Skip to content

armgit / solr

Public Imported from GitHub
0

Apache Solr open-source search software

Document cuvs module breakage in Solr 10.1 (#5182)

Rahul Goswami1bd538700363
diff --git a/solr/solr-ref-guide/modules/query-guide/pages/dense-vector-search.adoc b/solr/solr-ref-guide/modules/query-guide/pages/dense-vector-search.adoc
index 385ba4b3e76..8d742a68b47 100644
--- a/solr/solr-ref-guide/modules/query-guide/pages/dense-vector-search.adoc
+++ b/solr/solr-ref-guide/modules/query-guide/pages/dense-vector-search.adoc
@@ -1042,6 +1042,12 @@ Details about using the ReRank Query Parser can be found in the xref:query-guide
[NOTE]
This is feature is currently experimental.
+[WARNING]
+====
+Vector indexing with GPU acceleration does not work in Solr 10.1 (see SOLR-18556), and is expected to be fixed in a subsequent Solr 10.x release.
+See xref:upgrade-notes:major-changes-in-solr-10.adoc#cuvs-gpu-acceleration-broken[Major Changes in Solr 10] for details.
+====
+
Building HNSW graphs, esp. with high dimensions and cardinality, is usually slow. If you have a NVIDIA GPU, then building HNSW based indexes can be sped up manifold. This is powered by the https://github.com/rapidsai/cuvs-lucene[cuVS-Lucene] library, a pluggable vectors format for Apache Lucene. It uses the state of the art https://arxiv.org/abs/2308.15136[CAGRA algorithm] for quickly building a fixed degree connected graph, which is then serialized into a HNSW graph. https://developer.nvidia.com/cuda-downloads[CUDA 13.0+] and https://jdk.java.net/archive/[JDK 22] are required to use this feature.
You can know more about Nvidia's cuVS library here: https://developer.nvidia.com/cuvs
diff --git a/solr/solr-ref-guide/modules/upgrade-notes/pages/major-changes-in-solr-10.adoc b/solr/solr-ref-guide/modules/upgrade-notes/pages/major-changes-in-solr-10.adoc
index 2221fc7020b..37a82e00720 100644
--- a/solr/solr-ref-guide/modules/upgrade-notes/pages/major-changes-in-solr-10.adoc
+++ b/solr/solr-ref-guide/modules/upgrade-notes/pages/major-changes-in-solr-10.adoc
@@ -92,6 +92,18 @@ Streaming expressions and the cross-collection join query parser now restrict wh
If you use these features to query another cluster, add its connection details to the corresponding setting before upgrading.
+[#cuvs-gpu-acceleration-broken]
+=== GPU Acceleration (cuVS) Does Not Work in Solr 10.1
+
+[WARNING]
+====
+Vector indexing with the xref:query-guide:dense-vector-search.adoc#gpu-acceleration[GPU Acceleration] (cuVS) module (`CuVSCodecFactory` with `knnAlgorithm="cagra_hnsw"`) does not work in Solr 10.1 (refer SOLR-18556 for updates).
+Solr 10.1 upgrades to Lucene 10.4, and a cuVS-lucene (library which powers the GPU accelertion) library version compatible with Lucene 10.4 is not available as of Solr 10.1 release.
+
+If you use GPU acceleration for vector indexing, skip Solr 10.1 and stay on Solr 10.0.x.
+This is expected to be fixed in a subsequent Solr 10.x release.
+====
+
=== v2 API
Starting in Solr 10.1 it is no longer possible for users to disable the v2 API by use of the `solr.api.v2.enabled` system property, and the Solr server and tooling (`bin/solr`, Admin UI, etc.) will start using these APIs internally.