PXB-3862 : TEMPORARY: bundle the validation benchmark kit

https://perconadev.atlassian.net/browse/PXB-3862

Ships pxb-bench (the team's page-tracking benchmark tool, with numeric
gap modes enabled) and a from-bare-machine runbook inside the source
tree, so validating the measured bridge budget on a new machine needs
exactly one git clone. Runs the same 5..100% sweep as the original
PXB-3862 report plus a pin-10 column, and collects the calibration and
refusal log lines that explain every grouping decision.

NOT FOR MERGE: this commit is a distribution vehicle for the multi-
machine validation campaign and will be dropped from the branch before
the pull request.

Author: satya-bodapati

storage/innobase/xtrabackup/test/benchmarks/README-PXB-3862-benchmark.md

@@ -0,0 +1,178 @@
+# PXB-3862 read-request-cost validation — from a bare machine
+
+Everything needed is in this checkout: the xtrabackup source (this
+repo, branch PXB-3862-8.4) and the benchmark tool (pxb-test.tar.gz in
+this directory — jinyou's pxb-bench with numeric gap modes enabled;
+17 unit tests included). This file is the complete runbook.
+
+## Quick start — build, then one command
+
+```bash
+cmake -S . -B build -DBUILD_CONFIG=xtrabackup_release -DWITH_MAN_PAGES=OFF
+cmake --build build --target xtrabackup -j"$(nproc)"
+./storage/innobase/xtrabackup/test/benchmarks/run_benchmark.sh
+```
+
+run_benchmark.sh does everything below automatically: finds the built
+binary, deploys (or reuses) a dbdeployer sandbox named pxb3862_bench,
+installs the mysqlbackup component, loads test.sbtest1 with sysbench
+when missing (reused when present — reruns after a rebuild skip
+straight to the sweep), unpacks pxb-bench, generates config.toml,
+picks run-vs-restart mode from table size vs buffer pool, runs the
+sweep and prints the path of one results tarball to send back.
+Defaults: ROWS=10000000, PERCENTS=5:100:5, GAPS=none,0,auto,10,
+WORKLOADS=scattered. Override anything per invocation, e.g. on
+jinyou's rig:
+
+```bash
+SANDBOX=~/sandboxes/msb_8_4_8 ./run_benchmark.sh          # reuse 100M-row sandbox
+# or provision fresh from the JIRA layout:
+SANDBOX_BINARY=/opt/percona_server MYSQL_VERSION=8.4.8 ROWS=100000000 ./run_benchmark.sh
+```
+
+Only the packages in step 0 are prerequisites. The manual steps below
+remain the reference for what the script does (and for debugging it).
+
+Path variables used below — set them to taste:
+
+    SRC=~/WORK/pxb-3862-8.4        # this repo checkout
+    SB=~/sandboxes/msb_ps8_4_x     # the sandbox (created in step 3)
+
+## 0. Packages (once, ~5 min)
+
+```bash
+# Debian/Ubuntu (adjust for RHEL):
+sudo apt install -y build-essential cmake bison libssl-dev \
+    libncurses-dev libaio-dev libcurl4-openssl-dev libudev-dev \
+    libgcrypt-dev libev-dev libproc2-dev \
+    python3 python3-pip sysbench wget
+```
+
+dbdeployer (single static Go binary; assets are named
+dbdeployer-<ver>.linux.tar.gz):
+
+```bash
+VER=1.76.0    # or check https://github.com/datacharmer/dbdeployer/releases
+mkdir -p ~/bin
+wget -q https://github.com/datacharmer/dbdeployer/releases/download/v$VER/dbdeployer-$VER.linux.tar.gz
+tar xzf dbdeployer-$VER.linux.tar.gz
+mv dbdeployer-$VER.linux ~/bin/dbdeployer && chmod +x ~/bin/dbdeployer
+export PATH=$PATH:~/bin        # add to ~/.bashrc for permanence
+
+# one-time init: creates ~/opt/mysql (binaries), ~/sandboxes, and
+# fetches the tarball download index
+dbdeployer init
+```
+
+## 1. Source (one clone brings everything)
+
+```bash
+git clone -b PXB-3862-8.4 https://github.com/satya-bodapati/percona-xtrabackup.git $SRC
+```
+
+## 2. Build xtrabackup (~15 min)
+
+```bash
+cd $SRC
+cmake -S . -B build -DBUILD_CONFIG=xtrabackup_release -DWITH_MAN_PAGES=OFF
