ShinyProxy Operator

Easily run ShinyProxy on a Kubernetes cluster

(c) Copyright Open Analytics NV, 2020-2024 - Apache License 2.0

Why?

Deploying and managing ShinyProxy can get complex when many apps are used, especially when the configuration of ShinyProxy is often updated.

When restarting a running ShinyProxy instance (in order to update its configuration), users will face a disconnect from their running applications. The only solution to guarantee that users do not lose their connection to running apps, is to keep the current instance alive when updating ShinyProxy's configuration. However, manually keeping track of these instances would be too cumbersome and should therefore be automated.

The ShinyProxy operator for Kubernetes is able to manage multiple ShinyProxy instances and their configuration.

Example

Assume we have some ShinyProxy configuration config1 which contains one app called app1. When the operator starts working, it checks whether a ShinyProxy instance exists with that configuration. If not, it starts a ShinyProxy instance and all other required configuration. Users can now start using app1 on this instance.

Some time later, the need for a second app arises. Therefore, the administrator adapts the configuration of ShinyProxy to include a second app app2.

However, some users are still using app1 on the old instance. These apps may have some state, which should not be lost. Therefore, the operator starts a second ShinyProxy instance using configuration config2. The operator does not modify the original ShinyProxy server, therefore the existing apps continue to work (even if they are using Websocket connections).

All new HTTP (and Websocket) connections are forwarded to the new server, i.e. any new connection is handled by the new server. Therefore, if users go to the main ShinyProxy page, they will see that a new app is available. Every user ( also those still using the old application) can start the new app. The operator stops and removes the old server as soon as it has finished handling any existing (Websocket) connections.

Building from source

Clone this repository and run

mvn -U clean install

The build will result in a single .jar file: target/target/shinyproxy-operator-jar-with-dependencies.jar.

Running

The operator should be run in Kubernetes using the docker image. It can run in either clustered or namespaced mode. In the former the operator looks for ShinyProxy instances in all namespaces while in the latter it only manages ShinyProxy instances in its own namespace.

See the docs/deployment folder for more information.

Configuration

We try to keep the configuration of the Operator itself as minimum as possible. Furthermore, we want the operator to work without configuration in most cases. Nevertheless, for some specific cases some configuration options are available. For now these options are specified using environment variables. All variables start with the SPO prefix, meaning ShinyProxyOperator.

• SPO_MODE: can either be namespaced or clustered (default). This specifies whether the operator should only look in its own namespace for ShinyProxy configurations or in all namespaces.
• SPO_PROBE_INITIAL_DELAY: specifies the initial delay of the Readiness and Liveness probes. This is useful when the used Kubernetes version does not support startup probes.
• SPO_PROBE_FAILURE_THRESHOLD: specifies the failure threshold of the Readiness and Liveness probes. This is useful when the used Kubernetes version does not support startup probes.
• SPO_PROBE_TIMEOUT: specifies the timeout in seconds of the Readiness and Liveness probes. This is useful when the used Kubernetes version does not support startup probes.
• SPO_STARTUP_PROBE_INITIAL_DELAY: specifies the initial delay of the StartUp probe. By default, this is 60 seconds.
• SPO_LOG_LEVEL: configures the log level of the operator, may be one of the following:
  • OFF: disables logging
  • ERROR
  • WARN
  • INFO
  • DEBUG: default (may change)
  • TRACE
  • ALL: enables all logging

Note: in our deployments where startup probes aren't supported we have success with the following configuration:

• SPO_PROBE_INITIAL_DELAY to something between 40 and 140 depending on the performance of the cluster.
• SPO_PROBE_FAILURE_THRESHOLD to 2
• SPO_PROBE_TIMEOUT to 3

Supported Versions

The first stable release of the operator (1.0.0) is compatible with ShinyProxy 2.6.0. Older ShinyProxy versions are supported by running development snapshots of the operator, however, we strongly advice to upgrade to the latest version of ShinyProxy and the operator for the best experience.

Kubernetes versions

Note:

• ¹ requires a workaround, see below.
• ² version 2.0.0 supports these Kubernetes versions, but might stop syncing after some time, this issue is solved in version 2.1.0

Workaround for bug in Kubernetes

A bug affecting the operator was introduced in Kubernetes 1.20 and fixed in versions 1.20.10 and 1.21.3. However, some deployments (e.g. using EKS) are not able to use this version. When using the affected versions, Kubernetes stops sending events for the Service resources after a certain amount of time. Therefore, the Operator is unaware of any events happening on services and is therefore unable to fully configure a ShinyProxy server. The bug only occurs after the operator has been running for a random time between 30 minutes and two hours. Unfortunately, the only reasonable work-around is to regularly restart the Operator. Since version 0.1.0-SNAPSHOT-20210831.075527, it is possible to specify the SPO_PROCESS_MAX_LIFETIME environment variable. After the configured time (in minutes), the operator stops. The corresponding Docker container then automatically restarts the Java process.

Update to 2.0.0

Be aware of these changes when updating to version 2.0.0:

• the old mechanism where cookies were used to assign users to specific ShinyProxy servers is no longer used. Instead, as soon as a new server is started, all new requests will be handled by the new server, including requests for existing apps. Only existing websocket connections will stay open on the old servers. This has multiple benefits:
  • when a new server is started, users will immediately use and see the configuration of that new server. In other words, if a new configuration includes a new app, this app is immediately available to all users (even if they are using apps started on older servers)
  • there is no longer a process of transferring users to new servers. Both the forced method and the manual method (where users have to click a button) are removed. Users will immediately use the new configuration.
  • apps can be run for a (very) long time, even if frequently updating the configuration and without having many old servers. Old servers are removed as soon as no websocket connections are running on that server.
  • Skipper is no longer a dependency of the operator. There is no benefit in using with version two of the operator.
• the operator now requires ShinyProxy to store the active proxies in Redis. Therefore, since this release Redis takes a more critical role. When running Redis inside Kubernetes, it is therefore best practice to use Redis Sentinel. This way Redis runs in a High Available mode, using three replicas. Compared to running a single Redis server, this prevents a single point of failure on Redis and the node it is running on. This repository contains all manifests required to set up Redis Sentinel (based on the bitnamai Redis helm chart).

The best way to update to ShinyProxy 2.0.0 is by creating a fresh deployment of the operator and migrating users to this new deployment. The following changes need to be made to the ShinyProxy configuration file:

• add the property proxy.store-mode: Redis
• add the property proxy.stop-proxies-on-shutdown: false
• optionally add the property kubernetesIngressPatches in order to customize the ingress created by the operator.
• update the ShinyProxy image to openanalytics/shinyproxy:3.1.1

Update to 2.1.0

The ShinyProxy CRD has been updated in version 2.1.0, it is important to update the CRD in your cluster. Running the deployment commands is enough. The CRD can be updated while ShinyProxy and the ShinyProxy operator are running in the cluster.

Java Version

This project requires JDK 17.