Hydroland v1.2.6

Hydroland v1.2.6 is a stability + capability update for the mRM workflow. It delivers end-to-end ~5 km runs, first-class Bias Adjustment (BA) integration, a month-by-month execution mode, and safe restart from a previous Hydroland run.

No breaking changes. Monthly and hourly modes are unaffected.

🛠 Fixes

• Daily filename patterns
  Corrected output filename patterns in preprocess_forcings (and downstream steps) to prevent mismatches and missing-file lookups.

• OPA file retention
  Avoids removing OPA files that may be needed for later post-processing.

• Safe restart
  Fixed a bug that prevented clean resumption from a prior Hydroland execution.

✨ New & Improved

End-to-End ~5 km Runs

• Full workflow now supports ~5 km grid resolution.
• Compatible with parallel mRM orchestration and updated merge/mosaic stages.
• Safer temp/working-dir hygiene and more granular progress logging.

Heads-up: 5 km runs are heavier—expect increased I/O and memory. See Performance Notes below.

Integrated Bias Adjustment (BA)

• BA can be enabled from the CLI/config and is applied consistently across chunks.
• Clear logs and metadata tags indicate when BA is active.
• Keeps BA inside Hydroland for reproducibility (no external one-offs required).

mRM Workflow Orchestration

• ParallelRun – more robust job orchestration with safer temp/working-dir handling on large domains.
• write_run_parallel_mrm – reproducible launcher generation from parsed args (audit-friendly).
• generate_mrm_arg_list – clearer per-tile/per-chunk argument construction to aid debugging.

Flux Merging & Tiling

• mrm_fluxes_merge_batched (new fast path) and mrm_fluxes_merge_cdo (hardened)
  Deterministic, batched merges with retries for missing/empty tiles and stable output ordering.
• mosaic_all_tiles & _merge_pairs
  Better handling of borders/partials and uneven chunk counts.

⚙️ Performance Notes (5 km)

• Higher parallelism will amplify I/O; monitor filesystem hotspots.
• Memory footprint scales with tile/chunk size—tune chunking accordingly.
• Prefer local scratch for intermediates; cleanups are safer by default but still configurable.

Empirical timings

| run_step | grid_km | avg_total_min | avg_initial_preproc_min | avg_mhm_exec_min | avg_mrm_exec_min | |:--------:|:-------:|--------------:|------------------------:|-----------------:|-----------------:| | hourly | 10 | 1.094 | 0.196 | 0.329 | 0.569 | | daily | 10 | 0.451 | 0.127 | 0.155 | 0.170 | | hourly | 5 | 4.054 | 0.427 | 1.220 | 2.407 | | daily | 5 | 1.110 | 0.092 | 0.512 | 0.506 |

Quick takeaways

• Time composition (share of total):
  • Hourly 5 km: ~10.5% preproc, ~30.1% mHM, ~59.4% mRM.
  • Hourly 10 km: ~17.9% preproc, ~30.1% mHM, ~52.0% mRM.
  • Daily 5 km: ~8.3% preproc, ~46.1% mHM, ~45.6% mRM.
  • Daily 10 km: ~28.2% preproc, ~34.4% mHM, ~37.7% mRM.
• Scaling 5 km vs 10 km:
  • Hourly total time is ~3.71× (4.054 / 1.094).
  • Daily total time is ~2.46× (1.110 / 0.451).
• Hourly vs daily (same grid):
  • At 10 km, hourly is ~2.42× slower than daily (1.094 / 0.451).
  • At 5 km, hourly is ~3.65× slower than daily (4.054 / 1.110).

Thanks for using Hydroland! Feedback and contributions welcome—please open issues or PRs on GitHub.