Q: You need to run a Docker container with access to GPU. How?
Install NVIDIA Container Toolkit on the host. Run with --gpus all or --gpus '"device=0"'. In Docker Compose: deploy: resources: reservations: devices:. The container can then use CUDA APIs.
Want to master this scenario in a live sandbox? KodeKloud's Docker Certified Associate (DCA) Hands-On Lab Course covers this exact problem with hands-on terminal drills.
🛠️ Production Runbook & Step-by-Step Resolution
Direct Answer: "In an interview, I explain how we diagnosed container runtime failures without guessing. When addressing this question, I walk the interviewer through our production incident runbook: isolating the blast radius, checking diagnostic logs and metrics, and applying a safe fix."
Production Solution & Architecture
Install NVIDIA Container Toolkit on the host. Run with --gpus all or --gpus '"device=0"'. In Docker Compose: deploy: resources: reservations: devices:. The container can then use CUDA APIs. Verify: docker run --gpus all nvidia/cuda:11.0-base nvidia-smi.
- Immediate Triage: Install NVIDIA Container Toolkit on the host. Run with --gpus all or --gpus '"device=0"'. In Do
- Run targeted verification commands before modifying configuration.
- Automate permanent guardrails (CI check, alerts, IaC policy) to prevent recurrence.
Senior hiring managers and bar raisers evaluate how you prioritize during outages. Avoid these common anti-patterns:
-
Structuring Dockerfile instructions inefficiently by copying app source before package manifests when configuring need Docker container
Copying code before package manifests (
package.json,go.mod,requirements.txt) invalidates the build cache on every commit, forcing slow full reinstalls. -
Neglecting the PID 1 process reap problem and omitting proper SIGTERM signal forwarding when configuring need Docker container
If an application runs as PID 1 without an init system (like
tini), it fails to reap zombie child processes and ignoresSIGTERMsignals, causing unclean 10s timeouts. - Assembling single-stage images containing compilers, build SDKs, and header files in production when configuring need Docker container Single-stage builds bloat image size by gigabytes and expose extensive package vulnerabilities; multi-stage builds isolate artifacts into minimal distroless images.
- Confusing container port forwarding on bridge networks with unisolated host networking mode when configuring need Docker container Bridge networking utilizes iptables NAT translations which add latency overhead, whereas host networking bypasses the network namespace entirely.