloyd-Pottiger October 23, 2025 06:25

JaySon-Huang approved these changes Oct 23, 2025

View reviewed changes

ti-chi-bot bot added needs-1-more-lgtm Indicates a PR needs 1 more LGTM. approved labels Oct 23, 2025

Lloyd-Pottiger approved these changes Oct 23, 2025

View reviewed changes

ti-chi-bot bot added lgtm and removed needs-1-more-lgtm Indicates a PR needs 1 more LGTM. labels Oct 23, 2025

ti-chi-bot bot merged commit 8ac2a96 into pingcap:release-nextgen-20251011 Oct 23, 2025
4 checks passed

ti-chi-bot bot deleted the cherry-pick-10489-to-release-nextgen-20251011 branch October 23, 2025 08:19

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Storage: Improve small partition table read performance by limit concurrency (#10489)#10500

Storage: Improve small partition table read performance by limit concurrency (#10489)#10500
ti-chi-bot[bot] merged 1 commit intopingcap:release-nextgen-20251011from
ti-chi-bot:cherry-pick-10489-to-release-nextgen-20251011

ti-chi-bot commented Oct 23, 2025

Uh oh!

ti-chi-bot bot commented Oct 23, 2025

Uh oh!

ti-chi-bot bot commented Oct 23, 2025

Uh oh!

JaySon-Huang commented Oct 23, 2025

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

	size_t final_num_stream
	= enable_read_thread ? std::max(1, num_streams) : std::max(1, std::min(num_streams, tasks.size()));
	auto read_mode = getReadMode(db_context, is_fast_scan, keep_order, executor);
	const auto & final_columns_to_read
	= executor && executor->extra_cast ? *executor->columns_after_cast : columns_to_read;
	auto read_task_pool = std::make_shared<SegmentReadTaskPool>(
	extra_table_id_index,
	final_columns_to_read,
	executor,
	start_ts,
	expected_block_size,
	read_mode,
	std::move(tasks),
	after_segment_read,
	log_tracing_id,
	enable_read_thread,
	final_num_stream,
	dm_context->scan_context->keyspace_id,
	dm_context->scan_context->resource_group_name);
	dm_context->scan_context->read_mode = read_mode;

	if (enable_read_thread)
	{
	for (size_t i = 0; i < final_num_stream; ++i)
	{
	group_builder.addConcurrency(std::make_unique<UnorderedSourceOp>(
	exec_context,
	read_task_pool,
	final_columns_to_read,
	extra_table_id_index,
	log_tracing_id,
	runtime_filter_list,
	rf_max_wait_time_ms));
	}

	void add(PipelineExecGroupBuilder & group_builder)
	{
	RUNTIME_CHECK(group_builder.groupCnt() == 1);
	for (size_t i = 0; i < group_builder.concurrency(); ++i)
	{
	pool[pre_index++].push_back(std::move(group_builder.getCurBuilder(i)));
	if (pre_index == pool.size())
	pre_index = 0;
	}
	}

Conversation

ti-chi-bot commented Oct 23, 2025

What problem does this PR solve?

What is changed and how it works?

Manual test

Check List

Release note

Uh oh!

ti-chi-bot bot commented Oct 23, 2025

Uh oh!

ti-chi-bot bot commented Oct 23, 2025

[LGTM Timeline notifier]

Uh oh!

JaySon-Huang commented Oct 23, 2025

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants