perf(postgres): 回收突发连接并等待验收副作用
失败 Run 的未提交任务已取消，但退款 outbox 尚未完成时下一轮会继承 API 连接高水位和 canceled acquire 增量，污染容量验收。 新增 AI_GATEWAY_DATABASE_MAX_CONN_IDLE_SECONDS，生产配置 30 秒回收突发空闲连接；验收中止后等待旧任务、退款、回调和数据库连接全部收敛后才允许新负载启动。 验证：Go 全量测试、gofmt、bash -n、ShellCheck、kubectl kustomize 和 git diff --check 通过。
This commit is contained in:
@@ -164,7 +164,7 @@ AI_GATEWAY_DATABASE_URL=postgresql://easyai:easyai2025@localhost:5432/easyai_ai_
|
||||
|
||||
如果现有 `easyai-pgvector` 没有把 `5432` 映射到宿主机，就需要补端口映射，或者把 AI Gateway 后端容器化后接入同一个 `easyai` Docker network。
|
||||
|
||||
异步队列 worker 不使用固定业务并发。服务分别汇总启用平台模型和活跃用户组的有效 `concurrent` 策略，采用两者中更严格的集群容量，默认每 5 秒在线调整 River 执行容量。策略解析兼容历史 `platformLimits/modelLimits.max_concurrent_requests`，运行时统一转换为 `rules`。`AI_GATEWAY_ASYNC_WORKER_HARD_LIMIT` 默认 `2048`，限制策略推导出的集群目标；`AI_GATEWAY_ASYNC_WORKER_INSTANCE_HARD_LIMIT` 默认 `32`，按单 Worker 的内存安全容量限制实例分配，即使其他实例失活也不会突破。Worker 的 `AI_GATEWAY_DATABASE_MAX_CONNS` 必须高于实例执行容量，为心跳、选主、租约续期和健康检查保留连接；生产环境按每实例执行容量 `24` 配置连接池上限 `32`。`AI_GATEWAY_DATABASE_MIN_IDLE_CONNS` 控制启动预热连接数，默认 `0`，生产 K3s 配置为 `4`，避免把连接池上限全部常驻占用。平台模型和用户组的 PostgreSQL concurrency lease 仍是业务并发真值。可通过 `AI_GATEWAY_ASYNC_WORKER_REFRESH_INTERVAL_SECONDS` 调整刷新周期。
|
||||
异步队列 worker 不使用固定业务并发。服务分别汇总启用平台模型和活跃用户组的有效 `concurrent` 策略，采用两者中更严格的集群容量，默认每 5 秒在线调整 River 执行容量。策略解析兼容历史 `platformLimits/modelLimits.max_concurrent_requests`，运行时统一转换为 `rules`。`AI_GATEWAY_ASYNC_WORKER_HARD_LIMIT` 默认 `2048`，限制策略推导出的集群目标；`AI_GATEWAY_ASYNC_WORKER_INSTANCE_HARD_LIMIT` 默认 `32`，按单 Worker 的内存安全容量限制实例分配，即使其他实例失活也不会突破。Worker 的 `AI_GATEWAY_DATABASE_MAX_CONNS` 必须高于实例执行容量，为心跳、选主、租约续期和健康检查保留连接；生产环境按每实例执行容量 `24` 配置连接池上限 `32`。`AI_GATEWAY_DATABASE_MIN_IDLE_CONNS` 控制启动预热连接数，默认 `0`，生产 K3s 配置为 `4`；`AI_GATEWAY_DATABASE_MAX_CONN_IDLE_SECONDS` 可缩短突发连接的空闲回收时间，生产配置为 `30` 秒，避免把连接池高水位长期常驻。平台模型和用户组的 PostgreSQL concurrency lease 仍是业务并发真值。可通过 `AI_GATEWAY_ASYNC_WORKER_REFRESH_INTERVAL_SECONDS` 调整刷新周期。
|
||||
|
||||
## 迁移原则
|
||||
|
||||
|
||||
@@ -40,6 +40,7 @@ func main() {
|
||||
db, err := store.ConnectWithPoolOptions(ctx, cfg.DatabaseURL, store.PostgresPoolOptions{
|
||||
MaxConns: cfg.DatabaseMaxConns,
|
||||
MinIdleConns: cfg.DatabaseMinIdleConns,
|
||||
MaxConnIdleTime: time.Duration(cfg.DatabaseMaxConnIdleSeconds) * time.Second,
|
||||
IdleInTransactionTimeout: time.Duration(cfg.DatabaseIdleInTransactionTimeoutSeconds) * time.Second,
|
||||
LockTimeout: time.Duration(cfg.DatabaseLockTimeoutSeconds) * time.Second,
|
||||
})
|
||||
|
||||
@@ -60,6 +60,7 @@ type Config struct {
|
||||
ProcessRole string
|
||||
DatabaseMaxConns int
|
||||
DatabaseMinIdleConns int
|
||||
DatabaseMaxConnIdleSeconds int
|
||||
