Kaggle Kernel Execution Timeout — No Configurable Grace Period for State Persistence

Direct Answer

No, Kaggle kernels do not provide a configurable grace period or pre‑termination signal; they are terminated with SIGKILL once the runtime limit is reached.

Confirmed Facts

• Kernels are killed with SIGKILL after the platform‑enforced limit (default 60 min, up to 6 h for verified accounts).
• No SIGTERM, no shutdown hook, and the Kaggle API exposes no parameter to request a grace period or adjust the timeout.
• The kernels cancel endpoint also triggers an immediate SIGKILL.
• Dataset download via the Kaggle API uses a fixed ~60‑second HTTP timeout with no override.

Likely Explanation

There is no publicly known undocumented environment variable, metadata field, or hidden callback that grants a pre‑termination window. Feature requests for such functionality exist in Kaggle forums but have not been implemented as of the model’s knowledge cutoff (2024).

Recommended Steps

1. Implement periodic checkpointing in your kernel code (e.g., save model weights, metrics, or intermediate files to the working directory every N minutes or after each epoch).
2. Use a simple timer or epoch‑based trigger; write checkpoints with torch.save, joblib.dump, or plain file copies as appropriate.
3. Optionally, wrap the training loop in a try/except block to catch KeyboardInterrupt (though SIGKILL will not invoke it) and still perform a final save when the loop exits normally.
4. Monitor elapsed time (e.g., time.time()) and break the loop gracefully if the remaining time falls below a safety threshold, then perform a final checkpoint.

Diagnostic Question

Are you using a verified Kaggle account that grants the extended 6‑hour limit? Knowing this helps decide how aggressive your checkpoint interval should be.