# Full-scale run of record, derived from full.yaml.
#
# Two deliberate deviations from full.yaml, both budget-driven (subsample, and say so):
#
# 1. Churn cycles 10 -> 3. full.yaml's own header comment already flags
#    VACUUM as the dominant, non-linear cost (pgvector HNSW ambulkdelete
#    does not scale gently with dead-tuple count/graph size) and says
#    "budget VACUUM alone at 10s of minutes per cycle" at 1M rows -- 10
#    cycles could be hours by itself. 3 cycles matches the smoke run's own
#    cycle count for a like-for-like comparison and still shows a drift
#    trend, at a fraction of the cost.
# 2. hnsw ef_search 6 values -> 4 ([10,40,80,160], dropping 20 and 320) and
#    ivfflat probes 7 -> 5 ([1,2,4,8,16], dropping 32 and 64): the smoke run
#    already showed the interesting transitions (budget-capped plateau,
#    planner's index->seqscan flip) live inside this reduced range; the
#    dropped high values were seqscan/ceiling points carrying no new
#    information in the smoke run, at real per-point cost multiplied by the
#    1M-row table's slower per-query latency and full.yaml's 5s
#    throughput-pass floor (vs smoke's 1.5s).
#
# Everything else -- all 4 selectivities x both modes (8 filter combos),
# both query regimes for correlated filters, concurrency [1,8], SIFT1M --
# is unchanged from full.yaml: this is the full requested matrix, not a
# smaller one.
#
# Disk-safety addition (this machine has only ~17GB free on a 995GB/98%-full
# shared disk -- other checkouts and dev Postgres instances live here too):
# min_wal_size/max_wal_size capped well below the default 4GB/16GB so WAL
# growth can't threaten the shared disk; does not affect recall/latency
# measurements, only checkpoint frequency.
name: full_1m
seed: 42
k: 10
warmup_queries: 20
hardness_recall_threshold: 0.97
min_throughput_seconds: 5.0

postgres:
  maintenance_work_mem: 8GB
  min_wal_size: 1GB
  max_wal_size: 4GB

dataset:
  kind: synthetic
  n: 1000000
  dim: 768
  n_centers: 1000
  cluster_sigma: 1.1
  cluster_scale_mix: [0.6, 1.0, 1.6]
  background_frac: 0.02
  n_queries: 500

metric: l2

filters:
  selectivities: [0.001, 0.01, 0.10, 0.50]
  modes: [correlated, uncorrelated]
  run:
    - {selectivity: 0.001, mode: correlated}
    - {selectivity: 0.001, mode: uncorrelated}
    - {selectivity: 0.01, mode: correlated}
    - {selectivity: 0.01, mode: uncorrelated}
    - {selectivity: 0.10, mode: correlated}
    - {selectivity: 0.10, mode: uncorrelated}
    - {selectivity: 0.50, mode: correlated}
    - {selectivity: 0.50, mode: uncorrelated}

systems:
  - name: pgvector_hnsw
    type: hnsw
    build_params: {m: 16, ef_construction: 64}
    sweep_param: ef_search
    sweep_values: [10, 40, 80, 160]
    iterative_scan_values: [null]
    filtered_iterative_scan_values: ["off", "relaxed_order"]
    scan_budget_param: hnsw.max_scan_tuples
    filtered_scan_budget_values: [20000, auto]
  - name: pgvector_ivfflat
    type: ivfflat
    build_params: {lists: null}
    sweep_param: probes
    sweep_values: [1, 2, 4, 8, 16]
    iterative_scan_values: [null]
    filtered_iterative_scan_values: ["off", "relaxed_order"]

concurrency: [1, 8]

churn:
  enabled: true
  cycles: 3
  reembed_pct: 0.02
  delete_reinsert_pct: 0.02
  vacuum_after_each_cycle: true
  system: {name: pgvector_hnsw, type: hnsw, build_params: {m: 16, ef_construction: 64},
           sweep_param: ef_search, sweep_values: [80]}
  sweep_value: 80

sift1m:
  enabled: true
  sift_n_queries: 500
  sift_systems:
    - name: pgvector_hnsw
      type: hnsw
      build_params: {m: 16, ef_construction: 64}
      sweep_param: ef_search
      sweep_values: [10, 20, 40, 80, 160, 320]
    - name: pgvector_ivfflat
      type: ivfflat
      build_params: {lists: null}
      sweep_param: probes
      sweep_values: [1, 2, 4, 8, 16, 32, 64]