DatabaseIdleInTransactionTimeoutSeconds int
|
||||
DatabaseLockTimeoutSeconds int
|
||||
MediaRequestConcurrency int
|
||||
@@ -124,6 +125,7 @@ func Load() Config {
|
||||
ProcessRole: strings.ToLower(strings.TrimSpace(env("AI_GATEWAY_PROCESS_ROLE", "all"))),
|
||||
DatabaseMaxConns: envInt("AI_GATEWAY_DATABASE_MAX_CONNS", 0),
|
||||
DatabaseMinIdleConns: envOptionalIntValidated("AI_GATEWAY_DATABASE_MIN_IDLE_CONNS", 0),
|
||||
DatabaseMaxConnIdleSeconds: envOptionalIntValidated("AI_GATEWAY_DATABASE_MAX_CONN_IDLE_SECONDS", 0),
|
||||
|
||||
DatabaseIdleInTransactionTimeoutSeconds: envOptionalIntValidated(
|
||||
"AI_GATEWAY_DATABASE_IDLE_IN_TRANSACTION_TIMEOUT_SECONDS",
|
||||
@@ -153,6 +155,9 @@ func (c Config) Validate() error {
|
||||
(c.DatabaseMaxConns > 0 && c.DatabaseMinIdleConns > c.DatabaseMaxConns) {
|
||||
return errors.New("AI_GATEWAY_DATABASE_MIN_IDLE_CONNS must be between 0 and AI_GATEWAY_DATABASE_MAX_CONNS")
|
||||
}
|
||||
if c.DatabaseMaxConnIdleSeconds < 0 || c.DatabaseMaxConnIdleSeconds > 3600 {
|
||||
return errors.New("AI_GATEWAY_DATABASE_MAX_CONN_IDLE_SECONDS must be between 0 and 3600")
|
||||
}
|
||||
if c.DatabaseIdleInTransactionTimeoutSeconds < 0 || c.DatabaseIdleInTransactionTimeoutSeconds > 3600 {
|
||||
return errors.New("AI_GATEWAY_DATABASE_IDLE_IN_TRANSACTION_TIMEOUT_SECONDS must be between 0 and 3600")
|
||||
}
|
||||
|
||||
@@ -145,11 +145,13 @@ func TestProcessRolePrecedenceAndCompatibility(t *testing.T) {
|
||||
func TestValidateProcessRoleAndDatabasePool(t *testing.T) {
|
||||
cfg := Load()
|
||||
if cfg.DatabaseMinIdleConns != 0 ||
|
||||
cfg.DatabaseMaxConnIdleSeconds != 0 ||
|
||||
cfg.DatabaseIdleInTransactionTimeoutSeconds != 60 ||
|
||||
cfg.DatabaseLockTimeoutSeconds != 30 {
|
||||
t.Fatalf(
|
||||
"database pool minimum/transaction timeouts = %d/%d/%d, want 0/60/30",
|
||||
"database pool minimum/max-idle/transaction timeouts = %d/%d/%d/%d, want 0/0/60/30",
|
||||
cfg.DatabaseMinIdleConns,
|
||||
cfg.DatabaseMaxConnIdleSeconds,
|
||||
cfg.DatabaseIdleInTransactionTimeoutSeconds,
|
||||
cfg.DatabaseLockTimeoutSeconds,
|
||||
)
|
||||
@@ -174,6 +176,11 @@ func TestValidateProcessRoleAndDatabasePool(t *testing.T) {
|
||||
t.Fatalf("Validate() error = %v, want invalid non-integer database min idle conns", err)
|
||||
}
|
||||
t.Setenv("AI_GATEWAY_DATABASE_MIN_IDLE_CONNS", "4")
|
||||
cfg.DatabaseMaxConnIdleSeconds = 3601
|
||||
if err := cfg.Validate(); err == nil || !strings.Contains(err.Error(), "DATABASE_MAX_CONN_IDLE_SECONDS") {
|
||||
t.Fatalf("Validate() error = %v, want invalid database max connection idle seconds", err)
