rust/greptimedb
1
0
Fork 0
mirror of https://github.com/GreptimeTeam/greptimedb.git synced 2026-10-12 15:02:32 +00:00
Code Issues Packages Projects Releases Wiki Activity

test: preserve near-tie boundaries with separated-score join controls

Browse Source 
Signed-off-by: discord9 <55937128+discord9@users.noreply.github.com>
This commit is contained in:
discord9
2026-10-10 12:23:38 +08:00
parent 07ad633d38
commit 0421366827
2 changed files with 272 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/perf/query_cases/paper_ffx_ranked_balanced/case.toml
+136
View File
@@ -0,0 +1,136 @@
# WeightedAVG raw join versus host/area factorization; fixed direct-SST layout.
# Each measurement is timed separately. The query list deliberately puts raw
# first; compare with that ordering limitation in mind.
[case]
name = "paper_ffx_ranked_balanced"
description = "Raw versus factorized weightedAVG at a balanced join cardinality"
[scenario]
kind = "direct_readable_sst"
seed = 5204
[[scenario.tables]]
database = "public"
name = "paper_ffx_ranked_balanced_fact"
engine = "mito"
append_mode = true
sst_format = "flat"
primary_key = ["host", "instance"]
time_index = "ts"
[[scenario.tables.columns]]
name = "host"
type = "STRING"
semantic = "tag"
distribution = { kind = "cardinality", values = 16, prefix = "h" }
[[scenario.tables.columns]]
name = "instance"
type = "STRING"
semantic = "tag"
distribution = { kind = "cardinality", values = 80, prefix = "i" }
[[scenario.tables.columns]]
name = "metric_value"
type = "DOUBLE"
semantic = "field"
distribution = { kind = "deterministic_wave", min = 1.0, max = 500.0 }
[[scenario.tables.columns]]
name = "ts"
type = "TIMESTAMP(9)"
semantic = "timestamp"
[[scenario.tables]]
database = "public"
name = "paper_ffx_ranked_balanced_dim"
engine = "mito"
append_mode = true
sst_format = "flat"
primary_key = ["host", "area"]
time_index = "ts"
[[scenario.tables.columns]]
name = "host"
type = "STRING"
semantic = "tag"
distribution = { kind = "cardinality", values = 16, prefix = "h" }
[[scenario.tables.columns]]
name = "area"
type = "STRING"
semantic = "tag"
distribution = { kind = "cardinality", values = 5, prefix = "a" }
[[scenario.tables.columns]]
name = "weight"
type = "DOUBLE"
semantic = "field"
distribution = { kind = "deterministic_wave", min = 0.1, max = 10.0 }
[[scenario.tables.columns]]
name = "ts"
type = "TIMESTAMP(9)"
semantic = "timestamp"
[scenario.layout]
regions = 1
sst_count = 4
rows_per_sst = 1280
row_group_size = 128
series_count = 80
start_unix_nanos = 1704067200000000000
step_nanos = 1000000000
time_range_layout = "non_overlapping_per_sst"
series_layout = "round_robin"
# This positive ranking control selects the first 80 dimension rows across distinct area weights, separating top scores while preserving the weightedAVG inputs.
[[scenario.queries]]
name = "raw_wide"
kind = "sql"
query = "SELECT f.host, d.area, AVG(f.metric_value * d.weight) AS score FROM paper_ffx_ranked_balanced_fact f JOIN paper_ffx_ranked_balanced_dim d ON f.host = d.host WHERE f.ts >= TIMESTAMP '2024-01-01 00:00:00' AND f.ts < TIMESTAMP '2024-01-01 01:25:20' AND d.ts >= TIMESTAMP '2024-01-01 00:00:00' AND d.ts < TIMESTAMP '2024-01-01 00:01:20' GROUP BY f.host, d.area ORDER BY score DESC NULLS LAST, f.host, d.area LIMIT 10"
warmup = 3
iterations = 9
[scenario.queries.thresholds]
max_candidate_latency_regression_pct = 50
[[scenario.queries]]
name = "factor_wide"
kind = "sql"
query = "WITH ff AS (SELECT host, SUM(metric_value) AS fm, COUNT(metric_value) AS fn FROM paper_ffx_ranked_balanced_fact WHERE ts >= TIMESTAMP '2024-01-01 00:00:00' AND ts < TIMESTAMP '2024-01-01 01:25:20' GROUP BY host), dd AS (SELECT host, area, SUM(weight) AS dw, COUNT(weight) AS dn FROM paper_ffx_ranked_balanced_dim WHERE ts >= TIMESTAMP '2024-01-01 00:00:00' AND ts < TIMESTAMP '2024-01-01 00:01:20' GROUP BY host, area) SELECT ff.host, dd.area, CASE WHEN ff.fn = 0 OR dd.dn = 0 THEN NULL ELSE (ff.fm * dd.dw) / (CAST(ff.fn AS DOUBLE) * CAST(dd.dn AS DOUBLE)) END AS score FROM ff JOIN dd ON ff.host = dd.host ORDER BY score DESC NULLS LAST, ff.host, dd.area LIMIT 10"
warmup = 3
iterations = 9
[scenario.queries.thresholds]
max_candidate_latency_regression_pct = 50
# Full-group correctness output and EXPLAIN ANALYZE are diagnostic single executions and must be excluded from performance analysis.
