Files
new-api-custom/service/token_routing_test.go
T

214 lines
9.8 KiB
Go

package service
import (
"context"
"errors"
"fmt"
"github.com/QuantumNous/new-api/common"
"github.com/QuantumNous/new-api/constant"
"github.com/QuantumNous/new-api/model"
perfmetrics "github.com/QuantumNous/new-api/pkg/perf_metrics"
apiTypes "github.com/QuantumNous/new-api/relaykit/types"
"github.com/QuantumNous/new-api/setting/operation_setting"
"github.com/gin-gonic/gin"
"github.com/stretchr/testify/require"
"math"
"net/http/httptest"
"slices"
"testing"
"time"
)
func TestTokenRoutingSingleChannelRetryCooldownAndRecovery(t *testing.T) {
db := setupChannelSelectAutoGroupsTest(t)
createChannelSelectAutoGroupsChannel(t, db, 6911, "default", "retry-model")
createChannelSelectAutoGroupsChannel(t, db, 6912, "vip", "retry-model")
model.InitChannelCache()
ResetGroupAvailability()
t.Cleanup(func() { ResetGroupAvailability(); _ = poolCooldowns().Purge() })
makeCtx := func(token int) *gin.Context {
c := setupGroupPoolContext(t)
c.Request = httptest.NewRequest("POST", "/v1/responses", nil)
c.Set("id", 99)
c.Set("token_id", token)
c.Set(model.TokenRoutingContextKey, &model.TokenRoutingPolicy{Version: 1, Mode: "custom", Groups: []model.TokenGroupEntry{{Group: "default", Priority: 0}, {Group: "vip", Priority: 1}}})
return c
}
selectID := func(c *gin.Context) int {
t.Helper()
ch, _, err := selectTokenRoutingPool(&RetryParam{Ctx: c, ModelName: "retry-model", Retry: common.GetPointer(0)})
require.NoError(t, err)
require.NotNil(t, ch)
AppendUsedChannel(c, ch.Id)
return ch.Id
}
c := makeCtx(991)
require.Equal(t, 6911, selectID(c))
RecordPoolRetryableFailure(c, "retry-model", "default")
require.Equal(t, 6911, selectID(c), "a one-channel group must actually get its second attempt")
RecordPoolRetryableFailure(c, "retry-model", "default")
require.Equal(t, 6912, selectID(c))
require.Equal(t, 6912, selectID(makeCtx(991)))
require.Equal(t, 6911, selectID(makeCtx(992)))
key := poolCooldownKey(c, "retry-model", "default")
state, found, err := poolCooldowns().Get(key)
require.NoError(t, err)
require.True(t, found)
require.InDelta(t, int64((10*time.Minute)/time.Millisecond), state.Until-time.Now().UnixMilli(), 2000)
require.NoError(t, poolCooldowns().SetWithTTL(key, poolCooldown{Until: time.Now().Add(-time.Second).UnixMilli()}, 0))
trial := makeCtx(991)
require.Equal(t, 6911, selectID(trial))
require.Equal(t, 6912, selectID(makeCtx(991)), "concurrent request waits for recovery probe")
RecordPoolRetryableFailure(trial, "retry-model", "default")
require.Equal(t, 6912, selectID(trial), "failed probe switches without another primary attempt")
require.NoError(t, poolCooldowns().SetWithTTL(key, poolCooldown{Until: time.Now().Add(-time.Second).UnixMilli()}, 0))
recovery := makeCtx(991)
require.Equal(t, 6911, selectID(recovery))
RecordPoolSuccess(recovery, "retry-model", "default")
_, found, err = poolCooldowns().Get(key)
require.NoError(t, err)
require.False(t, found)
next := makeCtx(991)
require.Equal(t, 6911, selectID(next))
RecordPoolRetryableFailure(next, "retry-model", "default")
require.Equal(t, 6911, selectID(next), "recovered group returns to two-attempt behavior")
mixed := makeCtx(993)
require.Equal(t, 6911, selectID(mixed))
RecordPoolRetryableFailure(mixed, "retry-model", "default")
