06. Batching strategies — trade tiny waits for higher throughput only where the product allows it

~22 min read. A lead AI engineer does not optimize a single number; they make the tradeoff visible, measured, and reversible.

Builds on 00-eli5.md. The carpool lane groups work to raise utilization; the ETA call tells you whether the added wait is acceptable.

---

What previous chapters solved before this pressure appears

Streaming helped one user feel progress. Now traffic scale exposes a different pressure: the accelerator is expensive when it serves requests one by one with idle gaps. Latency anatomy taught us that queue time is visible, so batching cannot be a blanket win. This chapter shows where a tiny wait buys lower infrastructure cost and where it damages the product.

The accumulated lesson is already visible in the taxi fleet. meter ticks expose money, the ETA call exposes perceived wait, the dispatch board exposes route choice, and the fuel ledger keeps those choices honest. This file adds the next constraint without forgetting the earlier ones: every optimization relieves one pressure and creates another that some subsystem must absorb.

What this file solves

This file starts from GPU tokens/sec doubles after batching, but inline completions lose users because p95 TTFT crosses 900 ms. It shows the concrete design move a lead engineer makes, the artifact to inspect, and the signal that proves the optimization helped instead of merely moving pain elsewhere.

The opening failure shows up in a concrete artifact

The failure is not abstract: GPU tokens/sec doubles after batching, but inline completions lose users because p95 TTFT crosses 900 ms. Here is the early artifact a reviewer can inspect.

A batching review starts with endpoint budgets, not GPU utilization alone.

Batch policy comparison:

Endpoint	Batch window	p95 queue	Tokens/sec	Product result
inline completion	25 ms	240 ms	5,200	typing feels sticky
support draft	25 ms	260 ms	5,100	acceptable
nightly labeling	250 ms	410 ms	7,800	strong win
voice reply	10 ms	120 ms	4,100	still near limit

A smart team might try to fix the most visible line in that artifact. That is tempting, and it is incomplete. The root cause is product-blind scheduling: the serving layer optimized occupancy without respecting endpoint latency budgets. So how do we share hardware without making interactive users wait behind the carpool lane? This is the root-cause pivot: not a local metric problem, a boundary-and-pressure problem.

A tiny version exposes the whole mechanism

A toy request makes the pressure visible: one request looks fine in isolation, but after route mix, retries, context, or output length changes, the user-visible workflow changes. The smallest example is enough to show why the lever must be measured at the workflow boundary.

Rule: Batching buys throughput by spending queue time, and the endpoint decides how much queue time is legal.

Why this rule exists. The carpool lane works when riders can wait together; it fails when a rider is already at the curb for a live conversation. The fuel ledger matters because it shows whether the new pressure landed in cost, latency, memory, quality, or operator attention.

---

1) Batch only where queueing is cheaper than idle hardware

Start with the workflow, not the vendor feature. In Maya's review, the team takes one request and follows it from API ingress to model call, tool call, runtime behavior, and resource consequence. That cross-layer trace is the shortest path from symptom to lever. If the symptom is cost, the trace follows meter ticks. If the symptom is silence, it follows the ETA call. If the symptom is serving pressure, it follows the carpool lane and boot space.

user request
    │
    ▼
API gateway ── route/version ──► model/runtime
    │                              │
    │                              ├─ tokens / queue / KV / output
    │                              ▼
    └──────── outcome ◄──────── fuel ledger row

The counterintuitive part is that the most obvious metric can improve while the product gets worse. A smaller bill can hide more failed outcomes. Higher tokens/sec can hide longer queueing. A shorter prompt can hide missing evidence. The mechanism in this chapter is useful only when the trace keeps the relieved pressure and the newly created pressure in the same picture.

2) The occupancy picture before the queueing cost

Picture the chapter as a pressure transfer, not a free lunch.

Before optimization                    After optimization
┌──────────────────────┐              ┌──────────────────────┐
│ visible pain          │              │ relieved pain         │
│ cost / wait / memory  │──change──►   │ lower local metric    │
└──────────┬───────────┘              └──────────┬───────────┘
           │                                      │
           ▼                                      ▼
 hidden cause not named                 new pressure appears
 retries, route mix, context,           quality, queueing, cache,
 output, provider limits                memory, fallback, ops

The diagram is the reason this module keeps returning to the fuel ledger. The ledger is where the second box becomes visible instead of surfacing as an invoice surprise, p99 incident, or quality complaint weeks later.

3) Maya separates offline summaries from inline completions

Maya threads one workload through the design review: a production assistant with real traffic, route versions, prompt versions, and outcome labels.

Attempt A — optimize the visible line

The first attempt changes the local knob associated with the opening symptom. The local dashboard improves. The team celebrates too early because the request boundary is still broken: retries, quality loss, queueing, cache misses, or memory pressure move elsewhere.

Attempt B — optimize with the pressure chain

The second attempt keeps the artifact, the rule, and the guardrail together. Maya writes the expected improvement, the pressure that may worsen, the owner of that pressure, and the rollback trigger. The dispatch board may change a route, the memorized route may change a prompt prefix, or the carpool lane may change scheduling, but the same request ID proves whether the user outcome survived.