|
||||
}
|
||||
cfg.DatabaseMaxConnIdleSeconds = 30
|
||||
cfg.DatabaseIdleInTransactionTimeoutSeconds = 3601
|
||||
if err := cfg.Validate(); err == nil || !strings.Contains(err.Error(), "IDLE_IN_TRANSACTION") {
|
||||
t.Fatalf("Validate() error = %v, want invalid idle transaction timeout", err)
|
||||
|
||||
@@ -32,6 +32,7 @@ const (
|
||||
type PostgresPoolOptions struct {
|
||||
MaxConns int
|
||||
MinIdleConns int
|
||||
MaxConnIdleTime time.Duration
|
||||
IdleInTransactionTimeout time.Duration
|
||||
LockTimeout time.Duration
|
||||
}
|
||||
@@ -132,6 +133,9 @@ func postgresPoolConfigWithOptions(databaseURL string, options PostgresPoolOptio
|
||||
config.MaxConns = int32(options.MaxConns)
|
||||
}
|
||||
config.MinIdleConns = int32(min(max(options.MinIdleConns, 0), int(config.MaxConns)))
|
||||
if options.MaxConnIdleTime > 0 {
|
||||
config.MaxConnIdleTime = options.MaxConnIdleTime
|
||||
}
|
||||
return config, nil
|
||||
}
|
||||
|
||||
|
||||
@@ -61,6 +61,7 @@ func TestPostgresPoolConfigSetsBoundedTransactionTimeouts(t *testing.T) {
|
||||
PostgresPoolOptions{
|
||||
MaxConns: 32,
|
||||
MinIdleConns: 4,
|
||||
MaxConnIdleTime: 30 * time.Second,
|
||||
IdleInTransactionTimeout: 60 * time.Second,
|
||||
LockTimeout: 30 * time.Second,
|
||||
},
|
||||
@@ -77,6 +78,9 @@ func TestPostgresPoolConfigSetsBoundedTransactionTimeouts(t *testing.T) {
|
||||
if config.MaxConns != 32 || config.MinIdleConns != 4 {
|
||||
t.Fatalf("pool bounds max=%d minIdle=%d, want 32/4", config.MaxConns, config.MinIdleConns)
|
||||
}
|
||||
if config.MaxConnIdleTime != 30*time.Second {
|
||||
t.Fatalf("pool max idle time=%s, want 30s", config.MaxConnIdleTime)
|
||||
}
|
||||
}
|
||||
|
||||
func TestIsPostgresUnavailableClassifiesConnectivityFailures(t *testing.T) {
|
||||
|
||||
@@ -21,6 +21,7 @@ source "$config_file"
|
||||
: "${AI_GATEWAY_DATABASE_MAX_CONNS:=32}"
|
||||
: "${AI_GATEWAY_API_DATABASE_MAX_CONNS:=64}"
|
||||
: "${AI_GATEWAY_DATABASE_MIN_IDLE_CONNS:=4}"
|
||||
: "${AI_GATEWAY_DATABASE_MAX_CONN_IDLE_SECONDS:=30}"
|
||||
: "${AI_GATEWAY_MEDIA_MATERIALIZATION_CONCURRENCY:=24}"
|
||||
: "${AI_GATEWAY_MEDIA_REQUEST_CONCURRENCY:=24}"
|
||||
[[ $RELEASES_DIR == /* && $NAMESPACE =~ ^[a-z0-9-]+$ ]] || {
|
||||
@@ -34,6 +35,7 @@ for capacity_value in \
|
||||
"$AI_GATEWAY_ASYNC_WORKER_HARD_LIMIT" \
|
||||
"$AI_GATEWAY_DATABASE_MAX_CONNS" \
|
||||
"$AI_GATEWAY_API_DATABASE_MAX_CONNS" \
|
||||
"$AI_GATEWAY_DATABASE_MAX_CONN_IDLE_SECONDS" \
|
||||
"$AI_GATEWAY_MEDIA_MATERIALIZATION_CONCURRENCY" \
|
||||
"$AI_GATEWAY_MEDIA_REQUEST_CONCURRENCY"; do
|
||||
[[ $capacity_value =~ ^[1-9][0-9]*$ ]] || {
