Upgrading Seldon Core¶

This page provides with instructions on how to upgrade from previous versions. Each of these have to be run sequentially if it’s expected for previous running models to be upgraded without disruption (i.e if you are running version 0.4.2, you will first have to upgrade to 0.5.2 and then to 1.1, etc).

If you were running our Openshift 0.4.2 certified operator and are looking to upgrade to our 1.1 certified operator, you will also need to follow the “upgrading process” steps in the “Upgrading to 0.5.2 from previous versions” section.

Make sure you also read the CHANGELOG to see the detailed features and bug-fixes in each version.

Upgrading to 1.18¶

Please be advised that license under which Seldon Core is released changes to BSL 1.1 in Seldon Core 1.18.0. See the LICENSE file in our root folder for more details.

There’s no specific work needed for upgrade from 1.17 to 1.18.

Upgrading to 1.17¶

There’s no specific work needed for upgrade from 1.16 to 1.17.

Upgrading to 1.16¶

Kubernetes version upgrade¶

Seldon Core 1.16 bumps minimum Kubernetes version to 1.23. This is because as part of making Seldon Core compatible with Kubernetes 1.25 we moved from autoscaling/v2beta1 apiVersion of HorizontalPodAutoscaler to autoscaling/v2 (see this PR for further details).

Following table provides a summary of Seldon Core / Kubernetes version compatibility for recent version of Seldon Core.

Core Version K8s Version

1.21

1.22

1.23

1.24

1.25

1.26

1.27

1.11

âś“

1.12

âś“

âś“

âś“

âś“

1.13

âś“

âś“

âś“

âś“

1.14

âś“

âś“

âś“

âś“

1.15

âś“

âś“

âś“

âś“

1.16

âś“

âś“

âś“

âś“

âś“

1.17

âś“

âś“

âś“

âś“

âś“

1.18

âś“

âś“

âś“

âś“

âś“

It is always recommended to first upgrade Seldon Core to the latest supported version on your Kubernetes cluster and then upgrade the Kubernetes cluster.

For example, if you using Seldon Core 1.15 on Kubernetes 1.24, it is recommended to first upgrade Seldon Core to version 1.16 before upgrading your cluster to Kubernetes 1.25 and further.

Upgrading to 1.15¶

Flask upgrade¶

With this release of Seldon Core we upgrade Flask dependency to 2.x line. This includes Python wrapper and most pre-packaged model servers. In an unlikely scenario that this causes issues one can fallback to 1.14.1 images.

Services and Deployments Update¶

The label on services pointing to containers for each node in your inference graph has been changed to stop clashes in key-values for label selectors when you have multiple nodes in your inference graph. This will cause all your existing seldon deployments to be recreated as their service and deployment specs will have changed in the upgrade. This will cause some downtime in availability.

Upgrading to 1.14.1¶

OpenShift default storage initializer¶

In this patch release we fixed the default storage initializer used on OpenShift (both Community and Certified operators). The image used now is seldonio/rclone-storage-initializer:1.19.0-dev which is the same as one used for the non-OpenShift releases (*).

For Certified operator you will find this image defined as

- name: RELATED_IMAGE_STORAGE_INITIALIZER
  value: registry.connect.redhat.com/seldonio/rclone-storage-initializer@sha256:a0280c13136dcc870194af72630b9d2f7fc8bcff4edb54dd3bfbce36741af50c

in the CSV. This is the same image, just coming from the Red Hat registry.

If you prefer or need to keep using previous storage initializer you can set for Community Operator

- name: RELATED_IMAGE_STORAGE_INITIALIZER
  value: kfserving/storage-initializer:v0.6.1

or

- name: RELATED_IMAGE_STORAGE_INITIALIZER
  value: registry.connect.redhat.com/seldonio/storage-initializer@sha256:52f1e1901fc5de0734515f136847eeb698d48903cf7b6b1cbf273d303bb9029c

for the Certified one.

(*) Non-OpenShift releases defaulted to Rclone-based storage initializer in Seldon Core version 1.8.

Rclone Storage Initializer¶

Rclone storage initializer image is now based on Red Hat’s Universal Base Image (UBI8). This is not breaking change.

Upgrading to 1.14¶

CRD V1¶

Only the v1 versions of the CRD will be supported moving forward. The v1beta1 versions will remain in the Helm chart but will be not updated. Allowing the operator to create the CRDs will result in the v1 CRD being created so will only work on Kubernetes clusters >= 1.18.