4) Why endpoint-specific batching beats one global window

The plausible alternative is attractive because it is simpler to explain in a status update: change one knob, quote one percentage, and move on. That works for demos. It fails for lead-level ownership because it cannot answer which workload benefits and which workload pays.

Use this chapter's mechanism when the workload has the shape named in the opening artifact. Use the alternative when the product is small enough, stable enough, or low-risk enough that the extra machinery would cost more than it saves. The decision is not about elegance; it is about whether the signal-to-operator cost is worth it.

5) KV memory becomes the second limit after occupancy

Concrete numbers make the tradeoff review honest. The sample prices and memory figures below are illustrative; replace them with the provider, hardware, and workload numbers in your own stack.

Scenario	Fresh input	Cached input	Output	Extra condition	Lesson
One-by-one serving	0	0	0	2k tok/s	$0.556 per 1M tok
Dynamic 10 ms window	0	0	0	5k tok/s	$0.222 per 1M tok
Dynamic 50 ms window	0	0	0	6k tok/s	$0.185 per 1M tok, +50 ms queue
Continuous batching	0	0	0	8k tok/s	$0.139 per 1M tok
Oversized batch under long prompts	0	0	0	OOM risk	throughput collapses

The table teaches the design habit: every row says what improved and what might have worsened. If a row cannot name both, the proposal is not ready for production review.

6) The efficient GPU that made typing feel sticky

Walk the failure from top to bottom. The user action enters the API. The application builds a prompt or route. The runtime spends tokens, queue time, cache memory, or output steps. The dashboard records a local improvement. Then the user-visible metric moves the wrong way.

That failure is not bad luck. It is what happens when the optimization changes one layer and the observation stops one layer too early. In a review, Maya asks for the missing link: where did the pressure go after the local metric improved? If nobody can answer, the change ships behind a small canary or does not ship.

7) Signals that reveal whether batching strategies is healthy

• Healthy behavior: throughput improves while p95 TTFT remains inside endpoint budget.
• First degrading metric: queue wait rises before tokens/sec looks bad.
• Misleading beginner metric: GPU utilization alone, because it ignores user wait and memory pressure.
• Expert graph: queue wait, batch size, active sequences, KV memory, p95 TTFT, and abandonment by endpoint.

Mini-FAQ. "Why not watch the simplest metric?" Because the simplest metric is often the one the optimization directly manipulates. You need a paired guardrail that shows whether the system merely moved pain into another layer.

8) Boundaries where the chapter's lever works and where it turns pathological

• Strong fit: offline scoring, moderation queues, bulk summarization, and non-interactive generation.
• Pathology: voice, autocomplete, and live tutoring where tens of milliseconds matter.
• Scale or workload limit: when active sequences and KV cache fill memory before compute is saturated.

This boundary is not a disclaimer. It is a routing rule for engineering attention. The best optimization in one endpoint can be the wrong default for another endpoint with different latency tolerance, risk, context length, or outcome value.

9) Wrong mental model to replace

The wrong model is that higher utilization always means better serving. The better model is scheduling under a latency budget: idle hardware is waste, but user-visible queueing is also a cost.

The replacement model should change how you speak in design review. Do not say, "this reduces cost" or "this improves latency" without naming the request slice, expected magnitude, guardrail, and rollback trigger. Say which meter ticks, ETA call, carpool lane, or boot space pressure changed.

10) Other failure shapes you will recognize

1. One failure. one global batch window across all endpoints.
2. P95 failure. p95 queueing hidden by p50 TTFT.
3. Continuous failure. continuous batching enabled without KV memory limits.
4. Long failure. long prompts mixed with short completions causing head-of-line pain.
5. Cancellations failure. cancellations not removed from batches.
6. Offline failure. offline jobs sharing the live interactive pool.
7. Autoscaling failure. autoscaling triggered only by CPU/GPU utilization, not queue delay.

11) Cross-topic reinforcement — the same pressure shape returns

• Latency anatomy supplies the queue-time budget.
• Streaming can mask some waits but not a queue that delays first useful content.
• KV cache optimization is next because batches carry active memory.
• Capacity planning later forecasts arrival bursts and batch windows together.

12) Design-review questions that catch shallow plans

1. What is the maximum legal added wait per endpoint?
2. Does batching improve cost per token after memory overhead?
3. Can cancellations leave the batch promptly?
4. Do offline and interactive pools share capacity safely?

Where this shows up in production

• Enterprise support bot — turns route, token, cache, retry, and outcome rows into cost per resolved ticket rather than model spend per message.
• Coding assistant — separates inline completions from agentic edits because typing flow, repo context, and repair loops have different budgets.
• Search answer product — pays for rewriting, retrieval, reranking, synthesis, citations, and judge calls as one user-visible answer.
• Voice assistant — treats dead air, cancellation, and local fallback as product features because users notice 100 ms gaps.
• Back-office summarizer — uses larger queues and batches because humans care about daily throughput more than first-token immediacy.
• Commerce assistant — protects purchase-changing actions with stronger routes while letting read-only advice run cheaper.
• Internal data copilot — attributes spend by tenant, dataset, prompt version, and tool path so one team cannot hide another team's budget.
• Education tutor — spends tokens on safety and pedagogy rules, then watches whether shorter answers still teach well.
• Legal review workflow — keeps evidence and citation context even when compression pressure is high because unsupported claims are worse than cost.
• Healthcare intake helper — uses conservative routing and buffered streaming because safety checks are part of the latency path.
• Marketing content tool — controls output length and variant count because creative generation can silently explode spend.
• Incident-response copilot — prefers predictable latency and logs over clever savings during high-severity operations.

