bot-bottle/docs/research/firecracker-image-remote-store.md

# Firecracker Image Remote Store

**Date:** 2026-07-22  
**Context:** PR #459 (run-dir leak fix) surfaced that committed Firecracker snapshots
currently live only on the host machine. Once a host is wiped or a run-dir is
evicted, a user's preserved bottle is gone. This note investigates a secure remote
store so committed images survive host turnover and can be restored without the
user having to pre-flag which sessions to keep.

## Verdict

Backblaze B2 + Cloudflare CDN is the cost-optimal choice for most deployments.
Cloudflare R2 is the simpler zero-config option at slightly higher storage cost.
Self-hosted MinIO is the right call for air-gapped or on-premises installs.

On the image-size front, zstd-compressing the committed tar before upload
produces roughly a 60–70% reduction with negligible impact on restore latency.
OverlayFS (for in-flight working rootfs, not for the committed artifact) cuts
per-instance disk use to ~10–50 MB per extra bottle sharing the same base.

---

## What Gets Stored

The Firecracker backend produces two artifact types:

| Artifact | Created by | Size | Lifetime |
|----------|-----------|------|---------|
| `rootfs/agent-<digest>/` (dir) | `image_builder.py` → `mke2fs -d` | ~1 GB as ext4 | Cached per Dockerfile hash; evictable |
| `committed/<slug>/rootfs.tar` | `FirecrackerFreezer._freeze` via SSH tar | 500 MB–1 GB | User-preserved; must survive host wipe |

The committed artifact is a tar of the guest's live filesystem streamed out over
SSH (`freezer.py:58–90`). At resume time `launch.py` calls `mke2fs -d` to
rebuild a fresh ext4 from this tar. The tar — not the ext4 — is what needs to be
pushed to remote storage and pulled back at restore time.

The base image cache (`rootfs/agent-<digest>/`) is derivable from the Dockerfile
and can be rebuilt on demand; it is lower priority for remote storage.

---

## Storage Candidates

### Object storage

| Provider | Storage | Egress | Notes |
|----------|---------|--------|-------|
| **Backblaze B2** | $0.006/GB | Free (via Cloudflare Bandwidth Alliance) | Cheapest storage; pairs with Cloudflare CDN to eliminate egress |
| **Cloudflare R2** | $0.015/GB | $0 always | Zero-config egress; no lifecycle transitions (limitation) |
| **Wasabi** | $0.0069/GB | Free (1:1 ratio) | 90-day minimum retention; good for archival; lifecycle evaluated daily |
| **AWS S3** | $0.023/GB | $0.09/GB | Richest lifecycle support; expensive at scale; avoid unless already in AWS |
| **MinIO** (self-hosted) | Host cost only | None | S3-compatible; best for private/on-prem deployments |

**B2 + Cloudflare CDN** is effectively $0.006/GB with zero egress — about 18×
cheaper than S3 for restore-heavy workloads. **R2** is the zero-config choice
($0 egress by default, no Bandwidth Alliance pairing needed) at a slightly
higher storage rate.

**Cloudflare R2's missing lifecycle support** is the main caveat: auto-eviction
rules (evict images older than N days) cannot currently be expressed natively in
R2. Wasabi and S3 both support declarative lifecycle policies.

### Retention policy recommendation

The comment proposes:
- Retain images for ~1 week by default
- Warn when approaching a capacity threshold
- Auto-evict oldest images once threshold is exceeded

This maps cleanly to an application-level policy (not a provider lifecycle rule),
which avoids the R2 limitation and works consistently across providers:

1. On `commit`: upload tar, record `(slug, size_bytes, uploaded_at)` in a local
   or remote manifest file.
2. On startup / on `list`: scan the manifest, warn if total stored size exceeds
   e.g. 80% of the configured threshold.
3. On eviction run (CLI or cron): delete objects older than `retention_days`
   (default 7) that push total over `max_capacity`; oldest-first.

This keeps the policy logic in bot-bottle and the storage provider as a dumb
object store — no vendor-specific lifecycle API required.

