Skip to lesson
Exit
LLM Inference Fundamentals for the Routing Engineer1 / 2

2 min lesson

What this means for provider-backed routing

Consider this situation: "You route to a third-party provider whose internal batching you can't control. Which set of levers is genuinely yours at the gateway?" Start with the decision, then the evidence.

Step 1 of 2

What this means for provider-backed routingyou don't own their batcher, but you own the dials in front of it

When you route to a third-party provider, their serving stack does the batching and you can't see inside it. What you control is upstream: concurrency limits, queueing and admission, hedging and which provider gets a given request. Those decisions interact with the provider's batch composition even though you never touch it.

Dials you own at the gateway
Concurrency limit
How many in-flight requests you allow per provider/route
Admission / queueing
Whether a request enters now, waits or is shed under load
Hedging
Sending a duplicate to a second provider to cut tail latency, at extra cost
Provider selection
Which backend a request lands on, by health, latency, price, cache affinity
Interview move

When you reach a batching decision in a design round, say the words "this is the throughput-versus-latency tradeoff" and then anchor it to the workload: "For Tab I'd cap batch size and concurrency to protect p99; for bulk Agent generations I'd allow larger batches because throughput per dollar matters more than a few ms of ITL." Tying the dial to the specific surface is the senior signal.

Watch out

Don't claim continuous batching "removes" the latency cost of large batches. It removes the waiting and the head-of-line blocking of static batching; it does not repeal the fact that a bigger concurrent batch makes each decode step slower. Mixing those up is a common tell.