ROW
Row .nxb
Stream records as they arrive. O(1) seek via tail-index after seal. Sub-µs warm access to any field.
7 µs TTFR P50 (EPYC) · 37 µs Haswell
Virtual scroll · log explorers · APM traces · ticker updates
Browser-scale structured data
Stream, filter, and explore GB-scale structured data without JSON hydration bottlenecks, browser freezes, or memory blowups.
Scroll, search, and jump across millions of rows in the browser. The Explorer is the fastest way to feel the difference — no install, built-in fixtures up to 10M records.
.nxb bytes.nxb — before download finishesInteractive browser benchmarks · full methodology · why JSON breaks at scale
Modern observability and ops UIs ship multi-hundred-megabyte JSON exports. The pain is familiar before any wire format debate.
RangeError before JSON.parse() runs.2^53−1 lose precision with no error — only wrong analytics.Compare what users experience — not which ecosystem you must rip out.
| Workflow moment | JSON in the browser | Nyxis .nxb |
|---|---|---|
| Open a large export | Wait for full download + parse | Header + tail-index — stream rows as bytes arrive |
| First interaction | Often 10–15+ seconds on 100 MB+ | Scroll and search while the file is still loading |
| Filter / search | Scan in-memory object graph | Selective field reads + worker-backed index |
| Memory at 1M+ rows | Full dataset on the heap | Viewport pool — only visible rows decoded |
Nyxis moves parsing cost from query time to compile time. Human-readable .nxs compiles to memory-mapped .nxb with a tail-index for O(1) record seeks — readers use zero-copy selective access instead of materializing entire documents.
mmap-friendly wire cells; decode only the fields you touch. .nxs source; production .nxb for browsers and services. .nxsCompiler + drivers produce .nxbBrowser / BI / agents query without full decode Pick row, columnar, or pax at compile time — not one layout for every problem. Layout selection guide →
.nxbStream records as they arrive. O(1) seek via tail-index after seal. Sub-µs warm access to any field.
7 µs TTFR P50 (EPYC) · 37 µs Haswell
Virtual scroll · log explorers · APM traces · ticker updates
.nxbField buffers for charts and aggregates. No per-record traversal for column scans.
1.3× Arrow IPC on EPYC · 1.7× Apple Silicon
OLAP · export pipelines · CSV → chart demo
.nxbPage-oriented hybrid: scroll rows and scan columns from one sealed file (SPEC §4.5).
Workload E · page_size ≥ 32,768 on x86 server
Dashboards that mix virtual scroll with column charts
user { id = 42 name "ada_lovelace" score ~ 98.6 active ? true }
Every value in .nxs carries a sigil that declares its binary encoding. The source file is the schema — no separate IDL. Compile once; warehouses and UIs read aligned cells for the lifetime of the payload.
.nxs is plain text you can diff, review, and hand-edit..nxb on demand.nxs-mcp exposes typed tools so agents query .nxb without custom parsers. Producers emit aligned .nxb bytes while a segment is still open. Readers parse complete records as they arrive and only need the footer tail pointer once the writer seals the file.
.nxbPreamble TailPtr = 0 during ingest; sealing writes FooterTailPtr + MagicFooter at EOF..nxsw)Hot paths append NYXO rows without rewriting the tail-index every span.Workflow metrics first — time to interactive, filter latency, browser memory — then raw throughput vs Protobuf, FlatBuffers, Cap'n Proto, and Arrow. We publish losses where NXS is not the winner. Interactive charts · BENCHMARK.md.