Anthropic • 📍 San Francisco / US Remote •
$240k - $320k USD
You will debug multi-node NCCL/RDMA collective stalls, write lock-free RadixAttention prefix routers in Tokio, and interface directly with custom FP8 CUDA kernels via C FFI.
#io-uring • $197k-$267k USD Base
io_uring in Production Rust: Architecture & Open Roles
Linux kernel asynchronous ring-buffer I/O for NVMe & network sockets
1. Why Teams Deploy io_uring
Traditional POSIX file I/O in Linux requires thread pools because `epoll` does not support regular files, and Linux AIO suffers from severe quirks. `io_uring` solves this by sharing two lock-free ring buffers, the Submission Queue (SQ) and Completion Queue (CQ), in memory mapped between user space and the Linux kernel. Next-gen Rust databases like Turso's Limbo (SQLite rewrite) and Qdrant use `io_uring` to saturate modern NVMe drives.
2. io_uring Kernel & Memory Internals
With `io-uring` or thread-per-core runtimes like `glommio` and `monoio`, a storage engine queues dozens of NVMe read, write, `fsync`, and socket `recv` operations into the SQ ring and submits them in a single `io_uring_enter` syscall, or zero syscalls when `IORING_SETUP_SQPOLL` kernel thread polling is enabled. Registering fixed pre-allocated memory buffers (`IORING_REGISTER_BUFFERS`) eliminates per-I/O page pinning overhead inside the kernel.
3. io_uring Interview & Hiring Bar
Because `io_uring` operates on completion-based ownership rather than readiness-based borrowing (`Poll`), Rust engineers who understand buffer pinning, `tokio-uring`, and completion safety command top-decile storage salaries.
Open io_uring Engineering Positions (5)
Modal Labs • 📍 US / Americas Remote •
$190k - $260k USD
Optimize Linux cgroups v2 isolation, FUSE page-cache prefetching, and snapshot memory restoration across thousands of H100 nodes.
Qdrant • 📍 Europe / US Remote •
$175k - $235k USD
Design disk-backed HNSW graph traversal over Linux io_uring, tune Raft shard consensus rebalancing, and write hand-vectorized distance functions.
Astral • 📍 Worldwide Remote •
$190k - $255k USD
Optimize cross-platform filesystem syscalls on Linux, macOS, and Windows, tune PubGrub version solving for complex PyTorch/CUDA index trees, and reduce binary footprint.
Neon (Databricks) • 📍 Worldwide Remote •
$190k - $265k USD
Improve L0-to-L1 compaction throughput over `tokio-epoll-uring`, lower tail latency on cold page reconstructions, and harden Paxos WAL quorum recovery.