require.Equal(t, 6911, selectID(mixed))
require.Equal(t, 6912, selectID(mixed), "the one same-group retry must not repeat when a different error triggers another selection")
}
func TestAutomaticGroupPreferenceRanking(t *testing.T) {
original := []automaticCandidate{{group: "cheap", cost: .1, ttft: 900, success: 90, ratio: .2}, {group: "fast", cost: .2, ttft: 100, success: 95, ratio: .3}, {group: "stable", cost: .3, ttft: 400, success: 99, ratio: .4}, {group: "unknown", cost: math.Inf(1), ttft: math.Inf(1), success: -1, ratio: .01}}
for preference, want := range map[string]string{"cost": "unknown", "speed": "fast", "stable": "stable"} {
t.Run(preference, func(t *testing.T) {
items := slices.Clone(original)
sortAutomaticCandidates(items, preference)
require.Equal(t, want, items[0].group)
if preference != "cost" {
require.Equal(t, "unknown", items[len(items)-1].group)
}
})
}
}
func TestAutomaticGroupUsesAllAuthorizedModelGroups(t *testing.T) {
db := setupChannelSelectAutoGroupsTest(t)
for _, fixture := range []struct {
id int
group, name string
}{{6921, "default", "smart-model"}, {6922, "vip", "smart-model"}, {6923, "secret", "smart-model"}, {6924, "default", "another-model"}} {
createChannelSelectAutoGroupsChannel(t, db, fixture.id, fixture.group, fixture.name)
}
model.InitChannelCache()
groupModelCostSnapshot.Lock()
oldCosts, oldTime := groupModelCostSnapshot.values, groupModelCostSnapshot.updated
cheap, expensive := 0.1, 0.3
groupModelCostSnapshot.values = map[string]map[string]model.GroupModelCost{"default": {"smart-model": {USDPerMillion: &expensive}}, "vip": {"smart-model": {USDPerMillion: &cheap}}}
groupModelCostSnapshot.updated = time.Now()
groupModelCostSnapshot.Unlock()
automaticMetricCache.Set("smart-model", perfmetrics.QueryResult{})
t.Cleanup(func() {
groupModelCostSnapshot.Lock()
groupModelCostSnapshot.values, groupModelCostSnapshot.updated = oldCosts, oldTime
groupModelCostSnapshot.Unlock()
automaticMetricCache.Purge()
})
makeCtx := func() *gin.Context {
c := setupGroupPoolContext(t)
c.Request = httptest.NewRequest("POST", "/v1/responses", nil)
c.Set(model.TokenRoutingContextKey, &model.TokenRoutingPolicy{Version: 1, Mode: "smart", AutoPreference: "cost"})
return c
}
result := ResolveRequestPool(makeCtx(), "smart-model")
require.Equal(t, []model.TokenGroupEntry{{Group: "default", Priority: 0}, {Group: "vip", Priority: 1}}, result)
list := ResolveRequestPool(makeCtx(), "")
require.Len(t, list, 2)
for _, entry := range list {
require.NotEqual(t, "secret", entry.Group)
}
require.Empty(t, ResolveRequestPool(makeCtx(), "unknown-model"))
}
func TestPoolFailoverRespectsRetrySafetyBoundaries(t *testing.T) {
for _, code := range []int{400, 401, 429, 500, 502, 503, 504, 524} {
t.Run(fmt.Sprint(code), func(t *testing.T) {
c := setupGroupPoolContext(t)
c.Request = httptest.NewRequest("POST", "/v1/responses", nil)
c.Set(model.TokenRoutingContextKey, &model.TokenRoutingPolicy{Version: 1, Mode: "smart"})
err := apiTypes.NewErrorWithStatusCode(errors.New("test upstream"), apiTypes.ErrorCodeBadResponseStatusCode, code)
require.Equal(t, code == 500 || code == 502 || code == 503, IsPoolFailoverError(c, err))
ctx, cancel := context.WithCancel(c.Request.Context())
cancel()
c.Request = c.Request.WithContext(ctx)
