# Lane comparison at wiki_dpr 100k x 768d, trigger lane: pinecone.promote (triggers keep the namespace in sync, ids are the primary
# key) and pinecone.topk(table, query, k), next to remote_lanes_am_100k.yaml (same rows and queries). Its own serverless index,
# pg-pinecone-bench-vec-lane-topk; a promoted table searches the heap exactly until the backfill is done, so the gate waits for mode remote.
# topk sends topK = k, so k = 100 is one request. Two stages: --stage build, then --stage measure (deletes the index).
name: remote_lanes_topk_100k
seed: 42
k: 10
k_values: [10, 100]
warmup_queries: 10
hardness_recall_threshold: 0.97
min_throughput_seconds: 5.0
table: pg_pinecone_bench_vec_lane_topk

postgres:
  shared_preload_libraries: "'pg_pinecone'"
  pinecone.databases: "'bench'"
  pinecone.worker_naptime_ms: 2000
  log_min_error_statement: panic
  shared_buffers: 4GB
  maintenance_work_mem: 1GB

remote:
  rest_seconds: 20
  delete_index_after: true
  direct: {n: 1000, concurrency: [1, 4, 16], top_k: [10, 100]}

dataset:
  kind: real
  artifact: wikidpr_768d_1m.npz
  n: 100000
  dim: 768
  n_centers: 1000
  kmeans_sample: 100000
  kmeans_iters: 15
  n_queries: 500

metric: l2

filters:
  run: []

systems:
  - name: pg_pinecone_remote_topk
    type: pg_pinecone_remote_topk
    build_params: {index_name: pg-pinecone-bench-vec-lane-topk}
    sweep_param: none
    sweep_values: [0]

concurrency: [1, 4, 16]

churn:
  enabled: false

sift1m:
  enabled: false
