GitHub user karthikreddy-02 created a discussion: [GSoC 2026] TPU Support & SOTA Time Series Models for IoTDB-AINode- Potential Contributor
Hi all, I'm Karthik, a software engineer with a CS background currently working as a controls engineer, though I've been actively transitioning into backend/ML systems. I've been contributing to open source projects like Turso (libSQL) and SGLang, which is how I got comfortable with database internals and large codebases. I came across the AINode TPU project through the GSoC 2026 ideas list and it caught my attention immediately — honestly, I would've loved something like this at some of the cryogenic sites I was commissioning. The idea of keeping ML inference inside the database rather than shipping data out to an external platform just makes sense, especially for industrial IoT where latency and data security matter. I've been reading through the AINode codebase to understand how Timer and Chronos are registered and invoked through SELECT * FROM FORECAST(...). One thing I noticed is that the current runtime doesn't have an abstraction layer for swapping accelerators without touching model code. My current thinking is a DeviceManager class that handles detection priority (TPU → GPU → CPU) so model wrappers stay hardware-agnostic — curious if that's the direction you had in mind. On model integration, I'm leaning toward TimesFM as the primary candidate since it was built with TPU deployment in mind and has an official HuggingFace checkpoint. Moirai is my backup. Both support zero-shot forecasting, which matters here since users won't always have labeled training data for their sensor streams. Two quick questions: For Poetry packaging — would an optional extras group (poetry install --extras tpu) work, so the base install stays lean? For CI — are you open to using Google's TPU Research Cloud (TRC), or were you already thinking simulator-based? I have a preliminary proposal draft ready and would love any feedback on it. Also just subscribed to the dev mailing list and sent a note to Yongzao there as well. And shoutout to the other discussion posts — they nudged me to finally write this up! Thanks, looking forward to contributing! Karthik GitHub link: https://github.com/apache/iotdb/discussions/17252 ---- This is an automatically sent email for [email protected]. To unsubscribe, please send an email to: [email protected]
