AI 3.8-15.1.9 Release Notes
Summary
New Features
Bulk unmount support adds REST- and CLI-based unmount of multiple mount points, including wildcard and dry-run workflows.
Major Enhancements
Job Service load operations are the main focus of this release. Performance is significantly improved by moving load operations onto a unified streaming path, decoupling file listing from task dispatch through background prefetching, improving scheduler behavior on small clusters, and reducing large-file memory pressure so the system can sustain higher throughput across both many-file and large-file workloads.
Job Service load operations stability is also strengthened. Retries are better aligned with worker-loss behavior, offline-worker tasks can be reassigned more safely, and executor behavior is hardened so saturated workers are less likely to leave load operations stuck indefinitely.
S3 List API performance and compatibility are another major focus. Non-delimited
ListObjectsV2over large prefixes is faster, paging behavior is more robust, emptystart-afterhandling is safer for S3-compatible stores such as BOS, and S3 Express directory-bucket listing behavior is improved.Stale-cache cleanup is significantly safer and more usable. This release fixes an important issue where stale-cache cleanup could remove valid segments of segmented files, and it adds a safer scan-oriented mode so operators can inspect stale cache before taking cleanup actions. See Removing Data from the Cache.
Worker-side S3 read behavior is improved for bursty reads and slow or stalled UFS connections, helping reduce silent stalls in affected read paths.
Control-plane resilience improvements strengthen etcd recovery, membership handling, dynamic-config behavior, and repeated mount/update scenarios.
Bug Fixes
S3 request-path correctness fixes address event-loop blocking, HTTP state-machine corruption,
aws-chunkeddecode failures, list paging overruns, and idle-close issues on in-flight requests.Local page-store correctness fixes ensure page-write failures surface properly and prevent stale maintenance logic from deleting valid cache data.
Stale-cache cleanup correctness fixes prevent valid cache segments from being removed during cleanup workflows.
Job Service load operations recovery fixes prevent operations from stalling or losing files when workers go offline during a load.
Embedded JVM UFS loading and metadata fallback behavior are more reliable in affected environments.
Breaking Changes
No breaking changes in this release.
Upgrade Notes
No migration steps are required for this release.
Last updated