Model Health Checks¶

We have updated the health checks done by Seldon for the model nodes in your inference graph. If executor.fullHealthChecks is set to true then:

  • For Seldon protocol each node will be probed with /api/v1.0/health/status.

  • For the v2 protocol each node will be probed with /v2/health/ready.

  • For tensorflow just TCP checks will be run on the http endpoint.

By default we have set executor.fullHealthChecks to false for 1.14 release as users would need to rebuild their custom python models if they have not implemented the health_status method. In future we will default to true.

Request Logger¶

The Python request logger component example has been deprecated and removed as part of #4013.

Seldon Core Analytics¶

In this release we change status of Seldon Core Analytics Helm Chart to example only. This chart will not be further updated. We recommend to install and configure metrics monitoring using Prometheus Operator.

Upgrading to 1.13¶

Seldon Inference Payload Logging Changes¶

The seldon executor logs the prediction request and respose payload to a configured URL post during the inference at each stage of a inference graph. These request/response pairs are now matched correctly to reflect the component’s input/output at each node level as compared to pairing based on the tree structure of the inference graph previously. See relevant issue for more details.

Alibi Detect Server Event Update¶

The cloud event data that the alibi detect server sends to a configured replyURL post the outlier/drift detection was a JSON string. This is now updated to reflect the content type as JSON correctly.

Upgrading to 1.12¶

Support for Kubernetes 1.22¶

Seldon Core adds support for Kubernetes 1.22 by upgrading all ValidatingWebhookConfiguration resources to v1 intead of v1beta1. There are no specific breaking steps that were identified when upgrading using the helm charts for each respective version, but you need to make sure to check that this works as expected.

Updated Python wrapper folder configurations¶

  • The default running user in Seldon Core is 8888

  • Some servers like the MLFlow Server v1 installs installations in runtime

  • Access required to modify files in the local folder are required so the application folder should be writable

  • The default base image now changes the owner of the /microservice folder to user 8888

Updated executor request logger settings¶

The request logging from the executor now has a configurable queue size and write timeout. This will allow a tradeoff between pending request memory usage and failing requests when sending to various logging endpoints that may be slow. The write timeout will mean logging of requests will fail if waiting for more than the given time to be added to the work queue. The two settings are:

  • executor.requestLogger.workQueueSize (default 10000)

  • executor.requestLogger.writeTimeoutMs (default 2000)

It is also possible to update these values on a per SeldonDeployment basis with the annotations:

  • seldon.io/executor-logger-queue-size

  • seldon.io/executor-logger-write-timeout-ms

Upgrading to 1.11¶

Python S2I Wrapper¶

  • The default wrapper seldonio/seldon-core-s2i-python3 is now Python 3.8

  • Python 3.7 wrapper is still available as seldonio/seldon-core-s2i-python37

Upgrading to 1.10¶

Seldon Core Wrapper¶

  • With introduction of multi-processing in gRPC module the SO_REUSEPORT socket option is required. On certain Python distributions you may see AttributeError: module 'socket' has no attribute 'SO_REUSEPORT' error which would render gRPC endpoint non-operational. For Anaconda Python distributions we confirmed that upgrading to Python 3.7.10 or 3.8.10 removes the problem.

Server Updates¶

  • SKLearn server has been updated to use sklearn 0.24.2

  • XGBoost server has been updated to use xgboost 1.4.2

Alibi Server Updates¶

  • Alibi has been updated to 0.6.0

  • Alibi server python has been updated to 3.7.10

Upgrading to 1.8¶

Rclone Storage Initailizer¶

In Seldon Core 1.8 the rclone-based storage initializer becomes the default one.

The storage initailizer image that is being used is controlled by the helm value:

storageInitializer:
  image: seldonio/rclone-storage-initializer:1.19.0-dev

and can be customised on per-deployment basis as described in Prepackaged Model Servers documentation by setting value of storageInitializerImage variable in the graph definition.

This transition requires creation of the new secrets for the prepackaged model servers that will be compatible with the rclone configuration format as described here. Read more:

If you do not wish to configure these secrets now and wish to preserve prior behaviour you can opt for usage of previous storage initializer by using following helm value:

storageInitializer:
  image: kfserving/storage-initializer:v0.6.1

See further documentation here.

Request Logger¶

In Seldon Core 1.9 we will be moving seldon-request-logger to separate repository.

Legacy Java Engine Orchestrator¶

In Seldon Core 1.9 final deprecation of Java Engine will happen with removal of all the related code from the repository.

Upgrading to 1.7¶

Python Dependency Updates¶

Various CVEs were resolved via #2970, which included several packages upgrades which may affect applications that install packages which may not be compatible. This also includes the installation of pip==20.2, however this version of pip still uses the older resolver.

