Releases: broadinstitute/sparklespray
Release 3.8.1
bug fix: validating config caused bootDiskSizeGb to not work due to INI parser appears to drop capitalization
Release 3.8.0
Bug fix: when choosing the number of nodes via "sparkles watch --nodes ..." would not update the number of preemptive nodes allowed, resulting in all new nodes to not be preemptive.
Bug fix: When stopping the cluster, cancel all outstanding node requests as well.
New feature: Added a flag on clean: "sparkles clean --only-nodes JOBID" will now clean up completed node requests, but leave the job intact
Release 3.7.0
Changed 'sparkles reset' to also reset jobs which had a non-zero exit code
Bugfix for 'sparkles list': params (previously was showing all columns)
Changed 'sparkles status --stats' to report number of task preempts/restarts
Download Release 3.7.0
3.6.0
New features:
- Added a user message at job start with the job ID
- Abort if unknown parameters are listed in .sparkles config
- New flag on "sparkles status". Adding "--stats" will report summary stats of execution times and memory usage.
- Now displays whether workers are preemptible or not while job is running.
Release 3.5.6
Bug fix for problem where a worker node which is marked "failed" does not have it's claimed tasks reset. (Was resulting in the job hanging indefinitely if a node was unexpectedly killed)
To install, download: sparklespray-3.5.6.tar.gz
Release 3.5.5
Includes bug fix for:
- When running "watch" the return code from completed processes was not shown.
To install download: sparklespray-3.5.5.tar.gz
Various fixes to installation
Fixed problems with 'sparkles validate' and improved 'sparkles setup'
Added timeout to gprc called to prevent preemption from causing hangs
Improved printing logs from jobs:
- if job fails, always print logs from at least one task
- if we were streaming logs from a task and the process dies, fetch the remaining tail from gcs.
Also includes fixes for case where small number of tasks would occasionally be reported as "0 tasks succeeded and 0 tasks failed"
3.3.1
3.2.0
Added a new command "sparkles logs <job_id>" which prints the paths to the logs for each task. By default, only lists those tasks that failed, but you can report everything with the "--all" flag.
Improved "livelog" to stream the logs from the first running task it can find. Now shows logs even if there was more than one task submitted.
Livelog also reports memory usage stats.
Also fixed bug: if a submitting a job resulted in a crash before requesting nodes, the job could not automatically be cleaned up even though it wasn't actually running.