pypi genvarloader 0.43.0
v0.43.0

latest release: 0.43.1
one month ago

BREAKING CHANGE

  • datasets written by 0.43.0+ cannot be opened by earlier
    GenVarLoader, which fails with KeyError: 'vk_snp_range'. Reading older datasets
    is unaffected. Re-run gvl.write to shrink an existing dataset; there is no
    migration tool.

Feat

  • concat: merge svar2 range caches by region batch
  • svar2: project the range cache from realized fill
  • svar2: write the range cache sparsely
  • svar2: add the sparse region-CSR range layout
  • threads: let a dataset state its own parallelism policy

Fix

  • svar2: name the offending cell id, not the maximum beside it
  • harden svar2 sparse range cache against silent-error regressions
  • docs: correct svar2 accumulator GB figures and add genome-wide size
  • svar2: guard sparse writer preconditions that corrupted silently
  • svar2: cross-check dense vk shapes, align lookup input handling
  • svar2: validate CSR invariants, bound entries_for_regions, bind Protocol
  • ci: cache the 1kg Zenodo source and retry the download
  • svar2: account for indels in haplotype_lengths

Refactor

  • concat: clear final-review findings on analytic run planning
  • concat: dedupe region-only plan, tighten docstring/comments from review
  • svar2: read ranges through a layout protocol
  • haps: split the Haps role from its SVAR1 implementation
  • haps: dispatch track realignment through the reconstructor
  • haps: measure variant payload through the reconstructor
  • haps: query the reconstructor, not SVAR1 storage

Perf

  • svar2: consume genoray's sparse range stream in the writer
  • concat: yield source-dataset uniques from slot_batches
  • concat: plan every merge through RunPlan, dropping the provenance map
  • concat: build merged offsets from slot batches, cumsum in place
  • concat: add vectorized slot batches and analytic slot count to RunPlan
  • concat: derive merge runs from order without materializing slots
  • svar2: add the sparse range-lookup benchmark gate
  • svar2: stop allocating a dense n_variants at open
  • svar2: cut the ascending-cell-id guard from ~25n to ~n bytes
  • write: replace O(S^2) svar2 sample_cols with a hash lookup
  • svar2: stop re-doing work on the read-bound spliced path

Don't miss a new genvarloader release

NewReleases is sending notifications on new releases.