Storages: Optimize vector search in scenarios with updates#9597
Merged
ti-chi-bot[bot] merged 6 commits intopingcap:masterfrom Nov 13, 2024
Merged
Storages: Optimize vector search in scenarios with updates#9597ti-chi-bot[bot] merged 6 commits intopingcap:masterfrom
ti-chi-bot[bot] merged 6 commits intopingcap:masterfrom
Conversation
Contributor
|
Can you briefly describe what changes you've done in this PR? |
Contributor
Author
done |
Signed-off-by: Lloyd-Pottiger <yan1579196623@gmail.com> fix Signed-off-by: Lloyd-Pottiger <yan1579196623@gmail.com> fix Signed-off-by: Lloyd-Pottiger <yan1579196623@gmail.com>
f0ddea4 to
83c44a1
Compare
Comment on lines
+41
to
+42
| // Each ColomnFileTiny has its own vector index, rowid_start_offset is the offset of the ColmnFilePersistSet. | ||
| vec_index->get(rowid - rowid_start_offset, value); |
Member
|
Seem that what is changed is the order of these operations(the yellow and the green lines swapped their order). Could you explain more about why changing order can improve performance? I guess you can somehow reduce the vector index loaded from either ColumnFileTiny or Stable layer, but I am not so clear. |
Contributor
Author
Add an example in PR description. |
Signed-off-by: Lloyd-Pottiger <yan1579196623@gmail.com>
Contributor
Author
|
Flamegraph |
JinheLin
approved these changes
Nov 13, 2024
breezewish
approved these changes
Nov 13, 2024
Contributor
|
[APPROVALNOTIFIER] This PR is APPROVED This pull-request has been approved by: breezewish, JinheLin The full list of commands accepted by this bot can be found here. The pull request process is described here DetailsNeeds approval from an approver in each of these files:
Approvers can indicate their approval by writing |
Contributor
[LGTM Timeline notifier]Timeline:
|
Contributor
Author
|
/hold |
Contributor
Author
|
/unhold |
Contributor
Author
|
/cherry-pick release-8.5 |
Member
|
@Lloyd-Pottiger: new pull request created to branch DetailsIn response to this:
Instructions for interacting with me using PR comments are available here. If you have questions or suggestions related to my behavior, please file an issue against the ti-community-infra/tichi repository. |
12 tasks
Comment on lines
+3524
to
+3529
| auto * vector_index_stream = dynamic_cast<ConcatVectorIndexBlockInputStream *>(skippable_stream.get()); | ||
| auto stream = std::make_shared<BitmapFilterBlockInputStream>(columns_to_read, skippable_stream, bitmap_filter); | ||
| if (vector_index_stream) | ||
| { | ||
| // Squash blocks to reduce the number of blocks. | ||
| return std::make_shared<SquashingBlockInputStream>( |
Contributor
There was a problem hiding this comment.
Suggested change
| auto * vector_index_stream = dynamic_cast<ConcatVectorIndexBlockInputStream *>(skippable_stream.get()); | |
| auto stream = std::make_shared<BitmapFilterBlockInputStream>(columns_to_read, skippable_stream, bitmap_filter); | |
| if (vector_index_stream) | |
| { | |
| // Squash blocks to reduce the number of blocks. | |
| return std::make_shared<SquashingBlockInputStream>( | |
| auto stream = std::make_shared<BitmapFilterBlockInputStream>(columns_to_read, skippable_stream, bitmap_filter); | |
| if (auto * vector_index_stream = dynamic_cast<ConcatVectorIndexBlockInputStream *>(skippable_stream.get()); vector_index_stream) | |
| { | |
| // For vector search, there are more likely to return small blocks from different sub-streams. Squash blocks to reduce the number of blocks thus improve the performance of upper layer. | |
| return std::make_shared<SquashingBlockInputStream>( |
Comment on lines
+213
to
+222
| auto select_size = search_results.size() > topk ? topk : search_results.size(); | ||
| auto top_k_end = search_results.begin() + select_size; | ||
| std::nth_element(search_results.begin(), top_k_end, search_results.end(), [](const auto & lhs, const auto & rhs) { | ||
| return lhs.distance < rhs.distance; | ||
| }); | ||
| search_results.resize(select_size); | ||
| std::vector<UInt32> selected_rows; | ||
| selected_rows.reserve(search_results.size()); | ||
| for (const auto & row : search_results) | ||
| selected_rows.push_back(row.key); |
Contributor
There was a problem hiding this comment.
Suggested change
| auto select_size = search_results.size() > topk ? topk : search_results.size(); | |
| auto top_k_end = search_results.begin() + select_size; | |
| std::nth_element(search_results.begin(), top_k_end, search_results.end(), [](const auto & lhs, const auto & rhs) { | |
| return lhs.distance < rhs.distance; | |
| }); | |
| search_results.resize(select_size); | |
| std::vector<UInt32> selected_rows; | |
| selected_rows.reserve(search_results.size()); | |
| for (const auto & row : search_results) | |
| selected_rows.push_back(row.key); | |
| const auto select_size = std::min(search_results.size(), topk); | |
| auto top_k_end = search_results.begin() + select_size; | |
| std::nth_element(search_results.begin(), top_k_end, search_results.end(), [](const auto & lhs, const auto & rhs) { | |
| return lhs.distance < rhs.distance; | |
| }); | |
| std::vector<UInt32> selected_rows(select_size); | |
| for (size_t i = 0; i < select_size; ++i) | |
| selected_rows[i] = search_results[i].key); |
This could avoid re-allocate the vector size
Contributor
There was a problem hiding this comment.
Better add a check for select_size != 0. Because we can not ensure the length of index_stream->load() is always larger than 0 in this class
|
|
||
| UInt32 precedes_rows = 0; | ||
| std::vector<VectorIndexViewer::SearchResult> search_results; | ||
| for (size_t i = 0; i < stream->children.size(); ++i) |
Contributor
There was a problem hiding this comment.
Suggested change
| for (size_t i = 0; i < stream->children.size(); ++i) | |
| assert(stream->children.size() == index_streams.size()); // otherwise the `row.key` of the search result is not correct | |
| for (size_t i = 0; i < stream->children.size(); ++i) |
JaySon-Huang
added a commit
to JaySon-Huang/tiflash
that referenced
this pull request
Nov 14, 2024
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
6 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What problem does this PR solve?
Issue Number: ref #9600 , close #9599
Problem Summary:
What is changed and how it works?
Before, the read path of vector search is:
A simple example:
We have cf_0 (indexed, offset: 100-110), cf_1 (indexed, offset: 110-120), cf_3 (indexed, offset: 120-130), cf_4 (memtable, offset: 130-140), dmfile_0 (indexed, offset: 0-100, each 10 rows as a pack), to find top5:
So, in this example, we need to read 3 packs and 3 ColumnFileTinys.
This PR changes the read path to:
Still using the above example, now to find top5:
Now, we only need to read 2 packs and 2 ColumnFileTinys.
Benchmark
Run VectorDBBench, about 750k stable, 250k delta with vector index, 4k delta without vector index.
Note: should disable compact tiflash replica
Without this PR (after fix #9599)
With this PR
All data are in stable
Check List
Tests
Side effects
Documentation
Release note