The proxy trusted its health table, which the watcher refreshes every 10 s. A
replica that restarts — and a deploy restarts both — therefore produced a
window in which requests routed to the dead card got a 502 back, even though
the other card was serving the whole time.
A failed connection now demotes that replica immediately and retries the
request on another live one, before anything is written to the client.
vLLM's OpenAI endpoints are stateless, so re-sending a request that never
reached an engine is safe; a failure after streaming has begun is still not
retried. The failover is logged, because a silent one hides a card that is
down.
Measured with card 1 stopped mid-flight: six requests, six 200s, one failover
line in the log, no client-visible error. Before this commit that burst
returned 502s.
The two replicas behind :7770 were on plain round-robin, which throws away
prefix-cache hits: an agent loop re-sends its system prompt and file context
every turn, and alternating replicas means every other turn pays a cold
prefill. Measured with 90k-token contexts, warm-turn TTFT was 2676 ms.
Pinning everything to one replica is not the answer either — a cold batch
(the shape a benchmark or a burst of parallel subagents produces) then loses a
card's worth of prefill: 16 requests took 30.1 s pinned against 18.6 s spread.
So the policy routes by request shape. A continuation (more than one message —
the agent-loop signature) is pinned to the replica that served its prefix;
anything else spreads across the cards by counter. Measured: 19.97 s for the
burst and 1095 ms warm-turn TTFT — the best of both, without a regression on
either workload.
The policy lives in proxy/routing.py, separate from the HTTP plumbing in
proxy/affinity-proxy.py, so it can be tested without aiohttp, a network or a
GPU. tests/frank_medium_proxy.bats covers it, and it earned its keep: the
first version hashed fresh prefixes to pick a card, which sends every member of
a same-prefix burst to one card (29.8 s, measured), and a stale index in the
pin path returned 500 on every continuation. Both are now assertions.
ROUTING=roundrobin and ROUTING=affinity stay available to re-measure the
choice on this host; the default is adaptive.
Frank