[[scenario.queries]]
name = "raw_full_wide"
kind = "sql"
query = "SELECT f.host, d.area, COUNT(*) AS pairs, COUNT(f.metric_value * d.weight) AS valid, SUM(f.metric_value * d.weight) AS total, AVG(f.metric_value * d.weight) AS score FROM paper_ffx_ranked_balanced_fact f JOIN paper_ffx_ranked_balanced_dim d ON f.host = d.host WHERE f.ts >= TIMESTAMP '2024-01-01 00:00:00' AND f.ts < TIMESTAMP '2024-01-01 01:25:20' AND d.ts >= TIMESTAMP '2024-01-01 00:00:00' AND d.ts < TIMESTAMP '2024-01-01 00:01:20' GROUP BY f.host, d.area ORDER BY f.host NULLS LAST, d.area NULLS LAST"
warmup = 0
iterations = 1
[[scenario.queries]]
name = "factor_full_wide"
kind = "sql"
query = "WITH ff AS (SELECT host, SUM(metric_value) AS fm, COUNT(metric_value) AS fn, COUNT(*) AS fc FROM paper_ffx_ranked_balanced_fact WHERE ts >= TIMESTAMP '2024-01-01 00:00:00' AND ts < TIMESTAMP '2024-01-01 01:25:20' GROUP BY host), dd AS (SELECT host, area, SUM(weight) AS dw, COUNT(weight) AS dn, COUNT(*) AS dc FROM paper_ffx_ranked_balanced_dim WHERE ts >= TIMESTAMP '2024-01-01 00:00:00' AND ts < TIMESTAMP '2024-01-01 00:01:20' GROUP BY host, area) SELECT ff.host, dd.area, ff.fc * dd.dc AS pairs, ff.fn * dd.dn AS valid, ff.fm * dd.dw AS total, CASE WHEN ff.fn = 0 OR dd.dn = 0 THEN NULL ELSE (ff.fm * dd.dw) / (CAST(ff.fn AS DOUBLE) * CAST(dd.dn AS DOUBLE)) END AS score FROM ff JOIN dd ON ff.host = dd.host ORDER BY ff.host NULLS LAST, dd.area NULLS LAST"
warmup = 0
iterations = 1
[[scenario.queries]]
name = "diagnostic_explain_raw_wide"
kind = "sql"
query = "EXPLAIN ANALYZE VERBOSE SELECT f.host, d.area, AVG(f.metric_value * d.weight) AS score FROM paper_ffx_ranked_balanced_fact f JOIN paper_ffx_ranked_balanced_dim d ON f.host = d.host WHERE f.ts >= TIMESTAMP '2024-01-01 00:00:00' AND f.ts < TIMESTAMP '2024-01-01 01:25:20' AND d.ts >= TIMESTAMP '2024-01-01 00:00:00' AND d.ts < TIMESTAMP '2024-01-01 00:01:20' GROUP BY f.host, d.area"
warmup = 0
iterations = 1
[[scenario.queries]]
name = "diagnostic_explain_factor_wide"
kind = "sql"
query = "EXPLAIN ANALYZE VERBOSE WITH ff AS (SELECT host, SUM(metric_value) AS fm, COUNT(metric_value) AS fn FROM paper_ffx_ranked_balanced_fact WHERE ts >= TIMESTAMP '2024-01-01 00:00:00' AND ts < TIMESTAMP '2024-01-01 01:25:20' GROUP BY host), dd AS (SELECT host, area, SUM(weight) AS dw, COUNT(weight) AS dn FROM paper_ffx_ranked_balanced_dim WHERE ts >= TIMESTAMP '2024-01-01 00:00:00' AND ts < TIMESTAMP '2024-01-01 00:01:20' GROUP BY host, area) SELECT ff.host, dd.area, (ff.fm * dd.dw) / (CAST(ff.fn AS DOUBLE) * CAST(dd.dn AS DOUBLE)) AS score FROM ff JOIN dd ON ff.host = dd.host"
