# One namespace per tenant: a synthetic tenant_id column (100 tenants, each a group of 10 of the 1,000 k-means clusters, so a tenant's
# rows are one topical region) on the first 500,000 wiki_dpr vectors, pinecone.create_index(engine => 'remote', partition_column =>
# 'tenant_id'): namespace public.<table>:<tenant>, queried with WHERE tenant_id = $t (a query without it is refused).
# 500,000 rows keep the write units inside the budget (about 1.6M); the namespaces hold about 5,000 vectors each (15 MB).
# Queries: 20 held-out vectors from each tenant's own region (2,000, shuffled so concurrent clients meet different tenants), exact
# top-k among the tenant's rows only. Reports the backfill (requests are estimated from the chunking rule: one request per tenant per
# 5,000-row window), per-tenant latency (median of the tenants' p50) and the rejected share at 1, 4 and 16 clients.
# Two stages, as remote_1m.yaml: --stage build, then --stage measure (deletes the index).
name: remote_partition
seed: 42
k: 10
warmup_queries: 10
hardness_recall_threshold: 0.97
min_throughput_seconds: 5.0
table: pg_pinecone_bench_vec_part

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

partition:
  column: tenant_id
  n_tenants: 100
  queries_per_tenant: 20
  k_values: [10]

dataset:
  kind: real
  artifact: wikidpr_768d_1m.npz
  n: 500000
  dim: 768
  n_centers: 1000
  kmeans_sample: 100000
  kmeans_iters: 15
  n_queries: 500

metric: l2

filters:
  run: []

systems:
  - name: pg_pinecone_remote
    type: pg_pinecone_remote
    skip_unfiltered: true
    build_params: {partition_column: tenant_id}
    sweep_param: am_min_topk
    sweep_values: [50]
    record_scan_stats: true

concurrency: [1, 4, 16]

churn:
  enabled: false

sift1m:
  enabled: false