---

## Image Size Reduction

### Current artifact sizes

A typical committed tar for a Claude Code agent image is 500 MB–1 GB uncompressed.
The per-run ext4 (copy of the base, written at `start`) adds another ~1 GB of
local disk. The leak fix in this PR addresses the ext4 copies; the remote store
addresses the committed tars.

### Compression

zstd compression of the committed tar before upload is the highest-leverage
single change:

| Codec | Typical size (1 GB rootfs) | Compress speed | Decompress speed |
|-------|---------------------------|---------------|-----------------|
| gzip | 350–430 MB | ~100 MB/s | ~500 MB/s |
| **zstd (default)** | **330–360 MB** | **~400 MB/s** | **~2 GB/s** |
| xz | 290–320 MB | ~20 MB/s | ~200 MB/s |

**zstd is the best trade-off**: 65–67% size reduction, near-instantaneous
decompression. The `tar` call in `freezer.py` could pipe through `zstd` before
writing to disk and to the remote; `resume` decompresses on the way back. A
`.tar.zst` suffix marks compressed artifacts so old tars remain restorable
without the codec.

### SquashFS for the base image cache

The `rootfs/agent-<digest>/` directory (the buildah-exported tree) is rebuilt by
`image_builder.py` and turned into per-run ext4 by `mke2fs -d`. Storing the
base as a SquashFS image instead of a flat directory tree would reduce it from
~1 GB to ~330–360 MB and make the cache remote-friendly. Firecracker does not
directly boot SquashFS, but the existing `mke2fs -d` path reads a directory tree
— a SquashFS mount could serve as the source. This is a larger change and lower
priority than tar compression.

### OverlayFS for per-run rootfs

Multiple simultaneous bottles sharing the same agent image today each get a full
`mke2fs -d` copy (~1 GB). OverlayFS (read-only base + writable sparse overlay)
would reduce this to ~10–50 MB per instance beyond the first:

- Mount the base image directory as read-only lower layer
- Attach a sparse ext4 or tmpfs writable layer per bottle
- Pass the merged overlay to Firecracker as the block device

E2B's public write-up on Firecracker + OverlayFS confirms this approach works
at scale. The `launch.py` changes would be non-trivial (device mapper or
`fuse-overlayfs` plumbing), so this is a follow-up rather than a prerequisite
for the remote store.

---

## Recommended Approach

**Phase 1 — remote store with zstd (tight scope, actionable now)**

1. Add `--zstd` to the `tar` call in `FirecrackerFreezer._freeze`; name the
   artifact `rootfs.tar.zst`. Keep uncompressed restore path for legacy tars.
2. Add a `bb firecracker upload <slug>` / `bb firecracker pull <slug>` pair that
   pushes/fetches the compressed tar to the configured object store (S3-compatible
   API, so B2, R2, MinIO, and Wasabi all work with the same client).
3. Store a `manifest.json` in the bucket (or a local mirror) tracking slug →
   `{size, uploaded_at}`. Use it for threshold warnings and eviction.
4. Default retention: 7 days, configurable via `firecracker.image_retention_days`
   in `~/.config/bot-bottle/config.toml` (or equivalent).
5. Warn at 80% of `max_capacity` (default e.g. 50 GB); evict oldest on commit
   once at 100%.

**Storage recommendation:** Cloudflare R2 for hosted deployments (zero egress,
zero config), MinIO for private/on-premises.

**Phase 2 — base image cache compression**

Compress the `agent-<digest>` cache dir as a `.tar.zst` to save ~65% on repeated
image uploads. Low urgency since the base image is rebuildable.

**Phase 3 — OverlayFS per-run disk**

Replace the full per-run ext4 copy with an OverlayFS sparse layer. Largest disk
impact (~90–95% savings per concurrent bottle) but highest implementation
complexity. Track as a separate PRD.

---

## Open Questions

- Does the host have a configured object-store credential path, or should the
  remote store be an opt-in with an explicit `bb config set image-store.url ...`?
- Should `commit` automatically upload, or should upload be an explicit step to
  avoid surprise egress?
- What is the acceptable cold-start latency for a restore from remote? A 330 MB
  zstd tar at 100 Mbit/s takes ~26 s; at 1 Gbit/s, ~2.6 s. This bounds the
  retention strategy (evict from local after successful upload vs keep local copy).