Version 3.4.0 Release Notes
Release Date: 2 May 2022
This document outlines the main additions and changes to the Iguazio MLOps Platform ("the platform") in version 3.4.0.
Highlights
- Trino (formerly Presto) is upgraded to 370.
- New report that displays the consumer groups / v3io streams. See Consumer groups report.
- Real-time pipelines are displayed in the UI. See Real-time pipelines.
[Tech Preview] - Configuration of pod priority for the MLRun services and Nuclio functions in the UI. See Pod priority.
- Configuration of spot nodes for all functions, when creating a job, rerunning an existing job, and creating an ML functions. See Spot nodes.
- Configuration of CPU, GPU, and memory default limits for user jobs. See User job default limits.
- Improve service deployment upon changes. See Modifying a service.
Managed Application Services
Jupyter Notebook
- You can install packages on Jupyter, by using the iguazio user to perform
apt-getusing sudo. Installed packages are not persistent and are stored on the local container fs. Upon Jupyter restart/scale-down etc. those changes are lost. To make the packages persistent, add the package installation commands to yourinitscript.
MLRun
- The Feature store supports Spark engine ingestion from Snowflake.
- Enhanced masking of sensitive data in the UI.
- Setting owner and members have moved to a dedicated project Settings section in the UI.
- The dataset tab is now in the Projects main menu (was previously under the Feature store).
Nuclio Serverless Functions
Presto and Hive
Security and User Management
Backup, Recovery, and High Availability (HA)
- If a cluster goes offline, the system automatically tries to move it to online. (If the cluster was in a degraded state when it went offline, it remains in a degraded state when it comes online.)
Dashboard (UI)
Projects
- The
Project Monitoring report has a new tile namedConsumer groups (v3io streams) that shows the total number of consumer groups, with drill-down capabilities for more details.
- There is a new tab under
Projects | Models namedReal-time pipelines that displays the real time pipeline graph, with a drill-down to view the steps and their details.[Tech Preview]
Miscellaneous
- The Iguazio version in the
About window now includes the patch version. For example, 3.4.0, instead of the previous 3.4.
Services
- You can configure Pod priority for the MLRun services and Nuclio functions in the UI (previously it was configurable by SDK). For MLRun, it applies to the jobs created by MLRun. For Nuclio it applies to the pods of the Nuclio-created functions. See Pod priority.
- You can configure spot nodes for all functions, when creating a job, rerunning an existing job, and creating an ML function. See Preemption mode: Spot vs. On-demand nodes.
- You can configure default limits for user jobs for CPU, GPU, and memory. The default limits are set in the MLRun and Nuclio services. The values are applied to the service's user-created pods. You can overwrite the default when creating a job, or a function. See CPU, GPU, and memory limits for user jobs.
- When modifying a service in the
Service page, the system identifies the related services. You can choose to apply the change to all related services, or only to the service you initially modified.
- Services (in the
Services page) can be filtered by name, type, and running user, and sorted by CPU and memory.
- You can add Environment variables to a web shell service in the
Custom Parameters tab of the service. Variables defined in the service override the global environment variables.
Deployment Specifications
Cloud Deployments
- Supports external docker registries on ACR and GCR. Specify the off-cluster URL when you create the docker registry service in Provazio: In the App Cluster tab, select Use Custom User Docker Registry and type in the URL, the Registry username and the Registry password.
- Supports external load balancer for restricted global access mode while allowing node auto scaling groups in GKE.
- Supports accessing PyPi repos via certificate file. See Creating a PyPI certificate secret and loading it to Kubernetes.
[Tech Preview]