Staff / Principal ML Infrastructure Engineer

Summary

The Apple AI Platform team builds the foundational data storage, table format, and high-throughput data-loading systems powering Apple Foundation Models and large-scale ML training. We are designing and building an AI-native table format from the ground up: combining transactional snapshot isolation, concurrent updates, and multimodal schema evolution with built-in global indexing and sub-millisecond random access across multi-petabyte datasets on thousands of GPUs and TPUs.

Minimum Qualifications

10+ years of experience building storage engines, table or file formats, database internals, or high-performance distributed data and ML systems. Proven track record of designing and implementing table or file format internals (such as Apache Iceberg, Lance, Delta Lake, Parquet, Arrow, Vortex, Riegeli/ArrayRecord, RocksDB/LSM trees, or custom indexing and compaction engines), not just consuming them via high-level APIs. Production systems programming mastery in Rust or C++ (Rust strongly preferred for our core data plane), plus Python/PyO3 for ML framework bindings. Deep hands-on expertise in low-level I/O and memory performance engineering: mmap, zero-copy memory layouts, SIMD and vectorized decoding, RoaringBitmaps, async runtimes (tokio), and high-throughput S3 and GCS object-store client internals (range coalescing, connection pooling, tail-latency hedging). Strong understanding of transactional storage semantics: snapshot isolation, MVCC and optimistic concurrency control, manifest trees, and metadata scaling at millions of files and tens of billions of rows. Working knowledge of how large-scale distributed training (JAX/Grain, PyTorch) consumes data across GPU and TPU meshes (global shuffling, sequence packing, deterministic checkpoint and resume, host-to-device memory pipelines). Track record of setting technical direction across teams, shipping production systems at petabyte scale, and elevating senior engineers. B.S., M.S., or Ph.D. in Computer Science, Computer Engineering, or equivalent practical experience.

Preferred Qualifications

Core contributor or committer experience on Apache Iceberg, Apache Arrow, Parquet, Lance, Vortex, DataFusion, Velox, DuckDB, Polars, Daft, or RocksDB. Experience designing or scaling ML training data loaders and formats (PyGrain, ArrayRecord/Riegeli, Megatron-LM .idx/.bin, Mosaic StreamingDataset, Ray Data, or NVIDIA DALI/AIStore) on multi-hundred-host GPU and TPU clusters. Experience designing multimodal dataset storage (combining pre-tokenized sequences, audio, image, and video blobs, and evolving feature or label columns without full dataset rewrites). Experience with formal correctness or high-rigor verification practices (differential parity testing, property-based fuzzing, or state-machine modeling for concurrent commits).

Description

As a Staff/Principal ML Infrastructure Engineer on the team, you will architect the on-disk format, indexing structures, transactional metadata layer, and Rust data plane and training data loader that Apple's largest training workloads run on.

Responsibilities

Architect our next-generation AI-native table and indexing format: design the on-disk layout, snapshot isolation and time-travel metadata, optimistic concurrency control (OCC) for concurrent writers, zero-copy selection and deletion bitmaps, and multimodal column-group evolution across multi-petabyte datasets. Design and build compressed global indexing and random-access primitives: enable true global shuffling and single-RTT ranged object-store reads across tens of millions of files (ArrayRecord, Parquet, Lance) without exhausting host memory. Lead the core Rust data plane and high-performance training data loader: architect the off-GIL async transport, zero-copy Arrow and record decoding, range coalescing, and tail-latency request hedging that keep multi-thousand-chip GPU and TPU jobs compute-bound. Bridge storage engine internals with ML training frameworks: own the integration contracts between our storage and table layer and JAX (Grain), PyTorch, and Hugging Face for pre-training, mid-training, SFT, and RL workloads. Make system-level design calls backed by hardware and network math: benchmark and prove throughput, memory footprint, QPS scaling, and tail latency on real cluster workloads before shipping. Raise the engineering bar across the organization: mentor senior engineers, lead design and code reviews in Rust and Python, and partner with foundation-model research and data-platform teams across Apple.