RangeBitmap, SliceZ, and Kubernetes Metadata at Yandex
A concrete observability storage case where the right bitmap index turns time filtering from a linear scan into a production-scale metadata query path.
Links: Original source · Shared link · Related link · Related link 1 · Related link 2 · Related link 3
Logged at IST: 2026-08-19 14:28 IST
What it is: Richard Startin points to a Yandex Monium write-up using RangeBitmap, a range-encoded bit-sliced index he designed for Apache Pinot and which is available in RoaringBitmap. His follow-up links SliceZ, his newer Java successor for range and point queries over unsorted numeric data.
Gist: Yandex Monium stores billions of metrics, processes about 50 GB/s of logs, and serves around 16,000 internal users. Kubernetes broke an old metadata assumption: pod labels are short-lived, but historical label values accumulate. After a few weeks, selectors can hit "too many metrics" because the metadata service sees all historical pods, not just the currently active ones.
The practical fix was to add time filtering without physically segmenting the shard by time. Monium already had an inverted bitmap index over labels. They added RangeBitmap indexes over created_at_seconds and last_point_seconds, then used the inverted index result as a narrowing context for range queries. That made wide metadata filtering scale where a linear scan could not.
The reported result is typical speedups of 6-20x, with up to two orders of magnitude in favorable cases. The conclusion is also useful because it names the constraints: RangeBitmap rebuild costs need amortization, which Monium handles with LSM-like structures; batch loading and widened query ranges reduce rebuild pressure; binary search still wins when the data can be sorted; and small scans can be cheaper than the index overhead. With those trade-offs, Monium handles 650K-850K metadata queries/sec carrying time-range information.
Startin's follow-up is the next step: SliceZ is a Z-layout bit-sliced index for unsorted numeric data. It supports equality, ranges, between, in, counts, sums, means, and top/bottom-k queries, with persistence and memory mapping. He describes it as usually faster than RangeBitmap and as carrying the same filter-pushdown direction that made the Yandex use case work.
Newsletter angle: Good systems/data-structures piece. It shows bitmap indexes as a way to add logical time segmentation and filter pushdown to an observability metadata path without rebuilding the physical storage model.