Skip to content

CONCEPT Cited by 1 source

Offense / defense performance engineering

Definition

Offense / defense performance engineering is Meta's operational framing of hyperscale capacity efficiency as a two-sided problem that requires investment in both halves to keep megawatt delivery moving (Source: sources/2026-04-16-meta-capacity-efficiency-at-meta-how-unified-ai-agents-optimize-performance-at-hyperscale).

  • Offense — proactively find and ship optimizations: "searching for opportunities (proactive code changes) to make our existing systems more efficient, and deploying them."
  • Defense — proactively catch and mitigate regressions: "monitoring resource usage in production to detect regressions, root-cause them to a pull request, and deploy mitigations."

The framing's sharpest claim is architectural: "The breakthrough was realizing that both problems share the same structure" — same tool surface (profiling / code search / configuration history / documentation / experiment results), different skills — so one platform (systems/meta-capacity-efficiency-platform) can serve both.

Why both halves matter

Either half alone is insufficient:

  • Offense without defense leaks gains back through regressions. Every shipped optimization is eventually eroded by somebody else's change; without FBDetect-style continuous monitoring, the fleet drifts toward the pre-optimization baseline.
  • Defense without offense protects what you have but doesn't improve the operating point. Meta: "even a 0.1 % performance regression can translate to significant additional power consumption" at 3 B-user scale — staying flat is effectively losing if traffic grows.

Shape of each side

Defense pipeline

  1. Detect regression (FBDetect catches 0.005 % regressions in noisy production, thousands/week).
  2. Attribute to a root-cause pull request ("correlating regression functions with recent pull requests").
  3. Generate mitigation (AI Regression Solver: gather-context / apply-skill / create-PR-to-root-cause-author).
  4. Engineer reviews + merges.

Offense pipeline

  1. Identify optimization opportunity (proactive surveys, profiling, pattern-matching, suggestions from engineers).
  2. Engineer views opportunity → requests AI-generated fix.
  3. Generate candidate (patterns/opportunity-to-pr-ai-pipeline: gather-context / apply-skill / create-candidate-in-editor).
  4. Engineer reviews + applies + ships.

Meta names the parallel explicitly: "The pipeline mirrors the defensive AI Regression Solver."

Why this is a wiki concept

Previously the wiki had material on each half separately:

The 2026-04-16 post unifies them conceptually — offense and defense share a tool layer because they share a data shape — and structurally — the AI pipelines for both sides are three-phase (context → skill → resolution) with identical tool invocations. Canonical wiki statement of the symmetry.

Generalization beyond performance

The frame isn't specific to power/capacity. The same two-sided structure applies to:

  • Reliability — offense: proactive reliability work, chaos engineering, failure-mode surfacing; defense: incident response, RCA, post-incident work.
  • Security — offense: offensive security / pen-testing / pre-emptive hardening; defense: detection, IR, patch-lag reduction.
  • Cost — offense: FinOps forecasting + right-sizing; defense: billing-anomaly detection + chargeback enforcement.

All three share the shape "same data layer, different skills per side"; the Meta Capacity Efficiency Platform is the canonical wiki instance for performance, and a natural template for the others.

Seen in

Last updated · 319 distilled / 1,201 read