blocksnoop

Detect blocking calls in Python asyncio event loops using eBPF + Austin.

blocksnoop attaches to a running Python process (or launches one) and reports every time the event loop is blocked longer than a configurable threshold — with the Python stack trace that caused it.

How it works

eBPF (kernel)          Austin (userspace)
  │ monitors               │ samples Python
  │ epoll gaps              │ stacks continuously
  └──────────┐   ┌─────────┘
             ▼   ▼
          Correlator
             │
             ▼
       Reporter → sinks (console, JSON, file)

1. An eBPF probe hooks the epoll syscalls (epoll_wait, epoll_pwait, epoll_pwait2 — all of the family available on the kernel) and measures the time between returns (callback start) and the next entry (callback end). If the gap exceeds the threshold, it emits an event. Tracing every variant matters because which one a loop enters depends on its libc and implementation — glibc routes epoll_wait() through the epoll_pwait syscall, and uvloop/libuv call epoll_pwait directly.
2. A stack sampler (Austin) runs as a long-lived subprocess, continuously streaming Python stack traces into a ring buffer. Austin's pipe mode avoids per-sample subprocess overhead, enabling sub-10ms threshold detection.
3. The correlator enriches each blocking event with the closest matching Python stack.
4. The reporter fans out events to one or more output sinks.

Requirements

• Linux with eBPF support (the default Core backend requires kernel 5.7+ for PID-namespace filtering and a readable epoll syscall tracepoint pair)
• Root privileges (for eBPF and Austin)
• Native linux/amd64 or linux/arm64 Blocksnoop wheel, or the official Docker image (contains the default Core sidecar)
• BCC (BPF Compiler Collection) only for the explicit legacy --backend bcc compatibility path
• Austin
• austin-python (installed automatically as a dependency)
• Python 3.12+

Installation

pip install blocksnoop

Or for development:

git clone git@github.com:PaulM5406/blocksnoop.git
cd blocksnoop
uv sync --all-extras --dev

eBPF backends

Core is the default backend. Python launches the small blocksnoop-ebpf libbpf sidecar and consumes its versioned NDJSON event stream. Protocol v2 resolves the target PID namespace and its local PID/TID before attaching, so a hostPID collector can monitor a process in a private container namespace without broadening its filter.

The current program attaches only to syscall tracepoints and does not read kernel structures. Its object contains BTF metadata, but it has no kernel-layout relocations: a readable kernel BTF file is therefore useful diagnostic information, not a Core prerequisite. A real Core attach is the authoritative compatibility check.

On linux/amd64 and linux/arm64, PyPI selects a native wheel containing the sidecar and precompiled object, so pip install blocksnoop is sufficient for Core itself; Austin and the required privileges are still host prerequisites. Other platforms receive the portable compatibility wheel without native assets. A source checkout or unpacked source distribution can build the sidecar on Linux with make -C native and place native/blocksnoop-ebpf on PATH.

The official Docker image is intentionally Core-only: it installs the native wheel and Austin, but not BCC or kernel headers. There is no automatic fallback. Use --backend bcc only on a legacy host where you provisioned BCC yourself.

Before attaching, inspect the exact environment and optional target without loading BPF or spawning the sidecar:

blocksnoop doctor
blocksnoop doctor 1234
blocksnoop doctor 1234 --json
blocksnoop doctor --stats 1234  # eBPF-only pre-flight; Austin is not required

doctor exits non-zero when a required check fails and includes remediation in both human-readable and machine-readable output. Capture mode also checks Austin; doctor --stats intentionally does not. Target-specific output shows the host and namespace-local PID/TID plus whether the collector shares the PID namespace.

Legacy BCC compatibility

BCC is retained for hosts that cannot run the supported Core path, such as older kernels. It is never selected automatically and is not installed by pip install blocksnoop or the official image. After provisioning BCC and matching kernel headers yourself, select it explicitly:

sudo blocksnoop --backend bcc <PID>
sudo blocksnoop doctor --backend bcc <PID>

BCC is frozen as a compatibility backend through the pre-1.0 releases and any future 1.x series; new features target Core. Its removal would not happen before a future major release.

