LLMOps Intermediate — 009: Latency, throughput, and SLO design for LLM routes

Lesson 009: Latency, throughput, and SLO design for LLM routes

Focus

Prefer explicit failure rehearsals over aspirational wording. Token Latency, throughput, and SLO design for LLM routes|9 keeps neighbouring lessons differentiable.

Key ideas

• Thread: Latency, throughput, and SLO design for LLM routes · drill v9 · spin 756113.
• Habit: attach a trace_id to every completion you would paste into an ops dashboard.
• Guardrail: add one RACI bullet for prompt or index changes before tomorrow's standup.

Deep dive notebook

Canonical LLMOps lesson — long-form reference material (expanded for production readers; syllabus used only as structural inspiration).

Overview

Why this matters now

Design for tail latency—not average demo speed. Teams rarely fail because nobody read a paper—they fail because interfaces between data, models, and humans are underspecified. Use this page as a working document: paste links to your runbooks, ticket templates, and evaluation dashboards (as plain text descriptions if URLs are internal).

Stakeholder translation: If you must explain the same idea to leadership and engineers, prepare two paragraphs: one with outcomes and risk, one with system components and dependencies.

Learning outcomes (detailed)

1. Measure time-to-first-token and end-to-end latency separately.
2. Batching helps throughput but can harm interactive p95—publish both.
3. Right-size models per route; not every path needs the largest SKU.

Deep dive: applying this in production systems

Start by separating three clocks: model release cadence, data/index refresh cadence, and policy review cadence. When those drift, users experience “correct yesterday, wrong today” behavior even if accuracy metrics look flat. At Aurora Manufacturing, a common pattern is to snapshot evaluation sets per release and run them automatically against staging before any traffic shift. That sounds bureaucratic until the first time a tokenizer or retrieval change silently shifts answer style in a regulated workflow—then the audit trail pays for itself.

Second, write down interfaces between teams the way you would between services: who owns prompt text, who approves new tools/plugins, who gets paged when refusal rates spike, and where customer complaints land. At Riverbend Health, the breakthrough was not a better base model—it was a weekly 30-minute review where support brought verbatim failure cases and engineering classified them into “data gap,” “policy gap,” “model limitation,” and “user expectation mismatch.” That taxonomy turned random anecdotes into a prioritized backlog.

Design for tail latency—not average demo speed. Use this section as scratch space: paste identifiers (not secrets) from your systems so future readers know which deployment you meant.

Real-world scenario

Setting: You are a product manager for internal tools at Aurora Manufacturing. Design for tail latency—not average demo speed.

Tension: Budget is fixed for the quarter. Meanwhile, audit questions about data lineage and model updates, and executives asking for a demo timeline need a clear story—not only a model accuracy number.

What good looks like: Decisions are documented (what shipped, what was excluded), failures have owners, and the team can replay an incident with logs and prompts redacted appropriately. This lesson’s ideas apply even if your stack differs; translate nouns (vector DB, gateway, policy engine) to your internal services.

What would you measure first?

Pick one primary metric this week—not ten. Examples: P95 latency for first token, fraction of answers with a cited retrieval span, human escalation rate, or quantum job success rate vs queue depth. At Northwind Analytics, the team posted that metric in a shared dashboard with a threshold and a rollback plan when crossed. If you cannot graph it, you are not ready to argue you improved it.

Worked example (adapt freely)

Below is a template you can copy into your notes. Replace placeholders with your environment’s names so the example stays concrete.

# Example prompt skeleton (adapt to your policy)
Role: You are an assistant for {{DOMAIN}} analysts.
Context:
- User locale: {{LOCALE}}
- Retrieved excerpts (cite by [n]): {{CHUNKS}}
Task: Answer in {{FORMAT}}. If excerpts are insufficient, say what is missing.
Checks: List assumptions; flag uncertainty.

Visual reference

Policy and safety layers rarely live in one team—document handoffs.

Pitfalls teams actually hit

Tradeoff lens

Mini case study (fictional, composite)

Northwind Analytics ran a six-week pilot. Week 1–2 focused on instrumentation (latency, errors, human escalations). Week 3–4 tightened prompts and retrieval settings. Week 5–6 measured delta against the Week 1 baseline on the same tasks—avoiding “improvement” claims from a cherry-picked demo set. Their postmortem explicitly listed three refused or unsafe requests that surfaced, and how routing changed afterward. Copy that discipline: celebrate wins, but file the near-misses.

FAQ (short)

Q: Where should we start if we have only two weeks?
A: Pick one workflow, one metric, and one rollback story. Expand after you can demonstrate improvement on that slice.

Q: How do we avoid “slide-ware”?
A: Tie every recommendation to an observable: latency, cost, defect rate, or human review load—not generic “best practices.”

These answers are generic on purpose; replace them in your internal wiki with org-specific links.

Practice (from your catalog)

Set a realistic p95 target for one customer-facing route and how you would enforce it.

Try the exercise twice: once quickly, once after sleeping on it—often the second pass surfaces edge cases.

Before you close this lesson

Closing

Keep this lesson inside Quanta GenAI: add screenshots (as new static assets if your admins allow), links to internal tickets, and names of partners. The goal is not perfection on first read—it is repeatable improvement with evidence.

Bundled reference content for Quanta GenAI Learn. Extend with your organization’s specifics.

Practice

Practice Simulate degraded retrieval once; capture user-facing fallback copy. — 9 Bump 16.