Repository navigation
PRELUDE summary steps never cache on -resume (expensive on AWS Batch) #162
Description
Description of the bug
On -resume, the PRELUDE:PRELUDE_SUMMARY_* steps re-run every time, even when nothing upstream has changed.
It traces to subworkflows/local/prelude/main.nf, where the summary inputs are JSON files built on the fly inside the .map{} operators, e.g.
def f = file("$meta".md5().toString() + ".json")
f.write(new groovy.json.JsonBuilder(meta).toString())
The name and contents are stable, but the file gets rewritten on every run, so its last-modified time changes — and Nextflow's default input hashing keys on name+size+mtime, so it's a cache miss each time.
On local/SLURM this is harmless (the tasks take ~1s). On AWS Batch it isn't: each re-run has to be scheduled onto an instance (queue → scale-up → image pull), so a 1-second task shows a ~5-minute duration. Every resume then pays several minutes of pure scheduling overhead for work that didn't change, which eats into the whole point of resuming past the expensive ASHLAR/segmentation steps.
Fix would be to set cache 'lenient' on the PRELUDE summary processes (lenient hashing ignores mtime, so a stable name+size caches correctly), or to stop generating the input files inside the operators.
Command used and terminal output
Relevant files
No response
System information
No response
Activity