v1.48.1

Changed

• Cache sizes to be configurable
• Tracking of how rate limiting is affecting the queue

v1.48.0

Added

• Per-user per-pool job launch rate limiting
• Configurable checkpointing kill switch

Changed

• Dynamic compute cluster log from ERROR to WARN

Fixed

• Allowing for different rate limit for auth-bypass requests

v1.47.1

Changed

• Added warning log when jobs go unmatched for too long
• Added the ability to flush a rate limit from the cache

v1.47.0

Added

• Add rate limits per compute cluster

Changed

• Cache job-constant fields in defrecords of gpu-host-constraint and user-defined-constraint

v1.46.0

Added

• Workload fields to job application

v1.45.2

Changed

• Reduced excessive logging for k8s
• Added logging of offer and job resource percentiles

Fixed

• Missing compute cluster check
• Deleting unschedulable synthetic pods

v1.45.1

Added

• Dynamic cluster configuration support

Changed

• Improved logging for launching tasks, stop launching synthetic pods, and matching offers

v1.45.0

Added

• Support for EQUALS job constraints in k8s
• HOST_IP environment variable for k8s

Changed

• De-lazied the list of constraints to avoid locking in Fenzo
• Made job->previous-hosts-to-avoid use set instead of mapv -> distinct to reduce lock contention

Fixed

• Order of per-user and pool-global quota application

v1.44.0

Added

• Support for prefixed job labels to become k8s pod labels
• Attribution labels to k8s synthetic pods
• agent_id as a preferred alternative to slave_id on job instances
• /shutdown-leader admin-only API endpoint

Changed

• Improved matching log
• Improved per-user launch-rate-limit log
• Added log at start and end of job ranking
• Made "killing cancelled task" log INFO-level

Fixed

• Handling of preemption on k8s pod initialization

v1.43.2

• Using 1024*1024 (mebibytes) as the k8s memory multiplier
• Avoiding NPE due to missing resources when totaling resources for metrics