Update pgvector loading method to use binary format #488
Conversation
This file contains 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
Particularly on large vector types, the pgvector module was spending significant time on converting floating point values to ASCII before being transmitted to the PostgreSQL server. This changes keeps the format in binary, reducing overhead. One test demonstrated a 63% reduction in load time, which would have an impact on the overall "build" time as reported by this benchmark.
|
Noting that of the automated tests, the |
|
Thanks! |
jkatz
added a commit
to jkatz/VectorDBBench
that referenced
this pull request
Jun 14, 2024
PostgreSQL supports two methods of passing data from client to server: text and binary. While for many data types the difference may not be noticeable, we can see significant performance impact when converting a vector from binary => text => binary representation. See previous explanation here[1]. While the pgvector loading code accounts for this, the query code did not. This is due to the use of a list[float] type, which the pgvector-python adapter currently doesn't support. However, this adapter does support direct binary transfer if the data is represent as a Numpy array[2]. Testing shows that moving to a direct binary representation does have a significant impact on query results - my tests are showing a 3x impact -- and provides a more accurate representation for how this workload would execute. [1] erikbern/ann-benchmarks#488 [2] https://github.com/pgvector/pgvector-python?tab=readme-ov-file#psycopg-3
alwayslove2013
pushed a commit
to zilliztech/VectorDBBench
that referenced
this pull request
Jun 28, 2024
PostgreSQL supports two methods of passing data from client to server: text and binary. While for many data types the difference may not be noticeable, we can see significant performance impact when converting a vector from binary => text => binary representation. See previous explanation here[1]. While the pgvector loading code accounts for this, the query code did not. This is due to the use of a list[float] type, which the pgvector-python adapter currently doesn't support. However, this adapter does support direct binary transfer if the data is represent as a Numpy array[2]. Testing shows that moving to a direct binary representation does have a significant impact on query results - my tests are showing a 3x impact -- and provides a more accurate representation for how this workload would execute. [1] erikbern/ann-benchmarks#488 [2] https://github.com/pgvector/pgvector-python?tab=readme-ov-file#psycopg-3
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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.
Particularly on large vector types, the pgvector module was spending significant time on converting floating point values to ASCII before being transmitted to the PostgreSQL server. This changes keeps the format in binary, reducing overhead. One test demonstrated a 63% reduction in load time, which would have an impact on the overall "build" time as reported by this benchmark.