require.False(t, IsPoolFailoverError(c, err))
})
}
c := setupGroupPoolContext(t)
c.Request = httptest.NewRequest("POST", "/v1/responses", nil)
c.Set(model.TokenRoutingContextKey, &model.TokenRoutingPolicy{Version: 1, Mode: "smart"})
_, err := c.Writer.WriteString("already streaming")
require.NoError(t, err)
failure := apiTypes.NewErrorWithStatusCode(errors.New("test upstream"), apiTypes.ErrorCodeBadResponseStatusCode, 500)
require.False(t, IsPoolFailoverError(c, failure))
require.Equal(t, "stop", DecideRelayRetry(c, failure, 5).Action)
}
func TestTokenRoutingPoolPreferenceExclusionAndAffinityIsolation(t *testing.T) {
db := setupChannelSelectAutoGroupsTest(t)
const modelName = "pool-model"
createChannelSelectAutoGroupsChannel(t, db, 6101, "default", modelName)
createChannelSelectAutoGroupsChannel(t, db, 6102, "vip", modelName)
model.InitChannelCache()
ResetGroupAvailability()
t.Cleanup(ResetGroupAvailability)
config := operation_setting.GetChannelAffinitySetting()
previous := *config
*config = operation_setting.ChannelAffinitySetting{Enabled: true, DefaultTTLSeconds: 3600, SessionMode: "prefer", Rules: []operation_setting.ChannelAffinityRule{{Name: "pool-test", ModelRegex: []string{".*"}, KeySources: []operation_setting.ChannelAffinityKeySource{{Type: "request_header", Key: "Session-Id"}}, SessionMode: "prefer"}}}
t.Cleanup(func() { *config = previous; _ = poolBindings().Purge() })
ctx := func(token int, session string) *gin.Context {
c := setupGroupPoolContext(t)
c.Request = httptest.NewRequest("POST", "/v1/responses", nil)
c.Request.Header.Set("Session-Id", session)
c.Set("id", 100)
c.Set("token_id", token)
c.Set(model.TokenRoutingContextKey, &model.TokenRoutingPolicy{Version: 1, Mode: "custom", Groups: []model.TokenGroupEntry{{Group: "default", Priority: 0}, {Group: "vip", Priority: 1}}, ModelRules: []model.ModelRoutingRule{{Model: modelName, Groups: []model.TokenGroupEntry{{Group: "vip", Priority: 0}}}}})
return c
}
c := ctx(100, "session-a")
param := &RetryParam{Ctx: c, TokenGroup: "default", ModelName: modelName, Retry: common.GetPointer(0)}
first, group, err := SelectChannelForRequest(c, modelName, param)
require.Nil(t, err)
require.NotNil(t, first)
require.Equal(t, 6102, first.Id)
require.Equal(t, "vip", group)
AppendUsedChannel(c, first.Id)
second, group, err2 := CacheGetRandomSatisfiedChannel(param)
require.NoError(t, err2)
require.NotNil(t, second)
require.Equal(t, 6101, second.Id)
c.Set("channel_id", second.Id)
common.SetContextKey(c, constant.ContextKeyAutoGroup, group)
RequestPolicy(c).Successful = true
RecordPoolAffinity(c)
next := ctx(100, "session-a")
bound, boundGroup := PreferredPoolChannel(next, modelName)
require.NotNil(t, bound)
require.Equal(t, 6101, bound.Id)
require.Equal(t, "default", boundGroup)
isolated, _ := PreferredPoolChannel(ctx(101, "session-a"), modelName)
require.Nil(t, isolated)
isolated, _ = PreferredPoolChannel(ctx(100, "session-b"), modelName)
require.Nil(t, isolated)
isolated, _ = PreferredPoolChannel(ctx(100, "session-a"), "another-model")
require.Nil(t, isolated)
ClearPoolAffinity(next)
bound, _ = PreferredPoolChannel(ctx(100, "session-a"), modelName)
require.Nil(t, bound)
AppendUsedChannel(c, second.Id)
last, _, err2 := CacheGetRandomSatisfiedChannel(param)
require.NoError(t, err2)
require.Nil(t, last)
}