Usage

Attach to a running process

sudo blocksnoop <PID>
sudo blocksnoop -t 50 <PID>          # 50ms threshold (default: 100ms)
sudo blocksnoop --tid 1234 <PID>     # monitor specific thread
sudo blocksnoop -v <PID>             # enable debug logging

Launch and monitor a process

sudo blocksnoop -- python app.py
sudo blocksnoop -t 50 -- python app.py

Output modes

# Human-readable to stderr (default)
sudo blocksnoop -- python app.py

# Versioned NDJSON lifecycle to stdout (for piping to jq, etc.)
sudo blocksnoop --json -- python app.py

# Suppress individual event records; keep the start and final summary
sudo blocksnoop --json --summary-only -- python app.py

# Structured JSON to file (for Datadog/Fluentd/CloudWatch)
sudo blocksnoop --log-file /var/log/blocksnoop/events.json --service my-api --env production -- python app.py

# Combine: console to terminal + JSON to file
sudo blocksnoop --log-file /var/log/blocksnoop/events.json --service my-api -- python app.py

Stats mode

Use --stats to run only the eBPF detector (no Austin profiler, no stack traces) and see the distribution of all epoll gaps. This helps you pick the right --threshold before running a full profiling session.

# Capture all epoll gaps and display live statistics
sudo blocksnoop --stats <PID>

# Backward-compatible JSON stats snapshots (one record per second)
sudo blocksnoop --stats --json <PID>

# Only gaps above 10ms
sudo blocksnoop --stats -t 10 <PID>

Unlike normal capture mode, --stats --json is a stream of independent, backward-compatible statistics snapshots rather than a blocksnoop.events/v1 session lifecycle. Each line contains pid, elapsed_s, count, rate, and percentile fields once events exist.

Sample output (redrawn in place every second):

blocksnoop stats — PID 1234 — 12.3s — 4821 events (391/s)

  min          0.0ms
  avg          2.1ms
  p50          0.8ms
  p90          4.2ms
  p95          8.7ms
  p99         45.3ms
  max        302.1ms

Example output

Human-readable:

[   1.23s] #1   BLOCKED     302.1ms  tid=1234
  Python stack (most recent call last):
    app.py:7 in blocking_io
      time.sleep(0.5)
    app.py:13 in main
      blocking_io()

[   2.05s] #2   BLOCKED     298.5ms  tid=1234
  Python stack (most recent call last):
    app.py:7 in blocking_io
      time.sleep(0.5)
    app.py:13 in main
      blocking_io()

--- blocksnoop session ---
Duration: 8.0s
Blocking events detected: 2
Lost detector events: 0

Normal capture JSON (--json without --stats) is a versioned NDJSON lifecycle. Every line is independently parseable and shares schema, schema_version, type, and session_id. The three record types are session_start, blocking_event, and session_summary; filter events with jq 'select(.type == "blocking_event")'. The final summary makes a clean zero-event session distinguishable from a truncated stream and includes loss counts plus the top blocking call sites.

{"schema":"blocksnoop.events/v1","schema_version":1,"type":"session_start","session_id":"...","backend":"core","threshold_ms":100.0,"target_pid":5678,"target_tid":5678}
{"schema":"blocksnoop.events/v1","schema_version":1,"type":"blocking_event","session_id":"...","event_number":1,"timestamp_s":1.23,"duration_ms":302.1,"pid":5678,"tid":5678,"python_stacks":[[{"function":"blocking_io","file":"app.py","line":7,"source":"time.sleep(0.5)"}]],"level":"warning"}
{"schema":"blocksnoop.events/v1","schema_version":1,"type":"session_summary","session_id":"...","termination_reason":"clean","status":"completed","duration_s":8.0,"event_count":1,"lost_event_count":0,"total_blocked_ms":302.1,"max_blocked_ms":302.1,"top_signatures":[{"location":"app.py:7 in blocking_io","count":1,"total_blocked_ms":302.1,"max_blocked_ms":302.1}]}

