Prisma Streams Routing-Key Performance

Status: informational status note. Dedicated routing-key perf tooling is not currently maintained in this repo.

Current implementation status:

• Per-block bloom filters + segment footer index are implemented.
• Tier‑2 routing-key index runs (L0 + compaction to higher levels) are implemented.
• Keyed reads now populate the local segment cache on first touch and scan that local file, rather than issuing remote block-range reads against uploaded segments.
• Once cached locally, keyed reads use Bun.mmap() and a single forward pass over block headers / matching blocks instead of reopening the cached segment file for many tiny footer and block reads.
• If mmap is unavailable, keyed reads fall back to a single full-file byte buffer for the segment scan, not repeated small local file reads.
• When a routing-key tiered index candidate set exists, keyed reads no longer walk the full indexed sealed prefix segment-by-segment. They build a sealed read plan up front and scan only:
  • the candidate segments from the indexed uploaded prefix
  • the uncovered uploaded tail after indexed_through
• Timestamp seeks with key=... use the same sealed read plan, so routing-key cursor seeks also scale with candidate-set size instead of total sealed history.

That means routing-key reads should scale with the size of the candidate set and the uncovered tail, not with the total number of sealed uploaded segments in the stream.

For details, see:

• tiered-index.md
• overview.md (repository overview and benchmark entry points)
• segment-performance.md

If you need custom routing-key workload benchmarks, use experiments/bench/synth.ts or build a dedicated driver for your query pattern.