Upgrading to 1.6¶

Webhook Removal¶

As part of the 1.6.0 release we are removing the Seldon Core Mutating Webhook. This won’t cause any noticeable changes, but it is recommended that you manually remove the webhook once you upgrade to version 1.6.0

Upgrading to 1.5¶

REST and gRPC¶

To take advantage of the ability to handle both REST and gRPC on any deployed model python model images will need to be recreated using the 1.5 python wrapper. If they are not updated they will only expose the protocol they were originally wrapped for.

You can use and extend the backwards compatibility notebook to check your deployments will work if you do not intend to upgrade them.

Upgrading to Kubernetes version >= 1.18¶

If you have a Kubernetes cluster with Seldon Core installed, and you want to upgrade the Kubernetes cluster, you have to carry out a set of manual steps due to the more strict validation that this version of Kubernetes introduced.

To be more specific, we had to provide two versions of the CRD as part of the seldon core install helm chart. Similarly the CRD for Seldon Core in Kubernetes post-1.18 is actually differnt - namely, you can actually see that in version 1.3.0 we introduced new CRD changes in the helm chart via an IF statement to use a different CRD depending on the k8s version. Due to this, the path to upgrade from pre-1.18 to post-1.18 requires the following manual steps to be carried out:

  1. Start with kubernetes cluster pre 1.18 with seldon core pre-1.3.0

  2. Upgrade Kubernetes cluster to post 1.18 (seldon core CRD is now “invalid” but still installed as still in etcd)

  3. Manually add “spec.preserveUnknownFields”, to helm chart and install CRD (so it ignores invalid fields of now invalid CRD)

  4. Remove the “spec.preserveUnknownFields”, from helm chart manually again, and re-install now the current CRD

Upgrading to 1.3¶

Breaking Changes¶

The version of sklearn used by the default sklearn server will be 0.23.2. To use a different version you will need to follow the steps described in the sklearn server documentation.

Upgrading to 1.2.1¶

[NOTE] 1.2.0 has issue where all Seldon Deployments are marked as “NotReady” as there is a bug caused by a volumeName update. This can be resolved by following the 1.2.0 volume patch as outlined by this example. It is recommended to upgrade to version 1.2.1 directly instead.

All seldon-managed pods will be subject to a rolling update as part of this upgrade.

New Features / Breaking Changes¶

  • The helm value createResources has been renamed managerCreateResources.

  • To allow CRDs to be created by the manager. If managerCreateResources is true then extra RBAC to create CRDs is added from the previous versions RBAC which was to just list and get.

  • If upgrading the analytics helm chart then a kubectl delete deployment -n seldon-system -l app=grafana should be run first

  • All the prepackaged model servers are now created with RedHat UBI images. One consequence of this is that they will all run as non-root as it best practice.

Request Logger¶

The values.yaml for the seldon-core-operator helm chart has changed. The field defaultRequestLoggerEndpointPrefix is replaced by:

requestLogger:
  defaultEndpoint: 'http://default-broker'

This default value will find a broker in the model’s namespace. To point to another namespace it would be default-broker.anothernamespace.

Upgrading to 1.1 from previous versions¶

As we moved to 1.x+ there are several breaking changes that need to be considered. These are outlined below

New Features / Breaking Changes¶

Deployment Naming and Rolling Updates¶

The deployments created by Seldon Core have been changed to follow a fixed scheme. It will now be:

<seldondeployment name>-<predictor name>-<podspec idx>-<container names>

So for example:

apiVersion: machinelearning.seldon.io/v1
kind: SeldonDeployment
metadata:
  name: rest-seldon
spec:
  name: restseldon
  protocol: seldon
  transport: rest
  predictors:
  - componentSpecs:
    - spec:
        containers:
        - image: seldonio/mock_classifier_rest:1.3
          name: classifier
    graph:
      name: classifier
      type: MODEL
    name: model
    replicas: 1

For the above resource, one Deployment will be created with name:

rest-seldon-model-0-classifier

This will change how rolling updates are done. Now any change to the first PodSpec above will be updated via a rolling update as expected if the names of the containers are not changed. If however you changed “classifier” to “classifier2” you would get a new deployment created which would replace the old deployment when running.

New Service Orchestrator¶

From version 1.1 Seldon Core comes with a new service orchestrator written in Go which replaces the previous Java engine. Some breaking changes are present:

  • Metadata fields in the Seldon Protocol are no longer added. Any custom metata data will need to be added and exposed to Prometheus metrics by the individual components in the graph

  • All components in the graph must either be REST or gRPC and only the given protocol is exposed externally.

