Assign the dense integer mdl_id for a set of models

mdl_id is the compact partition key of the serving surface (serve/model_cell/mdl_id={id}/), so a titiler tile is a partition-pruned point read. The stable public identifier stays mdl_key: titiler resolves mdl_key -> mdl_id from the published model registry, so mdl_id never appears in a URL.

Usage

assign_mdl_id(mdl_key, published = NULL)

Arguments

mdl_key

character vector of model keys (need not be unique or sorted)

published

optional data frame of the already-published registry with columns mdl_key and mdl_id; ids for keys present here are preserved

Value

integer vector of mdl_id, parallel to mdl_key

Details

That indirection has a sharp edge. A plain dense_rank(mdl_key) is a deterministic function of the model set, which means adding any model renumbers every model sorted after it. Ingesting gm + nc into a released version's dist/ — registered but deliberately not merged — moved 45,499 of v8's 80,261 mdl_ids, and nothing would have failed: the registry and the serve/ partitions would simply disagree, and titiler would serve the wrong species' distribution for every model past ch_nmfs.

So ids are assigned against the published registry when one exists: a mdl_key already published keeps its id, and new keys are appended above the current maximum in sorted order. Deterministic given (published registry, model set), and it can never invalidate a published partition. With published = NULL (a version that has never shipped) it falls back to dense_rank over the sorted keys, which is what produced the ids of every release to date.

Examples

assign_mdl_id(c("b", "a", "c"))                                     # 2 1 3
#> [1] 2 1 3
assign_mdl_id(c("b", "a", "c"),
              data.frame(mdl_key = c("a", "c"), mdl_id = c(1L, 2L))) # 3 1 2
#> [1] 3 1 2