Browse questions, read quick answers, or expand full article breakdowns on demand. Filter by level, domain, or completion status using the sidebar dashboard.
A junior engineer says "a container is just a lightweight VM." Explain precisely why this analogy is wrong, and describe what a container actually *is* at the process level on the host.
You run `docker build` twice on the same Dockerfile with no changes and notice the second build finishes in under a second. Explain what mechanism makes this possible and what is actually being reused.
You run `docker build` twice on the same Dockerfile with no changes and notice the second build finishes in under a second. Explain what mechanism makes this possible and what is actually being reused.
A junior engineer says "a container is just a lightweight VM." Explain precisely why this analogy is wrong, and describe what a container actually *is* at the process level on the host.
A junior engineer says "a container is just a lightweight VM." Explain precisely why this analogy is wrong, and describe what a container actually *is* at the process level on the host.
A junior engineer says "a container is just a lightweight VM." Explain precisely why this analogy is wrong, and describe what a container actually *is* at the process level on the host.
You run `docker build` twice on the same Dockerfile with no changes and notice the second build finishes in under a second. Explain what mechanism makes this possible and what is actually being reused.
Your team's Dockerfile uses both `ENTRYPOINT` and `CMD`. A teammate wants to remove `CMD` because "it seems redundant." Explain why removing it would change the container's runtime behavior and give a scenario where this distinction matters operationally.
A developer mounts their local project folder into a container with `-v` during development, but in the production Dockerfile they use `COPY`. Explain why these two approaches exist and why you would never rely on a bind mount in a production deployment.
A container running a database is removed with `docker rm`, and the team is shocked their data disappeared. Diagnose what likely went wrong in their setup and explain the container filesystem lifecycle that caused this.
A service inside a container listens on port 8080, and it's exposed with `-p 80:8080`. Walk through what actually happens to a packet arriving at the host's port 80 to reach the process inside the container.
Two engineers build the same application — one from `ubuntu:latest`, another from `alpine`. Beyond image size, what practical engineering trade-offs should inform this base image decision?
A teammate adds application logging to a file inside the container instead of stdout/stderr, then complains `docker logs` shows nothing. Explain the underlying logging model Docker expects and why their approach breaks it.
Explain, namespace by namespace (PID, NET, MNT, UTS, IPC, USER), what specific kernel isolation each one provides to a container, and describe a real scenario where sharing one particular namespace between containers (e.g., `--pid=container:x`) is a deliberate and useful debugging technique.
A container set with `--memory=512m` gets OOM-killed even though `docker stats` shows it using only 300MB of RSS at the time. Explain what other memory accounting cgroups tracks (page cache, kernel memory) that could explain this, and how you'd investigate.
A build suddenly takes much longer after a teammate added a large `node_modules` folder to the project directory, even though it's never referenced in the Dockerfile. Explain what's happening and how to fix it.
A build suddenly takes much longer after a teammate added a large `node_modules` folder to the project directory, even though it's never referenced in the Dockerfile. Explain what's happening and how to fix it.
Diagram (in words) the full process chain from a `docker run` invocation down to the actual container process, explicitly naming the role of `dockerd`, `containerd`, `containerd-shim`, and `runc`, and explain why the shim's existence allows the Docker daemon to be restarted without killing running containers.
A container needs to bind to a privileged port and adjust system time as part of its function, but your security policy forbids `--privileged`. Explain the capability-based alternative, name the specific capabilities required, and articulate why granular capability grants are architecturally superior to the privileged flag.
Your application needs different database URLs in staging versus production, but you want to use the exact same image in both. Explain the mechanism that makes this possible without rebuilding the image, and why baking config into the image is considered an anti-pattern.
Your application needs different database URLs in staging versus production, but you want to use the exact same image in both. Explain the mechanism that makes this possible without rebuilding the image, and why baking config into the image is considered an anti-pattern.
You create a custom bridge network and notice new `iptables` chains and rules appear on the host without you touching `iptables` directly. Explain what Docker is doing to the NAT and FILTER tables to make inter-container and container-to-external routing work, and describe a scenario where a conflicting host firewall rule could silently break container connectivity.
In a multi-host overlay network (e.g., Swarm), two containers on different physical hosts communicate directly by container IP. Explain the encapsulation mechanism (VXLAN) that makes this possible at the packet level, and identify the MTU-related failure mode this commonly introduces.
A container running a simple shell script exits immediately after starting, even though the script "runs forever" when tested locally. What container lifecycle rule explains this behavior?
A container running a simple shell script exits immediately after starting, even though the script "runs forever" when tested locally. What container lifecycle rule explains this behavior?
A service configured with `--cpus=2` shows periodic latency spikes even though average CPU usage sits well under 200%. Explain how CFS (Completely Fair Scheduler) quota-based throttling within a fixed period can cause this, and why average utilization metrics can be misleading here.