|
||||
@@ -53,6 +55,7 @@ done
|
||||
AI_GATEWAY_API_DATABASE_MAX_CONNS <= 256 &&
|
||||
AI_GATEWAY_DATABASE_MIN_IDLE_CONNS <= AI_GATEWAY_DATABASE_MAX_CONNS &&
|
||||
AI_GATEWAY_DATABASE_MIN_IDLE_CONNS <= AI_GATEWAY_API_DATABASE_MAX_CONNS &&
|
||||
AI_GATEWAY_DATABASE_MAX_CONN_IDLE_SECONDS <= 3600 &&
|
||||
AI_GATEWAY_MEDIA_MATERIALIZATION_CONCURRENCY <= 256 &&
|
||||
AI_GATEWAY_MEDIA_REQUEST_CONCURRENCY <= 1024 )) || {
|
||||
echo 'worker capacity configuration exceeds the release safety bounds' >&2
|
||||
@@ -138,6 +141,7 @@ rollout_worker_site() {
|
||||
"AI_GATEWAY_ASYNC_WORKER_HARD_LIMIT=$AI_GATEWAY_ASYNC_WORKER_HARD_LIMIT" \
|
||||
"AI_GATEWAY_DATABASE_MAX_CONNS=$AI_GATEWAY_DATABASE_MAX_CONNS" \
|
||||
"AI_GATEWAY_DATABASE_MIN_IDLE_CONNS=$AI_GATEWAY_DATABASE_MIN_IDLE_CONNS" \
|
||||
"AI_GATEWAY_DATABASE_MAX_CONN_IDLE_SECONDS=$AI_GATEWAY_DATABASE_MAX_CONN_IDLE_SECONDS" \
|
||||
"AI_GATEWAY_MEDIA_MATERIALIZATION_CONCURRENCY=$AI_GATEWAY_MEDIA_MATERIALIZATION_CONCURRENCY" \
|
||||
"AI_GATEWAY_MEDIA_REQUEST_CONCURRENCY=$AI_GATEWAY_MEDIA_REQUEST_CONCURRENCY"
|
||||
"${kubectl[@]}" scale "deployment/easyai-worker-$site" -n "$NAMESPACE" \
|
||||
@@ -158,7 +162,8 @@ rollout_api_capacity_site() {
|
||||
"${kubectl[@]}" set env "deployment/easyai-api-$site" -n "$NAMESPACE" \
|
||||
"AI_GATEWAY_ASYNC_WORKER_HARD_LIMIT=$AI_GATEWAY_ASYNC_WORKER_HARD_LIMIT" \
|
||||
"AI_GATEWAY_DATABASE_MAX_CONNS=$AI_GATEWAY_API_DATABASE_MAX_CONNS" \
|
||||
"AI_GATEWAY_DATABASE_MIN_IDLE_CONNS=$AI_GATEWAY_DATABASE_MIN_IDLE_CONNS"
|
||||
"AI_GATEWAY_DATABASE_MIN_IDLE_CONNS=$AI_GATEWAY_DATABASE_MIN_IDLE_CONNS" \
|
||||
"AI_GATEWAY_DATABASE_MAX_CONN_IDLE_SECONDS=$AI_GATEWAY_DATABASE_MAX_CONN_IDLE_SECONDS"
|
||||
"${kubectl[@]}" rollout status "deployment/easyai-api-$site" -n "$NAMESPACE" --timeout=300s
|
||||
}
|
||||
|
||||
@@ -170,7 +175,7 @@ apply_capacity_config() {
|
||||
verify_site hongkong
|
||||
verify_site ningbo
|
||||
wait_for_url 'public readiness' "$PUBLIC_BASE_URL/api/v1/readyz" '"ok":true'
|
||||
echo "production_capacity=PASS ningbo_replicas=$AI_GATEWAY_WORKER_REPLICAS_NINGBO hongkong_replicas=$AI_GATEWAY_WORKER_REPLICAS_HONGKONG instance_limit=$AI_GATEWAY_ASYNC_WORKER_INSTANCE_HARD_LIMIT global_limit=$AI_GATEWAY_ASYNC_WORKER_HARD_LIMIT worker_database_pool=$AI_GATEWAY_DATABASE_MAX_CONNS api_database_pool=$AI_GATEWAY_API_DATABASE_MAX_CONNS database_min_idle=$AI_GATEWAY_DATABASE_MIN_IDLE_CONNS media_concurrency=$AI_GATEWAY_MEDIA_MATERIALIZATION_CONCURRENCY"
|
||||
