Skip to content

How to harden a Grafana Tempo container (and is tempo:2.6.1 safe for your traces?)

Tempo is where your distributed traces live: request spans, service dependencies, and the metadata that maps out exactly how your internal systems call each other. A stock docker run grafana/tempo:2.6.1 already runs as a non-root uid, which is a good start, but it still holds that data behind three weak boundaries. Graded on IronClaw's seven-dimension containment scale, the default configuration scores 63 of 100, grade C (partial). Higher is safer. Unlike a broker or a proxy, a trace store that only its co-located writer and query reader talk to can close every remaining dimension, including the network. A few runtime flags take the same image to a full 100 of 100, grade A. Here are the exact gaps and fixes from the scan data.

Every number here comes from a read-only docker inspect of grafana/tempo:2.6.1, the same data behind its isolation scorecard. No workload is executed. How scoring works →

Where the default configuration leaks

ironctl scan grades seven independent containment boundaries. On a default docker run grafana/tempo:2.6.1, two fail and one warns:

Dimension Verdict Score What the scan found
Non-root user (uid != 0) ✅ PASS 15/15 runs as 10001:10001 (uid != 0)
Dropped capabilities ❌ FAIL 4/20 default capability set retained (CAP_NET_RAW, CAP_MKNOD, and more)
Seccomp profile ✅ PASS 15/15 seccomp profile active
Network isolation / egress ⚠️ WARN 4/15 network=bridge: outbound egress is possible
Read-only root filesystem ❌ FAIL 0/10 root filesystem is writable
No docker.sock exposure ✅ PASS 15/15 no control socket mounted
No shared host namespaces ✅ PASS 10/10 no host PID/IPC/network sharing

Tempo already ships a non-root uid, so the two that should worry you most are egress and the default capability set. A trace store that can reach arbitrary destinations is one that can quietly ship your entire trace history, a live map of your internal topology, out the moment an ingest-parsing or plugin CVE lands code execution. The retained capabilities and writable rootfs widen and entrench that foothold.

Harden it: the exact --fix remediation

ironctl scan my-tempo --fix prints one remediation per failed dimension, then one hardened run. For grafana/tempo:2.6.1:

  • --cap-drop=ALL (Dropped capabilities, +16): drop every Linux capability; Tempo needs none of the default set to accept spans and serve its API on a high port.
  • --read-only --tmpfs /tmp (Read-only rootfs, +10): make the root filesystem read-only and mount /var/tempo as an explicit writable volume. Removes the persistence surface.
  • --user 65532:65532 (Non-root user, already PASS): keep the non-root uid and point the trace storage directory at a volume this uid owns.
  • --network=none (Network isolation, +11 to the full 15): this is the dimension a co-located store can actually max out. If the only writer and reader are on the same host or pod and reach Tempo over the loopback of a shared network namespace, cut the NIC entirely. Nothing external can connect, and the store cannot phone home.

When network=none is not honest

If remote agents push spans to Tempo over the network (a fleet of OpenTelemetry collectors, for example), you cannot use --network=none; the store has to accept those OTLP connections. In that case put it on a user-defined network scoped to just the collectors and the Grafana that queries it, with no default route out. That holds the network dimension at a WARN (4 of 15) and the honest ceiling becomes 89 of 100, grade B, the same as a broker. Use --network=none only for the single-writer, co-located case.

Before and after

# Before: 63/100, grade C
docker run -d --name tempo grafana/tempo:2.6.1

# After: 100/100, grade A (co-located store, no network needed)
docker run -d --name tempo-hardened \
  --user 65532:65532 \
  --cap-drop=ALL \
  --security-opt=no-new-privileges \
  --read-only --tmpfs /tmp \
  -v tempo-data:/var/tempo \
  --network=none \
  grafana/tempo:2.6.1

Rescan: ironctl scan tempo-hardened reports 100/100 grade A. A 37-point swing with no custom image build, just the right flags. Every dimension is closed because a co-located trace store does not need to talk to anything but the app on the other side of its loopback. That is the top grade, reserved for datastores whose clients live next to them.

Verify it on your own Tempo

# install (Homebrew)
brew install ironsecco/ironclaw/ironclaw

# grade your running container, then print the fixes
ironctl scan my-tempo
ironctl scan my-tempo --fix

ironctl scan also reads a docker-compose.yml service or a Kubernetes manifest, so you can grade the Tempo in your stack, not just a bare docker run.

Keep going