+cmake --build build --target xtrabackup -j"$(nproc)"
+# sanity - MUST print the option:
+build/runtime_output_directory/xtrabackup --help | grep page-tracking-merge-gap
+```
+
+## 3. Server sandbox (~5 min)
+
+```bash
+# download a Percona Server 8.4 tarball via dbdeployer's index
+# (or wget one from percona.com/downloads and skip this line):
+dbdeployer downloads get-by-version 8.4 --flavor percona --newest
+
+# make it deployable, then deploy one sandbox:
+dbdeployer unpack Percona-Server-8.4*.tar.gz --prefix ps
+dbdeployer deploy single ps8.4.<x>       # <x> from: dbdeployer versions
+SB=~/sandboxes/msb_ps8_4_<x>             # dbdeployer prints this path;
+                                         # port: grep ^port $SB/my.sandbox.cnf
+# recommended for large tables:
+printf 'innodb_buffer_pool_size=4G\ninnodb_redo_log_capacity=4G\n' >> $SB/my.sandbox.cnf
+$SB/restart
+```
+
+## 4. Load the table (100M rows ~ 24GB, ~1h; scale down freely)
+
+```bash
+PORT=$(awk -F= '/^port/{gsub(/ /,"");print $2;exit}' $SB/my.sandbox.cnf)
+$SB/use -e "CREATE DATABASE IF NOT EXISTS test"
+sysbench oltp_insert --tables=1 --table-size=100000000 \
+    --create_secondary=off --mysql-db=test --db-driver=mysql \
+    --mysql-host=127.0.0.1 --mysql-port=$PORT \
+    --mysql-user=msandbox --mysql-password=msandbox prepare
+```
+
+## 5. Install the benchmark tool (~2 min)
+
+```bash
+tar xzf $SRC/storage/innobase/xtrabackup/test/benchmarks/pxb-test.tar.gz -C $SB
+cd $SB/pxb-test && python3 -m pip install --user -e .
+```
+
+## 6. config.toml — edit ONLY what has no CLI flag (~1 min)
+
+In `$SB/pxb-test/config.toml` (marked CHANGE-ME):
+
+```toml
+[sandbox]    port = <the $PORT from step 4>
+[xtrabackup] full_binary        = "<absolute $SRC>/build/runtime_output_directory/xtrabackup"
+             incremental_binary = "<absolute $SRC>/build/runtime_output_directory/xtrabackup"
+[workload]   total_rows = <rows loaded in step 4>
+```
+
+Sweep parameters (percents/gaps/workloads) are NEVER edited here — they
+are command-line flags.
+
+## 7. Run the sweep (overnight at 100M rows)
+
+```bash
+pxb-bench --workdir $SB run \
+    --percents 5:100:5 --gaps none,0,auto,10 --workloads scattered
+```
+
+Same densities as the original PXB-3862 chart, so cells overlay it
+directly; the `10` column measures what bridging buys wherever auto
+declines.
+
+SMALL-TABLE caveat: if the table fits in the buffer pool, dirty pages
+never flush by eviction and the tracked set stays empty. Then drive the
+individual steps with a restart per density instead of `run`:
+
+```bash
+pxb-bench --workdir $SB full-backup
+for pct in 5 10 15 20; do
+  pxb-bench --workdir $SB change-data $pct --workload scattered
+  pxb-bench --workdir $SB restart
+  for gap in none 0 auto 10; do
+    pxb-bench --workdir $SB inc-backup $pct $gap --workload scattered
+  done
+done
+```
+
+## 8. Send back one tarball
+
+```bash
+cd $SB
+grep -h "calibrated storage\|typical gap\|request reduction" \
+    pxb-test/logs/scattered_*_gap_auto.log > pxb-test/results/decision_lines.txt
+tar czf /tmp/pxb3862_results.tgz pxb-test/results/ pxb-test/logs/
+```
+
+## 9. Predictions (split-disk, ~165us / ~1.5GB/s class storage)
+
+| density | old build (÷2, budget ~125KB) | this build (÷1.5, ~160KB) |
+|---|---|---|
+| 5%      | refused (correct, beats full scan) | identical |
+| **10%** | **refused -> 28.4s**               | **bridged -> ~17s = pin-10** |
+| 15-100% | bridged (wins/ties)                | identical |
+
+On fast shared-disk NVMe the 10% verdict is expected to be the
+OPPOSITE (refusal, with pin-10 measurably slower) — each machine
+correctly for its own storage; the calibration and refusal log lines
+state each decision's numbers.