Monitoring & Logs

Every deployment exposes logs, metrics, and health on the detail page.

Detail page layout

At deployer.zynd.ai/d/<id>:

Panel	Shows
Header	Name, slug, entity type, status badge, Stop button.
Info grid	Public URL, registry URL, entity ID, public key, host port, container ID, created time, started time, last exit code.
Live logs	SSE stream of stdout/stderr, tagged by stream.
Metrics	CPU %, memory MB, over last 3 days.

Log stream

Logs are served via Server-Sent Events at /api/deployments/<id>/logs/stream.

Every stdout/stderr line your agent prints is captured by docker logs -f, demuxed, and inserted into DeploymentLog with an incrementing lineNo. The browser subscribes and renders in real time.

Stream classes:

Stream	Color	Source
stdout	default	print(...) from your agent
stderr	red	exceptions, warnings, framework errors
system	blue	Deployer events — crash, health, unhealthy, stop

Live logs retain 7 days by default. System logs retain 30 days. Older lines are batch-deleted hourly.

Metrics

Sampled every 30 s while running.

Schema:

DeploymentMetric
├── deploymentId
├── sampledAt       timestamp
├── memUsedMb       container memory usage
├── memLimitMb      container memory limit (1536 default)
└── cpuPct          % of 1 CPU (100 = one full core)

The detail page plots both as time series.

Health

The worker polls http://127.0.0.1:<port>/health inside the container every 60 seconds.

State transitions:

• 0-2 failures → running (counter resets on any success).
• 3 consecutive failures → unhealthy. Status badge turns yellow. System log entry [UNHEALTHY] health probe failed 3 times.
• Health recovers → running again, status badge green.

The container is not restarted automatically — Docker's unless-stopped policy handles hard crashes, not slow /health.

Make sure your agent responds to GET /health quickly (< 2 s). The SDK handles this automatically; only custom implementations need to worry.

Crash detection

A persistent docker events subscription catches die and oom events. When a zynd-<id> container dies:

1. Exit code read via docker inspect.
2. Last 100 log lines captured into DeploymentLog.
3. Status → crashed.
4. System log entry: [CRASH] exit=<code> oom=<bool>.

If you set DEPLOYER_KEEP_CRASHED_CONTAINERS=true, the dead container is preserved for docker exec debugging. Otherwise it's removed.

Stopping

Click Stop. Worker:

1. Sets status=stopped.
2. drainStops loop finds the row on its next tick.
3. Sends SIGTERM to the container. Waits 10 s.
4. Sends SIGKILL if still alive.
5. docker rm the container.
6. Removes the Caddy route.
7. Releases the port.
8. Clears containerId and port on the row.

Failures at any cleanup step are logged as system entries but don't wedge the row.

Programmatic access

List deployments

curl https://deployer.zynd.ai/api/deployments

Single deployment

curl https://deployer.zynd.ai/api/deployments/<id>

Fetch logs

curl "https://deployer.zynd.ai/api/deployments/<id>/logs?since=<lineNo>&limit=500"

Stream logs (SSE)

curl -N https://deployer.zynd.ai/api/deployments/<id>/logs/stream

Stop

curl -X DELETE https://deployer.zynd.ai/api/deployments/<id>

Best practices

• Log structured JSON if you want to grep. python-json-logger is preinstalled in the agent base image.
• Don't use /health for heavy work. It's polled every minute — keep it a cheap return 200.
• Use external DBs for any state you care about. The container filesystem is ephemeral.
• Handle SIGTERM — your Flask server should exit cleanly on stop. The SDK does this by default.

Next

• Troubleshooting — what to do when things go wrong.
• Self-Host Deployer — tune retention, port ranges, limits.