mirror of
https://github.com/hanzoai/ai.git
synced 2026-11-20 14:13:23 +00:00
A 429 means "ask again shortly", and we are in a far better position to do
that than the client is: the client cannot pick a different provider, cannot
see which ones are healthy, and if every client retries at once they arrive
together and overload the provider again. A 429 that escapes to the client is
a retry loop we pushed onto something dumber than us.
failoverQueryText gave each provider exactly one attempt and never waited, so
a transient "Platform overloaded" -- which DO clears in seconds, measured --
went straight out as an error whenever the model had no fallback declared
(most of them).
The fix turns on a distinction the old code conflated. isRetryableError asks
"should we cascade?" and rightly says yes to both 429 and 403. But those want
opposite handling on the SAME provider:
transient (429/503/timeout) -- the provider is busy this second and will
serve us the next. Waiting works.
permanent (401/403/quota) -- it will never serve THIS request. Waiting
is pure latency; only another provider helps.
isTransientError splits them, and retryTransient rides out the first kind with
jittered exponential backoff before the failover loop cascades. Jitter is
load-bearing: un-jittered retries all return at t+delay and reassemble the herd
that caused the 429.
Bounded on purpose. Attempts are capped, backoff is capped, and the wait always
honours the caller's context -- "keep it open until it works" queues unbounded
in-flight requests, each pinning its prompt (250K tokens is not free), turning a
provider blip into an outage of our own. A provider still refusing after a few
seconds is not blipping; the cure is a different provider, not a longer sleep.
Retry only happens before the first byte reaches the client: once a stream is
committed it cannot be replayed without duplicating or corrupting it, so
errPartiallyWritten is non-transient and stops the loop immediately -- the same
rule cascading already obeyed via writerHasData.
The policy is config (models.yaml `retry:`), not a constant: the right numbers
are a property of the providers we use today, and baking them in means a rebuild
to tune a timeout. That is how the context table came to lie for a year.
173 lines
5.6 KiB
Go
173 lines
5.6 KiB
Go
// Copyright 2023-2025 Hanzo AI Inc. All Rights Reserved.
|
|
//
|
|
// Licensed under the Apache License, Version 2.0 (the "License");
|
|
// you may not use this file except in compliance with the License.
|
|
// You may obtain a copy of the License at
|
|
//
|
|
// http://www.apache.org/licenses/LICENSE-2.0
|
|
//
|
|
// Unless required by applicable law or agreed to in writing, software
|
|
// distributed under the License is distributed on an "AS IS" BASIS,
|
|
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
// See the License for the specific language governing permissions and
|
|
// limitations under the License.
|
|
|
|
package controllers
|
|
|
|
import (
|
|
"context"
|
|
"errors"
|
|
"testing"
|
|
"time"
|
|
)
|
|
|
|
// fast policy: exercise the logic without sleeping through a real backoff.
|
|
var testPolicy = retryPolicy{attempts: 3, base: time.Millisecond, max: 2 * time.Millisecond}
|
|
|
|
func TestIsTransientError_SeparatesWaitFromCascade(t *testing.T) {
|
|
// Waiting fixes these: the provider is busy now, fine in a moment.
|
|
transient := []string{
|
|
"429 Too Many Requests",
|
|
`{"message":"Platform overloaded. Please try again later."}`,
|
|
"503 service unavailable",
|
|
"context deadline exceeded",
|
|
}
|
|
for _, m := range transient {
|
|
if !isTransientError(errors.New(m)) {
|
|
t.Errorf("should be transient (worth waiting for): %q", m)
|
|
}
|
|
}
|
|
|
|
// Waiting fixes NONE of these — the provider will never serve this request.
|
|
// They must cascade immediately; sleeping first only burns the deadline.
|
|
permanent := []string{
|
|
"403 this model is not available for your account",
|
|
"401 unauthorized",
|
|
"insufficient_quota",
|
|
"402 payment required",
|
|
}
|
|
for _, m := range permanent {
|
|
err := errors.New(m)
|
|
if isTransientError(err) {
|
|
t.Errorf("must NOT be transient (waiting cannot help): %q", m)
|
|
}
|
|
// ...but they ARE worth trying a different provider for.
|
|
if !isRetryableError(err) {
|
|
t.Errorf("should still cascade to another provider: %q", m)
|
|
}
|
|
}
|
|
}
|
|
|
|
func TestRetryTransient_RidesOutABusyProvider(t *testing.T) {
|
|
// The exact case that sent a 429 to the client before: busy, then fine.
|
|
calls := 0
|
|
err := retryTransient(context.Background(), testPolicy, func() error {
|
|
calls++
|
|
if calls < 3 {
|
|
return errors.New("429 Platform overloaded. Please try again later.")
