[improve](ann index)Accumulate multiple small batches before training#57623
Merged
airborne12 merged 5 commits intoapache:masterfrom Nov 11, 2025
Merged
[improve](ann index)Accumulate multiple small batches before training#57623airborne12 merged 5 commits intoapache:masterfrom
airborne12 merged 5 commits intoapache:masterfrom
Conversation
uchenily
force-pushed
the
ann-optimize
branch
from
de23f84 to
88c3eca
Compare
Contributor
Author
|
run buildall |
ClickBench: Total hot run time: 27.76 s |
Contributor
BE UT Coverage ReportIncrement line coverage Increment coverage report
|
Contributor
Author
|
run buildall |
ClickBench: Total hot run time: 29.09 s |
Contributor
BE UT Coverage ReportIncrement line coverage Increment coverage report
|
Contributor
BE Regression && UT Coverage ReportIncrement line coverage Increment coverage report
|
| // VectorIndex should be weak shared by AnnIndexWriter and VectorIndexReader | ||
| // This should be a weak_ptr | ||
| std::shared_ptr<VectorIndex> _vector_index; | ||
| std::vector<float> _ann_vec; |
Contributor
There was a problem hiding this comment.
replace std::vector with DorisVector for memory safe.
|
|
||
| if (i > 0) { | ||
| vectorized::Int64 offset = i * dim; | ||
| std::copy(_ann_vec.begin() + offset, _ann_vec.end(), _ann_vec.begin()); |
Contributor
There was a problem hiding this comment.
cost of memory copy can be optimized by using std::list<std::shared_ptr<DorisVector>>
uchenily
force-pushed
the
ann-optimize
branch
from
d81e6ab to
bc19e89
Compare
Contributor
Author
|
run buildall |
TPC-H: Total hot run time: 34427 ms |
TPC-DS: Total hot run time: 187644 ms |
ClickBench: Total hot run time: 27.73 s |
BE UT Coverage ReportIncrement line coverage Increment coverage report
|
Contributor
BE Regression && UT Coverage ReportIncrement line coverage Increment coverage report
|
|
|
||
| size_t block_size = CHUNK_SIZE * build_parameter.dim; | ||
| // The array capacity will not change after resizing | ||
| _float_array.resize(block_size); |
Contributor
There was a problem hiding this comment.
reserve instead of resize
| size_t block_size = CHUNK_SIZE * build_parameter.dim; | ||
| // The array capacity will not change after resizing | ||
| _float_array.resize(block_size); | ||
| _array_offset = 0; |
Contributor
There was a problem hiding this comment.
_array_offset is not needed
zhiqiang-hhhh
approved these changes
Nov 10, 2025
Contributor
|
PR approved by anyone and no changes requested. |
github-actions
bot
added
the
approved
Contributor
|
PR approved by at least one committer and no changes requested. |
Contributor
Author
|
run buildall |
github-actions
bot
removed
the
approved
TPC-H: Total hot run time: 36144 ms |
TPC-DS: Total hot run time: 187633 ms |
ClickBench: Total hot run time: 27.84 s |
BE UT Coverage ReportIncrement line coverage Increment coverage report
|
Contributor
Author
|
run cloud_p0 |
Contributor
Author
|
run external |
Contributor
BE Regression && UT Coverage ReportIncrement line coverage Increment coverage report
|
github-actions
bot
added
the
approved
Contributor
|
PR approved by at least one committer and no changes requested. |
zhiqiang-hhhh
pushed a commit
to zhiqiang-hhhh/doris
that referenced
this pull request
Nov 12, 2025
…apache#57623) Accumulate multiple small batches to avoid the following error when training: `Error: 'nx >= k' failed: Number of training points should be at least as large as number of clusters`, and significantly reduce the time for faiss train/add.
zhiqiang-hhhh
pushed a commit
to zhiqiang-hhhh/doris
that referenced
this pull request
Nov 12, 2025
…apache#57623) Accumulate multiple small batches to avoid the following error when training: `Error: 'nx >= k' failed: Number of training points should be at least as large as number of clusters`, and significantly reduce the time for faiss train/add.
wyxxxcat
pushed a commit
to wyxxxcat/doris
that referenced
this pull request
Nov 13, 2025
…apache#57623) ### What problem does this PR solve? Accumulate multiple small batches to avoid the following error when training: `Error: 'nx >= k' failed: Number of training points should be at least as large as number of clusters`, and significantly reduce the time for faiss train/add.
wyxxxcat
pushed a commit
to wyxxxcat/doris
that referenced
this pull request
Nov 18, 2025
…apache#57623) ### What problem does this PR solve? Accumulate multiple small batches to avoid the following error when training: `Error: 'nx >= k' failed: Number of training points should be at least as large as number of clusters`, and significantly reduce the time for faiss train/add.
