Skip to content
This repository has been archived by the owner on May 10, 2024. It is now read-only.

riotkit-org/volume-syncing-controller

Repository files navigation

volume-syncing-controller

Notice: Sorry, we no longer have a possibility to maintain this project. Feel free to fork it

Docker container and Kubernetes controller for periodically synchronizing volumes to cloud-native storage, and restoring their state from cloud-native storage.

Example use cases

  • Service restored from cloud-storage in other data center with its own CNI
  • Kubernetes service state kept in cloud-storage, restored outside Kubernetes (e.g. in docker on desktop for development)
  • Snapshot type backup (does not support versioning on client side - the cloud remote storage could do this)
  • Ability to move services between machines in tiny K3s clusters, where network CNI is not available - only local storage is used

Roadmap

Features for first version:

  • Supports all storage kinds that are supported by Rclone
  • Rclone configuration using environment variables
  • End-To-End encryption support
  • Periodical synchronization with built-in cron-like scheduler
  • volume-syncing-controller sync-to-remote command to synchronize local files to remote
  • volume-syncing-controller remote-to-local-sync command to sync files back from remote to local
  • Support for Kubernetes: initContainer to restore files, and side-car to back up files to remote
  • Extra security layer preventing from accidental file deletion in comparison to plain rclone or rsync usage 💯
  • Non-root container
  • Allow to disable synchronization or restore in CRD
  • Jinja2 templating support inside kind: PodFilesystemSync to allow using single definition for multiple kind: Pod objects
  • Termination hook to synchronize Pod before it gets terminated
  • Allow to decide about the order of initContainer in CRD

v1.1:

  • Health check: If N-synchronization fails, then mark Pod as unhealthy (optionally)
  • Periodical synchronization using filesystem events instead of cron-like scheduler (both should be available)

v1.2:

  • Watch synchronization progress and update status field. Pods can notify controller using webhooks with a token granted on Pod creation
  • Multiple application replicas support: One Pod can be marked as primary, the rest will have sync-to-remote on hold
  • Expose Prometheus metrics at /metrics, add configuration for Prometheus and Victoria Metrics in Helm

Kubernetes controller architecture

The solution architecture is designed to be Pod-centric and live together with the application, not on the underlying infrastructural level.

Above means, that when Pod starts - the volume is restored from remote, then all data is synchronized to remote periodically during Pod lifetime to keep an external storage up-to-date.

There are added two containers:

  • init-volume-restore (initContainer - executes before application starts)
  • volume-syncing-sidecar (sidecar - lives together with application)

Security

  • The images are scanned for vulnerabilities: Open report
  • Distroless-type image reduces surface of attack
  • No bash scripts. The container contains only rclone and volume-syncing-controller binaries
  • Built-in E2E encryption automation
  • Semantic Versioning is used to protect end users from breaking changes as much as it is possible

Runtime compatibility

🐧 Platform Usage type
Bare metal ✔️ Environment variables
Docker ✔️ Environment variables
Kubernetes ✔️ Environment variables, Helm + operator (mutating kind: Pod by adding initContainer + side-car container)

How it works?

Container and binary are configured with environment variables that are translated into rclone configuration file.

Example usage:

export REMOTE_TYPE=s3
export REMOTE_PROVIDER=Minio
export REMOTE_ACCESS_KEY_ID=AKIAIOSFODNN7EXAMPLE
export REMOTE_SECRET_ACCESS_KEY=wJalrXUtnFEMI/K7MDENG/bPxRfiCYEXAMPLEKEY
export REMOTE_ENDPOINT=http://localhost:9000
export REMOTE_ACL=private

# synchronize to remote storage
volume-syncing-controller sync-to-remote -s ./ -d testbucket/some-directory

# synchronize back from remote storage to local directory
volume-syncing-controller remote-to-local-sync -v -s testbucket -d ./.build/testing-restore

Will translate into configuration:

[remote]
type = s3
provider = Minio
access_key_id = AKIAIOSFODNN7EXAMPLE
secret_access_key = wJalrXUtnFEMI/K7MDENG/bPxRfiCYEXAMPLEKEY
endpoint = http://localhost:9000
acl = private

And will run:

rclone sync ./ remote:/testbucket/some-directory

For more examples see our Makefile full of examples that could be run in context of this repository

⌚ Scheduling periodically

volume-syncing-controller has built-in crontab-like scheduler to optionally enable periodical syncing.

volume-syncing-controller --schedule '@every 1m'    # ...
volume-syncing-controller --schedule '0 30 * * * *' # ...

⚫ Safety valves

There are various "safety valves" that in default configuration would try to prevent misconfigured run from deleting your data.

sync-to-remote

⚫ Safety rule
Local directory cannot be empty (it would mean deleting all files from remote directory)

remote-to-local-sync

Validation of those rules can be intentionally skipped using commandline switch --force-delete-local-dir

⚫ Safety rule
Remote directory cannot be empty (it would mean deleting all local files)
/usr/bin, /bin, /, /home, /usr/lib cannot be picked as synchronization root due to risk of erasing your operating system
Local target directory cannot be owned by root
Local target directory must be owned by same user id as the current process runs on

Kubernetes example

Operator have to be installed using Helm first. It will react for every Pod labelled with riotkit.org/volume-syncing-controller: true and matching CRD of PodFilesystemSync kind.