Recall — rebuild batching strategies from memory

1. What concrete failure opened this chapter, and which artifact made it inspectable?
2. What root cause made the naive fix insufficient?
3. State the rule in one sentence without using vendor language.
4. Which pressure does the mechanism relieve, and which new pressure can it create?
5. Which operational signal degrades first when the mechanism is misapplied?
6. Where is the boundary where this lever becomes pathological?
7. How does this chapter reuse the fuel ledger or dispatch board from earlier chapters?
8. What would you put in the rollback trigger for this optimization?

Interview Q&A

Q: What problem does batching strategies actually solve?

A: It relieves the specific pressure from the opening failure: GPU tokens/sec doubles after batching, but inline completions lose users because p95 TTFT crosses 900 ms. The important answer names the workflow metric, the moved pressure, and the guardrail that keeps the optimization honest.

Common wrong answer to avoid: It is just a generic way to make LLMs cheaper or faster.

Q: Why is the naive fix dangerous?

A: Because the naive fix attacks the visible symptom while the root cause is the root cause is product-blind scheduling. It can improve a local metric and damage outcome quality, tail latency, memory, or operational clarity.

Common wrong answer to avoid: If the local metric improves, the product improved.

Q: What artifact would you inspect first?

A: The first artifact is the joined request trace or table for this chapter: route/version, token buckets, latency stages, outcome, and the topic-specific signal. Without it, the team debates guesses.

Common wrong answer to avoid: I would start by changing model parameters.

Q: How do you know the optimization worked?

A: The relieved pressure improves while the guardrail remains stable: throughput improves while p95 TTFT remains inside endpoint budget. You also check the first degrading metric: queue wait rises before tokens/sec looks bad.

Common wrong answer to avoid: The dashboard line for cost or latency went down.

Q: When should you avoid this lever?

A: Avoid it in the pathological case: voice, autocomplete, and live tutoring where tens of milliseconds matter. The workload shape must match the mechanism, or the optimization moves cost into quality, memory, or user wait.

Common wrong answer to avoid: Use it everywhere because it is a best practice.

Q: What is the common wrong mental model?

A: The wrong model is that higher utilization always means better serving. The better model is scheduling under a latency budget: idle hardware is waste, but user-visible queueing is also a cost.

Common wrong answer to avoid: The obvious intuition is good enough.

Q: How does this connect to earlier chapters?

A: It reuses the workflow boundary from cost anatomy, the stage trace from latency anatomy, and the fuel ledger discipline. The lever should be explained as pressure relief plus a new pressure, not as an isolated trick.

Common wrong answer to avoid: Earlier chapters are background only; this mechanism stands alone.

Apply now (10 min)

Step 1 — model the exercise. Use the modeled table in this chapter, then pick one feature you own or know. Fill in the same rows with rough numbers, change one assumption, and reproduce from memory which metric should improve, which guardrail could fail, and what rollback trigger you would set.

Step 2 — your turn. Pick a real LLM feature and write the same artifact with your own rough numbers. Name the pressure relieved, the pressure created, the owner, and the metric that would prove the change unsafe.

Step 3 — reproduce from memory. Close the file and redraw the two diagrams: request trace and pressure transfer. Then restate the rule and the first degrading metric without looking.

What you should remember

This chapter explained why the opening failure is not solved by changing one local knob. The useful move is to make the request boundary inspectable, apply the topic rule, and watch the paired guardrail so the optimization cannot hide its cost in another subsystem.

You learned to describe the lever as pressure movement: what it relieves, what it creates, and which team or resource absorbs the new cost. That is the difference between a trick and an operating practice.

Carry the diagnostic forward: if the dashboard cannot show the artifact, the route or version, the user outcome, and the first degrading signal in one place, the optimization is not yet reviewable.

Remember:

• Batching buys throughput by spending queue time, and the endpoint decides how much queue time is legal.
• The first artifact to inspect is the trace/table that exposes GPU tokens/sec doubles after batching, but inline completions lose users because p95 TTFT crosses 900 ms.
• The first degrading signal is: queue wait rises before tokens/sec looks bad.
• The misleading beginner signal is: GPU utilization alone, because it ignores user wait and memory pressure.
• Every optimization must name what pressure it relieves, what pressure it creates, and who owns the new cost.

---

Bridge. Batching keeps accelerators busy, but every active request carries memory while it waits and decodes. The next chapter opens the boot space to show why KV cache often becomes the real limit.

→ ./07-kv-cache-optimization.md