Enable directIO and bfloat16 for bbq and unquantized HNSW vector field types #135940
Conversation
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
thecoop
force-pushed
the
hnsw-directio-bfloat16-enabled
branch
from
8119a8c to
803dfc1
Compare
thecoop
force-pushed
the
hnsw-directio-bfloat16-enabled
branch
from
4f81b03 to
05f0b71
Compare
thecoop
changed the title
Member
Author
|
I've added a feature flag to enable the new formats - so we can turn them on all at once when ready |
phananh1010
added a commit
to phananh1010/elasticsearch
that referenced
this pull request
Nov 6, 2025
BASE=bcb861ae10000f94977487193d242c6f30443741 HEAD=eed749e3a6d09e03e0731f5c526fa49e15cd471a Branch=main
benwtrent
reviewed
Nov 7, 2025
...amlRestTest/resources/rest-api-spec/test/search.vectors/200_dense_vector_docvalue_fields.yml
Outdated
Show resolved
Hide resolved
benwtrent
reviewed
Nov 7, 2025
server/src/main/java/org/elasticsearch/index/mapper/vectors/DenseVectorFieldMapper.java
Outdated
Show resolved
Hide resolved
phananh1010
added a commit
to phananh1010/elasticsearch
that referenced
this pull request
Nov 7, 2025
BASE=bcb861ae10000f94977487193d242c6f30443741 HEAD=ab5ee251aa0c298a2ee2d980c54eb93af40a064a Branch=main
benwtrent
approved these changes
Nov 7, 2025
Member
benwtrent
left a comment
benwtrent
left a comment
There was a problem hiding this comment.
I think this is good. I have reviewed it many times.
My concerns are:
- I think we should run yaml tests all the time, so let's enable the feature flag for the yaml test runners?
- Let's not parse on_disk_rescore for BBQHnsw (I think thats the only one that was added?) unless the flag is enabled. My concern is that we accept the parameter and then do nothing (which we shouldn't do)
benwtrent
reviewed
Nov 7, 2025
Comment on lines
-1734
to
+1738
| Object onDiskRescoreNode = indexOptionsMap.remove("on_disk_rescore"); | ||
| Object onDiskRescoreNode = ES93GenericFlatVectorsFormat.GENERIC_VECTOR_FORMAT.isEnabled() | ||
| ? indexOptionsMap.remove("on_disk_rescore") | ||
| : false; |
Member
There was a problem hiding this comment.
I am not sure we can do this? Do we know if we have shipped a version of ES that allowed this option for diskbbq?
Member
Author
There was a problem hiding this comment.
9.2 didn't include it, so we can change it here
john-wagster
approved these changes
Nov 7, 2025
Contributor
john-wagster
left a comment
john-wagster
left a comment
There was a problem hiding this comment.
this is a lot of changes but I tried to go through it bit at a time today and I think it looks good. Nothing stood out to me as bad or missing. lgtm
Kubik42
pushed a commit
to Kubik42/elasticsearch
that referenced
this pull request
Nov 10, 2025
…ypes (elastic#135940) Currently behind a feature flag, pending further development
thecoop
added a commit
to thecoop/elasticsearch
that referenced
this pull request
Nov 10, 2025
thecoop
added a commit
that referenced
this pull request
Nov 12, 2025
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
4 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.
Enables direct IO rescoring and bfloat16 elements for the
bbq_hnswandbbq_diskvector types, and bfloat16 elements forhnswandbbq_flatelement types, using a new generic flat vector format that is element-type-agnostic.This adds new options
on_disk_rescoring: trueandelement_type: bfloat16to index options, allowing direct IO to be used for rescoring, and bfloat16 to be used as a raw vector type.element_type: bfloat16reduces disk usage for raw vectors (by far the dominating cost of storing vectors on disk) by 50%, at the cost of a ~5% runtime performance reductionon_disk_rescore: truedoes not load raw vectors into disk when rescoring vectors, significantly improving HNSW performance when the vectors do not all fit in memory. Due to architectural limitations, this option can only be set on newly-indexed vectors, requiring a reindex to enable on-disk rescoring for existing vectors.Int and flat vector types will be updated in a separate PR.
Specifing bfloat16 and using the new formats is currently behind a feature flag, pending further PRs in this area