echo "production_capacity=PASS ningbo_replicas=$AI_GATEWAY_WORKER_REPLICAS_NINGBO hongkong_replicas=$AI_GATEWAY_WORKER_REPLICAS_HONGKONG instance_limit=$AI_GATEWAY_ASYNC_WORKER_INSTANCE_HARD_LIMIT global_limit=$AI_GATEWAY_ASYNC_WORKER_HARD_LIMIT worker_database_pool=$AI_GATEWAY_DATABASE_MAX_CONNS api_database_pool=$AI_GATEWAY_API_DATABASE_MAX_CONNS database_min_idle=$AI_GATEWAY_DATABASE_MIN_IDLE_CONNS database_max_idle_seconds=$AI_GATEWAY_DATABASE_MAX_CONN_IDLE_SECONDS media_concurrency=$AI_GATEWAY_MEDIA_MATERIALIZATION_CONCURRENCY"
|
||||
}
|
||||
|
||||
rollout_site() {
|
||||
@@ -183,7 +188,8 @@ rollout_site() {
|
||||
"${kubectl[@]}" set env "deployment/easyai-api-$site" -n "$NAMESPACE" \
|
||||
"AI_GATEWAY_ASYNC_WORKER_HARD_LIMIT=$AI_GATEWAY_ASYNC_WORKER_HARD_LIMIT" \
|
||||
"AI_GATEWAY_DATABASE_MAX_CONNS=$AI_GATEWAY_API_DATABASE_MAX_CONNS" \
|
||||
"AI_GATEWAY_DATABASE_MIN_IDLE_CONNS=$AI_GATEWAY_DATABASE_MIN_IDLE_CONNS"
|
||||
"AI_GATEWAY_DATABASE_MIN_IDLE_CONNS=$AI_GATEWAY_DATABASE_MIN_IDLE_CONNS" \
|
||||
"AI_GATEWAY_DATABASE_MAX_CONN_IDLE_SECONDS=$AI_GATEWAY_DATABASE_MAX_CONN_IDLE_SECONDS"
|
||||
"${kubectl[@]}" set image "deployment/easyai-api-$site" -n "$NAMESPACE" \
|
||||
"api=$api_image"
|
||||
"${kubectl[@]}" rollout status "deployment/easyai-api-$site" -n "$NAMESPACE" --timeout=300s
|
||||
|
||||
@@ -12,5 +12,6 @@ AI_GATEWAY_ASYNC_WORKER_HARD_LIMIT=48
|
||||
AI_GATEWAY_DATABASE_MAX_CONNS=32
|
||||
AI_GATEWAY_API_DATABASE_MAX_CONNS=64
|
||||
AI_GATEWAY_DATABASE_MIN_IDLE_CONNS=4
|
||||
AI_GATEWAY_DATABASE_MAX_CONN_IDLE_SECONDS=30
|
||||
AI_GATEWAY_MEDIA_MATERIALIZATION_CONCURRENCY=24
|
||||
AI_GATEWAY_MEDIA_REQUEST_CONCURRENCY=24
|
||||
|
||||
@@ -56,6 +56,8 @@ spec:
|
||||
value: "64"
|
||||
- name: AI_GATEWAY_DATABASE_MIN_IDLE_CONNS
|
||||
value: "4"
|
||||
- name: AI_GATEWAY_DATABASE_MAX_CONN_IDLE_SECONDS
|
||||
value: "30"
|
||||
- name: AI_GATEWAY_MEDIA_REQUEST_CONCURRENCY
|
||||
value: "16"
|
||||
- name: AI_GATEWAY_MEDIA_MATERIALIZATION_CONCURRENCY
|
||||
@@ -187,6 +189,8 @@ spec:
|
||||
value: "64"
|
||||
- name: AI_GATEWAY_DATABASE_MIN_IDLE_CONNS
|
||||
value: "4"
|
||||
- name: AI_GATEWAY_DATABASE_MAX_CONN_IDLE_SECONDS
|
||||
value: "30"
|
||||
- name: AI_GATEWAY_MEDIA_REQUEST_CONCURRENCY
|
||||
value: "16"
|
||||
- name: AI_GATEWAY_MEDIA_MATERIALIZATION_CONCURRENCY
|
||||
@@ -318,6 +322,8 @@ spec:
|
||||
value: "32"
|
||||
- name: AI_GATEWAY_DATABASE_MIN_IDLE_CONNS
|
||||
value: "4"
|
||||
- name: AI_GATEWAY_DATABASE_MAX_CONN_IDLE_SECONDS
|
||||
value: "30"
|
||||
- name: AI_GATEWAY_ASYNC_WORKER_INSTANCE_HARD_LIMIT
|
||||
value: "24"
|
||||
- name: AI_GATEWAY_MEDIA_MATERIALIZATION_CONCURRENCY
|
||||
@@ -451,6 +457,8 @@ spec:
|
||||
