[BUG] fix multi collection log purge

chromadb/db/mixins/embeddings_queue.py

@@ -130,7 +130,7 @@ def delete_log(self, collection_id: UUID) -> None:
 
     @trace_method("SqlEmbeddingsQueue.purge_log", OpenTelemetryGranularity.ALL)
     @override
-    def purge_log(self) -> None:
+    def purge_log(self, collection_id: UUID) -> None:
         segments_t = Table("segments")
         segment_ids_q = (
             self.querybuilder()
@@ -140,6 +140,9 @@ def purge_log(self) -> None:
             # - > 1 has not never written to the max_seq_id table
             # In that case, we should not delete any WAL entries as we can't be sure that the all segments are caught up.
             .select(functions.Coalesce(Table("max_seq_id").seq_id, -1))
+            .where(
+                segments_t.collection == ParameterValue(self.uuid_to_db(collection_id))
+            )
             .left_join(Table("max_seq_id"))
             .on(segments_t.id == Table("max_seq_id").segment_id)
         )
@@ -255,7 +258,7 @@ def submit_embeddings(
             self._notify_all(topic_name, embedding_records)
 
             if self.config.get_parameter("automatically_purge").value:
-                self.purge_log()
+                self.purge_log(collection_id)
 
             return seq_ids
 

chromadb/ingest/__init__.py

@@ -42,7 +42,7 @@ def delete_log(self, collection_id: UUID) -> None:
         pass
 
     @abstractmethod
-    def purge_log(self) -> None:
+    def purge_log(self, collection_id: UUID) -> None:
         """Truncates the log for the given collection, removing all seen records."""
         pass
 

chromadb/test/db/test_log_purge.py

@@ -0,0 +1,26 @@
+from chromadb.api.client import Client
+from chromadb.config import System
+from chromadb.test.property import invariants
+
+
+def test_log_purge(sqlite_persistent: System) -> None:
+    client = Client.from_system(sqlite_persistent)
+
+    first_collection = client.create_collection(
+        "first_collection", metadata={"hnsw:sync_threshold": 10, "hnsw:batch_size": 10}
+    )
+    second_collection = client.create_collection(
+        "second_collection", metadata={"hnsw:sync_threshold": 10, "hnsw:batch_size": 10}
+    )
+    collections = [first_collection, second_collection]
+
+    # (Does not trigger a purge)
+    for i in range(5):
+        first_collection.add(ids=str(i), embeddings=[i, i])
+
+    # (Should trigger a purge)
+    for i in range(100):
+        second_collection.add(ids=str(i), embeddings=[i, i])
+
+    # The purge of the second collection should not be blocked by the first
+    invariants.log_size_below_max(client._system, collections, True)

chromadb/test/property/invariants.py

@@ -1,6 +1,5 @@
 import gc
 import math
-from chromadb.api.configuration import HNSWConfigurationInternal
 from chromadb.config import System
 from chromadb.db.base import get_sql
 from chromadb.db.impl.sqlite import SqliteDB
@@ -334,29 +333,29 @@ def _total_embedding_queue_log_size(sqlite: SqliteDB) -> int:
 
 
 def log_size_below_max(
-    system: System, collection: Collection, has_collection_mutated: bool
+    system: System, collections: List[Collection], has_collection_mutated: bool
 ) -> None:
     sqlite = system.instance(SqliteDB)
 
     if has_collection_mutated:
         # Must always keep one entry to avoid reusing seq_ids
         assert _total_embedding_queue_log_size(sqlite) >= 1
 
-        hnsw_config = cast(
-            HNSWConfigurationInternal,
-            collection.get_model()
-            .get_configuration()
-            .get_parameter("hnsw_configuration")
-            .value,
+        # We purge per-collection as the sync_threshold is a per-collection setting
+        sync_threshold_sum = sum(
+            collection.metadata.get("hnsw:sync_threshold", 1000)
+            for collection in collections
+        )
+        batch_size_sum = sum(
+            collection.metadata.get("hnsw:batch_size", 1000)
+            for collection in collections
         )
-        sync_threshold = cast(int, hnsw_config.get_parameter("sync_threshold").value)
-        batch_size = cast(int, hnsw_config.get_parameter("batch_size").value)
 
         # -1 is used because the queue is always at least 1 entry long, so deletion stops before the max ack'ed sequence ID.
         # And if the batch_size != sync_threshold, the queue can have up to batch_size - 1 more entries.
         assert (
             _total_embedding_queue_log_size(sqlite) - 1
-            <= sync_threshold + batch_size - 1
+            <= sync_threshold_sum + batch_size_sum - 1
         )
     else:
         assert _total_embedding_queue_log_size(sqlite) == 0