Overview
What the S3 integration does
The S3 integration backs up an S3 bucket, or one prefix of it, into encrypted, deduplicated Plakar snapshots, reading each object through the S3 API with list and get permissions only. It restores a snapshot, or a single prefix, into any bucket on any S3-compatible endpoint, and it can host the Plakar store itself in a bucket. Plakar runs on any host that reaches the endpoint, with an access key pair.
The s3 package provides three connectors. A source reads a bucket, a
destination writes objects back, and a store keeps the encrypted backup data of
any source, from a file server to a database, in a bucket with the storage class
you choose.
The connectors speak the plain S3 API, so they work with any service that speaks it, such as Amazon S3, Scaleway, OVHcloud, MinIO, Ceph RGW or Garage.
$ plakar pkg add s3
$ plakar source add assets \
s3://s3.fr-par.scw.cloud/acme-assets \
access_key=SCWXXXXXXXXXXXXXXXXX \
secret_access_key=xxxxxxxx-xxxx-xxxx
$ plakar at /var/backups backup "@assets"
e41b7c09: OK ✓ /exports/2026-11-13.parquet
e41b7c09: OK ✓ /exports/2026-11-14.parquet
e41b7c09: OK ✓ /exports
e41b7c09: OK ✓ /media/hero.webp
e41b7c09: OK ✓ /media
e41b7c09: OK ✓ /
e41b7c09: backup completed without errors in 9m03.518s
$ plakar at /var/backups ls
2026-11-13T03:00:02Z 7aa2f5d1 412 GiB 8m51s /
2026-11-14T03:00:03Z e41b7c09 415 GiB 9m3s /
$
Why Plakar
Why use Plakar for S3 bucket backup
Off-provider S3 backups without paying egress for every copy
Run Plakar next to the bucket and only new, compressed chunks leave the provider. A copy at another provider no longer costs a full transfer every night.
Read-only keys on the source bucket
A backup lists the bucket and reads its objects, nothing else. The source
credential needs s3:ListBucket and s3:GetObject, and the keys that write the
store are a different pair that the application never sees.
Back up S3 from any endpoint to any other
Source, destination and store are configured separately. Back up a bucket at one provider, keep the store at a second one, and restore into a third, a self-hosted MinIO or Ceph cluster included.
Restore one S3 prefix, not the whole bucket
Pick a prefix in a snapshot and only those objects are written back, under their original keys or under a new prefix next to the live data.
S3 backups verified apart from the provider
Every chunk of a snapshot is addressed by its content, and plakar check
verifies a snapshot end to end from the backup host, whatever the storage behind
the store reports.
Recovery scenarios
S3 backup and recovery scenarios
An AI agent empties an S3 prefix to cut storage costs
An ops agent asked to lower the storage bill runs
aws s3 rm s3://acme-assets/exports/ --recursive with the operator’s
credentials, then adds a lifecycle rule that expires noncurrent versions after
one day. The same keys purged the versions. With a snapshot taken before the
session, in a store whose passphrase the agent never saw,
plakar diff -recursive lists every removed key and a restore of /exports
writes them back. The agent’s other cleanups stay in place.
A leaked S3 key re-encrypts the bucket with SSE-C
A key with write access to the production bucket leaks from a CI log. The
attacker copies every object onto itself with CopyObject and an SSE-C key only
they hold. The objects are still there, unreadable. The Plakar store sits in
another bucket, ideally at another provider, under credentials the application
never held, and restores the objects as they were before the first unknown
request.
An S3 lifecycle rule deletes more than intended
A rule meant for tmp/ is saved with an empty filter and starts expiring
objects across the whole bucket. Once the noncurrent versions pass their window,
the bucket has nothing left to roll back to. Nightly snapshots keep each state
of the bucket for as long as your Plakar retention says, and deduplication
stores unchanged objects only once.
Deleted objects
How to restore deleted S3 objects
When objects are deleted, overwritten or encrypted with a key you do not hold, the snapshot from before the event brings them back. Restore them next to the live data first, then move them into place once checked.
Find the last good snapshot. plakar ls lists the snapshots with their
dates, and plakar locate shows which snapshots hold a given object name.
See what changed. plakar diff -recursive between that snapshot and the
latest one lists the keys that disappeared or changed.
