Chronoxide

Public, minimal extract of the Chronoxide (OTLP-native TSDB) workbench, focused on SymbolTable / string interning performance and memory behavior under high-cardinality label workloads.

This repo exists to accompany the "Arc<str> vs arena interning" write-up and to provide a small, reproducible codebase with:

• ArcSymbolTable: baseline HashMap<Arc<str>, SymbolId> interner.
• ArenaSymbolTable: arena-backed interner (single Vec<u8> + (offset,len) per symbol).
• TrackingAllocator: a cross-platform global allocator wrapper that tracks requested vs usable allocation sizes to approximate allocator internal fragmentation.

Layout

• chronoxide-core/
  • src/labels/symbol_table.rs: ArcSymbolTable + ArenaSymbolTable
  • src/alloc_tracking.rs: TrackingAllocator
  • benches/symbol_table.rs: Criterion benchmark comparing intern/lookup/resolve
  • examples/symbol_table_memory.rs: memory + fragmentation experiment using TrackingAllocator

Requirements

• Rust 1.92.0+ (workspace rust-version)
• No external services required (synthetic dataset generation)

Run the benchmark

cargo bench -p chronoxide-core --bench symbol_table -- --warm-up-time 5 --sample-size 400

The benchmark prints:

• wall-clock timings for intern/lookup/resolve
• best-effort size estimates (estimate_allocated_bytes, estimate_used_bytes)

Run the memory/fragmentation experiment

cargo run --release -p chronoxide-core --example symbol_table_memory -- 512 25000 75000

Arguments:

• 512: number of unique “keys” generated (plus __name__)
• 25000: number of “common” values (reused frequently)
• 75000: number of “rare” values (high-cardinality / mostly-new)

The output includes:

• req_current: total bytes requested by allocations still live
• usable_current: total bytes actually reserved by the allocator (includes rounding)
• internal_frag: usable_current - req_current and percentage
• allocation call counts (alloc_calls, realloc_calls)

Notes on dataset shape

The synthetic string generator is tuned to resemble real OTLP label workloads we observed during a 11M-message ingestion run:

• short keys (≈10–20 bytes typical, max ≈71)
• short-to-medium values with a long tail (up to ≈2048, P99 of Value Len Max/Series is 193)