[BLOG] Concurrent vector graph construction in OpenSearch (via jVector) #3864
Description
Describe the blog post
How we created a unique capability in OpenSearch Vector Indexing Engine to leverage the full power of concurrency for graph index construction.
With the introduction of the latest version of the jVector plugin to OpenSearch we introduced concurrent graph index construction.
It leverages a lock free non-blocking mechanism for nearly perfect linear scalability when constructing vector graph index. Up until now, the concurrency of graph construction was inherently limited and capped by the number of concurrent Lucene segments created within each shard. However, this implies that we are paying a significant penalty by being forced to create multiple small batches with the construction of each one capped by a single thread. This not only makes for a slower ingestion, but makes us use smaller batches that later force more costly merges.
In this blog post we are going to describe the concurrent lock free architecture of jVector and how we leverage it in OpenSearch to accelerate vector ingestion in the KNN engine.
We will try to answer important questions, such as:
- How concurrent graph building is even possible?
- How is it done in jVector?
- Are there any implications on graph accuracy?
- What additional resource management are important when dealing with concurrent construction?
Expected Title
Concurrent vector graph construction in OpenSearch (via jVector)
Authors Name
Samuel Herman
Authors Email
Target Draft Date
07/18/2025
Blog Post Category
technical
Target Publication Date
No response
Additional Info
No response