A self-improving RLM agent for coding workflows and long-running autonomous tasks.
Rust
[RSI, performance] perf(coding-agent): jitter provider retry backoff to de-synchronize retry waves (#2474)
Sebastian Müllercc55bb11a8d7
diff --git a/packages/coding-agent/.changes/retry-backoff-jitter.md b/packages/coding-agent/.changes/retry-backoff-jitter.mdnew file mode 100644index 0000000000..c3913ef9be--- /dev/null+++ b/packages/coding-agent/.changes/retry-backoff-jitter.md@@ -0,0 +1 @@+- Provider retries now add +/-25% jitter to their computed exponential backoff so concurrent sessions spread out during a shared outage instead of retrying in lockstep; server Retry-After waits are still honored exactly.diff --git a/packages/coding-agent/src/core/provider-retry.ts b/packages/coding-agent/src/core/provider-retry.tsindex 1f68acd786..e2040bc581 100644--- a/packages/coding-agent/src/core/provider-retry.ts+++ b/packages/coding-agent/src/core/provider-retry.ts@@ -82,19 +82,27 @@ export type ProviderRetryDelay = { kind: "wait"; delayMs: number } | { kind: "ex/** Node caps timers at 2^31-1 ms; longer delays overflow setTimeout and fire after ~1ms. */const MAX_TIMER_DELAY_MS = 2_147_483_647;-/** Delay before retry `attempt` (1-based), honoring a server-requested wait. */+/**+ * Delay before retry `attempt` (1-based). A server-requested wait at or above the backoff is used as-is;+ * otherwise the backoff is jittered so concurrent sessions do not retry in lockstep, floored at the server wait.+ */export function providerRetryDelay(attempt: number,retryAfterMs: number | undefined,policy: Pick<ProviderRetryPolicy, "baseDelayMs" | "maxRetryDelayMs">,+ rng: () => number = Math.random,): ProviderRetryDelay {if (retryAfterMs !== undefined && policy.maxRetryDelayMs > 0 && retryAfterMs > policy.maxRetryDelayMs) {return { kind: "exceeds-cap", retryAfterMs };}- return {- kind: "wait",- delayMs: Math.min(Math.max(policy.baseDelayMs * 2 ** (attempt - 1), retryAfterMs ?? 0), MAX_TIMER_DELAY_MS),- };+ const backoffMs = policy.baseDelayMs * 2 ** (attempt - 1);+ if (retryAfterMs !== undefined && retryAfterMs >= backoffMs) {+ return { kind: "wait", delayMs: Math.min(retryAfterMs, MAX_TIMER_DELAY_MS) };+ }+ const jitteredMs = providerWaitJitter(backoffMs, rng);+ const flooredMs = Math.max(jitteredMs, retryAfterMs ?? 0);+ const clampedMs = Math.min(flooredMs, MAX_TIMER_DELAY_MS);+ return { kind: "wait", delayMs: clampedMs };}/**@@ -247,7 +255,7 @@ export function providerWaitPingDelay(return Math.min(capped, MAX_TIMER_DELAY_MS);}-/** +/-25% jitter around a ping delay (avoids thundering-herd retries). */+/** +/-25% jitter around a delay (avoids thundering-herd retries). */export function providerWaitJitter(delayMs: number, rng: () => number = Math.random): number {const factor = 0.75 + 0.5 * Math.max(0, Math.min(1, rng()));return Math.max(0, Math.round(delayMs * factor));diff --git a/packages/coding-agent/test/provider-retry.test.ts b/packages/coding-agent/test/provider-retry.test.tsindex 4253146dac..7860ef2b78 100644--- a/packages/coding-agent/test/provider-retry.test.ts+++ b/packages/coding-agent/test/provider-retry.test.ts@@ -37,6 +37,14 @@ function providerError(kind?: string): AssistantMessage {}describe("completeWithProviderRetry", () => {+ it("jitters the computed backoff while honoring a server wait exactly", () => {+ const policy = { baseDelayMs: 2000, maxRetryDelayMs: 60_000 };+ expect(providerRetryDelay(1, undefined, policy, () => 0)).toEqual({ kind: "wait", delayMs: 1500 });+ expect(providerRetryDelay(1, undefined, policy, () => 1)).toEqual({ kind: "wait", delayMs: 2500 });+ expect(providerRetryDelay(1, 5000, policy, () => 1)).toEqual({ kind: "wait", delayMs: 5000 });+ expect(providerRetryDelay(3, 7000, policy, () => 0)).toEqual({ kind: "wait", delayMs: 7000 });+ });+it("returns an aborted result instead of the provider error when cancelled during backoff", async () => {const controller = new AbortController();setTimeout(() => controller.abort(), 10);