[dataset] name = "msmarco" size = "113m" table = "msmarco" expected_rows = 113520750 dims = 1024 vector_field = "emb" s3_base_path = "s3://paradedb-benchmark/datasets/msmarco/v1" # Both sections match the pdbench Compose service with CI's CPU/worker override. # `pdbench search` checks [postgres.search] before a sweep. CI builds the index with # pdbench-load, so [postgres.build] is not checked; the profile format requires it. [postgres.build] shared_buffers = "8GB" maintenance_work_mem = "8GB" max_parallel_maintenance_workers = 7 [postgres.search] shared_buffers = "8GB" max_parallel_workers = 8 max_worker_processes = 32 [index] target_segment_count = 8 vector_router_quantization = true [search] queries = 100 runs = 1 warmup_runs = 1 explain_sample = 100 [search.session] max_parallel_workers_per_gather = "4" jit = "off" plan_cache_mode = "force_custom_plan" "paradedb.vector_adaptive_probing" = "on" "paradedb.vector_router_rerank_factor" = "2" "paradedb.vector_router_min_beam_fraction" = "0.5" "paradedb.vector_router_aps" = "on" [search.sweep] param = "paradedb.vector_cluster_max_probe" min = 1 max = 100000 start = 100 integer = true targets = [0.9, 0.95, 0.99] tolerance = 0.000001 max_probes = 20