mirror of
https://github.com/GreptimeTeam/greptimedb.git
synced 2026-10-12 06:52:34 +00:00
test(perf): fix rate-fusion input and complete-result controls
Signed-off-by: discord9 <55937128+discord9@users.noreply.github.com>
This commit is contained in:
discord9
@@ -0,0 +1,50 @@
|
||||
# PromQL series rate-fusion regression fixture
|
||||
|
||||
This native Query Regression case isolates the candidate shape `SUM(Float64)` by
|
||||
`dc` over one direct `prom_rate` expression, projected as Float64 with a
|
||||
millisecond timestamp and zero offset. The immutable input uses the previously qualified dictionary-tag A direct-SST
|
||||
generator, schema, and layout. All parameters are specified in this case; no
|
||||
private or gitignored source file is required. It uses seed 25652, 4,096
|
||||
instances × 512 samples = 2,097,152 rows, 16 `dc` values, 4 devices, and two flat
|
||||
non-overlapping SSTs of 1,048,576 rows each, 8,192-row groups, 15-second
|
||||
millisecond timestamps, primary key `(instance, dc, device)`, and deterministic
|
||||
wave values from 0 through 100. It uses the same data generator, not the
|
||||
original 127-slope counter or an archived original input.
|
||||
|
||||
The range window is 1704069900–1704071715 seconds at 15-second steps (122
|
||||
evaluations), fully covered by fixture timestamps from 1704067200 through
|
||||
1704074865. The first SST ends at 1704071025; the 5-minute lookback begins at
|
||||
1704069600, so this range crosses the actual two-SST boundary and requests data
|
||||
from both files. Actual scanned row counts must be audited from runtime plans;
|
||||
the full 2,097,152-row fixture size is not an expected scan count for this query.
|
||||
The primary range query expects 16 `dc` output groups per evaluation (1,952
|
||||
result rows) and evaluates 4,096 series per evaluation (499,712
|
||||
evaluation-points). The instant query at 1704071715 expects 16 output rows.
|
||||
Primary `TQL EVAL` entries preserve complete response data and use 3 warmups/20
|
||||
range iterations and 3 warmups/9 instant iterations; the range EVAL is also the
|
||||
runner's untimed first-query validation. Separate one-shot `TQL ANALYZE VERBOSE`
|
||||
entries capture runtime gate, actual plan, scan/index, and work-count diagnostics
|
||||
without adding analyze overhead to timed samples. An untimed one-iteration SQL
|
||||
audit returns one row per `dc`, verifying 131,072 rows and 256 instances per
|
||||
group and min/max timestamps 1704067200 and 1704074865.
|
||||
|
||||
Controls are limited to grouped `sum_over_time` (wider non-rate path),
|
||||
grouped `min by (dc) (rate(...))` (a different aggregation operator and a
|
||||
negative fusion-gate control), and grouped `rate` with a 15-second offset
|
||||
(offset gate). MIN keeps the same range, rate-window work, and compact 1,952-row
|
||||
answer while not matching the SUM candidate. They reuse the same table and
|
||||
range. The offset query's parser acceptance must be confirmed on the pinned
|
||||
runtime. Each timed query retains the existing 10% candidate-latency-regression
|
||||
threshold: a pass only means that configured guard did not trip; it is not proof
|
||||
of candidate-gate activation or performance improvement. Do not loosen it to
|
||||
hide noise.
|
||||
|
||||
After a native run, audit the normalized plans and complete raw responses,
|
||||
including group labels, timestamps, numeric values with appropriate floating
|
||||
point tolerance, and expected row counts. Also audit source SHAs, compiler,
|
||||
features, fixture layout, the actual runtime gate, scans/indexes, and work counts.
|
||||
Official performance acceptance still needs clean base/candidate medians,
|
||||
control behavior, drift and tolerance review, and complete cross-arm answers.
|
||||
This fixture alone cannot establish a gain. Finite-pool behavior, drops, and
|
||||
spill behavior belong to code tests, not this CI case. No official workflow was
|
||||
run as part of adding this fixture.