|
|
}
|
|
return nil
|
|
})
|
|
if err != nil {
|
|
t.Fatalf("a provider that recovers must succeed, got %v", err)
|
|
}
|
|
if calls != 3 {
|
|
t.Errorf("want 3 attempts (2 refusals then success), got %d", calls)
|
|
}
|
|
}
|
|
|
|
func TestRetryTransient_DoesNotSleepOnPermanentFailure(t *testing.T) {
|
|
// A 403 must abort on the FIRST attempt: retrying an entitlement error is
|
|
// pure latency, and the client is waiting.
|
|
calls := 0
|
|
start := time.Now()
|
|
slow := retryPolicy{attempts: 5, base: 200 * time.Millisecond, max: time.Second}
|
|
err := retryTransient(context.Background(), slow, func() error {
|
|
calls++
|
|
return errors.New("403 this model is not available for your account")
|
|
})
|
|
if err == nil {
|
|
t.Fatal("expected the 403 to be returned")
|
|
}
|
|
if calls != 1 {
|
|
t.Errorf("a permanent failure must not be retried: %d attempts", calls)
|
|
}
|
|
if elapsed := time.Since(start); elapsed > 100*time.Millisecond {
|
|
t.Errorf("must not back off on a permanent failure, slept %v", elapsed)
|
|
}
|
|
}
|
|
|
|
func TestRetryTransient_GivesUpAfterAttempts(t *testing.T) {
|
|
calls := 0
|
|
err := retryTransient(context.Background(), testPolicy, func() error {
|
|
calls++
|
|
return errors.New("429 overloaded")
|
|
})
|
|
if err == nil {
|
|
t.Fatal("a provider that never recovers must surface its error")
|
|
}
|
|
if calls != testPolicy.attempts {
|
|
t.Errorf("want exactly %d attempts, got %d", testPolicy.attempts, calls)
|
|
}
|
|
}
|
|
|
|
func TestRetryTransient_StopsWhenClientIsGone(t *testing.T) {
|
|
// Holding a request open is bounded by the caller's deadline: when the
|
|
// client disconnects, we stop waiting instead of pinning a goroutine and a
|
|
// (possibly 250K-token) prompt in memory.
|
|
ctx, cancel := context.WithCancel(context.Background())
|
|
calls := 0
|
|
slow := retryPolicy{attempts: 5, base: 2 * time.Second, max: 5 * time.Second}
|
|
|
|
start := time.Now()
|
|
err := retryTransient(ctx, slow, func() error {
|
|
calls++
|
|
cancel() // client goes away during the first attempt
|
|
return errors.New("429 overloaded")
|
|
})
|
|
if err == nil {
|
|
t.Fatal("expected the last error")
|
|
}
|
|
if calls != 1 {
|
|
t.Errorf("must abandon the request once the client is gone, got %d attempts", calls)
|
|
}
|
|
if elapsed := time.Since(start); elapsed > time.Second {
|
|
t.Errorf("must not sleep out the backoff after cancellation, took %v", elapsed)
|
|
}
|
|
}
|
|
|
|
func TestRetryTransient_CannotRetryACommittedStream(t *testing.T) {
|
|
// Once bytes are on the wire the request is not replayable. errPartiallyWritten
|
|
// is deliberately non-transient so this stops at once rather than backing off.
|
|
calls := 0
|
|
err := retryTransient(context.Background(), testPolicy, func() error {
|
|
calls++
|
|
return errPartiallyWritten
|
|
})
|
|
if !errors.Is(err, errPartiallyWritten) {
|
|
t.Fatalf("want errPartiallyWritten, got %v", err)
|
|
}
|
|
if calls != 1 {
|
|
t.Errorf("a partially-written response must never be retried, got %d attempts", calls)
|
|
}
|
|
}
|
|
|
|
func TestRetryPolicy_BackoffGrowsAndIsCapped(t *testing.T) {
|
|
p := retryPolicy{attempts: 6, base: 100 * time.Millisecond, max: 400 * time.Millisecond}
|
|
for n := 0; n < 6; n++ {
|
|
d := p.backoff(n)
|
|
if d <= 0 || d > p.max {
|
|
t.Errorf("backoff(%d)=%v must be in (0, max=%v]", n, d, p.max)
|
|
}
|
|
}
|
|
// Jitter: the wait must not be identical every time, or every client that
|
|
// got a 429 together returns together and re-creates the overload.
|
|
seen := map[time.Duration]bool{}
|
|
for i := 0; i < 40; i++ {
|
|
seen[p.backoff(2)] = true
|
|
}
|
|
if len(seen) == 1 {
|
|
t.Error("backoff must be jittered — an un-jittered retry reassembles the thundering herd")
|
|
}
|
|
}
|