Skip to content

Conversation

@iverase
Copy link
Contributor

@iverase iverase commented Nov 6, 2025

This approach takes a more aggressive prefetching strategy by prefetching the next bulk before executing the current one.

Testing shown an improvement:

before:

index_name       index_type  visit_percentage(%)  latency(ms)  net_cpu_time(ms)  avg_cpu_count    QPS  recall    visited  filter_selectivity
---------------  ----------  -------------------  -----------  ----------------  -------------  -----  ------  ---------  ------------------  
wiki1024en.docs         ivf                 0.25        14.74              0.00           0.00  67.84    0.73    5828.14                1.00
wiki1024en.docs         ivf                 0.50        16.03              0.00           0.00  62.38    0.83   10854.43                1.00
wiki1024en.docs         ivf                 0.75        14.74              0.00           0.00  67.84    0.88   15880.75                1.00
wiki1024en.docs         ivf                 1.00        14.84              0.00           0.00  67.39    0.91   20883.03                1.00
wiki1024en.docs         ivf                 1.50        15.05              0.00           0.00  66.45    0.94   30841.79                1.00
wiki1024en.docs         ivf                 2.00        15.21              0.00           0.00  65.75    0.95   40852.71                1.00
wiki1024en.docs         ivf                 2.50        15.41              0.00           0.00  64.89    0.96   50842.37                1.00
wiki1024en.docs         ivf                 0.00        19.39              0.00           0.00  51.57    0.99  180853.42                1.00

after:

index_name       index_type  visit_percentage(%)  latency(ms)  net_cpu_time(ms)  avg_cpu_count    QPS  recall    visited  filter_selectivity
---------------  ----------  -------------------  -----------  ----------------  -------------  -----  ------  ---------  ------------------  
wiki1024en.docs         ivf                 0.25        11.74              0.00           0.00  85.18    0.73    5828.14                1.00
wiki1024en.docs         ivf                 0.50        11.82              0.00           0.00  84.60    0.83   10854.43                1.00
wiki1024en.docs         ivf                 0.75        11.51              0.00           0.00  86.88    0.88   15880.75                1.00
wiki1024en.docs         ivf                 1.00        11.13              0.00           0.00  89.85    0.91   20883.03                1.00
wiki1024en.docs         ivf                 1.50        11.65              0.00           0.00  85.84    0.94   30841.79                1.00
wiki1024en.docs         ivf                 2.00        11.74              0.00           0.00  85.18    0.95   40852.71                1.00
wiki1024en.docs         ivf                 2.50        12.33              0.00           0.00  81.10    0.96   50842.37                1.00
wiki1024en.docs         ivf                 0.00        15.98              0.00           0.00  62.58    0.99  180853.42                1.00

@iverase iverase requested a review from benwtrent November 6, 2025 11:28
@elasticsearchmachine elasticsearchmachine added the Team:Search Relevance Meta label for the Search Relevance team in Elasticsearch label Nov 6, 2025
@elasticsearchmachine
Copy link
Collaborator

Pinging @elastic/es-search-relevance (Team:Search Relevance)

@iverase iverase merged commit 2e8cbc9 into elastic:main Nov 6, 2025
34 checks passed
@iverase iverase deleted the prefetch-bulk branch November 6, 2025 16:15
Kubik42 pushed a commit to Kubik42/elasticsearch that referenced this pull request Nov 10, 2025
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

>non-issue :Search Relevance/Vectors Vector search Team:Search Relevance Meta label for the Search Relevance team in Elasticsearch v9.3.0

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants