Is it possible/safe to install Client Push App ?

[WiseMetalhead]

@andreasb said:

it looks like notify_push is up and running

I honestly don’t understand how you managed to get this working. I spent about 5 hours today trying to set it up and still didn’t get the result I expected.

1. @reboot cron job simply doesn’t work. It’s mentioned in the docs but I’ve never seen it run in the logs. @service does run but only once. Аfter subsequent Nextcloud restarts the job never triggers again.
2. My version of the script works fine if I run it manually from the terminal like this /app/data/notify_push.sh. The service starts and all checks pass.

root@71780ea5-011c-4c09-886e-5f7e17e0b670:/app/code# /app/data/notify_push.sh ; sleep 1 ; ps aux | grep notify_push ; ss -tlnp | grep 7867
✓ redis is configured
✓ push server is receiving redis messages
✓ push server can load mount info from database
✓ push server can connect to the Nextcloud server
✓ push server is a trusted proxy
✓ push server is running the same version as the app
  configuration saved
=> notify_push started on port 7867
root         265  0.3  0.0  25676  8796 ?        Ssl  15:57   0:00 /app/code/apps/notify_push/bin/x86_64/notify_push --allow-self-signed --nextcloud-url http://172.18.20.152 --port 7867 /app/code/config/config.php
root         286  0.0  0.0   4104  2100 pts/1    S+   15:57   0:00 grep --color=auto notify_push
LISTEN 0      128          0.0.0.0:7867       0.0.0.0:*    users:(("notify_push",pid=265,fd=10))

3. If I set up a cron job to run every 5 minutes.
   */5 * * * * echo "=> Enabling notify_push" && /app/data/notify_push.sh
   I can see in the logs that everything looks fine:

Apr 16 18:30:41 ✓ redis is configured
Apr 16 18:30:41 ✓ push server is receiving redis messages
Apr 16 18:30:41 ✓ push server can load mount info from database
Apr 16 18:30:41 ✓ push server can connect to the Nextcloud server
Apr 16 18:30:41 ✓ push server is a trusted proxy
Apr 16 18:30:41 ✓ push server is running the same version as the app
Apr 16 18:30:41 configuration saved
Apr 16 18:30:41 => notify_push started on port 7867

But in reality the process doesn’t survive after the script finishes. If I check manually with
ps aux | grep notify_push ; ss -tlnp | grep 7867 there’s nothing in the terminal.

Claude suggested this might be a cgroup-related issue in containers. The process started from cron gets attached to its cgroup and when the cron job exits, the process is killed along with it. I’m not sure how accurate that is.

Maybe most people who claimed to have notify_push working probably didn’t verify it properly, since the warning disappears from the Nextcloud admin panel. It looks like it’s working, even when it actually isn’t.

[andreasb]

sorry for replying late:
there's no cron job involved, maybe that's the issue?
Everything needs to be done manually:

• nginx-config whenever Cloudron box is updated, in an SSH session
• running the script to start notify_push whenever NC app is restarted, using a terminal from Cloudron UI to access the docker container

after running the start script, and using your ps-check command, I see a corresponding process. Additionally, running <occ notify_push:metrics> in a Cloudron terminal shows it works, all indicated stats are up-to-date and change over time.

[andreasdueren]

I think I got Nextcloud Client Push (notify_push) working persistently with the current Cloudron Nextcloud package.

Tested with:

• Cloudron Nextcloud package 5.8.5
• Nextcloud 34.0.2
• notify_push 1.3.5

This uses only /app/data; it does not modify read-only /app/code or /etc. I verified it across a full Cloudron app restart.

1. Install Client Push

Install Client Push from the Nextcloud Apps page, or run this in the Nextcloud app terminal:

sudo -u www-data php /app/code/occ app:install notify_push || true
sudo -u www-data php /app/code/occ app:enable notify_push

2. Create the persistent daemon runner

Run in the Nextcloud app terminal:

mkdir -p /app/data/notify_push

cat > /app/data/notify_push/runner.sh <<'EOF'
#!/bin/bash

set -u

binary="/app/data/apps/notify_push/bin/x86_64/notify_push"
config="/app/data/config/config.php"

if [[ ! -x "${binary}" ]]; then
    echo "notify_push binary is missing: ${binary}" >&2
    exit 1
fi

# Apache supplies a pipe on stdin. Preserve and drain it so requests cannot block.
exec 3<&0
cat <&3 >/dev/null &
reader_pid=$!

"${binary}" \
    --bind 127.0.0.1 \
    --port 7867 \
    --nextcloud-url http://127.0.0.1 \
    "${config}" &
push_pid=$!

cleanup() {
    kill -TERM "${push_pid}" "${reader_pid}" 2>/dev/null || true
    wait "${push_pid}" "${reader_pid}" 2>/dev/null || true
}

trap cleanup EXIT
trap 'exit 0' HUP INT TERM

wait "${push_pid}"
EOF

chown www-data:www-data /app/data/notify_push/runner.sh
chmod 0755 /app/data/notify_push/runner.sh

The daemon binds only to loopback. Port 7867 is not exposed publicly.

3. Add the internal callback configuration

The daemon must contact Nextcloud directly instead of looping back through Cloudron’s public proxy:

cat > /app/data/config/notify_push.config.php <<'EOF'
<?php

$CONFIG = [
    'trusted_domains' => [
        getenv('CLOUDRON_APP_DOMAIN'),
        '127.0.0.1',
    ],
    'trusted_proxies' => [
        getenv('CLOUDRON_PROXY_IP'),
        '127.0.0.1',
    ],
];
EOF

chown www-data:www-data /app/data/config/notify_push.config.php
chmod 0640 /app/data/config/notify_push.config.php

Do not add --glob-config to the daemon command. The Rust configuration parser cannot evaluate the getenv() calls in this extra PHP config file.

4. Add the persistent Apache proxy

First save the original configuration:

cp -n /app/data/apache/mpm_prefork.conf \
    /app/data/apache/mpm_prefork.conf.before-notify-push

Then append the integration once:

if ! grep -Fq "BEGIN CLOUDRON NOTIFY_PUSH" /app/data/apache/mpm_prefork.conf; then
cat >> /app/data/apache/mpm_prefork.conf <<'EOF'

# BEGIN CLOUDRON NOTIFY_PUSH
<IfModule !proxy_module>
    LoadModule proxy_module /usr/lib/apache2/modules/mod_proxy.so
</IfModule>
<IfModule !proxy_http_module>
    LoadModule proxy_http_module /usr/lib/apache2/modules/mod_proxy_http.so
</IfModule>
<IfModule !proxy_wstunnel_module>
    LoadModule proxy_wstunnel_module /usr/lib/apache2/modules/mod_proxy_wstunnel.so
</IfModule>

ProxyPass "/push/ws" "ws://127.0.0.1:7867/ws"
ProxyPass "/push/" "http://127.0.0.1:7867/"
ProxyPassReverse "/push/" "http://127.0.0.1:7867/"

GlobalLog "|/usr/local/bin/gosu www-data:www-data /app/data/notify_push/runner.sh" combined
# END CLOUDRON NOTIFY_PUSH
EOF
fi

chown www-data:www-data /app/data/apache/mpm_prefork.conf
chmod 0644 /app/data/apache/mpm_prefork.conf

GlobalLog gives Apache ownership of the daemon lifecycle. Apache starts the runner whenever the app starts and respawns it if it exits.

5. Validate and activate Apache

apache2ctl configtest

Only continue if it reports Syntax OK:

supervisorctl restart apache2

Check that the daemon is running:

ps -ef | grep '[n]otify_push'
curl -i "https://${CLOUDRON_APP_DOMAIN}/push/test/cookie"

The unauthenticated curl request should return HTTP 400 with Missing request header "token". That is expected and confirms the public proxy reaches the daemon.

6. Configure Nextcloud

sudo -u www-data php /app/code/occ \
    notify_push:setup "https://${CLOUDRON_APP_DOMAIN}/push"

All checks should pass:

✓ redis is configured
✓ push server is receiving redis messages
✓ push server can load mount info from database
✓ push server can connect to the Nextcloud server
✓ push server is a trusted proxy
✓ push server is running the same version as the app
  configuration saved

Run the self-test:

sudo -u www-data php /app/code/occ notify_push:self-test

7. Verify persistence

Restart Nextcloud from the Cloudron dashboard. Once it is healthy again, reopen the terminal and run:

sudo -u www-data php /app/code/occ notify_push:self-test
ps -ef | grep '[n]otify_push'

The saved endpoints should be:

https://your-nextcloud-domain.example/push
wss://your-nextcloud-domain.example/push/ws

I also confirmed that both, my Hermes and the official Nextcloud sync client connected and authenticated to the WebSocket after the restart.

Notes

This is an instance-level workaround because the current Cloudron package has no custom supervisor hook. It follows the official notify_push architecture: a background daemon plus an Apache reverse proxy.

Everything persistent is stored in:

/app/data/apache/mpm_prefork.conf
/app/data/notify_push/runner.sh
/app/data/config/notify_push.config.php

The original Apache configuration is retained at:

/app/data/apache/mpm_prefork.conf.before-notify-push

I would rerun notify_push:self-test after future Nextcloud or Cloudron package updates.

[sponch]

@andreasdueren Wow. thanks a lot! can you still recommend it?

[andreasdueren]

Yes, still recommend it with one caveat: the setup survives Cloudron package updates fine, but a Nextcloud major upgrade can silently take it out. Mine broke going from 34 to 35.

After the upgrade notify_push was gone from occ app:list, no process, nothing on port 7867, but oc_appconfig still held base_endpoint with enabled: "no". Config intact in the database, app code gone from disk. For the record, it was not the package overwriting anything: start.sh only copies mpm_prefork.conf when it is absent, and notify_push is not in apps_template, so both persistent hooks from my guide are genuinely durable. A second instance on the same Nextcloud version kept running throughout.

GlobalLog already respawns the runner, so process death was never the issue. A disabled app or missing binary was. I moved that check into runner.sh before the daemon launches: if the binary is not executable or occ app:list shows no version, it runs app:install (falling back to app:update) then app:enable. If the binary is still missing it sleeps 300s and exits, which matters because Apache restarts the pipe target immediately and a fast-failing runner turns into a hot loop. I tested it by deleting the app directory outright, and Apache had it reinstalled and serving again three seconds later, confirmed across a full app restart.

One pitfall: right after a self-heal reinstall, notify_push:self-test may report push server can't connect to the Nextcloud server / Client error: 404 Not Found. That is a stale route cache, not a broken setup. supervisorctl restart apache2 clears it. Still worth running occ notify_push:self-test after every Nextcloud upgrade, since the runner deliberately cannot paper over a release with no compatible notify_push at all.

[andreasdueren]

Here's the updated guide. Changes: hardened self-healing runner in step 2, new step 8 for failure injection, corrected facts about what actually survives updates, and refreshed versions.

1. Install Client Push

Install Client Push from the Nextcloud Apps page, or run this in the Nextcloud app terminal:

sudo -u www-data php /app/code/occ app:install notify_push || true
sudo -u www-data php /app/code/occ app:enable notify_push

2. Create the persistent daemon runner

Run in the Nextcloud app terminal:

mkdir -p /app/data/notify_push

cat > /app/data/notify_push/runner.sh <<'EOF'
#!/bin/bash
# Self-healing notify_push runner, started and respawned by Apache GlobalLog.
# Survives Nextcloud major upgrades that disable or remove the app.

set -u

binary="/app/data/apps/notify_push/bin/x86_64/notify_push"
config="/app/data/config/config.php"
occ="php /app/code/occ"
log="/app/data/notify_push/runner.log"

say() { echo "$(date -u +%FT%TZ) $*" >> "${log}"; }

# Apache supplies a pipe on stdin. Preserve and drain it so requests cannot block.
exec 3<&0
cat <&3 >/dev/null &
reader_pid=$!

cleanup() {
    kill -TERM ${push_pid:-} "${reader_pid}" 2>/dev/null || true
    wait ${push_pid:-} "${reader_pid}" 2>/dev/null || true
}

trap cleanup EXIT
trap 'exit 0' HUP INT TERM

# A major Nextcloud upgrade can mark the app incompatible, disable it, and
# leave the code directory gone while the DB config survives. Repair that here.
app_ver="$(${occ} app:list --output=json 2>/dev/null \
    | grep -o '"notify_push":"[^"]*"' | cut -d'"' -f4 || true)"

if [[ ! -x "${binary}" || -z "${app_ver}" ]]; then
    say "binary or app missing (app_ver='${app_ver}'); attempting repair"
    ${occ} app:install notify_push >>"${log}" 2>&1 \
        || ${occ} app:update notify_push >>"${log}" 2>&1 || true
    ${occ} app:enable notify_push >>"${log}" 2>&1 || true
fi

if [[ ! -x "${binary}" ]]; then
    say "repair failed, binary still missing; backing off 300s"
    sleep 300
    exit 1
fi

say "starting notify_push"
"${binary}" \
    --bind 127.0.0.1 \
    --port 7867 \
    --nextcloud-url http://127.0.0.1 \
    "${config}" &
push_pid=$!

wait "${push_pid}"
rc=$?
say "notify_push exited rc=${rc}; backing off 10s"
sleep 10
exit "${rc}"
EOF

chown www-data:www-data /app/data/notify_push/runner.sh
chmod 0755 /app/data/notify_push/runner.sh
touch /app/data/notify_push/runner.log
chown www-data:www-data /app/data/notify_push/runner.log
bash -n /app/data/notify_push/runner.sh && echo "syntax OK"

The daemon binds only to loopback. Port 7867 is not exposed publicly.

