Cole Halden

Low-jitter packet ingestion and storage engines for telemetry-heavy teams.

I design Rust and C23 data paths for robotics, sensor-fusion, and market-data teams that outgrow general-purpose message brokers. Most engagements replace heap-heavy ingestion queues with bounded ring buffers, deterministic replay logs, and compact Parquet or ClickHouse writers.

Seattle · Pacific Time Direct: cole@colehalden.com
NIC RX QUEUE -> UMEM FILL / COMPLEMENT RINGS 4096 DESC AF_XDP RX batch=64 SPSC Arena Ring 2 MiB hugepages io_uring WAL O_DIRECT NVMe INGEST TAIL DISTRIBUTION (US, ISOLATED CORES) p50 4.2 us p99 11.8 us
Pinned RX -> SPSC -> NVMe segment path Zero heap allocs in hot loop

Selected Work

Deterministic LiDAR and IMU Replay Store

Rust · io_uring · Zstd

Built a segment-indexed capture daemon for an autonomous perception fleet so firmware engineers can seek any 100 ms sensor window from S3 or local NVMe with bit-identical clock reconstruction.

Kernel-Bypass Multicast Recorder

C23 · AF_XDP · eBPF

Replaced a user-space socket collector dropping bursts during market open with an XDP filter and pre-faulted UMEM ring, holding zero packet drops across dual 25 GbE interfaces.

Parquet Compaction Worker for Edge Gateways

Rust · Arrow · ClickHouse

Wrote a bounded-memory streaming compactor that merges tiny sensor batches on 4 GB ARM gateways before uplink, cutting object-store PUT volume by an order of magnitude without OOM kills.

Field Notes

Date Topic Summary
2026-09 io_uring registered buffers on NVMe When IORING_OP_WRITE_FIXED beats buffered page-cache writes under sustained 1.4 GB/s append loads, and where alignment traps appear on ext4.
2026-06 Clock drift across PTP and host MONOTONIC_RAW Reconciling NIC hardware timestamps against host tracepoints without introducing backward time jumps in downstream Arrow tables.
2026-02 Sizing SPSC queues for bursty sensor frames Calculating ring depth from PCIe completion stalls and Zstd frame flush latency rather than average throughput.