A container running as an unprivileged process is still able to see and signal processes on the host when run with `--pid=host`. Explain precisely what isolation guarantee is being intentionally broken, and describe one legitimate production use case where this trade-off is justified.
Your team deploys using the `latest` tag and occasionally ends up with different code running on different hosts despite "not changing anything." Explain why tag-based deployment without immutable references is risky in production.
Your team deploys using the `latest` tag and occasionally ends up with different code running on different hosts despite "not changing anything." Explain why tag-based deployment without immutable references is risky in production.
A hardened production seccomp profile blocks an application from calling `clone()` with certain flags, causing an obscure runtime crash unrelated to any obvious security feature. Explain how seccomp filters operate at the syscall level, why default Docker seccomp profiles allow most syscalls but block dangerous ones, and how you'd methodically identify which specific syscall is being blocked.
Contrast what AppArmor's mandatory access control profile enforces versus what seccomp enforces for the same container, and describe a concrete attack scenario that seccomp alone would not stop but a correctly scoped AppArmor profile would.
Your Go application's final image is 900MB because it includes the full compiler toolchain. Design a multi-stage build that resolves this, and explain precisely what gets carried between stages and what gets discarded.
Your Go application's final image is 900MB because it includes the full compiler toolchain. Design a multi-stage build that resolves this, and explain precisely what gets carried between stages and what gets discarded.
The OOM killer selects a victim using a badness score derived from each process's memory footprint adjusted by oom_score_adj, and because PID 1 inside a container's PID namespace commonly receives special init-process treatment and a more negative adjustment, the kernel can end up killing a different, seemingly smaller process instead of the one actually consuming the most memory.
When a cgroup hits its memory limit (or the system as a whole is under global memory pressure), the kernel's OOM killer doesn't simply kill "the biggest process." It computes a badness score for every candidate process, primarily driven by RSS plus swap usage, and then applies a per-process adjustment: oom_score_adj, a value from -1000 to +1000 readable/writable at /proc/<pid>/oom_score_adj. A value of -1000 makes a process effectively immune to the OOM killer; +1000 makes it maximally likely to be chosen. The final decision picks the highest-scoring (most "killable") process among the candidates the kernel is allowed to consider for this event.
Cgroup-aware OOM killing (the default and more predictable behavior under cgroup v2, and available with memory.oom_group under specific configurations) changes the scope of that candidate pool: instead of comparing badness scores across the entire host, the kernel can be constrained to compare only among processes within the offending cgroup — the container that actually breached its own --memory limit. This is generally the more correct behavior for containerized workloads, since without it, a memory-hungry process in one container's cgroup could theoretically influence the kill decision for a completely unrelated container, or worse, the host's own processes could become collateral candidates for a limit that was never about the host at all.
Inside a container's PID namespace, the container's entrypoint process runs as PID 1 — and PID 1, in any PID namespace, inherits some of the same special kernel treatment that real host PID 1 (init/systemd) gets: default signal handling behaves differently (PID 1 ignores signals it hasn't explicitly registered a handler for, including SIGKILL's default disposition — though SIGKILL still forcibly terminates it, just without normal signal-delivery semantics applying), and many init systems and container runtimes explicitly set a lower (more negative) oom_score_adj for PID 1, on the reasoning that killing the container's init process is more disruptive than killing a single worker/child process — losing PID 1 tears down the entire container, whereas losing a child process might be recoverable.
This creates the exact scenario in the question: if a container runs a multi-process workload — say, an init/supervisor process (tini, or a custom entrypoint script) as PID 1, spawning several worker child processes — and one worker balloons in memory usage, the OOM killer's badness comparison isn't purely "who has the most RSS." PID 1's adjusted (more negative) score can keep it off the chopping block even if it happens to have non-trivial memory usage of its own, while among the remaining candidates, the kernel picks whichever child scores highest — which may or may not be the single largest consumer, depending on how oom_score_adj values were set across the process tree and whether any other process has an even less favorable adjustment.
/proc/<pid>/oom_score (the computed, current badness score) and /proc/<pid>/oom_score_adj (the adjustment) for every process in the container's tree before the next OOM event, not after — post-mortem, the killed process is already gone and its score history isn't retained.oom_score_adj explicitly higher (more positive) rather than relying on default kernel heuristics to make that call correctly.dmesg/kernel logs after an OOM-kill report the selected PID, its oom_score, and often a memory breakdown of the cgroup at kill time — cross-reference the killed PID against docker top <container> history (or application-level PID logging) to confirm which logical process it actually was, since PIDs are frequently reused quickly after a kill.tini-wrapped multi-worker setups) with awareness that PID 1 surviving an OOM event while a child dies can leave the container in a degraded-but-alive state rather than restarting cleanly — decide deliberately whether that's the desired failure mode or whether memory.oom_group (killing the whole cgroup together) is more appropriate for the workload.The OOM killer doesn't simply kill the largest process — it kills the highest-scoring one after oom_score_adj adjustments, and PID 1's typically favorable adjustment means the process most people assume should die (the biggest one) isn't always the one that actually does.
