Fixed issues in Cloudera AI on premises 1.5.5 SP3 CHF3

Previously, the in‑memory settings for RocksDB were too small for the livelog workload, which led to delays in accessing data on Ceph platform and slowed down all workspace sessions under load.

This issue is now fixed. The block cache is increased to 256 MB, the Memtable is increased to 256 MB and the RocksDB statistics collection is improved. When enabled, RocksDB logs internal counters, the enhanced statistics provide better visibility into cache behavior and background operations.

Previously, various operations on livelog topics used a single global lock that remained held during slow disk operations, causing active sessions to pile up behind one delayed request. Under heavy load, this created significant delays for a large number of topic lookups.

This issue is now fixed. Only requests for the same topic wait on each other. All other lookups continue independently, eliminating performance bottlenecks and improving overall responsiveness.

Previously, the livelog and livelog‑publisher pods used hardcoded operational parameters, preventing users from adjusting performance. In some cases, these defaults caused storage slowdowns that affected all workspace sessions.

This issue is now resolved. Performance and timeout settings are now fully configurable through Kubernetes, and memory tuning is improved to provide smoother, more reliable performance.

Previously, the livelog‑publisher could crash with a concurrent write to websocket error when its WebSocket connection dropped, resulting in temporary loss of container logs.

This issue is now fixed.

Previously, the livelog service deleted topic entries one key at a time, which made topic cleanup slow.

This issue is now fixed. Iterative deletion is replaced with a single bulk delete operation, making topic cleanup significantly faster and more reliable. Related cleanup code was simplified and corrected, resulting in smoother maintenance behavior without changing the external API.

Previously, Jupyter session UIs could freeze permanently when livelog was slow or unresponsive because the system stopped sending status updates and never recovered on its own.

The issue is now fixed. A timeout and safe retry mechanism is added so delayed updates fail quickly and automatically recover. The timeout is set to 30 seconds by default and can be adjusted as needed.

In Cloudera AI Workbench on premises, the team synchronization process could incorrectly assign owner permissions to users and service accounts. When a workspace administrator was processed during synchronization, the owner permission could persist for subsequent team members in the same synchronization operation. As a result, some service accounts lost their intended administrator or operator role and became unusable.

This issue is now fixed. Team synchronization now assigns permissions independently for each member. Administrators receive the correct administrator permission, while non-administrator users and service accounts retain their configured team roles.

This fix applies only to teams created after upgrading to Cloudera AI 1.5.5 SP3 CHF3 and does not change member permissions for existing teams.

In Cloudera AI on premises 1.5.5 SP3, updating an existing Workbench Job to change only the Runtime could fail with CPU and memory validation errors. The update request incorrectly sent 0 value for CPU and memory, even when the resource profile was unchanged. In addition, the Job > Settings page in the UI could display the wrong resource profile.

This issue was caused by a UI timing issue that reset the saved resource profile values before resource groups finished loading.

The issue is now fixed. Runtime updates now preserve the configured resource profile and correctly display the selected profile in Job > Settings.

In a quota-enabled Workbench with heterogeneous GPU mode active, editing per-accelerator quota values in Site Administration > Quotas → did not work correctly.

For example, when updating the quota value for an accelerator, such as NVIDIA-A100, the Maximum Accelerator Count did not reflect the entered value consistently. In addition, the Total Accelerator value was inconsistent because the individual accelerator quota values were not summed correctly. The Total Accelerator value must be the sum of the quota values for all individual accelerators.

This issue affected both the Default Quota and Custom Quota edit modals. CPU and Memory quota fields were not affected.

This issue is now fixed.

After upgrading to Cloudera AI 1.5.5 SP3, the AI Registry page could appear blank and prevent users from viewing registered models.

This issue was caused by incorrect client-side filtering and limited pagination, which excluded some models from display, including restored models, public models, and models owned by system or service accounts.

This issue is now fixed. The AI Registry page now correctly displays public and restored models and uses improved pagination to ensure models are loaded properly.

The Model Endpoint Details page failed to render for model endpoints hosted in older Cloudera AI Inference service versions and displayed an Unknown status for Cloudera AI 1.5.5 SP3. This issue has been resolved, and the UI now renders correctly regardless of your backend version.

This issue is now resolved, and the UI renders correctly regardless of your backend version.