CLI reference

blocksnoop [OPTIONS] [PID] [-- COMMAND ...]
blocksnoop doctor [OPTIONS] [PID]

Options:
  -t, --threshold FLOAT        Blocking threshold in ms (default: 100, or 0 with --stats)
  --stats                      eBPF-only mode: show epoll gap distribution (no Austin/stacks)
  --tid INT                    Thread ID to monitor (default: main thread)
  --backend {core,bcc}         eBPF backend to use (default: core; bcc is legacy)
  --json                       JSON lines output to stdout
  --summary-only               Suppress individual events; keep the session lifecycle
  --log-file PATH              Write structured JSON to file for log aggregators
  --service NAME               Service name for structured logs (default: blocksnoop)
  --env ENV                    Environment tag for structured logs
  --no-color                   Disable ANSI colors in terminal output
  -v, --verbose                Enable debug logging to stderr
  --error-threshold MS         Duration in ms above which events are errors (default: 500)
  --correlation-padding MS     Correlation time window padding in ms (default: 200)
  --fail-on {none,event,error} Exit 3 when a completed session violates the policy
  --fail-on-loss               Exit 3 when detector events were lost

Docker

blocksnoop requires kernel access, so Docker containers need --privileged and --pid=host:

# Pull from Docker Hub
docker pull oloapm/blocksnoop

# Check that the host kernel can run the Core backend
docker run --rm --privileged --pid=host \
  -v /sys/kernel/debug:/sys/kernel/debug \
  oloapm/blocksnoop blocksnoop doctor --json

# Attach to a process on the host with the precompiled libbpf backend
docker run --rm --privileged --pid=host \
  -v /sys/kernel/debug:/sys/kernel/debug \
  oloapm/blocksnoop blocksnoop -t 100 <PID>

# Launch and monitor a process
docker run --rm --privileged --pid=host \
  -v /sys/kernel/debug:/sys/kernel/debug \
  oloapm/blocksnoop blocksnoop -t 100 -- python app.py

For local development:

# docker-compose.yml
services:
  blocksnoop:
    build: .
    privileged: true
    pid: host

docker compose run --rm blocksnoop blocksnoop -t 100 -- python app.py

The image is published for linux/amd64 and linux/arm64. It has no BCC fallback: a Core prerequisite failure is reported directly. For a local image smoke before attaching to a workload:

docker build -t blocksnoop:local .
docker run --rm blocksnoop:local python -c '
from blocksnoop.core_backend import find_sidecar
assert find_sidecar()
'
docker run --rm blocksnoop:local sh -ec '
  austin --version >/dev/null
  python -c "import importlib.util; assert importlib.util.find_spec(\"bcc\") is None"
'

Kubernetes

blocksnoop uses eBPF which operates at the kernel level, so you run it on the node, not inside the application container. The target process just needs to be visible from the host PID namespace.

Note: On kernel 5.7+, both backends translate a host-visible target into its container-local PID/TID. A node-level collector still needs hostPID: true to see the target in /proc; an ephemeral container sharing the target process namespace can use its local PID directly.

Ephemeral debug container (recommended)

Attach directly to a running pod with an ephemeral container. --profile=sysadmin (K8s 1.28+) grants the privileged access required for eBPF:

# Find the pod
kubectl get pods -l app=my-api

# Attach an ephemeral debug container with eBPF privileges
kubectl debug -it my-api-pod-7b8c9d \
  --image=oloapm/blocksnoop:latest \
  --target=my-api \
  --profile=sysadmin \
  -- sh -c "mount -t debugfs debugfs /sys/kernel/debug 2>/dev/null; exec sh"

--target shares the process namespace with the app container, so you can see its PIDs. --profile=sysadmin enables privileged mode for eBPF and debugfs access.