Both sleep calls are deliberate. Apache respawns the pipe target immediately, so a runner that fails fast becomes a hot respawn loop. The runner executes under gosu www-data, so occ needs no sudo here.

3. Add the internal callback configuration

The daemon must contact Nextcloud directly instead of looping back through Cloudron's public proxy:

cat > /app/data/config/notify_push.config.php <<'EOF'
<?php

$CONFIG = [
    'trusted_domains' => [
        getenv('CLOUDRON_APP_DOMAIN'),
        '127.0.0.1',
    ],
    'trusted_proxies' => [
        getenv('CLOUDRON_PROXY_IP'),
        '127.0.0.1',
    ],
];
EOF

chown www-data:www-data /app/data/config/notify_push.config.php
chmod 0640 /app/data/config/notify_push.config.php

Do not add --glob-config to the daemon command. The Rust configuration parser cannot evaluate the getenv() calls in this extra PHP config file.

4. Add the persistent Apache proxy

First save the original configuration:

cp -n /app/data/apache/mpm_prefork.conf \
    /app/data/apache/mpm_prefork.conf.before-notify-push

Then append the integration once:

if ! grep -Fq "BEGIN CLOUDRON NOTIFY_PUSH" /app/data/apache/mpm_prefork.conf; then
cat >> /app/data/apache/mpm_prefork.conf <<'EOF'

# BEGIN CLOUDRON NOTIFY_PUSH
<IfModule !proxy_module>
    LoadModule proxy_module /usr/lib/apache2/modules/mod_proxy.so
</IfModule>
<IfModule !proxy_http_module>
    LoadModule proxy_http_module /usr/lib/apache2/modules/mod_proxy_http.so
</IfModule>
<IfModule !proxy_wstunnel_module>
    LoadModule proxy_wstunnel_module /usr/lib/apache2/modules/mod_proxy_wstunnel.so
</IfModule>

ProxyPass "/push/ws" "ws://127.0.0.1:7867/ws"
ProxyPass "/push/" "http://127.0.0.1:7867/"
ProxyPassReverse "/push/" "http://127.0.0.1:7867/"

GlobalLog "|/usr/local/bin/gosu www-data:www-data /app/data/notify_push/runner.sh" combined
# END CLOUDRON NOTIFY_PUSH
EOF
fi

chown www-data:www-data /app/data/apache/mpm_prefork.conf
chmod 0644 /app/data/apache/mpm_prefork.conf

GlobalLog gives Apache ownership of the daemon lifecycle. Apache starts the runner whenever the app starts and respawns it if it exits.

This file is safe to edit. start.sh only copies the package default when the file is absent:

[[ ! -f /app/data/apache/mpm_prefork.conf ]] && cp /app/pkg/mpm_prefork.conf /app/data/apache/mpm_prefork.conf

If you later tune MPM settings here, keep the NOTIFY_PUSH block.

5. Validate and activate Apache

apache2ctl configtest

Only continue if it reports Syntax OK:

supervisorctl restart apache2

Check that the daemon is running:

ps -ef | grep '[n]otify_push'
curl -i "https://${CLOUDRON_APP_DOMAIN}/push/test/cookie"

The unauthenticated curl request should return HTTP 400 with Missing request header "token". That is expected and confirms the public proxy reaches the daemon.

Raise MaxRequestWorkers

Important, courtesy of @sponch below: because the WebSocket is proxied through Apache, under prefork every connected client holds one worker for the whole lifetime of its push connection. The Nextcloud package default of MaxRequestWorkers 15 means roughly 15 connected clients can exhaust the pool and leave Nextcloud unresponsive, with no workers left for page loads, WebDAV or Talk.

