from-trajectories

The from-trajectories source lets a regular gridded recipe (driven by a dates: block) pull fields from a forecast archive by picking, for each validity time, a (basetime, step) pair that produces it.

It is the inverse of the trajectories layout: instead of materialising a 5-D (base_dates, variables, ensembles, steps, cells) array, it returns a plain 4-D gridded dataset whose fields happen to come from forecast runs.

Recipe shape

dates: {start: 2023-01-01, end: 2023-01-10, frequency: 6h}

input:
  from-trajectories:
    bases: "????-??-?? 00:00:00"     # optional fnmatch pattern on basetime
    steps: 6/to/24/by/6              # optional step spec
    source:
      mars:
        type: fc
        class: od
        expver: "0001"
        param: [t, q]
        levtype: pl
        level: [500, 850]

output:
  layout: gridded

Parameters

bases (optional) — an fnmatch-style wildcard pattern matched against the basetime formatted as "%Y-%m-%d %H:%M:%S". ? matches any single character and * matches any sequence. Omit to accept any basetime.
steps (optional) — a MARS-style step specification: an integer, a list, or a string such as "6/to/24/by/6". Omit to default to step 0 (analysis-like).
source (required) — a single-key dict describing the inner source. Today this must be a source that supports the forecast dispatch (mars:).

Resolution strategy

For each validity time requested by the recipe, candidate basetimes are computed as valid_time − step for every configured step, smallest step first. The first candidate that matches the bases pattern is kept. If none matches, a clear ValueError is raised at build time — so a misconfigured recipe fails loudly rather than producing a silently empty dataset.

The inner source is then invoked once with the resulting (valid_time, basetime) pairs, via the same forecast-dispatch path used inside a trajectory recipe.