value: "32"
|
||||
- name: AI_GATEWAY_DATABASE_MIN_IDLE_CONNS
|
||||
value: "4"
|
||||
- name: AI_GATEWAY_DATABASE_MAX_CONN_IDLE_SECONDS
|
||||
value: "30"
|
||||
- name: AI_GATEWAY_ASYNC_WORKER_INSTANCE_HARD_LIMIT
|
||||
value: "24"
|
||||
- name: AI_GATEWAY_MEDIA_MATERIALIZATION_CONCURRENCY
|
||||
|
||||
@@ -65,6 +65,7 @@ WebP 和 4K 图片。128 组输入组合避免只命中相同缓存；每四个
|
||||
AI_GATEWAY_ASYNC_WORKER_INSTANCE_HARD_LIMIT
|
||||
AI_GATEWAY_DATABASE_MAX_CONNS
|
||||
AI_GATEWAY_DATABASE_MIN_IDLE_CONNS
|
||||
AI_GATEWAY_DATABASE_MAX_CONN_IDLE_SECONDS
|
||||
AI_GATEWAY_MEDIA_MATERIALIZATION_CONCURRENCY
|
||||
AI_GATEWAY_MEDIA_REQUEST_CONCURRENCY
|
||||
AI_GATEWAY_WORKER_REPLICAS_NINGBO
|
||||
@@ -83,8 +84,10 @@ AI_GATEWAY_WORKER_REPLICAS_HONGKONG
|
||||
API Pod 使用独立发布配置 `AI_GATEWAY_API_DATABASE_MAX_CONNS`，生产同构验收默认使用 `64`；
|
||||
也可通过 `AI_GATEWAY_ACCEPTANCE_API_DATABASE_MAX_CONNS` 覆盖本次验收值，无需修改 Go 代码。
|
||||
所有 API/Worker Pool 的 `AI_GATEWAY_DATABASE_MIN_IDLE_CONNS` 固定为 `4`，仅预热必要连接，
|
||||
其余连接按负载弹性创建，避免四个 Pod 在空闲时就占满 PostgreSQL 连接预算。各档表中的
|
||||
数据库池仍表示单个 Worker 的最大连接数。
|
||||
`AI_GATEWAY_DATABASE_MAX_CONN_IDLE_SECONDS` 为 `30`，突发扩出的空闲连接会及时回落，
|
||||
避免四个 Pod 在空闲时占满 PostgreSQL 连接预算。各档表中的数据库池仍表示单个 Worker
|
||||
的最大连接数。中止旧 Run 后还必须等待未提交任务、退款 outbox、回调 outbox 和连接高水位
|
||||
全部收敛，才会启动下一轮负载。
|
||||
|
||||
每档依次运行全部五个模拟 Profile 三次。失败时恢复上一个已完整通过的档位，流量保持
|
||||
`validation`，不会自动放开正式请求。
|
||||
|
||||
@@ -226,21 +226,41 @@ wait_for_existing_tasks_to_drain() {
|
||||
|
||||
wait_for_terminal_acceptance_tasks_to_drain() {
|
||||
local deadline=$((SECONDS + 300))
|
||||
local active
|
||||
local state active side_effects connections max_connections
|
||||
while (( SECONDS < deadline )); do
|
||||
active=$(database_query "
|
||||
SELECT count(*)
|
||||
FROM gateway_tasks task
|
||||
JOIN gateway_acceptance_runs run ON run.id=task.acceptance_run_id
|
||||
WHERE run.status IN ('failed', 'aborted')
|
||||
AND task.status IN ('queued', 'running');")
|
||||
if [[ $active == 0 ]]; then
|
||||
state=$(database_query "
|
||||
SELECT
|
||||
(SELECT count(*)
|
||||
FROM gateway_tasks task
|
||||
JOIN gateway_acceptance_runs run ON run.id=task.acceptance_run_id
|
||||
WHERE run.status IN ('failed', 'aborted')
|
||||
AND task.status IN ('queued', 'running'))||','||
|
||||
(
|
||||
(SELECT count(*)
|
||||
FROM settlement_outbox settlement
|
||||
JOIN gateway_tasks task ON task.id=settlement.task_id
|
||||
JOIN gateway_acceptance_runs run ON run.id=task.acceptance_run_id
|
||||
WHERE run.status IN ('failed', 'aborted')
|
||||
AND settlement.status IN ('pending', 'processing', 'retryable_failed'))
|
||||
+
|