Inside the debug container, find the Python process and attach:

# Find the Python PID
ps aux | grep python

# Attach blocksnoop
blocksnoop -t 50 <PID>

# Or with structured logging
blocksnoop --json -t 50 <PID>

The Core image never compiles BPF at runtime and needs neither kernel headers nor BCC. If blocksnoop doctor fails, fix the reported host-kernel, tracefs, privilege, or target-visibility prerequisite instead of adding headers to the image.

Cross-container attach (different mount namespaces)

When blocksnoop runs from a sidecar, ephemeral debug container, or a privileged Job pinned to the target's node, it lives in a different mount namespace than the target — so the Python binary paths Austin reads from /proc/<pid>/maps (e.g. /usr/local/bin/python3.11) don't exist in blocksnoop's filesystem. Without help, Austin then logs 🔢 Cannot determine the version of the Python interpreter. and produces zero samples.

blocksnoop handles this automatically: when the target's mount namespace differs from blocksnoop's, it generates a thin wrapper that opens austin (and the musl linker) as file descriptors in its own namespace, then nsenters into the target and execs via /proc/self/fd/N. fds survive execve, so Austin loads from blocksnoop's rootfs while sampling against the target's filesystem view.

This means:

• The target image is never modified — no binaries copied, no files written under /proc/<TARGET>/root/.
• Works against hardened targets with readOnlyRootFilesystem: true.
• Works regardless of the target's libc (alpine/musl, debian/glibc, distroless).
• Requires CAP_SYS_ADMIN on the blocksnoop side (already granted by --profile=sysadmin or privileged: true).

See examples/reproduce-cross-ns.sh for a runnable Docker reproduction.

DaemonSet sidecar

For continuous monitoring, deploy blocksnoop as a DaemonSet that monitors processes on each node:

apiVersion: apps/v1
kind: DaemonSet
metadata:
  name: blocksnoop
spec:
  selector:
    matchLabels:
      app: blocksnoop
  template:
    metadata:
      labels:
        app: blocksnoop
    spec:
      hostPID: true
      containers:
        - name: blocksnoop
          image: oloapm/blocksnoop:latest
          # A DaemonSet needs an explicit target-selection policy. Set this
          # value from your workload controller; blocksnoop monitors one PID.
          env:
            - name: TARGET_PID
              value: "1234"
          command:
            - sh
            - -ec
            - >-
              exec blocksnoop --json
              --log-file /var/log/blocksnoop/events.json
              --service my-api --env production -t 100 "$TARGET_PID"
          securityContext:
            privileged: true
          volumeMounts:
            - name: logs
              mountPath: /var/log/blocksnoop
            - name: debugfs
              mountPath: /sys/kernel/debug
      volumes:
        - name: logs
          hostPath:
            path: /var/log/blocksnoop
        - name: debugfs
          hostPath:
            path: /sys/kernel/debug

The log file at /var/log/blocksnoop/events.json can be tailed by Datadog Agent, Fluentd, or any log collector running on the node.

Node shell (quick one-off)

For a quick check without building images:

# SSH into the node (or use a node shell tool)
kubectl node-shell <node-name>

# Install blocksnoop
pip install blocksnoop

# Find the Python process (hostPID shows all processes)
ps aux | grep python

# Attach
blocksnoop -t 50 <PID>

Development

# Install dependencies
uv sync --all-extras --dev

# Run unit tests
uv run --extra dev pytest tests/ -v --ignore=tests/integration

# Run integration tests (requires Docker)
uv run --extra dev pytest -m docker tests/integration/ -v

# Lint and format
ruff check blocksnoop/ tests/
ruff format blocksnoop/ tests/

# Type check
ty check blocksnoop/

License

GPL-3.0-or-later (due to the austin-python dependency)