ARCHIL
All comparisons

Archil vs Amazon S3 Files

Pick Amazon S3 Files for an AWS-native, IAM-governed NFS view of an existing S3 bucket, with S3 Versioning as history. Pick Archil for isolated, persistent agent workspaces with millisecond branches, a strongly consistent S3 API, and GCS or R2 as well as S3.

Where Archil stands out

  • An isolated workspace

    Each sandbox is a Firecracker microVM whose files survive stops and restarts.

  • Branches in milliseconds

    Checkpoint and branch a disk in tens of milliseconds, any size.

  • Any cloud, any bucket

    Disks mount anywhere and sync with S3, GCS, R2 or S3-compatible stores.

Archil and Amazon S3 Files, side by side

What it is
Archil: Managed POSIX file system with an S3 API, plus Firecracker sandboxes
Amazon S3 Files: Shared NFS for an S3 bucket, built on Amazon EFS
Compute
Archil: Any cloud or machine, Kubernetes, Archil sandboxes
Amazon S3 Files: AWS: EC2, ECS, EKS, Fargate, Batch, Lambda
Buckets
Archil: S3, GCS, R2 and S3-compatible
Amazon S3 Files: Amazon S3, with Versioning enabled
When file writes reach the bucket
Archil: In the background, after the write is durable on Archil
Amazon S3 Files: After 60 s without writes to the file, as a new object version
Reading fresh writes over S3
Archil: Disk's own S3 API is strongly consistent
Amazon S3 Files: Once the change is exported
Snapshots and branches
Archil: Checkpoint p50 13 ms, branch p50 33 ms
Amazon S3 Files: No native branching; S3 Versioning keeps object history
Running agent code
Archil: Firecracker sandboxes on a disk
Amazon S3 Files: Not included; mounts into your AWS compute

Measured on one EC2 host

npm ci took 12.4 s on Archil and 71.4 s on Amazon S3 Files; git clone took 19.8 s and 107.5 s. Local disk took 10.7 s and 15.5 s.

MethodOne AWS c7gn.2xlarge (arm64) in us-east-1, every file system mounted on the same host. Archil client 0.10.1 with an exclusive mount of a disk in aws-us-east-1. JuiceFS Community Edition 1.4.2 with data in Amazon S3 and metadata in ElastiCache for Redis 7.1 in the same Availability Zone, mounted with its defaults (1 s attribute cache, local disk cache) and separately with --writeback. Amazon S3 Files through amazon-efs-utils 3.0.1. Local disk is EBS gp3. Inputs sat on local disk, so GitHub and the npm registry aren't in the times: git clone of the Django repository from a local mirror, and npm ci --prefer-offline --ignore-scripts of a Next.js 15 app (18,602 files) from a warm local cache. Each run is timed until its writes are durable: archil sync on Archil, sync -f elsewhere, and for JuiceFS writeback also until every block staged on local disk had uploaded to S3. Medians of 3 runs. Run it yourself.

npm ci, 18,602 files, until durable

Lower is better
  • Archil12.4 s
  • Amazon S3 Files1 min 11 s
  • Local disk (EBS gp3), for reference10.7 s

git clone Django, until durable

Lower is better
  • Archil19.8 s
  • Amazon S3 Files1 min 48 s
  • Local disk (EBS gp3), for reference15.5 s
MethodOne AWS c7gn.2xlarge (arm64) in us-east-1 running two independent mounts of each file system. Archil client 0.9 with a disk in aws-us-east-1; Amazon S3 Files through amazon-efs-utils 3.0.1 (NFS 4.2), with a mount target in the same Availability Zone and a versioned bucket; Mountpoint for Amazon S3 1.22.3 with --allow-overwrite and --allow-delete. Medians of 10 to 200 operations per row; new-file-system rows are medians of 3 to 5 trials. Method.

New file system to first durable write

Lower is better
  • Archil0.34 s
  • Amazon S3 Files5 min 27 s

    about 5 minutes of that is creating the mount target

Mount an existing file system

Lower is better
  • Archil44 ms
  • Amazon S3 Files590 ms

Create a file, write 4 KiB, fsync, close

Lower is better
  • Archil4.2 ms
  • Amazon S3 Files8.5 ms

Second client opens and reads that file

Lower is better
  • Archil1.2 ms
  • Amazon S3 Files3.2 ms

Second client sees a newly created file

Lower is better
  • Archil30 s

    default client cache expiry

  • Amazon S3 Files30 s

    NFS directory attribute cache

Read 1 GiB sequentially

Higher is better
  • Archil902 MB/s
  • Amazon S3 Files753 MB/s

How they differ

Architecture

S3 Files turns an existing bucket into shared NFS without moving data. It's built on EFS, which loads the active working set onto fast storage, and it removes S3-to-EFS staging and sync scripts.

Archil serves disks from SSD replicated across Availability Zones, over FUSE, with your bucket or Archil-managed storage behind it.

Consistency between files and objects

S3 API changes show up in S3 Files. When both sides change a file, the bucket wins and the file system's version moves to lost-and-found.

On Archil, a write is durable when fsync returns. Readers that need the newest data use Archil's S3 API rather than the bucket.

Isolation for agent code

S3 Files inherits the isolation and IAM identity of the compute you mount it in.

An Archil sandbox is a microVM with its own kernel, private until you expose a port; files, packages and, when paused, running processes survive.

Sharing and branching

With S3 Files, parallel attempts are kept apart with separate paths.

An Archil disk mounts on up to 5,000 clients in shared or conditional mode, and each attempt can start from its own branch. Branching disks synced to your own bucket is an Enterprise feature; see plan features.

Multi-tenant workspaces

S3 Files access points can enforce a POSIX identity and confine a client to a root directory, alongside IAM and POSIX permissions.

Archil puts the tenant boundary at the disk, with its own credentials, and the sandbox. Billing on active data helps when workspace sizes vary.

Pricing model

S3 Files bills through AWS with the bucket. See Amazon S3 pricing.

Archil bills active data per GiB-month, with a lower archive rate on Archil-managed disks; on synced disks you pay your provider for the bucket. See Archil pricing.

Which one should you use?

Use Amazon S3 Files when

  • compute is on AWS and you want IAM-governed file access to existing S3 data
  • you want S3 Versioning as the history of every change
  • your compute already isolates the code, and you want no other vendor

Use Archil when

  • each agent needs an isolated sandbox with a persistent environment
  • agents need millisecond checkpoints and branches
  • data is in GCS or R2 too, or compute runs outside AWS
  • S3 readers must see file writes immediately

Questions people ask

More comparisons

Or see every comparison, grouped by sandboxes, file systems and object storage.