16 tasks
airborne12
pushed a commit
that referenced
this pull request
Dec 4, 2025
### What problem does this PR solve? Previous pr: #57623 The current granularity for index training and data ingestion is set to 1M and is hard-coded, which makes index construction unnecessarily slow in some scenarios. This should be made configurable and reduced when appropriate. For example, when having 1M vectors to add, and batch size of stream load is set to 0.3M, this means we will have 3 stream load requests. If it happens to make one request that having 0.3M to have 1 threads for adding, whole process of load will be very slow. A typical cpu usage will be like this: <img width="1902" height="552" alt="image" src="https://github.com/user-attachments/assets/65728e56-f333-4bd5-a54a-8c12d01668f1" /> We need to make batch size configurable so that we can modify them when we need to do it. For example, when we set batch size to 30K, we can have a more higher avg cpu usage when we like this: <img width="1890" height="554" alt="image" src="https://github.com/user-attachments/assets/7d664b0e-b017-4a2e-bed8-e40f56ff97b7" /> **Default value is still 1M, small batch size will do a damage to the recall of the hnsw.**
github-actions bot
pushed a commit
that referenced
this pull request
Dec 4, 2025
### What problem does this PR solve? Previous pr: #57623 The current granularity for index training and data ingestion is set to 1M and is hard-coded, which makes index construction unnecessarily slow in some scenarios. This should be made configurable and reduced when appropriate. For example, when having 1M vectors to add, and batch size of stream load is set to 0.3M, this means we will have 3 stream load requests. If it happens to make one request that having 0.3M to have 1 threads for adding, whole process of load will be very slow. A typical cpu usage will be like this: <img width="1902" height="552" alt="image" src="https://github.com/user-attachments/assets/65728e56-f333-4bd5-a54a-8c12d01668f1" /> We need to make batch size configurable so that we can modify them when we need to do it. For example, when we set batch size to 30K, we can have a more higher avg cpu usage when we like this: <img width="1890" height="554" alt="image" src="https://github.com/user-attachments/assets/7d664b0e-b017-4a2e-bed8-e40f56ff97b7" /> **Default value is still 1M, small batch size will do a damage to the recall of the hnsw.**
nagisa-kunhah
pushed a commit
to nagisa-kunhah/doris
that referenced
this pull request
Dec 14, 2025
…58645) ### What problem does this PR solve? Previous pr: apache#57623 The current granularity for index training and data ingestion is set to 1M and is hard-coded, which makes index construction unnecessarily slow in some scenarios. This should be made configurable and reduced when appropriate. For example, when having 1M vectors to add, and batch size of stream load is set to 0.3M, this means we will have 3 stream load requests. If it happens to make one request that having 0.3M to have 1 threads for adding, whole process of load will be very slow. A typical cpu usage will be like this: <img width="1902" height="552" alt="image" src="https://github.com/user-attachments/assets/65728e56-f333-4bd5-a54a-8c12d01668f1" /> We need to make batch size configurable so that we can modify them when we need to do it. For example, when we set batch size to 30K, we can have a more higher avg cpu usage when we like this: <img width="1890" height="554" alt="image" src="https://github.com/user-attachments/assets/7d664b0e-b017-4a2e-bed8-e40f56ff97b7" /> **Default value is still 1M, small batch size will do a damage to the recall of the hnsw.**
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?
Accumulate multiple small batches to avoid the following error when training:
Error: 'nx >= k' failed: Number of training points should be at least as large as number of clusters,and significantly reduce the time for faiss train/add.
Problem Summary:
Release note
None
Check List (For Author)
Test
Behavior changed:
Does this need documentation?
Check List (For Reviewer who merge this PR)