rust/greptimedb
1
0
Fork 0
mirror of https://github.com/GreptimeTeam/greptimedb.git synced 2026-10-12 23:12:28 +00:00
Code Issues Packages Projects Releases Wiki Activity

test(perf): align BSS reader benchmarks with prior study

Browse Source 
Signed-off-by: discord9 <55937128+discord9@users.noreply.github.com>
This commit is contained in:
discord9
2026-09-08 12:27:09 +08:00
parent 0f87a2e894
commit 40ef21b110
2 changed files with 19 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/perf/README.md
+15 -3
View File
@@ -407,11 +407,23 @@ query latency percentage is
The configured three warmups occur after the initial query validation and before
that query's 15 measured endpoint requests, so query latency is a warmed
frontend/cache measurement. After the datanodes stop, the case also runs seven
iterations of `parquetbench` for up to four inspected SST files and `scanbench`
for their scan groups. Their per-run output and aggregate
iterations of value-only `parquetbench` for up to four inspected SST files and
sequential `scanbench` over the corresponding regions, with parallelism one.
Their per-run output and aggregate
`parquetbench_median_average_ms` and `scanbench_median_average_ms` are under
`targets[].read_bench`; they are quiescent local-file read/scan diagnostics, not
warmed frontend-query latency measurements.
warmed frontend-query latency measurements. Bench averages include iteration one;
no OS cache is dropped, and the driver runs base before candidate.
Historical [PR #8548](https://github.com/GreptimeTeam/greptimedb/pull/8548)
reported storage savings alongside warm-read slowdowns. Its mixed counter/gauge
study used value-only and all-column projections, discarded the first iteration,
and alternated target order. This smaller integer-counter case is not an exact
reproduction: do not compare its inclusive averages directly with that study's
warm medians. For a matching reader comparison, reuse the recorded bench commands
against the stopped data directories, run both projections with eight iterations,
alternate target order, and compare medians of iterations 2–8 from raw output.
Keep post-flush and post-compaction measurements separate.
The case's `-5.0` storage target and `25` query-latency guardrail are experimental
acceptance targets, not observed-benefit claims; do not relax them if a run
tests/perf/query_cases/sst_float_bss/case.toml
+4
View File
@@ -51,6 +51,10 @@ enabled = true
parquetbench = true
scanbench = true
iterations = 7
projection = ["greptime_value"]
parquet_reader = "direct"
scan_scanner = "seq"
parallelism = 1
max_files = 4
[[scenario.queries]]