The riotkit.org/volume-syncing-controller: true is there for performance and cluster safety by limiting the scope of Admission Webhook.

---
apiVersion: riotkit.org/v1alpha1
kind: PodFilesystemSync
metadata:
    name: cloud-press
spec:
    # Follows K8s convention: https://kubernetes.io/docs/concepts/overview/working-with-objects/labels/#resources-that-support-set-based-requirements
    podSelector:
        matchLabels:
            my-pod-label: test

    localPath: /var/www/riotkit/wp-content
    remotePath: /example-org-bucket          # Can be also a JINJA2 template with access to Pod. Example: '/stalin-was-a-dickhead/{{ pod.ObjectMeta.Annotations["subdir"] }}'

    initContainerPlacement:
        containerReference: some-container-name # defaults to empty
        placement: before # before, after, first or last (first & last requires the `containerReference` to be empty). Defaults to "last"

    syncOptions:
        # NOTICE: every next synchronization will be cancelled if previous one was not finished
        method: "scheduler"  # or "fsnotify"
        schedule: "@every 5m"
        maxOneSyncPerMinutes: "15"  # When "fsnotify" used, then perform only max one sync per N minutes. Allows to decrease network/cpu/disk usage with a little risk factor

        # Optional "RunAs"
        permissions:
            # Can be overridden by Pod annotation `riotkit.org/volume-user-id`
            uid: "1001"
            # Can be overridden by Pod annotation `riotkit.org/volume-group-id`
            gid: "1001"
            
        # Optional
        cleanUp:
            # Decides if files are synchronized or copied. Synchronization means that redundant files are deleted. 
            # When in directory A you delete a file, then it gets deleted in directory B
            remote: true
            local: true

            # Disables "security valves"
            forceRemote: false    
            forceLocal: false

        # Optional
        allowedDirections:
            # Decides if a side-car container should be spawned to synchronize changes periodically TO REMOTE
            toRemote: true
            # Decides if an init container should be placed to restore files FROM REMOTE on startup
            fromRemote: true
            
        # Allows to decide on a case, when the directory is to be synchronized first time. 
        # Should the initContainer be placed and restore from remote (it may be potentially empty if is new?), or should it be skipped first time?
        # [!!!] This counts not for Pods, not for whole PodFilesystemSync but PER .spec.remotePath
        restoreRemoteOnFirstRun: true
    env:
        REMOTE_TYPE: s3
        REMOTE_PROVIDER: Minio
        REMOTE_ENDPOINT: http://localhost:9000
        REMOTE_ACL: private

        # Best practice is to move sensitive information into `kind: Secret`
        # and reference that secret in `envFromSecret`
        # to keep your secret in GIT you can try e.g. SealedSecrets or ExternalSecrets
        #REMOTE_ACCESS_KEY_ID: ...
        #REMOTE_SECRET_ACCESS_KEY: ...
    envFromSecrets:
        - name: cloud-press-secret-envs

    # Optional
    # Will generate a key, store it in `kind: Secret` and setup End-To-End encryption
    # if existing secret exists and is valid, then will be reused
    #
    # NOTICE: REMEMBER TO BACKUP THIS secret. End-To-End encryption means your data will be unreadable on the server
    automaticEncryption:
        enabled: true
        secretName: cloud-press-remote-sync

SIGTERM support

Pod's sidecar container named init-volume-restore is having implemented a signal handling. When Kubernetes wants to terminate our Pod, then a volume-syncing-controller interrupt command is invoked, next it sends a kill signal to the volume-syncing-controller main process.

volume-syncing-controller main process is stopping the cron-like scheduler and invokes last synchronization to remote before exit.

You may want to adjust your Pod's terminationGracePeriodSeconds to a value that makes sure the Kubernetes will wait longer before terminating containers.

Permissions

UID and GID can be specified in the PodFilesystemSync resource as well as in Pod annotations.

PodFilesystemSync example

# ...
spec:
    syncOptions:
        # ...
        permissions:
            # Can be overridden by Pod annotation `riotkit.org/volume-user-id`
            uid: 1001
            # Can be overridden by Pod annotation `riotkit.org/volume-group-id`
            gid: 1001

Pod example

# ...
metadata:
    annotations:
        riotkit.org/volume-user-id: 161
        riotkit.org/volume-group-id: 1312

The annotations have precedence over PodFilesystemSync resource settings.

Advanced usage notice

It's up to you to use this software with caution and design a proper architecture.

For multiple instances of same application you have to remember to enable synchronization only in one copy of your application, and then in all other replicas disable the synchronization and use only the restore mode.

In v1.1 there are plans to implement such logic in the Kubernetes controller, which may automate the process and introduce a leader election option, but still this issue will remain in bare-metal and docker usage.

PodFilesystemSync deletion

Deletion of a definition does not recursively delete secrets that were automatically created for encryption. In the other words - deleting and restoring PodFilesystemSync does not have impact on the automaticEncryption feature.

Triggering synchronization manually

To perform a synchronization in Kubernetes when automatic encryption is turned on, type:

kubectl exec -n my-namespace-name my-pod-name -c volume-syncing-sidecar -it -- /usr/bin/volume-syncing-controller sync-to-remote -s /path/to/local/directory -c /etc/volume-syncing-controller/rclone.conf

In case, when automatic encryption is not used, then you additionally need to specify the remote target directory.