|
||||
@@ -0,0 +1,144 @@
|
||||
# Minimal native Query Regression fixture for the narrow direct-prom-rate sum fusion candidate.
|
||||
# This deterministic wave is not the original 127-slope counter/input archive.
|
||||
# Retain complete raw responses for parity auditing; passing thresholds only means
|
||||
# the configured latency guard did not trip, not that fusion or a speedup was proven.
|
||||
|
||||
[case]
|
||||
name = "promql_series_rate_fusion"
|
||||
description = "PromQL rate-only grouped sum fusion regression fixture"
|
||||
|
||||
[scenario]
|
||||
kind = "direct_readable_sst"
|
||||
seed = 25652
|
||||
|
||||
[[scenario.tables]]
|
||||
database = "public"
|
||||
name = "promql_series_rate_fusion"
|
||||
engine = "mito"
|
||||
append_mode = true
|
||||
sst_format = "flat"
|
||||
primary_key = ["instance", "dc", "device"]
|
||||
time_index = "ts"
|
||||
|
||||
[[scenario.tables.columns]]
|
||||
name = "instance"
|
||||
type = "STRING"
|
||||
semantic = "tag"
|
||||
distribution = { kind = "cardinality", values = 4096, prefix = "instance" }
|
||||
|
||||
[[scenario.tables.columns]]
|
||||
name = "dc"
|
||||
type = "STRING"
|
||||
semantic = "tag"
|
||||
distribution = { kind = "cardinality", values = 16, prefix = "dc" }
|
||||
|
||||
[[scenario.tables.columns]]
|
||||
name = "device"
|
||||
type = "STRING"
|
||||
semantic = "tag"
|
||||
distribution = { kind = "cardinality", values = 4, prefix = "device" }
|
||||
|
||||
[[scenario.tables.columns]]
|
||||
name = "val"
|
||||
type = "DOUBLE"
|
||||
semantic = "field"
|
||||
distribution = { kind = "deterministic_wave", min = 0.0, max = 100.0 }
|
||||
|
||||
[[scenario.tables.columns]]
|
||||
name = "ts"
|
||||
type = "TIMESTAMP(3)"
|
||||
semantic = "timestamp"
|
||||
|
||||
[scenario.layout]
|
||||
regions = 1
|
||||
sst_count = 2
|
||||
rows_per_sst = 1048576
|
||||
row_group_size = 8192
|
||||
series_count = 4096
|
||||
start_unix_nanos = 1704067200000000000
|
||||
step_nanos = 15000000000
|
||||
time_range_layout = "non_overlapping_per_sst"
|
||||
series_layout = "timestamp_major"
|
||||
|
||||
# 122 evaluations * 16 dc groups = 1,952 output rows. Each range evaluates
|
||||
# 4,096 series at 122 timestamps (499,712 evaluated series points). This range
|
||||
# crosses the first-SST boundary (1704071025); its initial 5m lookback starts
|
||||
# at 1704069600, ensuring both SSTs are in the requested time span.
|
||||
[[scenario.queries]]
|
||||
name = "grouped_rate_sum_range"
|
||||
kind = "tql"
|
||||
query = "TQL EVAL (1704069900, 1704071715, '15s') sum by (dc) (rate(promql_series_rate_fusion[5m]))"
|
||||
warmup = 3
|
||||
iterations = 20
|
||||
|
||||
[scenario.queries.thresholds]
|
||||
max_candidate_latency_regression_pct = 10
|
||||
|
||||
# Same complete response shape at the latest timestamp: 16 groups.
|
||||
[[scenario.queries]]
|
||||
name = "grouped_rate_sum_instant"
|
||||
kind = "tql"
|
||||
query = "TQL EVAL (1704071715, 1704071715, '15s') sum by (dc) (rate(promql_series_rate_fusion[5m]))"
|
||||
warmup = 3
|
||||
iterations = 9
|
||||
|
||||
[scenario.queries.thresholds]
|
||||
max_candidate_latency_regression_pct = 10
|
||||
|
||||
# Separate diagnostics are intentionally not included in timed samples.
|
||||
[[scenario.queries]]
|
||||
name = "analyze_grouped_rate_sum_range"
|
||||
kind = "tql"
|
||||
query = "TQL ANALYZE VERBOSE (1704069900, 1704071715, '15s') sum by (dc) (rate(promql_series_rate_fusion[5m]))"
|
||||
warmup = 0
|
||||
iterations = 1
|
||||
|
||||
[[scenario.queries]]
|
||||
name = "analyze_grouped_rate_sum_instant"
|
||||
kind = "tql"
|
||||
query = "TQL ANALYZE VERBOSE (1704071715, 1704071715, '15s') sum by (dc) (rate(promql_series_rate_fusion[5m]))"
|
||||
warmup = 0
|
||||
iterations = 1
|
||||
|
||||
# Wider/non-rate path: does not match the direct rate-only fusion candidate.
|
||||
[[scenario.queries]]
|
||||
name = "grouped_sum_over_time_control"
|
||||
kind = "tql"
|
||||
query = "TQL EVAL (1704069900, 1704071715, '15s') sum by (dc) (sum_over_time(promql_series_rate_fusion[5m]))"
|
||||
warmup = 3
|
||||
iterations = 9
|
||||
|
||||
[scenario.queries.thresholds]
|
||||
max_candidate_latency_regression_pct = 10
|
||||
|
||||
# MIN is a distinct aggregation operator, so it must not match the SUM fusion
|
||||
# candidate. It retains the same rate-window work and compact grouped response.
|
||||
[[scenario.queries]]
|
||||
name = "grouped_rate_min_control"
|
||||
kind = "tql"
|
||||
query = "TQL EVAL (1704069900, 1704071715, '15s') min by (dc) (rate(promql_series_rate_fusion[5m]))"
|
||||
warmup = 3
|
||||
iterations = 9
|
||||
|
||||
[scenario.queries.thresholds]
|
||||
max_candidate_latency_regression_pct = 10
|
||||
|
||||
# Complete-table materialization audit: 16 rows, each dc has 131072 rows and
|
||||
# 256 instances, with timestamps spanning the entire two-SST fixture.
|
||||
[[scenario.queries]]
|
||||
name = "validate_materialized_data"
|
||||
kind = "sql"
|
||||
query = "SELECT dc, COUNT(*) AS row_count, COUNT(DISTINCT instance) AS series_count, MIN(ts) AS min_ts, MAX(ts) AS max_ts FROM promql_series_rate_fusion GROUP BY dc ORDER BY dc"
|
||||
warmup = 0
|
||||
iterations = 1
|
||||
|
||||
# Offset gate control, if accepted by this pinned parser; an offset is not fused.
|
||||
[[scenario.queries]]
|
||||
name = "grouped_rate_sum_offset_control"
|
||||
kind = "tql"
|
||||
query = "TQL EVAL (1704069900, 1704071715, '15s') sum by (dc) (rate(promql_series_rate_fusion[5m] offset 15s))"
|
||||
warmup = 3
|
||||
iterations = 9
|
||||
|
||||
[scenario.queries.thresholds]
|
||||
max_candidate_latency_regression_pct = 10
|
||||
Reference in New Issue
Block a user