Open agandra30 opened 2 weeks ago
alwayslove2013
commented
2 weeks ago @Sheharyar570 Could you please share any experiences or advice you have regarding the building performance of PGVectorScale's DiskANN? We would greatly appreciate it!
xiaofan-luan
commented
1 week ago @Sheharyar570 Could you please share any experiences or advice you have regarding the
buildingperformance ofPGVectorScale's DiskANN? We would greatly appreciate it!
Maybe we just need to give a user a timeout config?
xiaofan-luan
commented
1 week ago I ran multiple tests with PGvector scale using diskann and one of the biggest problem is that the vectordbbench just exits the execution with optmize timeout , this is very true for PGVECTOR SCALE and pgvectorRS.
I know we can increase the time in the scripts , but did anyone observed or recommend any configuration settings that can complete the execution with in the 5hours(default timeout) for 10M cohere 768 Dimension dataset ,we want to look at the cross comparison with out editing the default time outs for a large datastes , did any successfully completed with in that timeout (Milvus Yes , but other DBs ?)
Error message :
024-09-17 22:14:02,230 | WARNING: VectorDB optimize timeout in 18000 (task_runner.py:249) (3816719) 2024-09-17 22:14:02,274 | WARNING: Failed to run performance case, reason = Performance case optimize timeout (task_runner.py:191) (3816719) Traceback (most recent call last): File "/root/vectordbbench_runs/lib/python3.12/site-packages/vectordb_bench/backend/task_runner.py", line 247, in _optimize return future.result(timeout=self.ca.optimize_timeout)[1]Query:
`CREATE INDEX IF NOT EXISTS "pgvectorscale_index" ON public. "pg_vectorscale_collection" USING "diskann" (embedding "vector_cosine_ops" ) WITH ( "storage_layout" = "memory_optimized", "num_neighbors" = "50", "search_list_size" = "100", "max_alpha" = "1.2", "num_bits_per_dimension" = "2" ); (pgvectorscale.py:200) (3935818)
`
My Postgres server infra configuration :
- Installed on baremetal Ubuntu 22
- postgre 16.4 v(untuned)
- Memory available :
# free -mh total used free shared buff/cache available Mem: 1.0Ti 13Gi 986Gi 152Mi 7.3Gi 988Gi Swap: 0B 0B 0B4. Extensions used :` pgdiskann=# \dx; List of installed extensions Name | Version | Schema | Description -------------+---------+------------+------------------------------------------------------ plpgsql | 1.0 | pg_catalog | PL/pgSQL procedural language vector | 0.7.4 | public | vector data type and ivfflat and hnsw access methods vectorscale | 0.3.0 | public | pgvectorscale: Advanced indexing for vector data (3 rows)
`
BTW, I don't think it's a wise choice to run 10M or more data on PG vector, it is simply too slow.
KendrickChou
commented
1 week ago @Sheharyar570 Could you please share any experiences or advice you have regarding the
buildingperformance ofPGVectorScale's DiskANN? We would greatly appreciate it!Maybe we just need to give a user a timeout config?
A timeout config is exactly useful for me too!
alwayslove2013
commented
1 week ago @KendrickChou @xiaofan-luan I will consider adding a timeout setting in the UI.
Maybe we just need to give a user a timeout config?
A timeout config is exactly useful for me too!
We have set different default timeout based on the size of the dataset. Currently, the timeout config can be modified through code.
https://github.com/zilliztech/VectorDBBench/blob/b364fe316f72c86809d3203dc2b75437e9eabc90/vectordb_bench/__init__.py#L40-L56
Sheharyar570
commented
1 week ago @agandra30 Well, I've not tried running 10M cohere on PG Vector Scale. So I won't be able to suggest any specific configuration.
Although, I would suggest to make maintenance_work_mem larger than the index_size, but first you would need create the index to get index_size.
But you still may need to update the default timeout.
I ran multiple tests with PGvector scale using diskann and one of the biggest problem is that the vectordbbench just exits the execution with optmize timeout , this is very true for PGVECTOR SCALE and pgvectorRS.
I know we can increase the time in the scripts , but did anyone observed or recommend any configuration settings that can complete the execution with in the 5hours(default timeout) for 10M cohere 768 Dimension dataset ,we want to look at the cross comparison with out editing the default time outs for a large datastes , did any successfully completed with in that timeout (Milvus Yes , but other DBs ?)
Error message :
024-09-17 22:14:02,230 | WARNING: VectorDB optimize timeout in 18000 (task_runner.py:249) (3816719) 2024-09-17 22:14:02,274 | WARNING: Failed to run performance case, reason = Performance case optimize timeout (task_runner.py:191) (3816719) Traceback (most recent call last): File "/root/vectordbbench_runs/lib/python3.12/site-packages/vectordb_bench/backend/task_runner.py", line 247, in _optimize return future.result(timeout=self.ca.optimize_timeout)[1]Query:
`CREATE INDEX IF NOT EXISTS "pgvectorscale_index" ON public. "pg_vectorscale_collection"
USING "diskann" (embedding "vector_cosine_ops" ) WITH ( "storage_layout" = "memory_optimized", "num_neighbors" = "50", "search_list_size" = "100", "max_alpha" = "1.2", "num_bits_per_dimension" = "2" ); (pgvectorscale.py:200) (3935818)
`
My Postgres server infra configuration :
# free -mh total used free shared buff/cache available Mem: 1.0Ti 13Gi 986Gi 152Mi 7.3Gi 988Gi Swap: 0B 0B 0B` pgdiskann=# \dx; List of installed extensions Name | Version | Schema | Description
-------------+---------+------------+------------------------------------------------------ plpgsql | 1.0 | pg_catalog | PL/pgSQL procedural language vector | 0.7.4 | public | vector data type and ivfflat and hnsw access methods vectorscale | 0.3.0 | public | pgvectorscale: Advanced indexing for vector data (3 rows)
`