# Forced-seqscan exact-scan baseline series across the filter matrix.
#
# This is the honest incumbent the native index has to beat: every filtered
# configuration in the matrix (all 4 selectivities x both correlated/
# uncorrelated modes x both query regimes for correlated filters x 1 and 8
# connections) gets a companion measurement with `enable_indexscan = off`,
# `enable_bitmapscan = off`, `enable_seqscan = on` -- verified per point via
# EXPLAIN (plan_shape must come back `seqscan_forced`, or the point is
# recorded invalid, never silently blended in). Recall is asserted ==1.0 (an
# exact scan visits every qualifying row by construction; see
# runner.py:_assert_exact_recall) rather than just reported.
#
# The `exact` system type builds NO index (harness/workloads.py:build_index's
# kind=="exact" branch) -- every other secondary index is still dropped
# first, so the table carries zero indexes while this series runs. This is
# why it is CPU-cheap: no index builds, no ef/probes
# sweep (a single placeholder sweep_value stands in for "the knob doesn't
# exist here" so it can reuse run_sweep's grid machinery), only sequential
# scans.
#
# `dataset:` below is BYTE-FOR-BYTE identical to full_1m.yaml's synthetic
# corpus (same seed, n, dim, n_centers, cluster_sigma, cluster_scale_mix,
# background_frac, n_queries) and `warmup_queries`/`concurrency` match too --
# the dataset cache key and every filtered ground-truth cache key
# (harness/groundtruth.py) are pure functions of these values, so this run
# hits the SAME cached dataset .npz and the SAME cached brute-force ground
# truth full_1m.yaml already computed and paid for. No regeneration, no
# re-computation of ground truth: this config only adds new seqscan
# measurements on top of data that already exists on disk from the run of
# record. (If full_1m.yaml's dataset/filter params ever drift, this file
# must be updated in lockstep or the cache-reuse property silently breaks
# and this run pays the recompute cost instead -- not incorrect, just slow.)
#
# Deliberately excludes `churn:` and `sift1m:` -- this config's only job is
# the exact-scan companion series across the filter matrix, not a rerun of
# the full matrix.
name: baselines_exact
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: exact_seqscan
    type: exact
    sweep_param: none
    sweep_values: [0]

concurrency: [1, 8]