Restore the prefix next to the original. A destination whose location ends
with restore-2026-11-14 writes the objects under that prefix, and the live
exports/ is not touched.
Put the objects back in place. Once checked, restore the same snapshot path to a destination on the bucket root, and the objects return under their original keys.
$ plakar at /var/backups ls
$ plakar at /var/backups locate "2026-11-*.parquet"
$ plakar at /var/backups diff -recursive 7aa2f5d1 e41b7c09
$ plakar destination add assets-restore \
s3://s3.fr-par.scw.cloud/acme-assets/restore-2026-11-14 \
access_key=... secret_access_key=...
$ plakar at /var/backups restore -to "@assets-restore" 7aa2f5d1:/exports
$ plakar destination add assets-live \
s3://s3.fr-par.scw.cloud/acme-assets \
access_key=... secret_access_key=...
$ plakar at /var/backups restore -to "@assets-live" 7aa2f5d1:/exportsCross-provider
How to back up S3 to another provider
A copy at a second provider survives the loss of an account, a region or a set of credentials at the first. Plakar reads the bucket at one provider and writes encrypted snapshots to a store at another, and neither side holds the other’s keys.
Create the store at the second provider. A bucket there, with its own access key, holds the Plakar store. The passphrase stays on the backup host.
Add the production bucket as a source. A key limited to s3:ListBucket and
s3:GetObject is enough.
Back up on a schedule from next to the bucket. Run Plakar on a host or an edge in the same region as the production bucket.
Restore wherever you need. Into the original bucket, a new bucket at either provider, or a local disk.
$ plakar store add offsite \
s3://s3.eu-west-par.io.cloud.ovh.net/acme-backups \
access_key=... secret_access_key=...
$ plakar at "@offsite" create
$ plakar source add assets \
s3://s3.fr-par.scw.cloud/acme-assets \
access_key=... secret_access_key=...
$ plakar at "@offsite" backup "@assets"
$ plakar at "@offsite" checkEgress
Cutting egress costs when backing up S3 to another provider
Many teams keep their only backup inside the provider they protect, because copying every bucket out each night costs more in data transfer than in storage. Plakar cuts what leaves the provider at the source:
Reads stay inside the provider’s region.
Plakar runs on a host or an edge next to the bucket, so listing and reading the objects never crosses to the other provider.
Only new chunks cross.
Plakar splits objects into content-defined chunks and sends a chunk only when the store does not hold it yet. After the first run, a nightly backup transfers what changed since the night before.
What crosses is compressed and encrypted first.
Chunks are compressed with LZ4, then encrypted, before they leave the host. Even the first copy sends a duplicated object once, and text, logs or exports smaller than they sit in the bucket.
Native protections
S3 versioning, replication and Plakar snapshots compared
S3 versioning and replication answer the most frequent incidents inside the bucket’s own account. A Plakar snapshot adds a copy under other credentials.
| S3 versioning | S3 replication | Plakar snapshot | |
|---|---|---|---|
| Undo a single delete | YesRemove the delete marker | PartlyDepends on delete marker replication | YesRestore the key from any snapshot |
| Survives a purge of older versions | NoThe purged versions are gone | PartlyDepends on the replica’s own rules | YesThe store keeps its own retention |
| Credentials apart from the bucket | NoSame account and IAM | PartlyOften the same account | YesSeparate keys for the store |
| Copy at another provider | No | NoSame provider | YesAny S3-compatible endpoint, a disk or SFTP |
| Point-in-time view of a prefix | PartlyObject by object | No | YesBrowse, diff or mount a snapshot |
| Encrypted before it leaves you | PartlyServer-side encryption by the service | PartlyServer-side encryption by the service | YesEncrypted on the backup host |
Store on S3
Storing Plakar backups in S3: storage classes and Object Lock
Archival storage classes for long retention.
storage_class sets the class the provider offers for the objects of the store.
With an archival class, backup data goes to that class while the store
configuration and snapshot index stay in STANDARD, so plakar ls lists every
snapshot at once. Providers map the S3 class names to their own archival tiers,
such as GLACIER for Scaleway Glacier or DEEP_ARCHIVE for OVHcloud Cold
Archive.
Object Lock retention on the store bucket.