Check and raise it in the same file:

grep -E "MaxRequestWorkers" /app/data/apache/mpm_prefork.conf

Count one slot per device, not per user, since a phone and a laptop each hold their own connection. Add generous headroom on top for ordinary requests. Verify with apache2ctl configtest before restarting.

Do not simply copy a large number. Each prefork worker can grow to PHP's memory_limit, so the safe ceiling depends on your RAM and on any Cloudron memory limit for the app. Measure the marginal cost per worker rather than guessing:

for p in $(ps -eo pid,comm | awk '$2=="apache2"{print $1}'); do
    awk '/^Pss:/{s+=$2} END{print s+0}' /proc/$p/smaps_rollup 2>/dev/null
done | awk '{s+=$1; n++} END{printf "workers=%d avg PSS=%.1f MB\n", n, s/n/1024}'

Note that prefork also needs ServerLimit raised if you go above 256.

6. Configure Nextcloud

sudo -u www-data php /app/code/occ \
    notify_push:setup "https://${CLOUDRON_APP_DOMAIN}/push"

All checks should pass:

✓ redis is configured
✓ push server is receiving redis messages
✓ push server can load mount info from database
✓ push server can connect to the Nextcloud server
✓ push server is a trusted proxy
✓ push server is running the same version as the app
  configuration saved

Run the self-test:

sudo -u www-data php /app/code/occ notify_push:self-test

7. Verify persistence

Restart Nextcloud from the Cloudron dashboard. Once it is healthy again, reopen the terminal and run:

sudo -u www-data php /app/code/occ notify_push:self-test
ps -ef | grep '[n]otify_push'

The saved endpoints should be:

https://your-nextcloud-domain.example/push
wss://your-nextcloud-domain.example/push/ws

Note that a passing self-test alone does not prove persistence. The process must still be present after the restart, with new PIDs.

8. Verify self-healing

Simulate what a bad upgrade does, then confirm recovery:

mv /app/data/apps/notify_push /app/data/apps/.notify_push.bak
supervisorctl restart apache2
sleep 45

ls -la /app/data/apps/notify_push/bin/x86_64/notify_push
ps -ef | grep '[n]otify_push'
cat /app/data/notify_push/runner.log

Expected log output:

binary or app missing (app_ver=''); attempting repair
notify_push 1.4.1 installed
notify_push enabled
starting notify_push

Then clean up the injected directory:

rm -rf /app/data/apps/.notify_push.bak

Right after a self-heal reinstall, notify_push:self-test may report push server can't connect to the Nextcloud server with Client error: 404 Not Found. That is a stale Nextcloud route cache, not a broken setup; loopback status.php still returns 200. Run supervisorctl restart apache2 and re-test.

Notes

This is an instance-level workaround because the current Cloudron package has no custom supervisor hook. It follows the official notify_push architecture: a background daemon plus an Apache reverse proxy.

Everything persistent is stored in:

/app/data/apache/mpm_prefork.conf
/app/data/notify_push/runner.sh
/app/data/config/notify_push.config.php

The original Apache configuration is retained at:

/app/data/apache/mpm_prefork.conf.before-notify-push

For what it is worth, the Cloudron package itself does not remove any of this. notify_push is not in /app/pkg/apps_template, and the update loop only replaces template apps. The failure mode worth guarding against is the Nextcloud app being disabled or removed during a major Nextcloud upgrade, which is what step 2 now handles. I would still rerun notify_push:self-test after future Nextcloud or Cloudron package updates, and check runner.log if push goes quiet.

[sponch]

Hi @andreasdueren this is great and works perfect right now. Thanks a lot for your work!!
Do you see any chance this being integrated in the official package?

[andreasdueren]

@sponch No idea, I'm not maintaining it Gotta ping @girish

[girish]

We won't integrate plugins into the nextcloud package, if this is the question. Unfortunately, nextcloud is hard to maintain as it is with so many plugins.

[sponch]

@andreasdueren

After some digging, I found out that every connected Nextcloud client (desktop or phone) keeps one apache process busy for the push connection.
By default the app seems to allow only 15 of them (MaxRequestWorkers 15) in /app/data/apache/mpm_prefork.conf). Once about 15 clients were online NC was unresponsive.

I changed that line in /app/data/apache/mpm_prefork.conf to
"MaxRequestWorkers 150" and the problem was gone.

Just in case anyone else running over that. (maybe you can add it to your guide above).

[andreasdueren]

@sponch Yes, good point. I had raised the limit a while ago for different reason. Will update guide though with a hint to this.