warmup = 0
iterations = 1
tests/perf/query_cases/paper_ffx_ranked_hot/case.toml
+136
View File
@@ -0,0 +1,136 @@
# WeightedAVG raw join versus host/area factorization; fixed direct-SST layout.
# Each measurement is timed separately. The query list deliberately puts raw
# first; compare with that ordering limitation in mind.
[case]
name = "paper_ffx_ranked_hot"
description = "Raw versus factorized weightedAVG at a hot join cardinality"
[scenario]
kind = "direct_readable_sst"
seed = 5204
[[scenario.tables]]
database = "public"
name = "paper_ffx_ranked_hot_fact"
engine = "mito"
append_mode = true
sst_format = "flat"
primary_key = ["host", "instance"]
time_index = "ts"
[[scenario.tables.columns]]
name = "host"
type = "STRING"
semantic = "tag"
distribution = { kind = "cardinality", values = 1, prefix = "h" }
[[scenario.tables.columns]]
name = "instance"
type = "STRING"
semantic = "tag"
distribution = { kind = "cardinality", values = 64, prefix = "i" }
[[scenario.tables.columns]]
name = "metric_value"
type = "DOUBLE"
semantic = "field"
distribution = { kind = "deterministic_wave", min = 1.0, max = 500.0 }
[[scenario.tables.columns]]
name = "ts"
type = "TIMESTAMP(9)"
semantic = "timestamp"
[[scenario.tables]]
database = "public"
name = "paper_ffx_ranked_hot_dim"
engine = "mito"
append_mode = true
sst_format = "flat"
primary_key = ["host", "area"]
time_index = "ts"
[[scenario.tables.columns]]
name = "host"
type = "STRING"
semantic = "tag"
distribution = { kind = "cardinality", values = 1, prefix = "h" }
[[scenario.tables.columns]]
name = "area"
type = "STRING"
semantic = "tag"
distribution = { kind = "cardinality", values = 4, prefix = "a" }
[[scenario.tables.columns]]
name = "weight"
type = "DOUBLE"
semantic = "field"
distribution = { kind = "deterministic_wave", min = 0.1, max = 10.0 }
[[scenario.tables.columns]]
name = "ts"
type = "TIMESTAMP(9)"
semantic = "timestamp"
[scenario.layout]
regions = 1
sst_count = 4
rows_per_sst = 512
row_group_size = 128
series_count = 64
start_unix_nanos = 1704067200000000000
step_nanos = 1000000000
time_range_layout = "non_overlapping_per_sst"
series_layout = "round_robin"
# This positive ranking control selects the first 128 dimension rows across distinct area weights, separating top scores while preserving the weightedAVG inputs.
[[scenario.queries]]
name = "raw_wide"
kind = "sql"
query = "SELECT f.host, d.area, AVG(f.metric_value * d.weight) AS score FROM paper_ffx_ranked_hot_fact f JOIN paper_ffx_ranked_hot_dim d ON f.host = d.host WHERE f.ts >= TIMESTAMP '2024-01-01 00:00:00' AND f.ts < TIMESTAMP '2024-01-01 00:34:08' AND d.ts >= TIMESTAMP '2024-01-01 00:00:00' AND d.ts < TIMESTAMP '2024-01-01 00:02:08' GROUP BY f.host, d.area ORDER BY score DESC NULLS LAST, f.host, d.area LIMIT 10"
warmup = 3
iterations = 9
[scenario.queries.thresholds]
max_candidate_latency_regression_pct = 50
[[scenario.queries]]
name = "factor_wide"
kind = "sql"
query = "WITH ff AS (SELECT host, SUM(metric_value) AS fm, COUNT(metric_value) AS fn FROM paper_ffx_ranked_hot_fact WHERE ts >= TIMESTAMP '2024-01-01 00:00:00' AND ts < TIMESTAMP '2024-01-01 00:34:08' GROUP BY host), dd AS (SELECT host, area, SUM(weight) AS dw, COUNT(weight) AS dn FROM paper_ffx_ranked_hot_dim WHERE ts >= TIMESTAMP '2024-01-01 00:00:00' AND ts < TIMESTAMP '2024-01-01 00:02:08' GROUP BY host, area) SELECT ff.host, dd.area, CASE WHEN ff.fn = 0 OR dd.dn = 0 THEN NULL ELSE (ff.fm * dd.dw) / (CAST(ff.fn AS DOUBLE) * CAST(dd.dn AS DOUBLE)) END AS score FROM ff JOIN dd ON ff.host = dd.host ORDER BY score DESC NULLS LAST, ff.host, dd.area LIMIT 10"
warmup = 3
iterations = 9
[scenario.queries.thresholds]
max_candidate_latency_regression_pct = 50
# Full-group correctness output and EXPLAIN ANALYZE are diagnostic single executions and must be excluded from performance analysis.