Enable Object Lock with a default retention on the bucket that holds the store, and immutability comes from the provider. Plakar sends a Content-MD5 header with every write, so Object Lock buckets accept the store’s objects.
A second copy at another site.
plakar sync copies
snapshots from one store to another, such as a bucket at a second provider or an
SFTP host.
Store keys apart from production keys.
Give the store its own access key, allowed on the backup bucket only, and keep it and the store passphrase off the hosts and pipelines that use the production bucket.
Backups you have verified.
plakar check reads
a snapshot back and verifies every chunk against its content address.
Under the hood
How Plakar reads and writes S3 buckets
Path-style and virtual-hosted S3 URLs
Path style is the default. virtual_host=true with endpoint handles services
that put the bucket name in the hostname.
S3 buckets encrypted with SSE-C
For buckets that require customer-provided keys, sse_customer_key takes the
Base64 AES-256 key used to read and write objects.
Self-hosted S3 endpoints
port, region, use_tls and tls_insecure_no_verify reach MinIO, Ceph RGW
and Garage clusters on private networks with their own certificates.
Amazon S3
Amazon S3 backup: IAM roles, S3 Glacier and SSE-C
Back up Amazon S3 without stored keys
Leave access_key and secret_access_key out and Plakar uses the IAM role of
the EC2 instance, the ECS task or the EKS pod it runs in, through IRSA or Pod
Identity. In Plakar Control Plane, leave both credential fields empty to use the
IAM role of the edge that runs the task.
Keep Plakar backups in S3 Glacier
storage_class takes the Amazon S3 class names: STANDARD, STANDARD_IA,
ONEZONE_IA, INTELLIGENT_TIERING, GLACIER_IR, GLACIER, DEEP_ARCHIVE and
REDUCED_REDUNDANCY.
Amazon S3 buckets re-encrypted with SSE-C
AWS described this attack in a January 2025 security advisory: leaked keys used to copy every object onto itself with an SSE-C key only the attacker holds. The recovery scenario above shows the way back.
Comparison
S3 backup tools compared: Plakar, backup services and sync scripts
S3 buckets are usually protected in one of three other ways: an enterprise backup suite with an object storage module, the cloud provider’s own backup service, or a scheduled sync to a second bucket. Plakar keeps the open S3 API of the last one and adds a verified, encrypted history.
| Plakar | Enterprise backup suites | Cloud provider backup services | A scheduled bucket sync | |
|---|---|---|---|---|
| Any S3-compatible source | YesAny S3 API endpoint | PartlyDepends on the product | NoThat provider’s buckets | Yes |
| Copy at another provider | YesStore on any S3 endpoint, a disk or SFTP | PartlyDepends on the product | NoInside that provider | Yes |
| Read-only source credentials | Yess3:ListBucket and s3:GetObject |
PartlyDepends on the product | PartlyDepends on the service | Yes |
| Encrypted, deduplicated history | YesChecked with plakar check |
PartlyDepends on the product | PartlyDepends on the service | NoA sync mirrors deletions |
| Restore one prefix to any bucket | YesOn any S3-compatible endpoint | PartlyThrough the suite’s console | PartlyWithin that provider | PartlyCopied back by hand |
| File servers and databases in one tool | YesSame CLI and store | Yes | PartlyThe provider’s own services | No |
| Fleet-wide scheduling | YesPlakar Control Plane: discovery and policies | YesA central console | YesWithin that provider | NoYour own tooling |
| Free and open source | YesThe CLI | NoCommercial licence | NoA paid service | PartlyDepends on the tool |
Quickstart
How to back up an S3 bucket with the open-source Plakar CLI
The Plakar CLI is free and open source. Install the S3 package, create a store, add the bucket as a source, then restore a prefix into a bucket on another endpoint, here a MinIO lab.
The S3 CLI documentation covers every option, including addressing styles, ambient credentials and SSE-C.
$ plakar pkg add s3
$ plakar at /var/backups create
$ plakar source add prod-assets \
s3://s3.fr-par.scw.cloud/acme-assets \
access_key=... secret_access_key=...
$ plakar at /var/backups backup "@prod-assets"
$ plakar at /var/backups check
$ plakar destination add minio-restore \
s3://minio.lab.internal:9000/acme-assets-restore \
access_key=... secret_access_key=...
