schelk

---

name: schelk description: Use this skill when a user wants to set up, initialize, operate, or troubleshoot schelk: a Linux-only CLI for repeatable benchmarking on large stateful datasets using a virgin block device, a scratch block device, and dm-era metadata on a ramdisk. Use it for installation, prerequisite checks, init-new, init-from, mount, recover, promote, full-recover, status, and crash or reboot recovery.

schelk

schelk is a Linux-only benchmarking helper for large on-disk datasets such as database or reth datadirs. It keeps a pristine virgin block device, runs benchmarks on a scratch block device, and uses dm-era metadata on a ramdisk to learn which blocks changed. Recovery then copies only the changed blocks from virgin back to scratch, so the benchmark loop is fast without running on top of CoW or overlay machinery.

Use this skill when the user wants to install schelk, prepare baseline volumes, run the mount -> benchmark -> recover loop, promote a new baseline, or recover from an unsafe state.

Start From The URL

Do not assume the repository is already cloned locally.

If the user points you at this skill by URL, treat this as the entrypoint:

• schelk SKILL.md

If the repo is not available locally and you need the source or docs, clone it first:

git clone https://github.com/tempoxyz/schelk.git
cd schelk

Read These First

When behavior is unclear, prefer these project documents over guesswork:

• README.md: overview, prerequisites, example commands, limitations.
• principles.md: safety and UX principles.
• spec.md: intended command semantics.
• user-stories.md: representative workflows and failure cases.

Core Rules

• Treat init-new, init-from, full-recover, and promote as destructive.
• Do not guess device paths. Have the user confirm the exact --virgin, --scratch, --ramdisk, and --mount-point before destructive commands.
• Refuse to proceed on non-Linux hosts. This repository intentionally fails to build on macOS and Windows.
• All operational commands require root privileges.
• Do not mount or write to the virgin or scratch devices outside schelk.
• Use --no-copy only if the user explicitly states that virgin and scratch are already byte-identical.
• After a host reboot or power loss while mounted, assume incremental recovery is unsafe. Use full-recover.
• If the user wants multiple schelk instances in parallel, assign a unique --dm-era-name and a separate --state-path for each instance.
• Do not overpromise perfect device reset semantics. schelk restores logical blocks, not NVMe controller internal state.

Inputs You Need

Before setup, gather:

• Linux host with sudo or root access.
• Two equal-size block devices:
  • virgin: pristine benchmark baseline.
  • scratch: writable working copy used for benchmark runs.
• One ramdisk device for dm-era metadata, commonly /dev/ram0.
• A mount point such as /schelk.
• Filesystem type:
  • init-new always creates ext4.
  • init-from requires --fstype.
• Optional: mount options, non-default granularity, custom --state-path, custom --dm-era-name.

Defaults worth knowing:

• Granularity defaults to 4096 bytes and should only be changed deliberately.
• Granularity must be a multiple of 512 bytes.
• The default dm-era name is bench_era.

The default state file is /var/lib/schelk/state.json. Use --state-path or SCHELK_STATE only when the user needs an alternate state location.

Environment Checks

Before destructive work, verify the environment and explain any missing prerequisites before doing anything else.

uname -s
id -u
which cargo dmsetup era_invalidate mkfs.ext4
blockdev --getsize64 /dev/virgin /dev/scratch /dev/ram0
ls -l /dev/virgin /dev/scratch /dev/ram0

Interpretation:

• uname -s must be Linux.
• id -u should be 0 for operational commands.
• mkfs.ext4 is required for init-new.
• dmsetup and era_invalidate are required for mount, recover, and promote.
• Virgin and scratch must be different block devices of equal size.
• The ramdisk must be large enough for the chosen volume size and granularity.

Install

If the repository is already cloned locally, from the repository root:

cargo install --path .

If the repository is not cloned locally, install directly from GitHub:

cargo install --git https://github.com/tempoxyz/schelk.git

For development builds from a local checkout:

cargo build --release

If the current machine is not Linux, stop and tell the user they need a Linux host for both build and runtime.

