From 54f26ea81539370faed271e703977c5927de6f45 Mon Sep 17 00:00:00 2001
From: discord9 <55937128+discord9@users.noreply.github.com>
Date: Tue, 8 Sep 2026 13:38:47 +0800
Subject: [PATCH] test(perf): isolate BSS user SSTs and benchmark every file

Signed-off-by: discord9 <55937128+discord9@users.noreply.github.com>
---
 tests/perf/README.md                           | 14 +++++++++-----
 tests/perf/query_cases/sst_float_bss/case.toml |  2 +-
 2 files changed, 10 insertions(+), 6 deletions(-)

diff --git a/tests/perf/README.md b/tests/perf/README.md
index c22e7f8a546..cd129575508 100644
--- a/tests/perf/README.md
+++ b/tests/perf/README.md
@@ -406,11 +406,15 @@ query latency percentage is
 `(candidate_latency_ms_median - base_latency_ms_median) / base_latency_ms_median * 100`.
 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 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
+frontend/cache measurement. The case restricts storage inspection to
+`data/greptime/public`, excluding `greptime_private` SSTs. After the datanodes
+stop, it also runs seven iterations of value-only `parquetbench` for all
+inspected SST files and sequential `scanbench` over the corresponding regions,
+with parallelism one. The three explicit flushes yielded six SSTs in the observed
+run—three 1,105,920-row files and three 368,640-row files—because storage split
+each flush; this is an observation, not a guaranteed file layout. 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. Bench averages include iteration one;
 no OS cache is dropped, and the driver runs base before candidate.
diff --git a/tests/perf/query_cases/sst_float_bss/case.toml b/tests/perf/query_cases/sst_float_bss/case.toml
index c18dd97e840..63819158d24 100644
--- a/tests/perf/query_cases/sst_float_bss/case.toml
+++ b/tests/perf/query_cases/sst_float_bss/case.toml
@@ -41,6 +41,7 @@ pending_rows_flush_interval = "1s"
 max_batch_rows = 100000
 
 [scenario.remote_write.storage]
+root_suffix = "data/greptime/public"
 column = "greptime_value"
 min_files = 1
 min_files_with_column = 1
@@ -55,7 +56,6 @@ projection = ["greptime_value"]
 parquet_reader = "direct"
 scan_scanner = "seq"
 parallelism = 1
-max_files = 4
 
 [[scenario.queries]]
 name = "sum_all_values"