From 25c2ad15e596714df1f5ac6756aadfdc6adf1051 Mon Sep 17 00:00:00 2001
From: Sanket Kedia <sanket@trychroma.com>
Date: Tue, 22 Apr 2025 13:05:53 -0700
Subject: [PATCH 1/5] [ENH] Turn on spann by default
---
rust/frontend/sample_configs/distributed.yaml | 2 +-
rust/frontend/sample_configs/tilt_config.yaml | 2 +-
2 files changed, 2 insertions(+), 2 deletions(-)
diff --git a/rust/frontend/sample_configs/distributed.yaml b/rust/frontend/sample_configs/distributed.yaml
index ee42bffe86f..c1a3826ed39 100644
--- a/rust/frontend/sample_configs/distributed.yaml
+++ b/rust/frontend/sample_configs/distributed.yaml
@@ -46,4 +46,4 @@ scorecard:
- "collection_id:*"
score: 100
enable_span_indexing: true
-default_knn_index: "hnsw"
+default_knn_index: "spann"
diff --git a/rust/frontend/sample_configs/tilt_config.yaml b/rust/frontend/sample_configs/tilt_config.yaml
index 520bfd87bf9..9530ecfd521 100644
--- a/rust/frontend/sample_configs/tilt_config.yaml
+++ b/rust/frontend/sample_configs/tilt_config.yaml
@@ -54,4 +54,4 @@ scorecard:
circuit_breaker:
requests: 1000
enable_span_indexing: true
-default_knn_index: "hnsw"
+default_knn_index: "spann"
From 914d366278929717f0c1f9223a9791acbba7f9ae Mon Sep 17 00:00:00 2001
From: Sanket Kedia <sanket@trychroma.com>
Date: Thu, 24 Apr 2025 20:31:36 -0700
Subject: [PATCH 2/5] Take lock before doing hnsw.open()
---
rust/index/src/spann/types.rs | 5 +++++
1 file changed, 5 insertions(+)
diff --git a/rust/index/src/spann/types.rs b/rust/index/src/spann/types.rs
index 3b9f753fd91..9349e1f56f2 100644
--- a/rust/index/src/spann/types.rs
+++ b/rust/index/src/spann/types.rs
@@ -2022,6 +2022,11 @@ impl<'me> SpannIndexReader<'me> {
dimensionality: usize,
ef_search: usize,
) -> Result<HnswIndexRef, SpannIndexReaderError> {
+ // We take a lock here to synchronize concurrent open of the same index.
+ // Otherwise, we could end up with a corrupted index since the filesystem
+ // operations are not guaranteed to be atomic.
+ // The lock is a partitioned mutex to allow for higher concurrency across collections.
+ let _guard = hnsw_provider.write_mutex.lock(id).await;
match hnsw_provider.get(id, cache_key).await {
Some(index) => Ok(index),
None => {
From 49ad1f5b90e66e59665b742a0385609b5f78fd84 Mon Sep 17 00:00:00 2001
From: Sanket Kedia <sanket@trychroma.com>
Date: Thu, 24 Apr 2025 20:34:49 -0700
Subject: [PATCH 3/5] fix rust test
---
rust/frontend/src/impls/service_based_frontend.rs | 10 +++++++---
1 file changed, 7 insertions(+), 3 deletions(-)
diff --git a/rust/frontend/src/impls/service_based_frontend.rs b/rust/frontend/src/impls/service_based_frontend.rs
index 4d79e258b20..5c269164fa9 100644
--- a/rust/frontend/src/impls/service_based_frontend.rs
+++ b/rust/frontend/src/impls/service_based_frontend.rs
@@ -1438,9 +1438,13 @@ mod tests {
assert!(segments.iter().any(
|s| s.r#type == SegmentType::BlockfileMetadata && s.scope == SegmentScope::METADATA
));
- assert!(segments
- .iter()
- .any(|s| s.r#type == SegmentType::HnswDistributed && s.scope == SegmentScope::VECTOR));
+ assert!(
+ segments.iter().any(
+ |s| s.r#type == SegmentType::HnswDistributed && s.scope == SegmentScope::VECTOR
+ ) || segments
+ .iter()
+ .any(|s| s.r#type == SegmentType::Spann && s.scope == SegmentScope::VECTOR)
+ );
assert!(segments
.iter()
.any(|s| s.r#type == SegmentType::BlockfileRecord && s.scope == SegmentScope::RECORD));
From d05268fa2113d7dfb389290804e031b64d906a1d Mon Sep 17 00:00:00 2001
From: Sanket Kedia <sanket@trychroma.com>
Date: Thu, 24 Apr 2025 22:42:15 -0700
Subject: [PATCH 4/5] fix test
---
chromadb/test/property/invariants.py | 15 +++++++++------
1 file changed, 9 insertions(+), 6 deletions(-)
diff --git a/chromadb/test/property/invariants.py b/chromadb/test/property/invariants.py
index c48c5abd832..761ad3707ea 100644
--- a/chromadb/test/property/invariants.py
+++ b/chromadb/test/property/invariants.py
@@ -246,16 +246,19 @@ def fd_not_exceeding_threadpool_size(threadpool_size: int) -> None:
)
def get_space(collection: Collection):
+ # TODO: this is a hack to get the space
+ # We should update the tests to not pass space via metadata instead use collection
+ # configuration_json
+ space = None
if "hnsw:space" in collection.metadata:
- return collection.metadata["hnsw:space"]
+ space = collection.metadata["hnsw:space"]
if collection._model.configuration_json is None:
- return None
+ return space
if 'spann' in collection._model.configuration_json and collection._model.configuration_json.get('spann') is not None and 'space' in collection._model.configuration_json.get('spann'):
- return collection._model.configuration_json.get('spann').get('space')
+ space = collection._model.configuration_json.get('spann').get('space')
elif 'hnsw' in collection._model.configuration_json and collection._model.configuration_json.get('hnsw') is not None and 'space' in collection._model.configuration_json.get('hnsw'):
- return collection._model.configuration_json.get('hnsw').get('space')
- else:
- return None
+ space = collection._model.configuration_json.get('hnsw').get('space')
+ return space
def ann_accuracy(
collection: Collection,
From 7d79e57409e732c72e49dee68220543af0985d13 Mon Sep 17 00:00:00 2001
From: Sanket Kedia <sanket@trychroma.com>
Date: Thu, 24 Apr 2025 23:11:23 -0700
Subject: [PATCH 5/5] Modify get_space
---
chromadb/test/property/invariants.py | 3 ++-
1 file changed, 2 insertions(+), 1 deletion(-)
diff --git a/chromadb/test/property/invariants.py b/chromadb/test/property/invariants.py
index 761ad3707ea..58b780c80d1 100644
--- a/chromadb/test/property/invariants.py
+++ b/chromadb/test/property/invariants.py
@@ -257,7 +257,8 @@ def get_space(collection: Collection):
if 'spann' in collection._model.configuration_json and collection._model.configuration_json.get('spann') is not None and 'space' in collection._model.configuration_json.get('spann'):
space = collection._model.configuration_json.get('spann').get('space')
elif 'hnsw' in collection._model.configuration_json and collection._model.configuration_json.get('hnsw') is not None and 'space' in collection._model.configuration_json.get('hnsw'):
- space = collection._model.configuration_json.get('hnsw').get('space')
+ if space is None:
+ space = collection._model.configuration_json.get('hnsw').get('space')
return space
def ann_accuracy(