Ramdisk Setup

Typical setup:

sudo modprobe brd rd_size=4194304

rd_size is in KiB, so 4194304 is 4 GiB. This is only a starting point; trust schelk's runtime checks over rules of thumb.

Choose The Right Initialization

Use init-new when both volumes can be destroyed and a fresh ext4 baseline should be created:

sudo schelk init-new \
  --virgin /dev/virgin \
  --scratch /dev/scratch \
  --ramdisk /dev/ram0 \
  --mount-point /schelk

Use init-from when the virgin volume already contains the prepared dataset:

sudo schelk init-from \
  --virgin /dev/virgin \
  --scratch /dev/scratch \
  --ramdisk /dev/ram0 \
  --mount-point /schelk \
  --fstype ext4

Use --no-copy only when the user explicitly asserts that both volumes are already identical:

sudo schelk init-from \
  --virgin /dev/virgin \
  --scratch /dev/scratch \
  --ramdisk /dev/ram0 \
  --mount-point /schelk \
  --fstype ext4 \
  --no-copy

Do not add -y unless the user explicitly wants non-interactive execution or has already confirmed the destructive action with the exact device paths.

Normal Operating Loop

Start with status if state may already exist:

sudo schelk status

Mount the scratch volume with dm-era tracking:

sudo schelk mount

Run the system under test against the configured mount point.

Recover incrementally afterward:

sudo schelk recover

If unmount is blocked and the user accepts killing blockers:

sudo schelk recover --kill

Important behavior:

• recover is a safe no-op when schelk is not mounted.
• mount verifies the superblock hash of virgin and scratch before proceeding.
• Mount options and filesystem type come from the saved schelk state.

Promote A New Baseline

Use this when the current scratch state should become the new virgin baseline, such as after a schema migration or one-time snapshot load:

sudo schelk mount
# run the migration or snapshot load
sudo schelk promote

If unmount is blocked and the user accepts killing blockers:

sudo schelk promote --kill

promote is destructive. It permanently overwrites the virgin device.

Full-Recover Fallback

Use:

sudo schelk full-recover

Prefer full-recover when:

• the host rebooted or lost power while schelk was mounted;
• state says mounted but the dm-era device is gone;
• virgin or scratch integrity checks show the baseline is no longer trustworthy;
• a previous --no-copy assumption was wrong;
• the user wants to completely reset scratch from virgin.

This overwrites all data on scratch.

Decision Guide

• The user needs a fresh empty baseline: use init-new.
• The user already prepared the baseline volume: use init-from.
• The user wants fast repeatable benchmark runs: use mount, run the workload, then recover.
• The user wants the current scratch state to become future baseline: use promote.
• The machine rebooted while mounted: use full-recover, then continue with mount.
• recover says nothing is mounted: treat it as success, not failure.

Agent Behavior

• If the user only gives you the skill URL, start from that URL and do not assume a local checkout exists yet.
• If you need source context and the repo is absent locally, clone https://github.com/tempoxyz/schelk.git.
• Explain the exact command before any destructive step.
• Prefer schelk status first when state may already exist.
• Keep the benchmark workflow transparent: tell the user whether you are initializing, mounting, recovering, promoting, or doing a full reset.
• Preserve schelk's assumptions: the virgin and scratch devices are dedicated to schelk once initialized.
• If a command fails because prerequisites are missing, stop and surface the missing tool or environmental issue directly instead of improvising around it.

Example User Requests

This skill should trigger for prompts like:

• "Use schelk to set up two NVMe volumes and /dev/ram0 for benchmarking."
• "Adopt this prepared reth datadir as the schelk baseline."
• "Mount schelk, run the benchmark, and recover afterward."
• "Promote the migrated scratch volume so future runs start from the new baseline."
• "schelk says state is inconsistent after a reboot; recover it safely."