Your Go application's final image is 900MB because it includes the full compiler toolchain. Design a multi-stage build that resolves this, and explain precisely what gets carried between stages and what gets discarded.
You need to perform a zero-downtime Docker Engine upgrade on a host running stateful production containers. Explain how `live-restore` allows containers to keep running through a daemon restart, what it does *not* protect against, and what happens if the containerd/shim architecture itself needs an upgrade at the same time.
A production incident requires live debugging of a distroless container that has no shell, no package manager, and no coreutils. Walk through at least two distinct real-world techniques (e.g., ephemeral debug containers sharing namespaces, sidecar attach) to inspect its running process, filesystem, and network state without modifying the original image.
A Dockerfile does `COPY . .` before `RUN npm install`. Every single code change forces a full dependency reinstall during CI, adding 4 minutes per build. Explain the caching mechanic causing this and restructure the instructions to fix it.
A Dockerfile does `COPY . .` before `RUN npm install`. Every single code change forces a full dependency reinstall during CI, adding 4 minutes per build. Explain the caching mechanic causing this and restructure the instructions to fix it.
A high-throughput edge service running in containers starts silently dropping new connections under load, and `dmesg` shows conntrack table full errors. Explain how Docker's iptables-based NAT interacts with the kernel's connection tracking table, why containerized workloads are especially prone to exhausting it, and the tuning levers available.
Standard CPU/memory utilization metrics look healthy, but application-level latency at the edge is degrading under load. Explain how Pressure Stall Information (PSI) in cgroups v2 exposes resource contention that traditional utilization metrics miss, and how you'd use it to diagnose whether the bottleneck is CPU, memory, or I/O contention.
Even after reordering Dockerfile instructions for cache-friendliness, your team's CI runners (ephemeral, ban `--cache-from` layer reuse) still reinstall dependencies every run. What BuildKit feature addresses this specific problem, and how does it differ from ordinary layer caching?
Even after reordering Dockerfile instructions for cache-friendliness, your team's CI runners (ephemeral, ban `--cache-from` layer reuse) still reinstall dependencies every run. What BuildKit feature addresses this specific problem, and how does it differ from ordinary layer caching?
Explain, at a mechanistic level, how a historical runc container-breakout vulnerability (such as the `/proc/self/exe` file-descriptor overwrite class of CVEs) allowed a malicious container to overwrite the host `runc` binary, and what defense-in-depth layers (user namespaces, read-only host binaries, monitoring) would have limited the blast radius even if the specific CVE were unpatched.
During a rolling deployment, containers are hard-killed after a 10-second timeout, dropping in-flight requests, even though the application has SIGTERM handling implemented correctly. Diagnose the most likely architectural cause involving PID 1, process supervision, and signal propagation inside the container, and describe the fix.
Two containers on the default `bridge` network can't resolve each other by container name, but two containers on a user-defined bridge network can. Explain the underlying difference in how Docker handles DNS resolution between these two network types.
Two containers on the default `bridge` network can't resolve each other by container name, but two containers on a user-defined bridge network can. Explain the underlying difference in how Docker handles DNS resolution between these two network types.
A latency-sensitive service performs noticeably better under `--network host` than the default bridge network. Explain the actual network path difference that causes this performance gap, and identify the operational trade-off the team is accepting by using host networking.
A `docker-compose.yml` uses `depends_on` to ensure the database container starts before the API container, but the API still crashes on startup trying to connect. Explain why `depends_on` alone doesn't solve this problem and what actually needs to happen.
A container runs as root by default, and a file it writes to a bind-mounted volume ends up owned by `root` on the host, breaking the host user's ability to edit it. Explain the UID mapping reality behind this, and describe two distinct strategies to prevent it.
Your security team mandates that no production container may run as UID 0. Walk through what changes are required in the Dockerfile and what operational issues (port binding, file permissions, package installs) commonly break as a result — and how to resolve each.
A container intermittently fails to resolve an external hostname under load, though `curl` works fine most of the time. Explain how Docker's embedded DNS resolver works and a plausible root cause for intermittent resolution failures.
A Compose stack reports all containers as "running," yet the application is non-functional because the API started before the database finished initializing. Design a `HEALTHCHECK`-based solution and explain how it changes container state reporting versus a plain process check.
Your CI logs show the build context being sent to the daemon is 1.2GB despite a small application. Explain the mechanism by which this bloat occurs and how `.dockerignore` interacts with the build process to prevent it.
A team migrating from a single Docker host to a small Swarm/multi-host setup discovers that named volumes don't "follow" a rescheduled container to another node. Explain why this happens and what class of solution is required.