  • The metric names placed in Prometheus have changed to include the executor name rather than engine : see the analytics docs

The new service orchestrator comes with several advantages including ability to handle Tensorflow REST and gRPC protocols and full metrics and tracing support for both REST and gRPC.

For those wishing to use the deprecated Java engine service orchestrator see the service orchestrator docs for details.

Ambassador Retries¶

Ambassador retries has been removed from the previous hardwired value of 3. Retries is now available via an annotation for Ambassador.

Python Wrapper Tag Update¶

The Python Wrapper was using naming convention in the format 0.1 … 0.18. In this release we have renamed the version of the Python Wrapper tag to match the same convention as the Executor, Operator, etc. This means that the Python Wrapper tag for this release is 1.1, and the snapshot would be 1.1.1-SNAPSHOT

Dated SNAPSHOTS¶

Whenever a new PR was merged to master, we have set up our CI to build a “SNAPSHOT” version, which would contain the Docker images for that specific development / master-branch code.

Previously, we always had the SNAPSHOT tag being overridden with the latest. This didn’t allow us to know what version someone may be trying out when using master, so we wanted to introduce a way to actually get unique tags for every image that gets landed into master.

Now every time that a PR is landed to master, a new “dated” SNAPSHOT version is created, which pushes images with the tag "<next-version>-SNAPSHOT_<timestamp>". A new branch is also created with the name "v<next-version>-SNAPSHOT_<timestamp>", which contains the respective helm charts, and allows for the specific version (as outlined by the version in version.txt) to be installed.

You can follow the instructions in the installation page to install the snapshot version.

Wrapper compatibility table¶

To verify if Seldon Core v1.0 and v.1.1 is compatible with older s2i wrapper versions we conducted a simple test with a one-node model. The model has been deployed both with REST and GRPC API with both new orchestrator and the deprecated Java engine (v1.0 only with Java Engine). Test verifies if model can successfully serve inference requests.

NOTE: Full support of custom metrics and tags with new orchestrator is only available from Python wrapper version 0.19. If you need to use older version of Python wrapper you can continue to use Java engine as described above until the next release.

Language Wrapper

Version

API Type

New Orchestrator

Deprecated Java engine

Notes

Python

0.19

both

yes

yes

full support of custom metrics and tags

Python

0.11 … 0.18

both

yes

yes

.

Python

0.10

REST

no

yes

.

Python

0.10

GRPC

yes

yes

.

Python

< 0.10

GRPC

?

?

.

Java

0.2 & 0.1

REST

yes

yes

minor difference in request format

Java

0.2 & 0.1

GRPC

yes

yes

.

Example of request format difference with Java wrapper deployed with REST API:

  1. Using new orchestrator:

    curl -s -X POST \
     -d 'json={"data": {"names": ["a", "b"], "ndarray": [[1.0, 2.0]]}}' \
     localhost:8003/seldon/seldon/compat-rest-java-02-executor/api/v1.0/predictions
    
  2. Using deprecated Java engine:

    curl -s -X POST -H 'Content-Type: application/json' \
     -d '{"data": {"names": ["a", "b"], "ndarray": [[1.0, 2.0]]}}' \
     localhost:8003/seldon/seldon/compat-rest-java-02-engine/api/v1.0/predictions
    

Upgrading to 0.5.2 from previous versions¶

This version included significant improvements and features, including the addition of pre-packaged model servers, fixing several critical bugs.

This was the version that also dropped support for kubernetes 1.11, and added changes in the mutating and validating webhooks that you need to make sure are transitioned as outlined below.

Upgrading process¶

In order to upgrade, the main requirement is to make sure that the Kubernetes cluster is updated to 1.12 or higher.

Once this is done, it’s necessary to delete the old webhooks. This can be done with the following commands (you need to make sure that the commands are executed in the namespace in which seldon core was installed).

Upgrading to 0.2.8 from previous versions.¶

Upgrading process¶

Installation process now with Helm¶

The helm charts to install Seldon Core have changed. There is now a single Helm chart seldon-core-operator that installs the CRD and its controller. Ingress options are now separate and you need to choose between the available options which are at present:

  • Ambassador - via its official Helm chart

  • Istio

For more details see the install docs.

The Helm chart seldon-core-operator will require clusterwide RBAC and should be installed by a cluster admin.

Dropping support for KSonnet¶

Ksonnet is now deprecated. You should convert to using Helm to install Seldon Core.