Jupyter notebook kernel keeps dying: DeadKernelError, crash vs out of memory (tested)

When a Jupyter kernel keeps dying, the Python process was killed or crashed, so there is no traceback: running out of memory (the OS kills it) and a crash in compiled code look exactly the same. Below both are reproduced, plus how a memory limit turns the death into a catchable MemoryError.

1. Reproduce: a crash, a hard kill, and a memory limit

run.sh builds three small notebooks and executes them headless with nbconvert, which reports a dead kernel as DeadKernelError (the Notebook/Lab UI shows its own "kernel died / restarting" notice for the same event). os.abort() stands in for a crashing C extension; SIGKILL is what the Linux out-of-memory killer sends.

cd /tmp/kd408
export PATH=/tmp/jn408/ve/bin:$PATH
mk() { python -c "
import json,sys
nb={'cells':[{'cell_type':'code','execution_count':None,'id':'c%d'%i,'metadata':{},'outputs':[],'source':[s]} for i,s in enumerate(sys.argv[2:])],
 'metadata':{'kernelspec':{'name':'python3','display_name':'Python 3','language':'python'}},'nbformat':4,'nbformat_minor':5}
json.dump(nb,open(sys.argv[1],'w'))" "$@"; }
mk crash.ipynb "print('cell 1 ok')" "import os; os.abort()   # stands in for a crashing C extension"
mk kill.ipynb "print('cell 1 ok')" "import os, signal; os.kill(os.getpid(), signal.SIGKILL)   # what the Linux OOM killer does"
mk limit.ipynb "import resource; resource.setrlimit(resource.RLIMIT_AS, (1500 * 1024**2,) * 2)" "x = bytearray(3000 * 1024**2)" "print('kernel still alive')"
echo '$ jupyter nbconvert --to notebook --execute crash.ipynb'
jupyter nbconvert --to notebook --execute crash.ipynb 2>&1 | grep -E "DeadKernelError|Kernel died|^\[NbConvertApp\]" | head -3
echo
echo '$ jupyter nbconvert --to notebook --execute kill.ipynb'
jupyter nbconvert --to notebook --execute kill.ipynb 2>&1 | grep -E "DeadKernelError|Kernel died" | head -2
echo
echo '$ jupyter nbconvert --to notebook --execute --allow-errors --stdout limit.ipynb | grep -E "ename|kernel still"'
jupyter nbconvert --to notebook --execute --allow-errors --stdout limit.ipynb 2>/dev/null | grep -E '"ename"|kernel still alive\\n"$'

Output (nbconvert 7.17.2, ipykernel, Python 3.13 on Linux):

$ jupyter nbconvert --to notebook --execute crash.ipynb
[NbConvertApp] Converting notebook crash.ipynb to notebook
[NbConvertApp] ERROR | Kernel died while waiting for execute reply.
    raise DeadKernelError("Kernel died") from None

$ jupyter nbconvert --to notebook --execute kill.ipynb
[NbConvertApp] ERROR | Kernel died while waiting for execute reply.
    raise DeadKernelError("Kernel died") from None

$ jupyter nbconvert --to notebook --execute --allow-errors --stdout limit.ipynb | grep -E "ename|kernel still"
     "ename": "MemoryError",
      "kernel still alive\n"

• A native crash (abort/segfault in compiled code) and a hard kill look identical from Jupyter's side: Kernel died while waiting for execute reply. There is no Python traceback, because the process is gone, so the message alone doesn't tell you which it was.
• An out-of-memory kill gives no Python MemoryError. When the process has a memory limit instead (resource.setrlimit(RLIMIT_AS, ...) at 1.5 GB in limit.ipynb), the 3 GB allocation raised MemoryError as a normal exception and the next cell still ran: the kernel survived.

2. Finding the cause

1. Memory: watch memory while the cell runs; on Linux check the system log (dmesg or journalctl -k) for an out-of-memory kill right after the death. Fix by loading less (chunks, smaller dtypes, del big objects) or, as above, set a memory limit so you get a catchable MemoryError instead of a dead kernel.
2. Native crash: run the same code as a plain script (jupyter nbconvert --to script, then python -X faulthandler file.py) to see which line and library crashes; mismatched binary packages (e.g. two copies of a library from pip and conda, or a wheel built for another NumPy) are the usual suspects. Reinstall them in one environment.
3. Wrong environment: if the kernel dies right at startup, check jupyter kernelspec list and the Python path in the kernel's kernel.json (see kernel not found).

Related: Jupyter "Error Starting Kernel", jupyter nbconvert is not installed.

Written with AI assistance. Measured on 2026-10-23 with nbconvert 7.17.2, ipykernel and Python 3.13 on Linux. The script ran verbatim and the output was pasted unedited. PlainTables is our extension; Jupyter is free.