Checkpoints: a requeued job resumes from its saved directories issue #77
open andymccall opened this · 0 comments
andymccall
A job may declare checkpoint "/src/target" (and more): directories saved by content (chunked, deduplicated, kept for the run) on drain and every few minutes. A requeued job restores them before running the same exec again, so incremental tools (cargo, gradle, make) skip what was done. The partial run's receipt signs the checkpoint; the resumed job's receipt names it among its inputs, so provenance says which runner did which part; policy decides whose checkpoints a job accepts (issue #79). No VM migration (docs/handover.md §3). Design: §7.
A job may declare
checkpoint "/src/target"(and more): directories saved by content (chunked, deduplicated, kept for the run) on drain and every few minutes. A requeued job restores them before running the same exec again, so incremental tools (cargo, gradle, make) skip what was done. The partial run's receipt signs the checkpoint; the resumed job's receipt names it among its inputs, so provenance says which runner did which part; policy decides whose checkpoints a job accepts (issue #79). No VM migration (docs/handover.md §3). Design: §7.