[[scenario.queries]]
name = "raw_full_wide"
kind = "sql"
query = "SELECT f.host, d.area, COUNT(*) AS pairs, COUNT(f.metric_value * d.weight) AS valid, SUM(f.metric_value * d.weight) AS total, AVG(f.metric_value * d.weight) AS score FROM paper_ffx_ranked_hot_fact f JOIN paper_ffx_ranked_hot_dim d ON f.host = d.host WHERE f.ts >= TIMESTAMP '2024-01-01 00:00:00' AND f.ts < TIMESTAMP '2024-01-01 00:34:08' AND d.ts >= TIMESTAMP '2024-01-01 00:00:00' AND d.ts < TIMESTAMP '2024-01-01 00:02:08' GROUP BY f.host, d.area ORDER BY f.host NULLS LAST, d.area NULLS LAST"
warmup = 0
iterations = 1
[[scenario.queries]]
name = "factor_full_wide"
kind = "sql"
query = "WITH ff AS (SELECT host, SUM(metric_value) AS fm, COUNT(metric_value) AS fn, COUNT(*) AS fc FROM paper_ffx_ranked_hot_fact WHERE ts >= TIMESTAMP '2024-01-01 00:00:00' AND ts < TIMESTAMP '2024-01-01 00:34:08' GROUP BY host), dd AS (SELECT host, area, SUM(weight) AS dw, COUNT(weight) AS dn, COUNT(*) AS dc FROM paper_ffx_ranked_hot_dim WHERE ts >= TIMESTAMP '2024-01-01 00:00:00' AND ts < TIMESTAMP '2024-01-01 00:02:08' GROUP BY host, area) SELECT ff.host, dd.area, ff.fc * dd.dc AS pairs, ff.fn * dd.dn AS valid, ff.fm * dd.dw AS total, CASE WHEN ff.fn = 0 OR dd.dn = 0 THEN NULL ELSE (ff.fm * dd.dw) / (CAST(ff.fn AS DOUBLE) * CAST(dd.dn AS DOUBLE)) END AS score FROM ff JOIN dd ON ff.host = dd.host ORDER BY ff.host NULLS LAST, dd.area NULLS LAST"
warmup = 0
iterations = 1
[[scenario.queries]]
name = "diagnostic_explain_raw_wide"
kind = "sql"
query = "EXPLAIN ANALYZE VERBOSE SELECT f.host, d.area, AVG(f.metric_value * d.weight) AS score FROM paper_ffx_ranked_hot_fact f JOIN paper_ffx_ranked_hot_dim d ON f.host = d.host WHERE f.ts >= TIMESTAMP '2024-01-01 00:00:00' AND f.ts < TIMESTAMP '2024-01-01 00:34:08' AND d.ts >= TIMESTAMP '2024-01-01 00:00:00' AND d.ts < TIMESTAMP '2024-01-01 00:02:08' GROUP BY f.host, d.area"
warmup = 0
iterations = 1
[[scenario.queries]]
name = "diagnostic_explain_factor_wide"
kind = "sql"
query = "EXPLAIN ANALYZE VERBOSE WITH ff AS (SELECT host, SUM(metric_value) AS fm, COUNT(metric_value) AS fn FROM paper_ffx_ranked_hot_fact WHERE ts >= TIMESTAMP '2024-01-01 00:00:00' AND ts < TIMESTAMP '2024-01-01 00:34:08' GROUP BY host), dd AS (SELECT host, area, SUM(weight) AS dw, COUNT(weight) AS dn FROM paper_ffx_ranked_hot_dim WHERE ts >= TIMESTAMP '2024-01-01 00:00:00' AND ts < TIMESTAMP '2024-01-01 00:02:08' GROUP BY host, area) SELECT ff.host, dd.area, (ff.fm * dd.dw) / (CAST(ff.fn AS DOUBLE) * CAST(dd.dn AS DOUBLE)) AS score FROM ff JOIN dd ON ff.host = dd.host"
warmup = 0
iterations = 1