tini vs dumb init

readme.md

I test images. Sometimes they use tini, sometimes they use dumb-init. I have noticed a case where the tini doesn't forward SIGTERM all the way to the app, but dumb-init does.

I want to understand what's different about how these tools are handling signals.

The problematic case arises when I insert a shim which runs the actual app as a subprocess. The goal is to add a custom signal handler to the shim so that I don't have to modify the app. Ideally this will work without me having to think about whether the image uses tini or dumb-init.

In that handler I intend to export telemetry data, but for now it just does echo after so I can see if it ran.

The problem is that when I do docker stop, the underlying app doesn't get the signal. Instead, docker just waits 10 seconds and kills the app--which means that my handler doesn't run.

Dockerfile

FROM ubuntu
RUN apt update
RUN apt install -y tini dumb-init python3-pip
RUN pip install apache-airflow

# place the entrypoint script
COPY entrypoint.sh /entrypoint
RUN chmod +x /entrypoint

# a script that calls airflow (and does test-relevant things too)
COPY shim.sh /root/shim

# rename airflow and put the shim in its place
RUN DIR=$(dirname $(which airflow)) \
    && mv $DIR/airflow $DIR/actual-airflow \
    && mv /root/shim $DIR/airflow \
    && chmod +x $DIR/airflow

# works
ENV INITIALIZER=dumb-init

# breaks
#ENV INITIALIZER=tini

ENTRYPOINT [ "/entrypoint" ]

entrypoint.sh

#!/usr/bin/env bash
set -ex

# We can't call the initializer in the entrypoint because
# of something to do with wanting to run it as a different user.

# I have simplified that part away in this case because the problem persists
# even without different users involved.

if [[ -z "$ALREADYINIT" ]]; then
ALREADYINIT=1 exec $INITIALIZER -- "$0" "$@"
fi

exec "$@"

session.dumb-init

❯ docker run experiment bash -c 'airflow db init && airflow scheduler'

    + [[ -z '' ]]
    + ALREADYINIT=1
    + exec dumb-init -- /entrypoint bash -c 'airflow db init && airflow scheduler'
    + [[ -z 1 ]]
    + exec bash -c 'airflow db init && airflow scheduler'
    before (PID 8)
    DB: sqlite:////root/airflow/airflow.db
    ...
    Initialization done
    
    after   <------!!!!! that's my handler for `airflow db init`
    
    before (PID 7)
      ____________       _____________
     ____    |__( )_________  __/__  /________      __
    ____  /| |_  /__  ___/_  /_ __  /_  __ \_ | /| / /
    ___  ___ |  / _  /   _  __/ _  / / /_/ /_ |/ |/ /
     _/_/  |_/_/  /_/    /_/    /_/  \____/____/|__/
     ...
    [2022-07-08 20:41:47 +0000] [15] [INFO] Booting worker with pid: 15
    
    # `docker stop` executed here
    
    [2022-07-08 20:41:53 +0000] [11] [INFO] Handling signal: int
    ...
    [2022-07-08 20:41:55,139] {scheduler_job.py:768} INFO - Exited execute loop
    
    after  <------!!!!! that's my handler for `airflow scheduler`

session.tini

❯ docker run experiment bash -c 'airflow db init && airflow scheduler'

    + [[ -z '' ]]
    + ALREADYINIT=1
    + exec tini -- /entrypoint bash -c 'airflow db init && airflow scheduler'
    + [[ -z 1 ]]
    + exec bash -c 'airflow db init && airflow scheduler'
    before (PID 8)
    DB: sqlite:////root/airflow/airflow.db
    ...
    Initialization done
    
    after    <------!!!!! that's my handler for `airflow db init`
    
    before (PID 7)
      ____________       _____________
     ____    |__( )_________  __/__  /________      __
    ____  /| |_  /__  ___/_  /_ __  /_  __ \_ | /| / /
    ___  ___ |  / _  /   _  __/ _  / / /_/ /_ |/ |/ /
     _/_/  |_/_/  /_/    /_/    /_/  \____/____/|__/...
    [2022-07-08 20:38:29 +0000] [15] [INFO] Booting worker with pid: 15
    
    # `docker stop` killed the container, but no more logs were shown, and the handler was not called

shim.sh

#!/bin/bash
echo "before (PID $$)"

# run this on sigterm
function finished()
{
    echo after
    exit $1
}
trap finished SIGTERM SIGINT

actual-airflow $@

# also run it when not interupted
finished $?