Skip to content

Fix p95 fallback in speed-optimized routing #7128

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
diegosouzapw merged 3 commits into from
Jul 16, 2026
Merged

Fix p95 fallback in speed-optimized routing #7128

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
4d80e0a
fix(auto): use p95 fallback in speed factors
KooshaPari Jul 14, 2026
7a74eb4
Merge remote-tracking branch 'origin/release/v3.8.49' into babysit-71…
diegosouzapw Jul 15, 2026
a45350e
Merge branch 'release/v3.8.49' into codex/agent-auto-routing-audit
diegosouzapw Jul 16, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
29 changes: 24 additions & 5 deletions open-sse/services/autoCombo/__tests__/speedRanking.test.ts
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -165,15 +165,34 @@ describe("rankBySpeed — factor breakdown", () => {
});

it("falls back to 0.5 per missing metric so new providers are not crushed", () => {
const ranked = rankBySpeed([candidate({ provider: "fresh", model: "m" })]);
const ranked = rankBySpeed([
candidate({
provider: "fresh",
model: "m",
p95LatencyMs: undefined,
latencyStdDev: undefined,
}),
]);
expect(ranked).toHaveLength(1);
// No telemetry at all → weighted sum lands near 0.5 with reliability multiplier 1
expect(ranked[0].factors.reliability).toBe(1);
expect(ranked[0].factors.health).toBe(1);
expect(ranked[0].factors.ttft).toBe(0.5);
expect(ranked[0].factors.tps).toBe(0.5);
});
});
expect(ranked[0].factors.tps).toBe(0.5);
});

it("uses p95 latency when TTFT and E2E telemetry are unavailable", () => {
const ranked = rankBySpeed([
candidate({ provider: "slow-tail", model: "m", p95LatencyMs: 4000 }),
candidate({ provider: "fast-tail", model: "m", p95LatencyMs: 1000 }),
]);
const fast = ranked.find((entry) => entry.provider === "fast-tail");
const slow = ranked.find((entry) => entry.provider === "slow-tail");

expect(fast?.factors.ttft).toBeGreaterThan(slow?.factors.ttft ?? 1);
expect(fast?.factors.e2e).toBeGreaterThan(slow?.factors.e2e ?? 1);
});
});

describe("rankBySpeed — weight overrides", () => {
it("respects caller weight overrides (e.g. heavy TTFT bias)", () => {
Expand Down Expand Up @@ -223,4 +242,4 @@ describe("pickFastest", () => {
const winner = pickFastest([slow, fast]);
expect(winner?.provider).toBe("fast");
});
});
});
10 changes: 8 additions & 2 deletions open-sse/services/autoCombo/speedRanking.ts
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -211,9 +211,15 @@ function speedFactorsFor(
failureRate: number
): SpeedFactors {
return {
ttft: lowerIsBetter(positiveFinite(candidate.avgTtftMs), maxima.ttft),
ttft: lowerIsBetter(
positiveFinite(candidate.avgTtftMs) ?? positiveFinite(candidate.p95LatencyMs),
maxima.ttft
),
tps: higherIsBetter(positiveFinite(candidate.avgTokensPerSecond), maxima.tps),
e2e: lowerIsBetter(positiveFinite(candidate.avgE2ELatencyMs), maxima.e2e),
e2e: lowerIsBetter(
positiveFinite(candidate.avgE2ELatencyMs) ?? positiveFinite(candidate.p95LatencyMs),
maxima.e2e
),
p95: lowerIsBetter(positiveFinite(candidate.p95LatencyMs), maxima.p95),
health: healthScoreFor(candidate.circuitBreakerState),
reliability: clamp01(1 - failureRate),
Expand Down
Loading