v0.20.0

@kitcn/resend

Published @kitcn/resend@0.20.0.

kitcn

Minor Changes

• #337 5d3172b Thanks @MikeyZhang75! - ## Breaking changes

  • Range-filtered count() and aggregate() now stop at aggregateWorkBudget
    work units and throw COUNT_FILTER_UNSUPPORTED /
    AGGREGATE_FILTER_UNSUPPORTED naming the index, instead of reading the whole
    equality prefix and failing on Convex's transaction read limit. Bucket scans
    cost one unit; _min and _max also reserve one extrema read per matching
    bucket. One budget covers every IN prefix and extrema metric sharing the
    range plan. Raise it cautiously if a wide range scan is intentional.

  // Before
  export default defineSchema({ runs });
  
  // After
  export default defineSchema(
    { runs },
    // Keep headroom below Convex's 32,000-document transaction ceiling.
    { defaults: { aggregateWorkBudget: 20_000 } }
  );

  Features

  • kitcn aggregate rebuild and index pruning clear stored aggregate state in
    scheduled batches, so rebuilding or dropping an aggregateIndex / rankIndex
    on a large table no longer has to fit in a single Convex mutation. Indexes
    report a CLEARING status while draining, and kitcn aggregate prune reports
    how many removed indexes are still being cleared in the background.
  • ORM writes targeting a declared index in CLEARING fail before the document
    write. Retry after the index advances to BUILDING or READY; this prevents
    concurrent writes from being erased by a multi-mutation clear.
  • Automatic pruning follows canonical aggregate lifecycle state instead of
    reverse-scanning every distinct backing-table index. Exact tableName and
    indexName handler arguments retain bounded recovery for state-less storage.

  Patches

  • Reduce database reads per rank-index write: the btree descent no longer
    re-queries the tree document at every node, re-reads nodes it just wrote, or
    re-reads the child it descends into.
  • Reduce database reads for rank().min(), rank().max() and rank().random()
    by dropping a redundant count scan per call.
  • Reduce document writes during aggregate backfill: a page of rows sharing a key
    tuple now writes its bucket once instead of once per row.
  • Stop writing aggregate_member rows when a mutation changes no aggregated
    field.
  • Improve count() and aggregate() latency for multi-value IN filters and
    for _min / _max by issuing independent bucket reads through a bounded pool.
  • Keep a partially cleared index in CLEARING when an error interrupts a
    rebuild, so a retry resumes clearing instead of building over stale buckets.