$ plakar at /var/backups restore -to "@minio-restore" <snapshot_id>:/exportsControl Plane
S3 bucket backups across accounts with Plakar Control Plane
Plakar Control Plane schedules bucket backups across every account you connect. Managed inventories discover the buckets of each cloud account, and other S3-compatible endpoints are registered in a self-managed inventory. Tasks run on an edge.
See how Plakar Control Plane works, read the Control Plane S3 documentation, or contact us to plan a deployment.
Bring the buckets into an inventory. A managed inventory lists the buckets
of a cloud account. Other endpoints are added as resources of class
Object Storage, subclass S3.
Choose the credentials. Enter an access key pair for the endpoint.
Create a store. Set its passphrase and storage class. A store on an archival class is marked Cold so policies account for it.
Let a policy schedule it. A policy sets the frequency, the retention and the target store, and every bucket that matches it is backed up automatically.
FAQ
S3 backup FAQ
How do I back up an S3 bucket?
Install the package with plakar pkg add s3, add the bucket with
plakar source add assets s3://<endpoint>/<bucket> and a read-only key, then
run plakar backup "@assets" on a schedule. Each run is a full, deduplicated
snapshot of the bucket.
Is S3 versioning a backup?
Versioning keeps older copies of each object inside the same bucket, under the same account and permissions. A backup is a copy held elsewhere, with its own credentials and retention. Plakar provides that copy as encrypted snapshots.
How do I restore deleted S3 objects?
Find the snapshot taken before the deletion with plakar ls, then restore the
prefix with plakar restore -to @dest <snapshot_id>:/prefix. Restoring under a
new prefix first leaves the live objects untouched.
How do I back up an S3 bucket to another provider?
Add the bucket as a source and create the Plakar store on a bucket at the other provider, or on a disk or SFTP host. Each side gets its own credentials.
Where can I restore an S3 backup?
Into the original bucket, a new bucket that the destination creates, a bucket at another S3-compatible provider, or a local directory. A restore can cover the whole snapshot or one prefix.
Can I back up only part of an S3 bucket?
Yes. Put a prefix in the location, such as s3://endpoint/bucket/media, or set
root. Each source covers one bucket.
Which permissions does the S3 backup user need?
s3:ListBucket on the bucket and s3:GetObject on its objects. Write
permissions belong to the destination and the store, with other credentials.
What does a snapshot of an S3 bucket contain?
The content of every object at backup time, with its key, size and last modified date. The series of snapshots, kept by your retention policy, is the history of the bucket.
Does Plakar work with S3-compatible storage?
Yes. The connectors use the S3 API, so any service that speaks it, such as Amazon S3, Scaleway, OVHcloud, MinIO, Ceph RGW or Garage, works as source, destination or store.
Can Plakar back up Amazon S3 without stored keys?
Yes. With no access key configured, Plakar uses the ambient credentials of Amazon S3: the instance profile of an EC2 instance, an ECS task role, or an EKS pod role through IRSA or Pod Identity.
Can I keep Plakar backups in Amazon S3 Glacier?
Yes. Set storage_class on the store to GLACIER_IR, GLACIER or
DEEP_ARCHIVE. With GLACIER or DEEP_ARCHIVE, backup data goes to the
archival class while the snapshot index stays in STANDARD, so every snapshot
remains listed. Other providers’ archival tiers use the same option.
How does Plakar reduce egress costs for S3 backups at another provider?
Run Plakar next to the bucket, inside the provider’s region: it reads the objects there and sends to the other provider only the chunks the store does not hold yet, compressed with LZ4 and encrypted. After the first run, each backup transfers what changed.
How do I protect S3 backups from ransomware?
Keep the store in another bucket or at another provider, with keys absent from
the application, enable Object Lock with a default retention on that bucket, and
keep a second copy with plakar sync.
How much storage does a nightly S3 backup use?
After the first run, each snapshot stores only new or changed content, so a nightly backup of a mostly stable bucket adds little to the store.
Should I use the free CLI or Plakar Control Plane?
The free, open-source CLI suits a handful of buckets. Plakar Control Plane adds discovery, policies and one view across accounts and teams, with the same S3 connectors.