Comparison
Mountpoint for Amazon S3 vs Training Pipes
AWS's Mountpoint is a faster, more stable FUSE client for S3 than s3fs. Training Pipes is a different architecture entirely — a regional caching NFS gateway. Here's when each wins.
Short answer
Mountpoint for Amazon S3 is AWS's official FUSE-based filesystem client for S3. It's significantly better than s3fs for read-heavy sequential workloads, but it still runs as a per-node FUSE process with limited POSIX support (no random writes, no atomic rename in general, no shared cache across nodes). Training Pipes takes a different approach: a regional gateway with a shared NVMe cache exposes your S3 bucket over standard NFSv4 or SMB, giving you full POSIX semantics and a cache that scales across your cluster.
Feature-by-feature comparison
| Feature | Mountpoint | Training Pipes |
|---|---|---|
| Protocol | FUSE → S3 API | NFSv4 / SMB → gateway → S3 |
| Read support | Yes | Yes |
| Write support Checkpoints often need atomic rename, which Mountpoint doesn't support. | Sequential, single writer, no modify | Full POSIX writes |
| Atomic rename | No | Yes |
| Random writes | No | Yes |
| Shared cache across nodes Per-node cache fragmentation wastes bandwidth at multi-node scale. | No | Yes |
| Multi-cloud | No | Yes |
| Container support | Via CSI, privileged pods | Standard NFS CSI |
| SMB access | No | Yes |
| Transport encryption | TLS to S3 | WireGuard + TLS |
| Pricing | Free (+ S3 request costs) | Plan tier + storage |
When to use Mountpoint
- Read-only, sequential-access workloads on AWS.
- Single-node jobs that don't need shared caching.
- Analytics or batch jobs where POSIX writes aren't required.
- When cost is the overriding concern and GPU idle time is cheap.
When to use Training Pipes
- ML training that writes checkpoints needing atomic rename.
- Multi-node training where cache sharing pays for itself.
- Non-AWS clouds or mixed environments.
- Workloads where predictable tail latency matters.
- Any case where FUSE container plumbing has caused you pain.
The verdict
Mountpoint is a fine choice for read-only, single-node batch jobs on AWS when you don't need POSIX writes. It's free and AWS-native. Training Pipes wins for multi-node training, workloads that write (checkpoints!), multi-cloud environments, and any scenario where per-node FUSE caches waste bandwidth and money. Cost-wise, the tradeoff is free-but-limited (Mountpoint) vs paid-but-full-featured (Training Pipes).
Frequently asked questions
- Can I write checkpoints to Mountpoint?
- With caveats. Mountpoint supports sequential writes of new files, but doesn't support modifying existing files or atomic rename. Most checkpointing code uses temp-file + rename for consistency, which breaks on Mountpoint. Training Pipes supports the full POSIX pattern.
- Is Mountpoint really free?
- The Mountpoint software itself is free and open source. You still pay full S3 request and egress costs, and for training workloads with high read counts, those bills can exceed Training Pipes's all-in plan price.
- Does Training Pipes work only with AWS S3?
- No. Training Pipes works with AWS S3, Google Cloud Storage, Azure Blob, Cloudflare R2, Wasabi, Backblaze B2, MinIO, or any S3-compatible endpoint. Mountpoint is AWS-only.
- How does cache sharing actually help?
- With Mountpoint, every node mounting the bucket has its own cache. Eight GPU nodes training on the same dataset = eight independent fetches of every shard. With Training Pipes, the gateway has one cache shared across all eight nodes, so each shard is fetched from S3 once and served from NVMe to every client.
Related reading
Ready to try Training Pipes?
Spin up a regional NFS file system in five minutes. Free tier available.