gedih3.doctor.diagnoses.tmp_partitions_health

tmp_partitions_health diagnosis — post-build forensics on tmp/partitions/.

Audit-only check that NEVER iterates healthy partitions (Pillar 1 — no driver-side O(N) GPFS scans). It acts strictly on the signals the build already persisted:

• tmp_dir/_merge_failures/*.fail sentinels (one per failed merge)

• tmp_dir/_granule_failures.jsonl (one record per failed granule)

• tmp_dir/_merge_progress.txt vs. the DB _manifest.txt line count

The cost of the check is O(N_failures) + O(1) on the manifest — never O(N_partitions). All work runs on the driver because the inputs are narrow (a single sentinel directory + two flat files).

The --fix variant calls gh3builder.preclean_merge_failures() to remove zero-byte / unreadable parquets + .tmp siblings under every failed-merge partition. Granule failures are reported only — recovery belongs in gh3_update --recover-missing-vars. As a hard guard, the fix refuses to act when a fresh gh3_build process appears to be running (recent gh3_build.log mtime + matching pgrep PID).

Functions

tmp_partitions_health_check(→ gedih3.doctor.report.Report)
tmp_partitions_health_fix(→ gedih3.doctor.report.Report)

Module Contents

gedih3.doctor.diagnoses.tmp_partitions_health.tmp_partitions_health_check(ctx: gedih3.doctor.report.DoctorContext) → gedih3.doctor.report.Report

gedih3.doctor.diagnoses.tmp_partitions_health.tmp_partitions_health_fix(ctx: gedih3.doctor.report.DoctorContext, report: gedih3.doctor.report.Report) → gedih3.doctor.report.Report