||||
(SELECT count(*)
|
||||
FROM gateway_task_callback_outbox callback
|
||||
JOIN gateway_tasks task ON task.id=callback.task_id
|
||||
JOIN gateway_acceptance_runs run ON run.id=task.acceptance_run_id
|
||||
WHERE run.status IN ('failed', 'aborted')
|
||||
AND callback.status <> 'delivered')
|
||||
)||','||
|
||||
(SELECT count(*) FROM pg_stat_activity WHERE backend_type='client backend')||','||
|
||||
(SELECT setting FROM pg_settings WHERE name='max_connections');")
|
||||
IFS=',' read -r active side_effects connections max_connections <<<"$state"
|
||||
if [[ $active == 0 && $side_effects == 0 ]] &&
|
||||
(( connections * 2 < max_connections )); then
|
||||
return 0
|
||||
fi
|
||||
echo "waiting_for_terminal_acceptance_tasks=$active"
|
||||
echo "waiting_for_terminal_acceptance_tasks=$active side_effects=$side_effects database_connections=$connections"
|
||||
sleep 2
|
||||
done
|
||||
echo "terminal acceptance tasks did not drain within 5 minutes: active=$active" >&2
|
||||
echo "terminal acceptance state did not drain within 5 minutes: active=$active side_effects=$side_effects database_connections=$connections" >&2
|
||||
return 1
|
||||
}
|
||||
|
||||
@@ -675,6 +695,7 @@ apply_capacity_profile() {
|
||||
local slots pool media global
|
||||
local api_pool=$AI_GATEWAY_ACCEPTANCE_API_DATABASE_MAX_CONNS
|
||||
local min_idle=4
|
||||
local max_idle_seconds=30
|
||||
case $profile in
|
||||
P24) slots=24; pool=32; media=24; global=48 ;;
|
||||
P28) slots=28; pool=36; media=28; global=56 ;;
|
||||
@@ -688,6 +709,7 @@ apply_capacity_profile() {
|
||||
"AI_GATEWAY_ASYNC_WORKER_HARD_LIMIT=$global" \
|
||||
"AI_GATEWAY_DATABASE_MAX_CONNS=$pool" \
|
||||
"AI_GATEWAY_DATABASE_MIN_IDLE_CONNS=$min_idle" \
|
||||
"AI_GATEWAY_DATABASE_MAX_CONN_IDLE_SECONDS=$max_idle_seconds" \
|
||||
"AI_GATEWAY_MEDIA_MATERIALIZATION_CONCURRENCY=$media" \
|
||||
"AI_GATEWAY_MEDIA_REQUEST_CONCURRENCY=$media" >/dev/null
|
||||
remote_kubectl rollout status "deployment/easyai-worker-$site" \
|
||||
@@ -695,7 +717,8 @@ apply_capacity_profile() {
|
||||
remote_kubectl set env "deployment/easyai-api-$site" -n "$namespace" \
|
||||
"AI_GATEWAY_ASYNC_WORKER_HARD_LIMIT=$global" \
|
||||
"AI_GATEWAY_DATABASE_MAX_CONNS=$api_pool" \
|
||||
"AI_GATEWAY_DATABASE_MIN_IDLE_CONNS=$min_idle" >/dev/null
|
||||
"AI_GATEWAY_DATABASE_MIN_IDLE_CONNS=$min_idle" \
|
||||
"AI_GATEWAY_DATABASE_MAX_CONN_IDLE_SECONDS=$max_idle_seconds" >/dev/null
|
||||
remote_kubectl rollout status "deployment/easyai-api-$site" \
|
||||
-n "$namespace" --timeout=300s
|
||||
done
|
||||
|
||||
Reference in New Issue
Block a user